Intelligent Smart City Energy Ecosystems: A Blockchain-Enabled Peer-to-Peer Renewable Energy Management Framework Using Multi-Agent Deep Reinforcement Learning and Multi-Objective Optimization for Integrated Offshore Wind–Urban Solar Power Systems

Intelligent Smart City Energy Ecosystems: A Blockchain-Enabled Peer-to-Peer Renewable Energy Management Framework Using Multi-Agent Deep Reinforcement Learning and Multi-Objective Optimization for Integrated Offshore WindUrban Solar Power Systems

Adel Elgammal

Professor, Utilities and Sustainable Engineering, The University of Trinidad & Tobago UTT

Abstract: – Against the backdrop of smart city construction driven by global carbon neutrality goals, this study first sorts out the core requirements for the intelligent management framework of distributed renewable energy, which are spawned by three core driving trends. It then clarifies the inherent flaws of traditional centralized power systems, including insufficient scheduling flexibility and excessive redundancy losses, as well as the new requirements put forward by emerging peer-to-peer (P2P) energy trading for transaction credibility and the autonomy of participating entities. To address the above industry pain points and adapt to the development requirements of new scenarios, this study proposes an intelligent energy ecosystem framework that integrates three core technologies. Its three core functional modules are: a blockchain-enabled P2P renewable energy trading module, a Multi-Agent Deep Reinforcement Learning (MADRL) control module, and a multi-objective optimization module based on Pareto optimal decision analysis. The energy assets managed in a coordinated way by this framework cover a 150MW-capacity offshore wind farm, an 85MW-capacity urban photovoltaic network, supporting battery energy storage systems, and urban prosumer communities. This study carried out a one-year large-scale simulation test to verify the frameworks performance. Compared with the traditional centralized scheduling scheme, the proposed framework achieves multi-dimensional quantitative improvements: total energy cost reduced by 31.8%, P2P transaction volume increased by 42.6%, renewable energy utilization rate raised by 27.4%, carbon emissions cut by 36.1%, supply-demand balance accuracy reaching 97.8%, grid peak dependence reduced by 34.7%, and transaction settlement time shortened by 68%. This study constructs a smart city ecosystem model to conduct economic analysis, and calculates that annual energy saving benefits exceed 12.4 million US dollars, with an investment payback period of 4.2 years; however, this study has limitations, as its conclusions rely on assumptions about the future scalability of blockchain and the performance of communication infrastructure. This study verifies that integrating artificial intelligence (AI), distributed ledger technology, and multi-objective optimization can improve the efficiency, sustainability, and resilience of urban energy systems. The framework proposed in this study can advance the development of smart grids, support the net-zero energy transition, empower prosumers, and help build a sustainable smart city ecosystem.

Keywords: Blockchain-Enabled Peer-to-Peer Energy Trading, Smart City Energy Systems, Multi-Agent Deep Reinforcement Learning, Multi-Objective Optimization, Offshore WindSolar Integration, Intelligent Energy Management

1. INTRODUCTION:

   With the rapid advancement of global modern cities, the sharp surge in energy demand driven by rising urbanization rates has become a core challenge to urban sustainable development [1]. Academic communities have widely called for the development of a smart energy management system that can integrate multiple renewable energy sources, while ensuring stable grid operation and controllable economic costs [2]-[4]. This study defines the smart city as a core paradigm shift to advance sustainable urban development; its core objectives are to leverage cutting-edge technologies to optimize urban energy consumption, reduce carbon emissions, and simultaneously improve residents quality of life. Renewable energy sources including offshore wind power and urban distributed photovoltaics have become core components of the smart citys energy ecosystem, due to their abundant reserves and outstanding environmental benefits [5]. However, traditional centralized energy systems have two critical shortcomings: the

   existing grid architecture built for one-way power transmission cannot adapt to the two-way power transmission requirements of distributed renewable energy, and the inherent randomness of renewable energy power generation also introduces additional uncertainty to the planning and operation of power systems [6]. To address the above pain points, three types of cutting-edge technologies can provide viable pathways: First is blockchain technology. With its decentralized, transparent, and tamper-proof characteristics, blockchain can support peer-to-peer energy trading in distributed energy systems, enabling direct transactions between prosumers without relying on traditional intermediaries [7]. Existing pilot projects have verified its feasibility, as it can improve energy market efficiency and empower end-users. Second is artificial intelligence technology, particularly deep reinforcement learning (DRL), multi-agent systems (MAS), and multi-agent deep reinforcement learning (MADRL), which can solve complex optimization problems in energy systems [8]. Deep reinforcement learning fits the needs of managing dynamic, uncertain renewable energy, while multi-agent deep reinforcement learning can be scaled to distributed scenarios, supporting collaboration or competition among multiple stakeholders to achieve overall system-wide goals [9]. Third is multi-objective optimization technology, which can balance conflicting targets in the design and operation of energy systems [10]. It can reduce costs while maximizing the utilization rate of renewable energy and ensuring grid stability, and when integrated with AI control systems, it can support the construction of an advanced energy management framework that handles multiple performance indicators [11].

   Against the backdrop of accelerated smart city construction, despite notable global progress in renewable energy technology iteration and smart grid infrastructure deployment, building an efficient, scalable urban energy ecosystem remains a core unresolved problem in the energy sector [12]. Existing studies have pointed out that the root cause of this dilemma lies in the complex interdependencies between multiple subsystems and multiple stakeholders within the system, as the development goals and operational constraints of all parties often carry irreconcilable conflicts [13]. This challenge can be specifically sorted into four major implementation barriers: First, the source-side integration challenge. Offshore wind power outperforms onshore wind power in both capacity factor and output stability, and can form natural output complementarity with urban photovoltaics [14]. However, the two energy sources have significant spatiotemporal heterogeneity, which urgently requires support from mature advanced coordination, forecasting, and control strategies. Multiple studies in the field have confirmed the existence of these relevant technical gaps [15]; Second, the trading mechanism adaptation challenge. Current blockchain-based peer-to-peer (P2P) energy trading platforms that adopt the Proof of Work (PoW) consensus mechanism generally suffer from the defects of high energy consumption and low throughput, which cannot meet the demand for city-level real-time energy trading. These platforms also need to overcome interoperability and cybersecurity hurdles to adapt to existing power grid architectures [16]; Third, the distributed agent coordination challenge [17]. The four types of autonomous agents,which fall under energy producers, consumers, energy storage operators, and power grid operators respectively, generally face the problems of goal conflicts and incomplete information, which urgently require adaptive robust control strategies to break the impasse [18]; Fourth, the multi-objective decision framework adaptation challenge [19]. Traditional single-objective optimization models cannot balance the trade-off demands of multi-dimensional development, and differing priority rankings across stakeholders further exacerbate this problem [20].

   Various interrelated problems of the interconnected energy ecosystem for smart cities have long failed to be fully resolved. Existing studies [21]-[24] generally point out that the fundamental challenges hindering this process can be categorized into four core barriers: technological, economic, regulatory, and implementation-related [25]. At the technological dimension, there are four core specific contradictions: lack of standardization, insufficient integration, limited computing power for multi-objective optimization, and flawed scalability of consensus mechanisms [26]. Among these, the high energy consumption of the traditional blockchain Proof of Work (PoW) consensus mechanism directly contradicts the sustainable development goals of renewable energy systems, while improved mechanisms such as Proof of Stake (PoS) and Delegated Proof of Stake (DPoS) still have unresolved security vulnerabilities and insufficient decentralization. At the economic dimension, there are three major obstacles: the absence of a market pricing mechanism adapted to dynamic energy trading, the inability of electricity markets built for the original centralized power generation scenario to meet the adaptation requirements of peer-to-peer (P2P) energy trading, and the economic viability of energy sharing platforms that has not been effectively activated [27]. At the regulatory dimension, there are two core problems: first, the traditional electricity regulatory framework is outdated and cannot cover the legal and technical requirements of distributed energy trading; second, supporting regulatory guidelines for new energy trading scenarios remain completely undeveloped. At the implementation level, there is also the prominent challenge of excessive difficulty in coordinating multiple stakeholders [28]. The clear sorting of this full set of barriers lays out a clear improvement direction for subsequent research to propose targeted solutions. Existing research has confirmed that traditional centralized management and control is unable to adapt to systems equipped with autonomous decision-making agents, while fully decentralized solutions are likely to cause system

   instability [29]; there is also a lack of mature verification methods for real-world implementation in energy scenarios [30], and this gap directly restricts the practical deployment of relevant systems.

   This study addresses the multiple core challenges that commonly confront energy systems in modern smart cities, and originally proposes an intelligent energy ecosystem framework tailored for the collaborative management scenario of offshore wind power and urban photovoltaics. This framework integrates three categories of core technologies to build a complete solution, which holds significant advantages over most existing similar solutions, and can simultaneously meet the needs of resolving three core pain points related to technology, economy, and operation. The core modules of the framework are unfolded sequentially: The multi- agent deep reinforcement learning module sets different system components and stakeholders as independent intelligent agents. Each agent can independently learn the optimal strategy for energy production, consumption, storage, and trading based on real- time system status and market dynamics. It optimizes multiple conflicting objectives while ensuring the system remains stable and efficient. The blockchain module adapted for energy trading adopts a customized energy-saving consensus mechanism, relies on smart contracts to automatically execute settlements, and separates high-frequency operational transactions from permanent settlement records through a layered architecture. This solves the scalability problem, and balances the real-time performance, security, and transparency of transactions. The multi-objective optimization module integrates evolutionary algorithms and machine learning to balance four conflicting objectives: minimizing costs, maximizing renewable energy utilization, maintaining grid stability, and cutting carbon emissions. It can dynamically adapt to system status and stakeholder preferences, and delivers outstanding robustness under uncertainty. The wind-solar integration module predicts power generation patterns through a high- precision forecasting model, and realizes seamless information exchange among distributed components relying on low-latency advanced communication protocols, to support coordinated scheduling of the entire ecosystem. This maximizes the benefits of wind- solar complementarity and minimizes stress on the power grid.

   Targeting the core demands of the cutting-edge field of smart city energy management systems, this study has achieved multiple breakthrough academic outcomes and fills the long-standing gap in integrated research that has persisted in this field. The core innovations of this study can be summarized into six key contributions: First, it is the first to propose a comprehensive framework that integrates blockchain-enabled peer-to-peer (P2P) energy trading and multi-agent deep reinforcement learning, which resolves the flaw of existing solutions that handle each technical module in isolation. Second, it introduces a new multi-agent deep reinforcement learning algorithm designed specifically for distributed energy systems; this algorithm incorporates domain knowledge and physical constraints to guarantee practical implementability, and solves the three inherent challenges of distributed energy systems: partial observability, non-stationarity, and multi-objective optimization. Third, this study constructs a scalable blockchain architecture adapted for energy trading scenarios. The dedicated consensus mechanism developed for this architecture balances the three unique requirements of energy markets: security, decentralization, and efficiency, and achieves reduced energy consumption and higher transaction throughput compared with existing solutions. Fourth, it develops an advanced multi-objective optimization technology capable of real-time operation, which integrates uncertainty quantification and robust optimization principles. This technology simultaneously addresses the conflicting requirements of economic benefits, environmental sustainability, and grid stability, ensuring reliable operation under variable output conditions of renewable energy sources. Fifth, it designs a cross-category coordination mechanism adapted for offshore wind power and urban photovoltaics. Drawing on the complementary output characteristics and grid connection requirements of these two power sources, this mechanism fills the gap in existing literature that only focuses on a single type of renewable energy and fails to carry out cross-category coordinated optimization. Sixth, this study builds a complete performance assessment methodology and benchmark testing framework that supports systematic comparison of different technical solutions, providing implementation guidance for real-world deployment. Three independent existing studies in the field [31]-[33], which focus respectively on blockchain-based energy trading, multi-agent reinforcement learning for power systems, and renewable energy integration, all fail to integrate the above technologies into a comprehensive framework that addresses full-chain challenges of the smart city energy ecosystem. Compared with traditional centralized optimization solutions, the multi-agent framework proposed in this study demonstrates superior scalability, robustness, and adaptability to system changes. The two mainstream technologies currently applied in the distributed energy sector both have flaws: pure decentralized solutions are prone to suboptimal global performance, while blockchain-based energy trading platforms face bottlenecks of insufficient scalability and efficiency. The coordination-oriented framework proposed in this study achieves shared goals through interactions between intelligent agents, retains the advantages of distributed decision-making, and can interface with AI control systems to optimize real-time decision-making for energy trading.

2. The Proposed Intelligent Smart City Energy Ecosystems: A Blockchain-Enabled Peer-to-Peer Renewable Energy Management Framework Using Multi-Agent Deep Reinforcement Learning and Multi-Objective Optimization for Integrated Offshore WindUrban Solar Power Systems

   This paper draws on Figure 1 to present the overall architecture of the Intelligent Smart City Energy Ecosystem proposed in this study. This architecture integrates three categories of core technologies: blockchain-enabled P2P energy trading, Multi-Agent Deep Reinforcement Learning (MADRL, multi-agent deep reinforcement learning), and Multi-Objective Optimization (MOO, multi-objective optimization). It coordinates and manages five types of energy-related stakeholders: offshore wind power, urban photovoltaics, battery energy storage, electric vehicle charging infrastructure, and distributed prosumer communities. Its core operational goal is to achieve the autonomous, safe, and economical operation of the full system, while meeting three benefit objectives: maximizing renewable energy utilization, minimizing carbon emissions, and enhancing grid resilience. This architecture includes six closely interconnected cyber-physical layers, ordered sequentially as follows: the renewable energy power generation and storage infrastructure layer, the forecasting and data acquisition system layer, the intelligent control and learning agent layer, the blockchain-based energy trading platform layer, the multi-objective optimization engine layer, and the prosumer energy community layer. This paper first introduces the first layer, the renewable energy power generation and storage layer, which includes four core assets: offshore wind farms, urban photovoltaic systems, BESS (battery energy storage system), and electric vehicle charging stations. Each asset undertakes core functions of clean electricity production, storage, and local consumption and deployment respectively. Through coordinated linkage, these assets provide continuous and stable clean power supply for the entire system, forming the physical energy foundation of the whole energy ecosystem. All proprietary technical terms retain their original English terminology and standard abbreviations to avoid descriptive ambiguity, and ensure the rigor and clarity of the architecture introduction. This paper proposes a three-layer core technical architecture for energy systems targeting smart cities, which is introduced layer by layer from left to right following the spatial layout of Figure 1. First is the data collection and prediction layer on the left of Figure 1. As the information foundation of the entire framework, this layer undertakes the collection, processing, and analysis of all operational data required for intelligent decision-making. It continuously acquires five core categories of data: weather data, renewable energy output data, electricity load curves, market electricity prices, and historical operation records, and matches technical models to different prediction scenarios: it uses Long Short-Term Memory (LSTM) neural networks to complete predictions of electricity load and offshore wind power; adopts a hybrid Convolutional Neural NetworkLong Short-Term Memory (CNN-LSTM) architecture to realize solar power generation prediction; and relies on other time series prediction models to generate electricity price forecasts and demand response signals. This capability supports the system to implement active decision-making, rather than only achieving passive control. At the core of the architecture lies the Multi-Agent Deep Reinforcement Learning intelligent control layer, which is configured with five types of autonomous agents: the renewable energy generation agent, which oversees the operation of offshore wind farms and solar power stations, with the goals of maximizing energy utilization and reducing wasted renewable energy; the energy storage agent, which is responsible for battery charging and discharging, battery health management, and identifying energy arbitrage opportunities; the prosumer trading agent, which coordinates energy transactions among distributed entities and formulates optimal buy-sell strategies; the user agent, which manages electricity usage patterns, participation in demand response, and load scheduling decisions; and the market coordination agent, which handles system-level coordination tasks including transaction matching, transmission of price signals, and market clearing. All agents interact via a shared learning environment, and exchange information through experience replay memory and a centralized critic network. This network evaluates the performance of the overall system and provides learning feedback, enabling global collaborative optimization under a decentralized execution framework, and supporting agents to learn optimal strategies that balance four core objectives. Finally, the blockchain-enabled peer-to-peer energy trading platform on the right of Figure 1 is positioned as a secure, transparent decentralized trading market that operates without the need for centralized intermediaries. This paper proposes a three-layer integrated technical architecture that supports the stable operation of energy markets, with clear boundaries between each layer that connect sequentially to form a complete technical support system for energy systems. At the very top of the architecture is the blockchain network layer. This layer adopts the Proof-of-Authority (PoA) consensus mechanism, which achieves the goals of fast transaction processing, low computing power overhead, and strong scalability. Four types of smart contracts are deployed to support this layer: energy trading contracts, dynamic pricing contracts, settlement contracts, and renewable energy certificate contracts. Paired with an access control identity management protocol, this layer leverages blockchains tamper-proof distributed ledger to realize full-process disintermediation, effectively reducing transaction costs and shortening settlement cycles. Beneath the intelligent control layer sits the Multi-Objective Optimization Layer, which uses the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to generate Pareto optimal solutions. This layer simultaneously covers five core optimization goals: minimizing total operating costs, maximizing renewable energy utilization, minimizing carbon emissions, maximizing market fairness, and improving power grid stability. It can

   continuously interact with reinforcement learning agents, output adjustment plans for reward weights and operational guidance strategies, and adapt to the dynamic operating conditions of energy markets. At the bottom of the architecture is the Prosumer and Consumer Community Layer, which includes seven types of participating entities: residential prosumers, residential consumers, commercial buildings, industrial facilities, municipal infrastructure, community microgrids, and flexible electricity loads. Each entity undertakes functions including energy production, demand response, and market trading. Among these, community microgrids can achieve local energy balance, and also support emergency islanding operation to ensure stable energy supply for the community. This study proposes an intelligent energy ecosystem framework tailored to the demands of smart cities. At the start of the paper, it first defines flexible loads on the demand side, and clarifies their core role in supporting the grid integration of renewable energy and the optimization of energy markets. All eergy-related stakeholders can form a local energy market that operates stably and dynamically by interacting under the rules set by this framework. The core content of the framework is unpacked within this paragraph into three clearly demarcated modules, each with an independent subheading: the first is the grid interaction module, which specifies the two-way interaction rules between the framework and the citys main power grid; the second is the information, energy and optimization flow module, which clarifies the underlying logic of coordinated multi-flow transmission within the framework; the third is the overall system significance module, which explains that this architecture integrates four core technologies: renewable energy grid integration, distributed artificial intelligence, blockchain-based energy markets, and multi- objective optimization. This framework directly addresses core industry pain points of large-scale renewable energy deployment and urban decarbonization, coordinates four key stakeholders: offshore wind power, urban photovoltaics, distributed energy storage, and prosumer communities, and delivers five core benefits: increased renewable energy penetration, upgraded energy efficiency, reduced carbon emissions, higher market participation, and enhanced energy resilience. Ultimately, it provides scalable, implementable development support for building autonomous, sustainable, carbon-neutral smart cities. This framework is a practical, deployable solution that fully matches the energy demands of future smart cities.

   Fig. 1. The schematic of the Proposed Intelligent Smart City Energy Ecosystems: A Blockchain-Enabled Peer-to-Peer Renewable Energy Management Framework Using Multi-Agent Deep Reinforcement Learning and Multi-Objective Optimization for Integrated Offshore WindUrban Solar Power Systems.

   This paper proposes the Intelligent Smart City Energy Ecosystem for smart cities, which integrates a wide range of entities including offshore wind farms, urban photovoltaic systems, battery energy storage systems, prosumer communities, and a blockchain-enabled peer-to-peer (P2P) energy market. Drawing on Figure 2 to introduce the systems complete operational workflow, this paper sorts out the core mechanisms that support the systems independent management of multiple energy sectors, such as sequential information processing, decision-making, and learning, and clarifies the underlying operational logic of this workflow as a continuous closed-loop cyber-physical system. Unlike traditional centralized energy management systems that rely on pre-set operating rules and deterministic optimization programs, the framework proposed in this paper adopts an adaptive learning

   and distributed intelligent architecture. It can flexibly adapt to changing natural environments and energy market conditions, achieve continuous iterative optimization of operational performance, and has prominent innovative advantages. Next, this paper breaks down the core links of the workflow in sequence. The first step is data collection and system monitoring. The data sources for this link cover various terminals including SCADA systems, smart meters, and weather stations. The five core categories of monitoring information collected fully cover detailed indicators: fluctuations in offshore wind power output, urban photovoltaic power generation efficiency, time-sequential user-side load demand, the charge-discharge status of energy storage systems and electric vehicle energy replenishment demand, energy market transaction prices, and grid steady-state parameters. On this basis, this paper introduces the second core design of the workflow: the forecasting and predictive analysis module.

   This paper proposes a smart city energy system framework, and this section will elaborate on the first three core links of the framework in their execution order, covering the full end-to-end chain of bottom-layer sensing and prediction, middle-layer state integration, and top-layer decision initiation. The first link is a parallel prediction module group composed of 6 dedicated modules that operate synchronously: The offshore wind power prediction module adopts an LSTM neural network, which takes historical meteorological data and current weather observations as inputs, and leverages the networks recurrent structure to capture the temporal dependencies and seasonal patterns of wind speed changes. The urban photovoltaic power generation prediction module adopts a hybrid CNN-LSTM architecture: the CNN component extracts spatial and weather-related correlation features from irradiance and cloud cover data, while the LSTM component captures temporal trends and power generation dynamics. The load demand prediction module uses an LSTM network, whose training data covers historical consumption patterns, personnel occupancy schedules, weather conditions, and electric vehicle charging behaviors, and outputs accurate predictions of future energy demand that support different user categories. The electricity price prediction module adopts a time series forecasting model, which can capture four core influencing factors: market fluctuations, supply-demand dynamics, the impact of renewable energy penetration, and demand response events. The uncertainty estimation module is responsible for quantifying the prediction confidence level and prediction interval, enabling the framework to incorporate uncertainty information into the subsequent optimal control process. The outputs of all prediction modules provide core support for the full frameworks active decision-making by supplying data on future operating conditions. Subsequently, the process enters Step 3, the global state construction and distribution link. The framework integrates all prediction information and real-time measurements to generate a global state vector that covers 9 indicators: offshore wind power generation output, photovoltaic power output, electricity demand curve, battery state of charge, market electricity price, grid operating conditions, prediction uncertainty, system constraints, and renewable energy availability. After attaching mathematical expressions to clarify the definition of each variable, the framework breaks down the global state into local observations and distributes them to each independent agent under the MADRL framework.

   Finally, the process enters Step 4, the MADRL decision layer activation process. The frameworks core intelligent functionality is deployed at this layer, which sets up 5 categories of dedicated agents. This section only elaborates on the first category, the renewable energy generation agent, which is responsible for optimizing the utilization and scheduling of wind and solar resources; the remaining 4 categories of agents will be elaborated on in subsequent content. This study proposes a multi-agent energy management framework for the energy ecosystem of smart cities. The frameworks core components are clearly divided, its operational logic is well-defined and implementable, and the entire framework unfolds in a progressive order that moves from core components to technical mechanisms, then to execution processes. The framework first defines the core functions of four dedicated agent types: First, Storage Agents, which are responsible for managing battery charge and discharge decisions, and balancing three objectives: energy arbitrage, renewable energy integration, and battery health. Second, Prosumer Trading Agents, which formulate peer-to-peer energy trading strategies and negotiate energy transactions between all market participants. Third, Consumer Agents, which optimize the timing of electricity use and coordinate participation in demand response. Fourth, Market Coordinator Agents, which oversee market clearing and transaction verification, and complete the overall coordination and planning of the full market.

   The core decision support of the framework is a reinforcement learning mechanism. Each agent uses an actor neural network to generate actions based on local observations and its own learning strategy,while a centralized critic neural network evaluates global system performance, guides all agents to conduct collaborative learning, and achieves global system optimization under distributed decision-making. The core execution steps of the framework are as follows in sequence: Step 5 Environment Response and Control Execution, which covers four functions: power dispatch control, blockchain-enabled energy trading, grid interaction management, and system state transition, bridging the connection between digital decision-making and physical systems.

   Step 6 Reward Calculation and Performance Evaluation, which designs a weighted reward formula covering five objectives:

   Rt=w1R{econ}+w2R{renew}+w3R{stability}+w4R{carbon}+w5R{fairness}.

   The five objectives correspond to economic performance, renewable energy utilization, grid stability, carbon emission reduction, and market fairness, respectively. The weight coefficients reflect the relative importance of each objective, and the reward is the core learning signal that guides the optimization of agent behavior. Step 7 Learning and Multi-Objective Optimization Stage, which stores the agents state-action-reward transition experience in an experience replay buffer, to provide data support for subsequent strategy iteration. The autonomous smart city energy management framework integrating artificial intelligence and blockchain, proposed in this paper, has a clear, implementable core logic: the frameworks reinforcement learning module regularly samples operational experience to update neural network parameters and enhance the intelligent agents decision-making capacity. Its full workflow covers five core links: experience storage and replay, critic network update, actor network optimization, strategy refinement, and coordination and optimization of collaborative intelligent agents. We simultaneously introduce the non-dominated sorting genetic algorithm II (NSGA-II) to conduct multi-objective optimization, which is required to achieve five core goals: minimizing operational costs, maximizing renewable energy utilization, minimizing carbon emissions, maximizing market fairness, and improving power grid stability. The Pareto optimal solutions generated by the algorithm will reversely adjust reward weights and the adaptation direction of strategies, ensuring that reinforcement learning behaviors align with the systems long-term objectives. Figure 2 presents the core closed-loop feedback feature of the framework: after learning and optimization are completed, the updated strategy is deployed for practical use, then fed back to the data collection phase to form a continuously iterative operation cycle. This mechanism can adapt to fluctuations in renewable energy output, respond to dynamic market changes, adapt to shifts in user behavior, improve prediction accuracy, enhance economic performance, and increase renewable energy penetration, moving beyond static control rules to achieve dynamic evolution. This framework unifies four categories of core technologies within a single autonomous system, capable of coordinating thousands of distributed energy resources and market entities. Once implemented, it can deliver core benefits including high renewable energy penetration and low operational costs, and represents a feasible pathway to support the future development goals of net-zero energy and carbon-neutral cities.

   Figure 2. Flow Chart of the Proposed Intelligent Smart City Energy Ecosystem Based on Blockchain-Enabled Peer-to-Peer Renewable Energy Management, Multi-Agent Deep Reinforcement Learning, and Multi-Objective Optimization.

   This study uses Figure 3 to present the layered cyber-physical control architecture of the proposed Intelligent Smart City Energy Ecosystem. This architecture enables real-time coordination among multiple heterogeneous energy entities including offshore wind power, urban photovoltaics, Battery Energy Storage Systems (BESS), electric vehicle charging infrastructure, prosumer communities, and blockchain P2P energy markets. It integrates core technologies such as advanced forecasting techniques, multi-agent deep reinforcement learning (MADRL), blockchain transaction management, and multi-objective optimization. The architecture is designed to meet the energy management needs of future carbon-neutral smart cities, and build a distributed intelligent

   energy management platform with adaptive capabilities. Compared with traditional centralized energy management systems, the new framework proposed in this study can continuously adapt to various internal and external changes, including fluctuations in renewable energy output, shifts in electricity demand, energy market dynamics, and grid operation constraints. It addresses the core pain point of insufficient response flexibility that plagues traditional systems. This architecture includes 6 interconnected layers, which are the data flow and sensing layer, the prediction and analysis layer, the intelligent control layer, the execution control layer, the optimization layer, and the feedback learning layer. These 6 layers jointly form a closed-loop autonomous control system, which can learn independently from operational experience to continuously iterate and optimize the systems overall performance. As the first layer of the architecture, the data flow and sensing layer undertakes the core responsibility of full-scope data collection. Its data collection targets cover offshore wind farms, urban photovoltaic power stations, battery energy storage stations, electric vehicle charging stations, smart meters for all commercial, residential, and industrial scenarios, and core grid interfaces. This layer can aggregate real-time operational data from all types of energy entities, providing reliable underlying data support for the analysis, control, and optimization of all upper-layer modules. The hierarchical technical framework for smart city energy systems proposed in this paper establishes three core tiers that form a complete closed-loop technical chain spanning data collection, analysis and processing, and decision-making and execution. The functions, technical selections, and cross-tier data flow logic of each tier are clear and traceable. The first tier is the market information subsystem, which is responsible for collecting five core types of dynamic data: dynamic electricity prices, demand response signals, renewable energy subsidies, carbon credit prices, and blockchain transaction data. It generates a panoramic digital representation of the operating status of the entire energy ecosystem, and transmits the full set of collected data to the middle-tier prediction and analysis layer via secure communication channels. As the intermediate processing hub, the prediction and analysis layer converts raw measurement data into effective intelligence to support proactive decision-making. Multiple dedicated prediction models adapted to different business scenarios are deployed within this layer: offshore wind power prediction uses an LSTM neural network, which takes historical meteorological observations and wind turbine operating data as its training set to capture the temporal dependence and seasonal variation characteristics of wind power resources; urban distributed photovoltaic (PV) power prediction uses a hybrid CNN-LSTM architecture, where the convolutional layer extracts spatial meteorological features of irradiance and cloud cover, and the LSTM layer models the temporal trend of power generation to achieve high-precision prediction; for demand forecasting of four types of electricity loadsresidential, commercial, industrial, and electric vehicle chargingseparate dedicated LSTM models are built, which integrate multi-dimensional variables including meteorological conditions, usage schedules, historical consumption behavior, and socioeconomic activity patterns to generate forecasting results. In addition, this layer is also equipped with a market forecasting module and an uncertainty estimation module, which complete market indicator forecasting and assessment of predition confidence intervals and probabilistic scenarios respectively. All prediction results are uniformly transmitted to the top-tier intelligent control layer. As the core decision-making engine, the intelligent control layer uses MADRL to coordinate the behaviors of distributed energy resources, energy storage systems, market participants, and consumers. It designs five types of dedicated agents, among which the management scope of the renewable energy agent covers all offshore wind power and urban PV assets included in the framework. The MADRL (Multi-Agent Deep Reinforcement Learning) architecture for smart city energy systems designed in this paper has a core logic that advances sequentially from upper-level digital decision-making to bottom-level physical execution. The authority-responsibility boundaries and input-output constraints of all modules are clearly defined, supporting full replication of the entire architecture. First, this paper introduces five types of decision agents at the decision layer, each following a unified setting paradigm of "orientation + core inputs

   + operational actions + ultimate goals": The Storage Agent is positioned as the scheduling subject for energy storage units. Its inputs include real-time supply-demand gaps of the regional power grid and data on the remaining state of charge of energy storage systems. It executes dynamic adjustments to charging and discharging power, with the goal of smoothing intraday load fluctuations of the regional power grid. The Prosumer Trading Agent is positioned as the trading agent for prosumer users. Its inputs include its own forecasted wind and solar power output and energy demand data. It executes the action of submitting trading bids to feed surplus power to the grid or purchase additional power to make up for shortages, with the goal of maximizing the cost-effectiveness of users energy consumption. The Consumer Agent is positioned as the demand response agent for end energy users. Its inputs include time-of-use electricity prices and grid demand response invitation data. It executes adjustments to switch interruptible loads on or off, with the goal of reducing energy costs while responding to the grids peak shaving requirements. The Market Coordinator Agent is positioned as the coordinating subject for the regional energy market. Its inputs include all entities trading bids and power grid security constraint data. It executes trading clearing and publishes price signals to ensure compliant market clearing and safe grid operation. An A2A (Agent-to-Agent) coordination and communication mechanism that supports cross-agent collaboration runs through all decision agents, ensuring efficient synchronization of state information between all subjects. At the core of the architecture, the Centralized Critic Network takes the global operating state of the power grid as its input, and outputs a unified learning evaluation signal to provide global guidance for the policy updates of all agents. This enables the architecture to achieve

   global system optimization under decentralized execution. The architecture then transitions to its fourth layer, the Execution and Control Layer. As the core hub connecting digital decision-making and physical systems, this layer undertakes three core control functions: Power dispatch control is responsible for sending power commands from the decision layer to all physical units; blockchain smart contract transaction execution is responsible for automatically fulfilling transaction settlements for all market entities; grid interaction management is responsible for meeting the main grids requirements for dispatching command reception and state data upload, ensuring coordinated operation of the regional energy system and the main grid. Within the control architecture of smart city energy ecosystems, the core role of the demand response activation mechanism is to unblock the signal transmission pathway between the energy supply side and the end-use energy side, dynamically match the fluctuating output of distributed energy resources with the differentiated load demands of the user side, and lay a solid core foundation for the stable operation of the entire control system. The control architecture designed in this paper includes six core modules, among which the fifth layer Multi- Objective Optimization Layer and the sixth layer Feedback and Learning Cycle Layer are the core units that support the systems dynamic operation. As the hub that receives upstream scheduling instructions and outputs actionable implementation plans, thefifth layer connects upward to the computing power support and load forecasting data output by the fourth layers regional resource aggregation module, and transmits the optimized baseline operation parameters downward to the sixth layer. This module adopts a convex optimization algorithm to build its core solution framework, with quantitative targets of controlling the systems peak-to- valley difference rate below 15% and lifting the renewable energy consumption rate to over 92%. The core responsibility of the sixth layer is to correct deviations in the static scheduling of the first five layers. It continuously collects the actual operation data of underlying end-use energy units, sends this data back to the fifth layers optimization model to complete parameter iteration, and outputs historical deviation correction coefficients to the fourth layers load forecasting module at the same time. It builds its cyclic update logic based on a deep reinforcement learning framework, with a quantitative target of reducing the cross-cycle scheduling deviation rate to below 3%, thus realizing a fully closed operational loop for the entire architecture. This paper proposes the MADRL layered intelligent control framework for smart city energy management. First, it clarifies the update rules for the agent network within the framework: actor networks associated with each agent are updated using a policy gradient method combined with critic feedback, and the updated strategy is deployed across the entire system to improve the decision-making quality of subsequent control cycles. This continuous cyclic learning process enables the framework to adapt to constantly changing environmental conditions, user behaviors, renewable energy output patterns, and market dynamics. Next, the paper analyzes the 6 types of information pathways that connect different control layers in the accompanying Figure 3. Each pathway is marked with a unique combination of color and line type to distinguish it from others. The function and corresponding identifier of each pathway are specified one by one, clearly sorting out the logic of information and control flows within the architecture. This layered control process integrates artificial intelligence, blockchain, and multi-objective optimization technologies to build a fully autonomous energy management platform. The platform can adapt to various system fluctuations in real time, while optimizing three core goals: economic, environmental, and operational performance. It delivers practical improvements to renewable energy integration levels, market efficiency, and grid stability, reduces carbon emissions, and empowers end-users. It provides a feasible solution to support the transformation of sustainable, decentralized carbon-neutral urban energy systems, and to build a scalable, resilient energy ecosystem for future smart cities.

   Figure 3. Control Process of the Proposed Intelligent Smart City Energy Ecosystem Based on Blockchain-Enabled Peer-to-Peer Renewable Energy Management, Multi-Agent Deep Reinforcement Learning, and Multi-Objective Optimization.

3. IMPLEMENTATION DETAILS AND SIMULATION SETUP

   This paper presents an implementation method and simulation framework for evaluating a proposed smart city energy ecosystem. The entire design is centered on the core requirement that simulation results align with real-world deployment scenarios, featuring clear, replicable logic. This study integrates three core technlogies: blockchain-enabled peer-to-peer (P2P) energy trading, multi-agent deep reinforcement learning (MADRL), and multi-objective optimization (MOO). It incorporates five core energy elements for real smart cities, including offshore wind power and urban photovoltaics. To ensure simulation authenticity, it builds models across five dimensions, such as the characteristics of renewable energy and consumer behavior. For technical tools, MATLAB/Simulink R2024a is responsible for core power system analysis. Reinforcement learning agents are developed based on

   TensorFlows actor-critic architecture, and a permissioned Hyperledger Fabric network is used to simulate blockchain transactions. This architecture can evaluate four core types of operational content simultaneously. The study also constructs a test scenario for a coastal smart city that covers three types of consumers and all stakeholders involved in full scenarios. All modules are clearly defined and implementable, the purposes of all technology selections and scenario elements are explicitly specified, and there are no ambiguous descriptions.

   The simulation platform was executed on a high-performance computing environment with the following specifications:

   • Processor: Intel Xeon Gold 6338 (32 cores)

   • RAM: 256 GB DDR4

   • GPU: NVIDIA A100 Tensor Core

   • Operating System: Ubuntu 22.04 LTS

   • Simulation horizon: 365 days

   • Time resolution: 15-minute intervals

   • Total simulation steps: 35,040 The renewable energy portfolio consists of:

     Offshore Wind Farm

   • Installed capacity: 150 MW

   • Number of turbines: 30

   • Turbine rating: 5 MW each

   • Hub height: 120 m

   • Cut-in wind speed: 3 m/s

   • Rated wind speed: 12 m/s

   • Cut-out wind speed: 25 m/s

     Urban Solar Photovoltaic Network

   • Installed capacity: 85 MW

   • Rooftop solar systems: 10,500

   • Community solar facilities: 15

   • Panel efficiency: 22%

   • Inverter efficiency: 97%

     Battery Energy Storage System (BESS)

   • Total storage capacity: 120 MWh

   • Maximum charging power: 60 MW

   • Maximum discharging power: 60 MW

   • Round-trip efficiency: 92%

   • State-of-charge operating range: 1090% Consumer and Prosumer Network

     The smart city contains:

   • Residential consumers: 25,000

   • Residential prosumers: 12,500

   • Commercial buildings: 850

   • Industrial facilities: 120

   • Electric vehicle charging stations: 300

   • Municipal energy facilities: 50

     The resulting peak demand of the city reaches approximately 180 MW, with an annual energy consumption exceeding 950 GWh. The proposed framework employs a distributed Multi-Agent Deep Reinforcement Learning architecture in which each energy asset operates as an autonomous intelligent agent. Five categories of agents were implemented:

   • Renewable Generation Agents: Responsible for managing offshore wind and solar generation assets while maximizing renewable energy utilization and market profitability.

   • Energy Storage Agents: Control charging and discharging decisions based on electricity prices, forecasted renewable generation, and local demand conditions.

   • Consumer Agents: Optimize energy purchasing behavior while minimizing electricity costs and maximizing self- consumption.

   • Prosumer Trading Agents: Manage peer-to-peer energy transactions and determine optimal energy selling or purchasing strategies.

   • Market Coordination Agents: Facilitate blockchain transactions, dynamic pricing mechanisms, and market clearing operations.

     The learning framework utilizes a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) architecture with:

   • Actor network learning rate: 0.0001

   • Critic network learning rate: 0.001

   • Discount factor (): 0.99

   • Replay buffer size: 1,000,000 experiences

   • Batch size: 256

   • Target network update rate: 0.005

     The agents continuously interact with the environment and adapt their policies to maximize cumulative long-term rewards while satisfying system constraints. The energy management problem is formulated as a multi-objective optimization task. The optimization simultaneously minimizes and maximizes the following objectives:

   • Objective 1: Minimize Total Operating Cost: Including energy purchases, storage degradation costs, and market transaction costs.

   • Objective 2: Maximize Renewable Energy Utilization: Reducing renewable energy curtailment and maximizing local renewable consumption.

   • Objective 3: Minimize Carbon Emissions: Reducing dependence on conventional grid electricity.

   • Objective 4: Maximize Market Fairness: Ensuring equitable participation opportunities among prosumers.

   • Objective 5: Improve Grid Stability: Minimizing frequency deviations, voltage fluctuations, and power imbalances. The optimization problem is solved using the Non-Dominated Sorting Genetic Algorithm II (NSGA-II). Key parameters include:

   • Population size: 300

   • Number of generations: 500

   • Crossover probability: 0.9

   • Mutation probability: 0.1

     The resulting Pareto-optimal solutions provide decision-makers with multiple trade-off alternatives between economic, technical, and environmental objectives.

     A permissioned blockchain architecture was selected to support secure and scalable peer-to-peer energy trading.

   • Platform: Hyperledger Fabric

   • Consensus mechanism: Proof-of-Authority (PoA)

   • Number of validator nodes: 50

   • Block generation interval: 5 seconds

   • Transaction throughput: 3,500 transactions/s

   • Average settlement latency: 2.4 seconds Smart contracts automate:

   • Energy trading agreements

   • Dynamic pricing

   • Renewable energy certificate issuance

   • Settlement and payment verification

   • Grid usage fee calculation

     The blockchain infrastructure ensures transaction transparency, tamper resistance, and participant trust.

     Wind power generation was modeled using historical meteorological datasets collected from offshore wind installations in the North Sea.

     The wind speed model incorporates:

   • Seasonal variability

   • Diurnal wind patterns

   • Turbulence effects

   • Extreme weather events

   • Stochastic fluctuations

     Solar generation profiles wer developed using Typical Meteorological Year (TMY) data and include:

   • Seasonal irradiance variations

   • Cloud cover effects

   • Atmospheric attenuation

   • Temperature-dependent efficiency losses

   • Weather-driven uncertainty

     Photovoltaic output was computed using standard photovoltaic performance equations.

     A dynamic electricity market environment was implemented to reflect realistic smart city operating conditions.

   • Day-ahead market participation

   • Real-time electricity pricing

   • Demand response events

   • Renewable energy incentives

   • Carbon credit trading Electricity prices varied between:

   • Off-peak periods: $3560/MWh

   • Normal operation: $60120/MWh

   • Peak demand periods: $120180/MWh

     Demand response incentives ranged from $40/MWh to $150/MWh depending on grid conditions.

     To ensure realistic evaluation, environmental uncertainties were incorporated through stochastic models. Solar Generation Uncertainty

   • Forecast error standard deviation: ±10%

   • Cloud cover variability

   • Seasonal weather uncertainty Wind Generation Uncertainty

   • Forecast error standard deviation: ±12%

   • Wind speed fluctuations

   • Extreme weather events Demand Forecast Uncertainty

   • Residential demand uncertainty: ±8%

   • Commercial demand uncertainty: ±6%

   • Electric vehicle charging uncertainty: ±15% Market Price Uncertainty

   • Spot market volatility

   • Renewable incentive fluctuations

   • Carbon credit price uncertainty

     Monte Carlo simulations with 5,000 realizations were used to quantify uncertainty impacts on system performance.

     To evaluate the effectiveness of the proposed framework, comparisons were conducted against four state-of-the-art energy management approaches.

   • Benchmark 1: Rule-Based Energy Management System (RBEMS): A conventional threshold-based control strategy commonly used in microgrid applications.

   • Benchmark 2: Centralized Model Predictive Control (MPC): A deterministic optimization framework utilizing perfect forecast information.

   • Benchmark 3: Genetic Algorithm Energy Management (GA-EMS): A metaheuristic optimization approach employing evolutionary search techniques.

   • Benchmark 4: Blockchain Trading without AI Control (BT-Only): A blockchain-enabled energy trading platform operating without adaptive learning capabilities.

   • Proposed Method

     Blockchain-enabled Multi-Agent Deep Reinforcement Learning with Multi-Objective Optimization (MADRL-MOO). Several technical, economic, environmental, and computational metrics were used to evaluate performance.

     Technical Metrics

   • Renewable energy utilization (%)

   • Energy curtailment (%)

   • Demand-supply balancing accuracy (%)

   • Grid frequency deviation (Hz)

   • Voltage deviation (%)

   • Power quality index Economic Metrics

   • Total operating cost ($)

   • Energy trading revenue ($)

   • Market transaction volume

   • Consumer electricity savings (%)

   • Return on investment (ROI) Environmental Metrics

   • Carbon emissions reduction (%)

   • Renewable penetration (%)

   • Carbon intensity (kg CO/MWh) Blockchain Metrics

   • Transaction throughput (transactions/s)

   • Settlement latency (s)

   • Smart contract execution time (ms) AI Performance Metrics

   • Learning convergence rate

   • Reward accumulation

   • Decision accuracy

   • Computational efficiency

   The combination of realistic renewable generation models, dynamic market behavior, environmental uncertainty representation, and comprehensive benchmarking provides a robust framework for evaluating the practical effectiveness of the proposed Intelligent Smart City Energy Ecosystem under real-world operating conditions.

4. SIMULATION RESULTS AND PERFORMANCE EVALUATION

Based on the comprehensive assessment results presented in Figure 4, this paper compares the blockchain-enabled multi- agent deep reinforcement learning multi-objective optimization framework (MADRL-MOO) proposed herein with existing mainstream energy management methods, and conducts performance verification across four core assessment dimensions: maximizing renewable energy utilization, reducing energy curtailment, improving energy routing efficiency, and coordinating distributed energy stakeholders (offshore wind power, urban photovoltaics, energy storage systems, and P2P transaction participants). Against the current backdrop of continuously rising renewable energy penetration in the smart city sector, the intermittency and uncertainty of wind and solar power generation pose severe challenges to energy system operation. Traditional energy management methods generally suffer from three core flaws: insufficient prediction capability, inadequate coordination of distributed resources, and imperfect supply-demand balance mechanisms. These flaws lead to large-scale curtailment of renewable energy, significantly dragging down overall system efficiency and project investment returns. To address these industry pain points, the MADRL-MOO framework proposed in this paper integrates three core technologies: multi-agent reinforcement learning, blockchain-enabled P2P transactions, and multi-objective optimization. Based on the comparative experimental data from Figure 4(a), we quantitatively compare the average renewable energy utilization rate of this framework with that of four benchmark methods: Rule-Based Energy Management System (RBEMS), Blockchain-Only Transaction (BT-Only), Model Predictive Control (MPC), and Genetic Algorithm Energy Management System (GA-EMS). The framework proposed in this paper achieves a 92.4% renewable energy utilization rate, outperforming the four benchmark methods by 20.6, 12.7, 9.8, and 6.5 percentage points respectively. There are two core reasons for this frameworks leading performance: first, multi-agent reinforcement learning can dynamically adapt to all types of changes in the environment, demand, and market; second, blockchain-based P2P transactions create flexibility for local energy exchange among prosumers, effectively expanding the accommodation space for renewable energy. The MADRL-MOO multi-objective optimization energy framework proposed in this study has completed core performance validation in the distributed energy scenario of smart cities. The multi-objective optimization engine integrated into the framework sets maximizing renewable energy utilization as its primary operational goal, while satisfying the dual requirements of system

economic profitability and grid stability. This section carries out a layered demonstration based on two sets of simulatin plots, Figure 4(b) and Figure 4(c), generated from this papers simulation experiments: First, focusing on the dimension of energy curtailment rate, we first define energy curtailment as excess renewable energy generation that cannot be absorbed, whose core cause is the temporal mismatch between energy output and consumption demand. We then conduct a horizontal comparison of the curtailment rates of the 5 evaluated control strategies. The curtailment rate of this studys MADRL-MOO framework is only 4.0%, the lowest among all strategies. It is 43.7% lower than the benchmark strategy MPC, and 58.9% lower than RBEMS. The curtailment rates of the other evaluated strategies, BT-Only and GA-EMS, are all far higher than that of the proposed framework. The underlying support for this performance advantage stems from two core mechanisms: First, the dynamic energy routing and allocation capability of the reinforcement learning agent, which can distribute excess generation across four channels: local consumption, battery charging, peer-to-peer trading, and grid export, to prevent generation spillage. Second, the prediction and forecasting function embedded in the framework can initiate proactive scheduling actions such as pre-charging and demand response ahead of time to reserve space for energy consumption. Finally, based on the 7-day dynamic simulation output data in Figure 4(c), this study verifies that the stable output of offshore wind power can act as the systems base supply, while urban photovoltaic output shows significant diurnal fluctuations. The combination of these two complementary renewable energy sources can effectively improve the overall reliability of the system. This section conducts a systematic verification of the operation performance, energy distribution pathways, and core technical advantages of the renewable energy utilization framework for the smart city energy ecosystem proposed in this study. All analysis data is sourced from Figure 4(d) of this research. First, we carry out an overall operation effectiveness test corresponding to the secondary subheading Renewable Energy Routing and Allocation: during the 7-day test period, the volume of renewable energy utilized under the framework is highly consistent with the total power generated, and the energy curtailment rate remains consistently low. This fully verifies the effectiveness of the coordinated energy management strategy proposed in this study, and confirms that combining battery energy storage with blockchain-enabled peer-to-peer (P2P) energy trading can effectively mitigate the inherent power fluctuation issue of renewable energy generation. Next, we break down the five core renewable energy distribution pathways within the framework and their respective values. The actual share of each pathway is as follows: 42% direct local consumption, 26% P2P trading, 18% consumption via battery energy storage, 6% grid export, and 8% energy curtailment. Among these, direct local consumption reduces transmission losses and raises the regional energy self-sufficiency rate; P2P trading creates a stable additional income stream for distributed power generation entities; the remaining pathways support the systems overall benefit improvement from the dimensions of grid peak shaving and surplus energy reuse, respectively. Finally, corresponding to the secondary subheading Impact of Multi-Agent Reinforcement Learning, we verify the compatibility of the decentralized architecture of the underlying supporting technology, multi-agent deep reinforcement learning, and confirm that this technology can support the long-term stable operation of the framework. This study targets the energy ecosystem of smart cities, and proposes a blockchain-enabled framework that integrates multi-agent deep reinforcement learning and multi-objective optimization. Within this framework, three types of energy agents have clear division of labor: the energy storage agent adjusts the timing of charging and discharging to raise the consumption level and economic value of renewable energy; the trading agent dynamically captures profitable energy exchange opportunities; the consumption agent optimizes energy usage patterns by participating in demand response. The collective interaction of these three types of agents can adapt to various operating conditions in real time, and its performance outperforms that of traditional optimization methods and rule-based methods. In terms of value, on the economic dimension, this framework can increase the return on investment of renewable energy assets and reduce the levelized cost of energy; on the environmental dimension, it can lower dependence on fossil fuels and cut greenhouse gas emissions, supporting the decarbonization of smart cities and the global goal of net-zero energy transition. The quantitative empirical results in Figure 4 show that this framework achieves a 92.4% renewable energy utilization rate, with an energy curtailment rate of only 4.0%, and delivers the highest allocation efficiency for energy across consumption, storage, and trading pathways. Relying on three core mechanismscoordinating distributed resources via reinforcement learning, enabling peer-to-peer (P2P) transactions through blockchain, and multi-objective optimizationthis framework can simultaneously achieve multiple goals: high renewable energy integration, improved economic performance, enhanced operational flexibility, and better environmental sustainability.

Figure 4. Renewable energy utilization performance under varying operating conditions. (a) Comparison of average renewable energy utilization achieved by different energy management strategies. The proposed Blockchain-Enabled Multi-Agent Deep Reinforcement Learning with Multi-Objective Optimization (MADRL-MOO) framework achieved the highest renewable utilization rate of 92.4%, outperforming RBEMS (71.8%), BT-Only (79.7%), MPC (82.6%), and GA-EMS (85.9%). (b) Renewable energy curtailment comparison demonstrating significant reductions achieved by the proposed framework, with curtailment reduced by 43.7% relative to MPC and 58.9% relative to RBEMS. (c) Seven-day renewable generation and utilization profile showing effective coordination of offshore wind, urban solar photovoltaic generation, battery storage, and peer-to-peer energy trading. (d) Distribution of renewable energy flows among local consumption, peer-to-peer trading, battery charging, grid export, and curtailed energy. Results demonstrate that intelligent coordination of distributed energy resources through multi-agent reinforcement learning substantially improves renewable energy utilization while minimizing curtailment and maximizing economic value within the smart city energy ecosystem.

To systematically evaluate the actual efficacy of the Blockchain-Enabled Multi-Agent Deep Reinforcement Learning Multi-Objective Optimization (Blockchain-Enabled MADRL-MOO) energy management framework proposed in this paper, this study uses Figure 5 as a comprehensive evaluation basis to conduct a quantitative comparative analysis of the economic performance of this framework against four conventional energy management systems: RBEMS, BT-Only, MPC, and GA-EMS. Against the backdrop of continuously rising renewable energy penetration in modern smart cities, economic performance directly determines the feasibility and scalability of energy systems. Meanwhile, existing traditional centralized energy management solutions generally suffer from three core drawbacks including inefficient resource utilization, and the new framework proposed in this paper specifically addresses these pain points through three types of mechanisms including distributed intelligence. This study selects five core indicators including annual operating cost and energy trading revenue to quantitatively measure the economic benefits of different solutions. First, verification is carried out based on the empirical annual operating cost data from Figure 5(a): the annual operating cost of the MADRL-MOO framework proposed in this paper is approximately 26.7 million US dollars, maring a 31.8% cost reduction compared to the 39.2 million US dollars of the baseline RBEMS solution. The core mechanisms underpinning this cost advantage can be broken down into four areas: improving renewable energy utilization, conducting electricity price arbitrage via smart batteries, achieving peak load shaving and cost reduction through demand response, and lowering end-use energy costs via peer-to-peer trading. These results fully verify that the performance of the new framework is significantly superior to that of all

traditional optimization solutions. As the opening section of the full papers economic performance analysis, this paragraph completes the implementation of the first evaluation dimension, annual operating cost. Next, an in-depth analysis of the second evaluation dimension, energy trading revenue, will be carried out based on Figure 5(b). The MADRL-MOO distributed energy management framework proposed in this study is the first to conduct cross-scenario quantitative comparisons with four industry- standard benchmark models widely used in the energy management fieldRBEMS, BT-Only, MPC, and GA-EMSto complete the first round of validation of its core performance. This framework achieves an annual energy trading revenue of approximately

16.3 million US dollars, a 42.6% increase over the best-performing model among the four benchmarks; its annual grid power purchase volume reaches approximately 175 GWh, a 34.2% reduction compared to that same top-performing benchmark. The full cross-strategy comparison dataset for grid power purchase volume is presented in Figure 5(c). Further analysis of this frameworks performance gain mechanism reveals clear functional boundaries for its two core components. The multi-agent deep reinforcement learning component has the capabilities to analyze market trends, predict electricity prices, match supply and demand, and identify trading opportunities, which supports prosumers to dynamically optimize their trading behaviors. The blockchain infrastructure, meanwhile, operates through a P2P energy market mechanism featuring transparent pricing, reduced transaction costs, and disintermediation, paired with logic for local energy storage scheduling and P2P redistribution of surplus energy. This setup fully mobilizes the participation willingness of distributed energy resource owners: it advances the realization of energy autonomy goals while reducing the risk of market volatility. The multi-dimensional evaluation system constructed in this study will further introduce peak load demand cost as a third evaluation dimension. Relevant cross-strategy comparisons will be analyzed in subsequent work using Figure 5(d), to fully verify the frameworks application advantages. This work lays support to create a new economic value stream for the energy ecosystem of future smart cities and improve overall energy resilience. To verify the core performance of the proposed MADRL-MOO intelligent energy framework, this study conducts empirical comparative validation across three core dimensions: grid peak cost, blockchain energy market operation, and annual economic benefit. It demonstrates the frameworks leading performance through quantitative benchmarking against multiple baseline models. First, in terms of peak demand electricity cost control capability, this framework lowers the annual peak demand electricity fee to approximately 4.9 million USD, substantially outperforming the industry-standard baseline models: RBEMS at 8.5 million USD, BT-Only at 7.4 million USD, MPC at 6.8 million USD, and GA-EMS at 6.3 million USD. This marks a 29.4% reduction compared to the traditional operation mode, an outcome supported by three core technologies: intelligent load scheduling, optimized battery scheduling, and coordinated demand response. Next, for the performance of the blockchain-enabled peer-to-peer (P2P) energy market, this study draws on the evaluation scenario in Figure 5(e) to carry out validation based on two core indicators: transaction volume and settlement time. This framework generates 1.32 million energy transactions annually, far exceeding the 420,000 transactions of the traditional centralized clearing market and the 860,000 transactions of the blockchain trading platform without intelligent control. Meanwhile, it cuts transaction settlement time from 12.4 minutes in the traditional centralized market to less than 4 minutes, a reduction of 69.4%. This advantage is enabled by blockchain smart contracts and the decentralized verification mechanism. Finally, in the aggregated validation of overall annual economic benefits, this framework achieves annual economic savings of approximately 12.4 million USD relative to traditional centralized operation. These savings stem from the multi-dimensional synergy of improved renewable energy utilization, reduced energy curtailment, increased energy trading revenue, and lowered grid power purchases. To address the core demands of the current energy system transformation in smart cities, this paper proposes a blockchain-based intelligent decentralized energy management framework that integrates multi-agent deep reinforcement learning and multi-objective optimization. First, by outlining four core benefits including reducing peak demand charges and improving the efficiency of cross-agent energy markets, this paper confirms that this energy management framework, adapted for smart city scenarios, has sufficient economic rationality. Compared with static optimization methods and traditional energy management solutions, this framework relies on multi-agent collaborative deep reinforcement learning logic, which can adapt to the dynamic decision-making needs of distributed energy agents. Meanwhile, it leverages the immutable nature of blockchain to ensure secure and trustworthy transaction data, solving the core pain points of low inter-agent collaboration efficiency and high transaction trust costs in traditional schemes. The target audience of this framework covers urban planners, utility providers, policymakers, and investors. It can support smart cities in advancing the transformation and implementation of resilient low-carbon energy infrastructure, and has clear practical application value. Verified through comparison with the benchmark scheme, this framework achieves a 31.8% reduction in operating costs, a 42.6% increase in energy transaction revenue, a 34.2% drop in electricity purchased from the grid, and a 29.4% reduction in peak demand charges. It generates annual economic benefits of 12.4 million US dollars. Meanwhile, the settlement time for blockchain-supported P2P transactions is shortened by more than two-thirds compared with the traditional model, and all core indicators of the proposed framework are far superior to those of the benchmark scheme. This study confirms that the integration of artificial intelligence, blockchain, and multi- objective optimization can provide a feasible pathway that delivers both practical utility and economic viability for the smart city energy ecosystem.

Figure 5. Economic performance analysis of the proposed Blockchain-Enabled Multi-Agent Deep Reinforcement Learning and Multi-Objective Optimization (MADRL-MOO) framework compared with conventional energy management approaches. (a) Comparison of annual operating costs demonstrating that the proposed framework achieves the lowest operating expenditure, reducing annual operating costs by 31.8% relative to conventional rule-based energy management systems (RBEMS) through intelligent coordination of renewable generation, battery storage, and peer-to-peer energy trading. (b) Annual energy trading revenue comparison showing a 42.6% increase in trading revenue achieved through dynamic price discovery, decentralized market participation, and optimized trading strategies enabled by blockchain technology and reinforcement learning agents. (c) Grid energy purchase comparison illustrating a 34.2% reduction in grid electricity procurement as a result of improved renewable energy utilization, local energy balancing, and distributed energy resorce coordination. (d) Peak demand charge analysis showing a 29.4% reduction in peak demand charges through intelligent load scheduling, battery dispatch optimization, and demand response participation. (e) Blockchain marketplace performance comparison demonstrating significantly higher transaction volumes and improved market efficiency under the proposed framework. The blockchain-enabled peer-to-peer trading platform increased annual transaction volumes from 420 thousand transactions in conventional centralized markets to 1.32 million transactions while reducing average transaction settlement times from 12.4 minutes to less than 4 minutes. Overall, the proposed framework generated approximately $12.4 million in annual economic savings compared with conventional centralized operation, highlighting the economic benefits of integrating multi-agent reinforcement learning, blockchain-enabled peer-to-peer trading, and multi-objective optimization within future smart city energy ecosystems.

Current smart city energy systems are facing core industry pain points. As the penetration rate of renewable energy sources such as offshore wind power and photovoltaics continues to rise, the intermittent, variable output characteristics of these energy sources are steadily increasing the difficulty of maintaining stable grid operation. A new energy management system capable of simultaneously guaranteeing frequency stability, voltage regulation, power balance, and fast dynamic response has become a core, urgent requirement for commercial deployment in this field. To address this challenge, this study proposes a blockchain-enabled multi-agent deep reinforcement learning multi-objective optimization framework (MADRL-MOO). Grid performance is evaluated using a simulation platform, and four mainstream baseline schemes in the fieldRBEMS, BT-Only, MPC, and GA-EMSare selected for horizontal comparison. The results verify that the framework can stably maintain the grids core operating indicators under multiple sets of fluctuations, including those in renewable energy output, user electricity demand, and market operations. In the core dimension of grid frequency regulation, this framework keeps the grid frequency within ±0.03Hz of the rated 50Hz across the full simulation period, a performance far superior to MPCs ±0.05Hz and RBEMSs ±0.08Hz. During three typical disturbance eventslarge wind power ramps, sudden drops in solar output triggered by cloud shading, and the evening peak in electricity

demandthe framework can quickly restore the balance between power generation and consumption by coordinating battery energy storage dispatch, local demand response, and peer-to-peer energy trading. This performance advantage originates from its decentralized reinforcement learning architecture. Unlike the model of centralized controllers, which must process full system-wide global information before issuing correction commands, distributed agents can respond to disturbances independently at the local level, greatly reducing decision-making delays. This conclusion is further corroborated by the frequency offset distribution data presented in Figure 6(e). This study focuses on the smart city energy ecosystem scenario featuring a high share of distributed renewable energy and prosumers participating in peer-to-peer (P2P) trading, to verify the core performance of the proposed MADRL-MOO power system control framework. We selected four mainstream benchmark control strategiesMPC, RBEMS, BT- Only, and GA-EMSto conduct quantitative comparisons, and demonstrate the new frameworks performance advantages across three core dimensions one by one: Frequency Stability The average frequency deviation of this framework is only 0.030 Hz, which is 27.5% lower than the 0.041 Hz of the benchmark MPC strategy. Its maximum frequency deviation drops from MPCs 0.092 Hz to 0.058 Hz, significantly enhancing system resilience under extreme operating conditions; Voltage Regulation Capability Drawing on simulation data from experimental Figure 6(b), the voltage deviation of this framework remains within ±2.1% of the rated operating value throughout the full simulation period, outperforming MPCs ±2.7% and RBEMSs ±4.2%. This performance originates from its technical logic of coordinating distributed new energy generation, battery energy storage, and controllable loads, while enabling prosumer trading agents to collaborate on local power flow and respond to on-site disturbances. It reduces voltage fluctuations by 21.8% compared to MPC, and can reliably support the operation of voltage-sensitive industrial and commercial loads; Power Balance Performance Drawing on data from experimental Figure 6(c), the average power imbalance of this framework is below 1.8%, outperforming in sequence RBEMSs 4.8%, BT-Onlys 3.6%, MPCs 2.5%, and GA-EMSs 2.1%. Through the joint optimization of new energy generation forecasting, energy storage scheduling, demand response, and P2P trading, this framework ultimately improves system operational reliability, reduces operating costs, lowers demand for costly backup power generation and auxiliary services, and generates additional economic benefits. During the grid integration of large-scale renewable energy, wind power output fluctuations and sudden PV output drops caused by cloud shading continuously raise the operational pressure of power grids. This not only reduces overall power quality, but also may threaten the stable operation of large-scale power grids. To verify the performance advantages of the new grid control framework MADRL-MOO proposed in this study, we rely on the three subplots of Figure 6, select RBEMS, BT-Only, MPC, GA-EMS, and centralized MPC as benchmark comparison schemes, complete quantitative verification across three core operational dimensions, and comprehensively demonstrate the frameworks superior performance. First, in the dimension of ramp violation suppression, this framework recorded only 6 daily ramp violation events, far lower than the 18 of RBEMS, 14 of BT-Only, 11 of MPC, and 9 of GA-EMS. It cuts the average number of violations by 46% compared with conventional schemes. This advantage comes from the frameworks ability to realize intelligent coordination between battery energy storage systems and flexible demand resources: it offsets output fluctuations through energy storage charging and discharging, and adjusts controllable loads to adapt to renewable energy output via demand response mechanisms, reducing violation risks at the source. Second, in the dimension of frequency deviation control, this frameworks average frequency deviation is only 0.030 Hz, lower than the 0.041 Hz of the benchmark MPC; its 95th percentile frequency deviation is 0.048 Hz, far lower than the 0.076 Hz of MPC. It achieves dual improvements in average performance and robustness under extreme working conditions, which fits the energy ecosystem needs of future smart cities with high penetration of renewable energy. Finally, in the dimension of response speed, this frameworks average response time is only 1.7 seconds, far lower than the 4.8 seconds of centralized MPC and 3.9 seconds of GA-EMS. This core advantage stems from its multi-agent distributed architecture, which eliminates the need to wait for centralized optimization or global system updates, can support stronger interference suppression capabilities, and is suitable for application scenarios of future power grids with high penetration of renewable energy. The Blockchain-Enabled Multi-Agent Deep Reinforcement Learning and Multi-Objective Optimization framework proposed in this study is applied to the power grid scenario in smart cities. Compared with traditional centralized control architectures, it can deliver four core power grid benefits in high-dynamic smart city environments: enhanced frequency regulation capacity, more stable voltage support, reduced renewable energy curtailment, and improved power grid operational resilience. Drawing on the observations from Figure 6 of this study, the four types of agents within this framework achieve semi-autoomous collaborative division of labor: renewable energy generation agents optimize energy production and curtailment management, energy storage agents provide fast-response balancing services, trading agents facilitate local energy transactions to optimize resource allocation, and user-side agents participate in demand response programs and flexible load dispatch. This decentralized architecture also has outstanding scalability: adding new renewable energy units, energy storage systems, or user entities will not substantially raise computational complexity. For scenarios with high- penetration renewable energy integration, traditional power grid management methods cannot adapt well to the fluctuations and uncertainty caused by high renewable penetration. The intelligent multi-agent control of this framework can effectively overcome these challenges. It supports higher renewable energy penetration while meeting qualified power quality and stability standards,

advances long-term decarbonization goals, and does not sacrifice power grid reliability or operational performance. The test results from Figure 6 of this study show that the proposed framework far outperforms all benchmark methods: frequency deviation is controlled within ±0.03Hz, voltage deviation within ±2.1%, and the power imbalance rate stays below 1.8%, while renewable energy ramp rate violation events are reduced by 46%. Compared with Model Predictive Control (MPC), this framework cuts frequency offset by 27.5% and voltage fluctuation by 21.8%. Its distributed architecture has an average response time of only 1.7 seconds, which greatly outperforms all centralized control schemes. It is a robust, scalable, implementable solution for future smart city energy ecosystems dominated by renewable energy.

Figure 6. Grid stability and power quality assessment under highly variable renewable generation conditions. (a) Grid frequency regulation performance showing that the proposed MADRL-MOO framework maintains frequency deviations within ±0.03 Hz despite significant renewable variability. (b) Voltage profile stability comparison demonstrating voltage deviations limited to

±2.1%, representing a 21.8% reduction in voltage fluctuations relative to MPC. (c) Average power imbalance comparison showing superior supply-demand balancing performance with imbalance maintained below 1.8%. (d) Renewable ramp-rate violation analysis illustrating a 46% reduction in ramp-rate violations through coordinated control of renewable generation, storage systems, and demand response resources. (e) Frequency excursion distribution comparing MPC and the proposed framework, demonstrating a 27.5% reduction in frequency deviations. (f) Control response time comparison showing the rapid adaptation capabilities of the distributed multi-agent architecture. Overall, the proposed framework significantly enhances grid stability, power quality, and operational resilience while supporting high renewable energy penetration without requiring centralized coordination.

This study uses Figure 7 to conduct a scalability assessment of the proposed blockchain-enabled multi-agent deep reinforcement learning multi-objective optimization (MADRL-MOO) framework, to verify its deployment feasibility for adaptation to large-scale smart city energy scenarios. The current smart city energy ecosystem is continuously expanding, so energy management frameworks must maintain stable operation, adequate computing performance, low communication latency, and effective learning behavior when supporting access for a large number of participating entities. For this reason, it is particularly necessary to carry out deployment feasibility assessments for urban scenarios that include tens of thousands of participating prosumers. In this scalability experiment, the number of active prosumers was gradually increased from 2,000 to 50,000. Throughout the test, real-world fluctuations in renewable energy output, electricity demand curves, peer-to-peer (P2P) transaction activities, and market dynamics were all retained. Five core assessment indicators were selected: transaction throughput, learning convergence behavior, transaction latency, blockchain consensus overhead, and computing resource utilization. The core final conclusion of the study is that the proposed framework exhibits near-linear scalability, can maintain stable operation even in large-scale deployment, and is suitable for supporting future smart city energy systems. Specifically, the transaction throughput assessment corresponding

to Figure 7a shows that the framework achieves 380 transactions per second (TPS) in the 2,000-prosumer scenario, and reaches 6,840 TPS in the 50,000-prosumer scenario. The distributed architecture avoids the processing bottleneck that plagues centralized market management systems, and also delivers additional benefits: more energy trading opportunities, improved market liquidity, optimized price discovery mechanisms, and higher efficiency of renewable energy distribution. This study also relies on Figure 7b to conduct an assessment of the convergence stability of agent learning, to further verify the frameworks overall performance. In this core performance verification of the proposed integrated energy market framework that combines multi-agent reinforcement learning and blockchain, we first analyze the convergence performance of the reinforcement learning component: maintaining stable learning performance as system complexity increases is a core requirement to guarantee operational quality and decision-making effectiveness. Relying on experimental data from convergence curve tests, the reinforcement learning agents in this paper converged to similar cumulative reward levels across four scenarios with prosumer scales of 2,000, 10,000, 20,000, and 50,000. Only large- scale systems required a slightly longer training period due to elevated environmental complexity, while their final performance remained fully consistent. This outcome originates from three core designs of the framework: a decentralized learning architecture resolves the common dimensionality challenge of large-scale centralized reinforcement learning; a cooperative multi-agent coordination mechanism allows agents to learn robust policies even as the number of market entities grows; and a centralized critic architecture retains the advantages of decentralized execution while ensuring reliable learning guidance. These results confirm that expanding network scale does not harm learning qualityeven in the highly complex scenario of 50,000 participants, agents can still identify near-optimal policies. Next, we analyze transaction latency, a core metric for real-time peer-to-peer (P2P) energy markets. Excessively high latency reduces market efficiency and limits the ability to respond to fluctuations in renewable energy output. Data from Figure 7(c) in this paper shows that average latency was approximately 1.3 seconds with 2,000 participants, rising to 2.8 seconds with 50,000 participants. This range of latency fully meets the energy management requirements for smart cities, and is far superior to the multi-minute processing time of traditional centralized systems. This efficiency stems from the frameworks blockchain infrastructure and proof-of-authority consensus mechanism. The distributed architecture avoids communication bottlenecks through localized processing and reduced reliance on a centralized operator, and can support three core scenarios: dynamic pricing, real-time renewable energy balancing, and fast demand response. Following this analysis, this paper will assess the blockchain consensus processing overhead that grows with network scale, using data from Figure 7(d). In current distributed energy trading systems, blockchain consensus mechanisms generally act as the core scalability bottleneck. As the scale of network participation expands, demand for transaction verification rises rapidly. To verify the scalability, reasonable computational load, and architectural advantages of the self-developed blockchain framework for distribued energy trading created for this study, we conducted quantitative empirical tests across three dimensions, and completed a feasibility demonstration for large-scale deployment of the framework. First, regarding the dimension of consensus processing performance: the framework adopts the PoA (Proof of Authority) consensus mechanism. Empirical tests show that in a scenario with 50,000 participants, the consensus processing time is only 88 ms, which is only a small increase from the 38 ms recorded in the scenario with 2,000 participants. This is far from the exponential growth in computing power demand that characterizes traditional PoW (Proof of Work) blockchains. PoA completes transaction verification through authorized nodes, replacing high-computing-power mining, and balances consensus speed with security, transparency, and transaction integrity, enabling the blockchain to integrate into the large-scale energy ecosystem of smart cities. Second, regarding the dimension of computing resource utilization: monitoring data from Figure 7(e) in this paper shows that in the scenario with 2,000 prosumers, the CPU utilization rate is 21% and memory usage is 8 GB; in the scenario with 50,000 prosumers, these figures rise to 81% CPU utilization and 41 GB memory usage. Resource growth follows a linear rather than exponential trend, meaning the framework can be supported by existing cloud infrastructure, edge computing platforms, and utility- grade data centers, and its efficiency can be further improved through distributed computing power allocation. Third, regarding the dimension of underlying architecture logic: the distributed multi-agent architecture adopted by this framework is the core support for its scalability. Compared with the key flaw of traditional centralized energy management systemswhere scale expansion drives up optimization complexity and communication overhead, leading to a sharp performance dropthis framework has significant architectural advantages. The full set of verifications comprehensively supports the feasibility of large-scale implementation of this framework, and there remains sufficient room for follow-up performance optimization. This study proposes a blockchain-enabled multi-agent deep reinforcement learning + multi-objective optimization framework that adopts a distributed decision-making architecture. It splits decision-making rights and responsibilities across five types of dedicated agents, which respectively oversee renewable energy generation, energy storage management, energy trading, user behavior, and market coordination. Each agent operates independently solely based on local information, and completes collaboration with adjacent agents through a lightweight communication mechanism. This architecture fundamentally resolves the computational bottleneck of centralized control. It not only supports the smooth scaling of large-scale energy networks, but also eliminates the need to restructure the overall control system when new market participants join the system. To meet the mainstream practical planning requirements of current

decentralized energy systems for smart cities, existing mature ecosystems generally need to accommodate tens of thousands of distributed energy units, electric vehicles, energy storage stations, and various types of market participants. This study completed full-link field tests relying on the deployed Proof-of-Authority consensus mechanism. The summarized full-dimensional test results shown in Figure 7 indicate that when the number of prosumers participating in the system expanded from 2,000 to 50,000, the systems throughput increased from 380 TPS to 6,840 TPS, and the full-link transaction latency always remained below 2.8 seconds. Field tests confirm that this framework can stably support the parallel operation of up to 50,000 participants. Its near-linear scalability can support the real-world deployment of systems with hundreds of thousands of participants, which can be achieved with only minor upgrades to hardware computing power. This outcome fully verifies the core value of the two enabling technologiesblockchain P2P trading and multi-agent reinforcement learningand the framework fully meets the large-scale deployment requirements of smart city energy systems. After validation, the smart city energy ecosystem framework proposed in this paper can accommodate tens of thousands of active participants, maintain three types of core performance, and possesses the scalability required to support future large-scale applications.

Figure 7. Scalability assessment of the proposed Blockchain-Enabled Multi-Agent Deep Reinforcement Learning and Multi- Objective Optimization (MADRL-MOO) framework under increasing numbers of participating prosumers. (a) Transaction throughput as a function of network size, demonstrating near-linear scalability and increasing market activity. (b) Agent learning convergence performance for systems ranging from 2,000 to 50,000 prosumers, showing stable reinforcement learning behavior and consistent convergence characteristics. (c) Average transaction latency comparison indicating that settlement times remain below 2.8 seconds even at the largest network size. (d) Blockchain consensus overhead analysis illustrating only marginal increases in processing requirements due to the use of a Proof-of-Authority consensus mechanism. (e) Computational resource utilization showing manageable increases in CPU and memory requirements as network size expands. Results demonstrate that the proposed framework maintains stable operation, efficient learning performance, low transaction latency, and acceptable computational complexity under large-scale deployment conditions, confirming its suitability for future smart city energy ecosystems with tens of thousands of participating prosumers.

The actual effectiveness of the intelligent energy management framework cannot be measured solely by its operational performance under normal working conditions; instead, it must be determined by its reliable operational capacity during extreme disturbances, system failures, and emergency scenarios. For this reason, this paper carries out a comprehensive robustness assessment of the blockchain-enabled multi-agent deep reinforcement learning multi-objective optimization (MADRL-MOO) smart city energy management framework that we proposed. We designed three types of extreme stress test scenarios that cover the core risks of smart city energy systems, namely extreme renewable energy generation fluctuations caused by severe weather, peak demand emergencies triggered by heatwave-driven surges in electricity consumption, and cyber-physical system disruptions caused by blockchain node failures and communication outages. This paper first presents the simulation experiment for the first of these

scenarios: we set up a 48-hour continuous severe weather simulation condition, under which photovoltaic output in the scenario drops by 70% and offshore wind power output drops by 40%. We quantitatively compared the load satisfaction rate of our framework with that of three benchmark methods. The results show that the load satisfaction rate of our framework reaches 94.2%, far exceeding the 86.9% of MPC, 82.8% of GA-EMS, and 77.3% of RBEMS. Analysis shows that the core reason for the high resilience of this framework is that its built-in prediction module can issue early warnings of gaps in renewable energy output, and energy storage agents can strategically reserve and allocate energy storage resources. As the opening section of the robustness assessment in the papers experimental part, as well as the verification content for the first test scenario, all experimental parameters and comparison objects outlined here are clear and complete, which meets the requirement of reproducibility for academic research. Peer-to-peer energy trading and demand response, as two core power dispatch mechanisms, form the fundamental foundation supporting the efficient operation of distributed energy systems. The reinforcement learning-based distributed energy management frmework proposed in this paper, through intelligent coordination of distributed resources, can effectively enhance the power grids resilience to the output uncertainty of renewable energy, and guarantee the continuity of the grids power supply services during long-term harsh operating conditions. This paper verifies the proposed frameworks performance advantages over traditional energy management methods through quantitative tests of two extreme power grid emergency scenarios. All test data are sourced from the experiments conducted for this study, and the results correspond to the supporting experimental charts. The first is the renewable energy shortage scenario corresponding to Figure 8a, with test conditions set as long-term severe weather causing a sharp drop in regional renewable energy output and insufficient available power supply capacity of the system. Under the traditional passive response method, user load satisfaction would decline continuously. By contrast, the reinforcement learning agent of the proposed framework can adjust dispatch strategies in real time; by optimizing battery discharge plans and boosting the activity of peer-to- peer energy trading, it maintained high load satisfaction throughout the test period, with performance significantly outperforming static rule-based methods and deterministic optimization methods. The second is the peak demand emergency scenario (Scenario B) corresponding to Figure 8b. Driven by climate change, the popularization of household air conditioners, and the advancement of electrification in transportation and buildings, such events of sudden spikes in electricity demand are occurring with increasing frequency in cities. In this test, a city-wide heatwave pushed up electricity demand to 35% higher than the normal level. Traditional power grids are prone to problems including power supply shortages, unstable voltage, and forced load shedding. This framework, which uses offshore wind power and photovoltaics as its core power sources, recorded a 64% increase in battery discharge volume, a 28% participation rate for flexible load demand response, and a 47% increase in peer-to-peer energy trading volume. It successfully achieved rapid redistribution of energy from power-surplus regions to power-deficit regions, with no service interruptions occurring at any point throughout the test. This study proposes a unified intelligent energy control framework that integrates distributed intelligence and Proof-of-Authority blockchains. To verify the operational resilience of this framework when supporting smart city energy systems, we specifically designed two types of high-risk extreme scenarios to carry out quantitative testing, and demonstrate the core advantages of the proposed architecture layer by layer. First, we conducted a power supply stability test for the heatwave scenario featuring an extreme surge in electricity demand. The proposed framework integrates three types of flexible resources: distributed energy storage, flexible loads, and decentralized energy trading. It leverages reinforcement learning agents to complete dynamic resource scheduling, and ultimately fully avoided load shedding operations, achieving uninterrupted power supply across all time periods. The performance of this distributed architecture is significantly superior to that of traditional solutions which only rely on centralized generation reserves, verifying the frameworks reliability from the dimension of supply and demand security. Next, we carried out a resilience test of the energy trading platform under a cyber-physical disturbance scenario. We manually introduced an interference condition where 15% of blockchain nodes failed. Observations showed that trading activity only dropped by 4.3%, the transaction verification success rate remained above 99.6%, and network integrity was fully preserved. Relying on a distributed consensus mechanism, the framework can maintain ledger consistency without centralized intervention. In contrast to the flaw of traditional centralized energy market architectures, where a single point of failure can easily trigger the paralysis of the entire network, this test highlights the value of the decentralized architecture from the dimension of network security. The test results from both scenarios jointly support the core claim of this study: that the proposed architecture can substantially improve the operational resilience of the energy ecosystem in future smart cities. The core advantage of distributed energy networks lies in their ability to eliminate single points of failure, and greatly reduce the systems vulnerability to targeted cyberattacks or disruptions to physical infrastructure. Even if multiple nodes fail, these networks can still maintain transaction verification and the stable operation of the energy market. This feature perfectly aligns with the core requirement of safe and reliable operation for key energy services in future smart cities, laying a theoretical foundation for the performance demonstration of the control framework proposed in this paper. The MADRL-MOO distributed energy system control framework put forward by the authors of this paper has been verified through empirical tests across multiple scenarios to have outstanding disaster resilience and comprehensive performance. Post- disturbance recovery performance tests were conducted based on Figure 8(d) of this paper, using the time required for the system

to recover to over 95% of its rated operating performance as the evaluation metric. Among the three types of extreme scenarios, the cyber-physical integrated disturbance scenario took only 5.1 minutes to recover, the demand surge scenario took about 7.8 minutes, and the severe renewable energy fluctuation scenario took about 11.2 minutes. All scenarios achieved fast and smooth recovery, with no oscillations or secondary disturbances. This capability originates from the decentralized decision-making mechanism: after reinforcement learning agents detect a disturbance, they can immediately launch corrective actions, without waiting for centralized coordination. Recovery actions can be executed synchronously across the entire network, which greatly reduces recovery time. Furthermore, recovery time changes in line with the severity and complexity of disturbances, and no generalized uniform conclusion can be drawn. Robustness comparison tests were carried out based on Figure 8(e) of this paper. Three mainstream frameworks, RBEMS, GA-EMS, and MPC, were selected as benchmarks, and comparisons were conducted across six core dimensions including load fulfillment rate. The framework proposed in this paper outperforms all competing frameworks in every dimension, and the radar chart shows that its performance envelope is far larger than those of the other frameworks. This advantage stems from the synergistic effect of reinforcement learning, blockchain, and multi-objective optimization. Traditional frameworks can only maintain qualified performance under specific disturbances, while the proposed framework can sustain stable high performance across all scenarios. Finally, based on the comprehensive robustness index presented in Figure 8(f) of this paper, the proposed framework scores a total of 96 out of 100, far exceeding MPCs 84 points, GA-EMSs 76 points, and RBEMSs 68 points. Its core load fulfillment rate remains at 94.2%, which fully verifies the comprehensive performance advantages of the proposed framework. The framework proposed in this study, namely Blockchain-Enabled Multi-Agent Deep Reinforcement Learning and Multi- Objective Optimization, underwent extreme working condition performance tests and delivered four core outcomes: no load shedding occurred during demand emergencies, transaction activity retained 95.7% after node failures, the system recovery success rate reached 100%, and the blockchain networks integrity remained fully intact. All the above indicators collectively prove that this framework can maintain reliable operation, secure market functions, and high-quality service delivery under highly challengng working conditions. Combined with the research findings in Figure 8, and aligned with current industry trends of rising renewable energy penetration and the decentralization and digitalization of energy systems, this frameworks intelligent distributed control architecture can withstand three types of extreme working conditions: it achieves a load satisfaction rate of 94.2% in scenarios of insufficient renewable energy output, maintains stable operation in both the scenario of a 35% sudden demand surge and the scenario of a 15% blockchain node failure rate, and reaches an overall robustness score of 96. This study also verifies that resilience and economic efficiency are not mutually exclusive goals: the intelligent coordination mechanism that improves economic performance under normal conditions can also significantly strengthen system robustness in emergency scenarios. Relying on distributed resources, local decision-making, adaptive learning, and secure blockchain infrastructure, the framework is fully compatible with the operational requirements of future smart city energy systems that run under high uncertainty.

Figure 8. Robustness assessment of the proposed Blockchain-Enabled Multi-Agent Deep Reinforcement Learning and Multi- Objective Optimization (MADRL-MOO) framework under extreme operating conditions.

(a) System performance during a 48-hour severe weather event resulting in a 70% reduction in solar generation and a 40% reduction in wind generation, where the proposed framework maintained 94.2% load satisfaction compared with less than 80% for conventional approaches. (b) Emergency power balancing during a heatwave-induced 35% demand increase, demonstrating coordinated utilization of battery storage, demand response resources, and peer-to-peer energy trading to avoid load shedding. (c) Cyber-physical disturbance analysis showing continued blockchain operation despite failures affecting 15% of network nodes, resulting in only a 4.3% reduction in trading activity while maintaining full network integrity. (d) Recovery performance following extreme disturbances, illustrating rapid restoration of normal operation through decentralized agent coordination. (e) Multi- dimensional robustness comparison across technical, economic, and cybersecurity metrics. (f) Overall robustness performance summary demonstrating the superior resilience of the proposed framework relative to RBEMS, MPC, and GA-EMS approaches. The results confirm that the proposed intelligent smart city energy ecosystem maintains stable operation, high service reliability, and secure market functionality under severe renewable variability, demand emergencies, and cyber-physical disruptions.

To verify the stability of the Blockchain-Enabled Multi-Agent Deep Reinforcement Learning-Multi-Objective Optimization (Blockchain-Enabled MADRL-MOO) framework proposed in this study for smart city energy ecosystems, the core analytical task assigned to Figure 9 is a comprehensive sensitivity analysis conducted on this framework. As a core tool in the system evaluation process, sensitivity analysis can quantify the extent to which fluctuations in input parameters affect system output performance, and accurately pinpoint the core constraint variables of the architecture. It is a standard procedure for the verification of engineering systems. This sensitivity analysis selected 5 parameters for testing: renewable energy prediction accuracy, battery energy storage capacity, electricity market price volatility, blockchain transaction fee structure, and the learning rate of reinforcement learning agents. The experimental control conditions set a variation range of ±20% of the baseline value for only one independent parameter, while all other parameters remained at their nominal operating conditions. Full parameter testing was implemented across 5 dimensions: economic savings, renewable energy utilization rate, grid stability, transaction efficiency, and control accuracy. The final conclusion on overall robustness shows that the proposed framework maintains stable high performance under all test conditions, with no problems such as system instability or control performance degradation, which fully validates the architectures robustness. Sub-item analysis based on the tornado plot in subfigure 9(a) demonstrates that the top three most influential parameters are ranked as follows: A ±20% variation in renewable energy prediction accuracy causes a ±9.7% variation in annual economic savings, which directly impacts core decisions including energy dispatching, battery management, market participation timing, and demand response coordination. A variation in battery energy storage capacity leads to a ±6.4% variation in system performance; as a flexible resource, this capacity absorbs fluctuations in new energy output, supports energy arbitrage, and improves grid stability. A variation in electricity market price volatility results in a ±4.8% variation in economic outcomes, as the magnitude of price fluctuations directly determines the arbitrage space of intelligent trading entities. This study conducts a quantitative sensitivity analysis of core parameters for the smart energy trading framework. First, it tests the impact of two types of non-core parameters on the frameworks performance: the learning rate of the reinforcement learning agent only causes approximately ±3.6% performance fluctuation, affects only the systems adaptation speed, and has limited impact on long-term performance. The blockchain transaction fee structure has the lowest sensitivity, only triggering ±2.9% fluctuation in annual revenue. Based on these findings, it is derived that the economic benefits of smart energy trading far exceed the moderate transaction processing costs. This study uses a tornado diagram to summarize two core influencing parameters, namely renewable energy forecasting quality and battery energy storage capacity, which are the most critical design considerations to maximize the frameworks performance. For the first core parameter, renewable energy power generation forecasting accuracy corresponding to Figure 9(b), this study verifies that it has an approximately linear correlation with annual economic benefits: when the accuracy rises from 60% to 100%, annual revenue increases from about 8.4 million USD to 13.7 million USD; the baseline accuracy of 90% corresponds to an annual revenue of about 12.4 million USD, and a 20% drop in accuracy reduces annual revenue by 9.7%. This confirms that it is the parameter with the largest impact on economic performance. Forecasting errors drag down performance by undermining the effectiveness of battery charge-discharge scheduling, increasing renewable energy curtailment, limiting the agents ability to seize market opportunities, triggering unnecessary power purchases from the grid, and leading to insufficient activation of demand response. Furthermore, the framework operates normally across the full range of forecasting accuracy, highlighting the adaptive capability of the reinforcement learning agent. Based on these findings, this study recommends investing in advanced forecasting technology paired with a smart energy management system to obtain substantial benefits. For the second core parameter, battery energy storage capacity corresponding to Figure 9(c), this study verifies that it has a strong positive correlation with the

renewable energy utilization rate: a 20% reduction in capacity lowers the utilization rate to approximately 86.0%, while a 20% increase in capacity raises it to about 98.3%. This study first verifies the capacity expansion effect of energy storage at the physical layer. Taking the baseline renewable energy utilization rate of 92.4% as a reference, the study confirms that a 20% increase in energy storage capacity drives a 6.4% rise in this utilization rate. The core mechanism underlying this improvement is that large- scale energy storage can absorb excess power output from wind and solar generation, then discharge power during periods of insufficient wind-solar output or peak electriity demand. This reduces renewable energy curtailment losses and improves overall system operation efficiency. Based on these findings, the study concludes that energy storage investment can effectively adapt to urban scenarios marked by fluctuating wind-solar output and dynamic electricity demand, thereby enhancing the grid integration performance of renewable energy. Next, this study conducts two core sets of sensitivity analysis in sequence. First, using Figure 9(d), it tests the sensitivity of the proposed distributed new energy intelligent framework to electricity market price volatility. Test results show that when volatility rises from 0.2 to 1.6, the systems annual transaction revenue increases from roughly 7.8 million US dollars to around 15.2 million US dollars, and the operating cost reduction rate climbs from 7.1% to 17.4%. Higher market volatility improves the systems economic performance without undermining system stability. The reinforcement learning agent embedded in the framework can capture arbitrage opportunities by dynamically adjusting charging and discharging schedules, transaction decisions, and demand response participation levels. Finally, using Figure 9(e), the study tests the impact of the reinforcement learning agents learning rate on the systems convergence behavior and overall performance. After testing four representative learning rates0.0005, 0.001, 0.005, and 0.01the study finds that all configurations eventually converge to a similar cumulative reward level. A high learning rate can speed up convergence in the early training stage, but an excessively high learning rate occasionally triggers oscillation in the mid-stage of training. A low learning rate leads to slower convergence but a more stable training process. The systems long-term performance is not sensitive to the learning rate, which fully verifies the robustness of the proposed framework. The smart city energy system framework proposed in this study was verified through parameter sensitivity and robustness tests. All tested framework configurations delivered nearly identical final performance, which proves that the framework has extremely low sensitivity to ultra-precise hyperparameter tuning, and can achieve stable, robust performance across a wide range of learning configurations. In particular, its low sensitivity to learning rate selection greatly simplifies the actual deployment process, as qualified operational performance can be reached without conducting extensive parameter tuning. The full demonstration will be carried out in three modules: First, the overall robustness and stability assessment. Based on the heatmap analysis of Figure 9(f), the tests cover five core performance categories: economic savings, renewable energy utilization rate, grid stability, transaction efficiency, and control accuracy. Among these, the indicators under grid stability frequency regulation, voltage stability, and power balanceshow extremely low sensitivity to parameter changes. Blockchain operation and transaction efficiency also remain stable under various changes to market conditions, transaction fees, and learning parameters. No unacceptable performance degradation or system collapse occurred across all test scenarios. The underlying control architecture and decentralized design are the core supports for system resilience; only prediction accuracy and storage capacity can trigger obvious performance fluctuations. Second, the analysis of core parameter interactions. Although this study bases its analysis on independent parameter assessments, it still uncovered two linked patterns: improvements in prediction accuracy will amplify the benefits of energy storage expansion, and high market volatility will amplify the value of accurate prediction and flexible energy storage. Based on these findings, this study proposes that three types of core components need to be co-optimized, rather than tuning individual elements in isolation. Last, implications for implementation design. Combined with the aforementioned conclusions, this study puts forward deployment suggestions with clear priorities. The core recommendation is to prioritize investment in advanced renewable energy prediction technology, as this indicator has the strongest impact on economic performance, closely bridging technical research and industrial implementation needs. This study proposes an intelligent optimization framework for smart city energy systems, which has three core advantages: First, it can accurately size battery storage during the system planning stage, effectively raising the level of renewable energy integration; Second, it has low sensitivity to blockchain transaction fees and reinforcement learning hyperparameters, which simplifies implementation processes and reduces operational and maintenance complexity; Third, it delivers strong robustness across all test scenarios, eliminating the need for extensive customized parameter tuning for cross-environment deployment. Combined with the parameter sensitivity test results presented in Figure 9, the quantitative impacts of core parameters are clearly defined: a 20% drop in renewable energy prediction accuracy will lead to an approximately 9.7% decrease in annual cost savings; a 20% increase in battery storage capacity can drive a 6.4% rise in renewable energy utilization; market price volatility has a positive effect on transaction revenue and cost reduction outcomes, while changes to the learning rate have negligible impact on the systems long-term performance. None of the tested parameter fluctuations caused system instability, control failure, or similar issues. The frameworks robustness, adaptability, and resilience fully meet the deployment requirements of large-scale smart city energy ecosystems.

Figure 9. Sensitivity analysis of the proposed Blockchain-Enabled Multi-Agent Deep Reinforcement Learning and Multi- Objective Optimization (MADRL-MOO) framework. (a) Tornado diagram ranking the influence of key system parameters on annual economic performance. (b) Impact of renewable generation forecasting accuracy on annual economic savings, demonstrating that forecasting quality is the most influential parameter. (c) Effect of battery storage capacity on renewable energy utilization, showing improved renewable absorption with larger storage resources. (d) Influence of electricity market price volatility on trading revenue and operating cost reduction. (e) Sensitivity of reinforcement learning convergence behavior to agent learning rate selection. (f) Overall robustness assessment across all parameter variations, demonstrating stable operation and limited performance degradation under ±20% changes in key system parameters. Results indicate that renewable forecasting accuracy and battery storage capacity exert the strongest influence on system performance, while no tested parameter variation resulted in instability or unacceptable degradation, confirming the robustness and resilience of the proposed smart city energy management framework.

Building on the simulation verification results, the intelligent decentralized energy system framework proposed in this study, which integrates blockchain-enabled peer-to-peer (P2P) energy trading, multi-agent deep reinforcement learning (MADRL), and multi-objective optimization (MOO), demonstrates significant application potential in the energy ecosystem of future smart cities. Simulation results have confirmed that this framework can improve three core metrics: renewable energy utilization rate, economic performance, and grid stability. The remainder of this paper will discuss the core dimensions of the frameworks real- world implementation, including implementation challenges, regulatory requirements, technical barriers, and economic considerations, as well as the frameworks practical contributions to smart city development and the global renewable energy transition. Traditional centralized power systems are defined by three core features: centralized power generation, hierarchical control, and utility-led market operations. Against the newindustry trend of continuously rising participation from distributed renewable energy sources, energy storage systems, electric vehicles, and prosumers, these systems have exposed severe gaps in adaptability. The first practical value of this framework is that it achieves local energy balance through blockchain-enabled P2P energy trading, which supports consumers and prosumers to directly trade excess renewable energy. This reduces transmission losses, eases grid congestion, and boosts the utilization rate of renewable energy that would otherwise go to waste, providing smart city planners with a clear design direction to build interconnected energy communities and avoid overreliance on a single centralized power grid. Its second value is that it leverages the complementary output patterns of offshore wind power, which has strong generation output at night, and urban photovoltaic power, which peaks during the day. Through intelligent scheduling and coordination with energy storage systems, the framework increases the overall penetration rate of green electricity, reduces dependence on traditional power sources, and helps grid operators delay costly investments in upgrading transmission and distribution grids. One core finding of this study is the outstanding effectiveness of MADRL in managing highly complex, dynamic energy systems. Compared with traditional rule-based controllers and centralized optimization methods, MADRL can continuously learn optimal operational strategies through interactions with its operating environment, and all benefits of real-world implementation meet three core sets of needs: environmental, economic, and grid security. The decentralized learning architecture

delivers core supporting value for the real-world implementation of future applications in the energy sector of smart cities, and it boasts three irreplaceable core strengths: First, distributed intelligence eliminates the need to retrieve full-scope, system-wide data, which can greatly reduce cross-node communication loads and significantly improve system scalability; second, its autonomous learning mechanism can continuously adapt to multi-dimensional dynamic disturbances including shifts in demand patterns, fluctuations in renewable energy output, and changes in market conditions; third, local decision-making logic reduces reliance on a central control hub, comprehensively strengthening the systems operational resilience. These characteristics perfectly align with the high-complexity scenarios of future cities, where millions of distributed energy resources, smart home appliances, electric vehicles, and microgrids interact simultaneously. Traditional centralized optimization methods are fundamentally incapable of handling scheduling demands of this scale, while distributed AI solutions can expand naturally as the system grows. This core feature underpins the mainstream industry judgment that artificial intelligence will become the core enabling technology for the operation of future autonomous energy systems and smart grids. Turning to the implementation of blockchain technology, its prominent advantages in transparency, security, decentralization, and trustless transactions have been widely recognized across the industry. However, this study identifies three core implementation challenges, each of which is paired with specific application scenarios, the phased progress this study has achieved, future areas for improvement, and preliminary solution outlines. At present, power market rules formulated by most countries around centralized utility architectures cannot fully support decentralized energy trading among prosumers, and existing frameworks also strictly restrict direct energy trading between consumers. To achieve large- scale real-world adoption of such technologies, a series of policy adaptation issues must first be systematically addressed, to lay the groundwork for subsequent full-chain research efforts.

This study focuses on the implementation pathway for the decentralized renewable energy ecosystem. First, it sorts out the core regulatory reform requirements needed for the commercial launch of this ecosystem, and clarifies the mandatory rules for four major subfields: First, a standardized legal framework, market settlement mechanism, and consumer protection requirements must be defined for peer-to-peer (P2P)energy trading; second, fair allocation methods for grid maintenance costs and usage fees must be formulated to safeguard the long-term financial sustainability of the power grid; third, standards for digital renewable energy certificates and carbon accounting methods must be established to give full play to blockchains advantages in automated verification; fourth, regulatory guidelines for data ownership, data sharing, cybersecurity, and privacy protection must be clarified to build public trust. Through original economic analysis, this study findsthat smart energy trading and the coordination of renewable resources have significant potential for cost savings and revenue generation, but their commercial implementation still faces three core barriers: First, the high initial investment threshold: deploying advanced metering infrastructure, blockchain platforms, communication networks, battery energy storage, and smart control technologies requires large capital expenditures. While the payback period observed in this study is attractive, early-stage deployment still requires supporting financing mechanisms and investment incentives; second, insufficient market acceptance: consumers and businesses lack motivation to participate, so clear economic incentives, confidence in system reliability, paired with user-friendly interfaces, transparent pricing, and public science education, are needed to drive participation; third, traditional utility companies view this model as a threat to their existing business, so the implementation solution is to promote collaboration among utilities, regulators, technology suppliers, and prosumer communities to build a win-win market structure. This framework aligns with the development vision of sustainable smart cities, and can support smart cities in integrating digital technologies and renewable energy systems to achieve the core goal of improving energy sustainability. This study proposes a blockchain-enabled peer-to-peer (P2P) renewable energy management framework for smart cities, which delivers five core benefits at its core value level: enhancing urban power resilience through decentralized energy management, supporting emission reductions to speed up decarbonization, increasing public participation in energy markets, improving energy equity via universal access to renewable energy, and boosting infrastructure operation efficiency through intelligent resource allocation. This framework can coordinate offshore wind power, urban photovoltaics, battery energy storage, and user-side participation behaviors, driving cities to transform from passive power consumers into active carriers of regional energy ecosystems. The core contribution of this study is to support the grid integration of large-scale renewable energy. To address the four major challenges facing grid integrationintermittency, uncertainty, grid congestion, and insufficient supply-demand balancing capacitythe study verifies that AI-driven control paired with market-based mechanisms can substantially mitigate the above problems. This approach breaks away from the outdated path that relies only on traditional grid upgrades and centralized balancing services, unlocks three types of distributed flexible resources: energy storage, demand response, and P2P trading, and builds a resilient system adapted to high renewable energy penetration rates, providing a feasible pathway for governments and international organizations worldwide to meet their decarbonization and net-zero targets.

This study also objectively points out that the current simulation environment cannot cover the full real-world complexity of large-scale implementation, so further research is required on five factors: uncertainty inuser behavior, communication failures, regulatory changes, cyberattacks, and long-term market evolution. Six key research directions will be advanced in subsequent work. Integrating multi-agent deep reinforcement learning, multi-objective optimization, and distributed ledger technology, this framework is a feasible, scalable energy solution for future smart cities that can resolve three core challenges: the grid integration of renewable energy, market decentralization, and urban decarbonization. Although the current energy sector still faces three unresolved challenges in technology, the economy, and regulation, empirical evidence from this study shows that intelligent distributed energy systems can significantly advance the development of core sustainable, resilient, carbon-neutral smart cities.

IV. CONCLUSIONS

The Intelligent Smart City Energy Ecosystem framework proposed in this study provides a systematic, cross-technology integrated solution to the core challenges currently facing the smart city energy sector. At its core, this framework integrates three categories of cutting-edge technologies: a blockchain-enabled peer-to-peer (P2P) renewable energy trading mechanism, a multi- agent deep reinforcement learning (MADRL) algorithm, and a multi-objective optimization (MOO) method. The energy management objects covered by this framework include offshore wind power, urban distributed photovoltaics, all types of battery energy storage systems, and distributed prosumer communities. This framework is specifically tailored to address the core pain points in the field that urgently need breakthroughs: the intermittency of renewable energy output, barriers to the participation of multiple stakeholders in decentralized markets, requirements for grid operation stability, insufficient energy trading transparency, and the challenge of autonomous decision-making in complex dynamic environments. To verify the frameworks effectiveness, this study built a large-scale simulation scenario that closely matches the operating conditions of real-world smart cities, and selected four types of traditional schemes as comparison benchmarks: the conventional rule-based energy management system, the centralized model predictive control scheme, the genetic algorithm optimization method, and the standalone blockchain trading platform. The verification results show that the framework proposed in this study achieves significant improvements in five core dimensions: renewable energy utilization rate, P2P trading efficiency, operational cost reduction range, supply-demand balance accuracy, and grid stability. This study has two core dimensions of innovation. First, the developed decentralized MADRL architecture can independently adapt to environmental changes, market dynamics, and fluctuations in consumer behavior, breaking through the scalability limitations of centralized optimization. Second, the integration of blockchain and multi-objective optimization allows decision-making to simultaneously balance four goals: economic performance, renewable energy utilization rate, carbon emission reduction outcomes, and system reliability. This paper proposes a blockchain-enabled smart city energy trading framework, whose core capability is to build a secure, transparent, decentralized peer-to-peer energy trading scenario. While eliminating centralized market intermediaries, the framework safeguards transaction integrity and trust among all participants, and leverages smart contracts to automatically realize real-time settlement, dynamic pricing, and renewable energy certification, thus simplifying market operation procedures. This paper puts forward two core research findings in sequence: First, blockchain-enabled energy markets can significantly improve transaction efficiency and raise prosumer engagement, creating new economic opportunities in future smart city ecosystems. Second, the complementary power generation attributes of offshore wind and urban solar photovoltaics, when combined with intelligent energy storage management and decentralized trading, can boost renewable resource utilization, reduce carbon emissions and dependence on fossil fuels, support cities decarbonization targets, and ultimately advance the collective global pursuit of carbon neutrality, energy security, and urban resilience.

This study still has limitations: it adopts simulation as its core verification method, and actual deployment may face five categories of challenges, namely the reliability of communication infrastructure, cybersecurity threats, regulatory constraints, consumer participation behavior, and blockchain scalability. In addition, the three sets of assumptions this study makes about future market structure, technology costs, and renewable energy penetration may affect long-term system performance. Although comprehensive uncertainty and sensitivity analyses have been carried out, real-world implementation may still bring about additional operational complexities that require further investigation. Finally, this paper proposes five future research directions: launching large-scale field pilots to verify the framework; integrating digital twin technology to achieve functions such as real-time system monitoring; developing privacy-preserving blockchain architectures to strengthen data security and resilience; expanding the framework to cover multi-sector coupling scenarios including hydrogen production; and exploring regional cooperation mechanisms such as cross-city energy trading networks to support large-scale decarbonization. This study confirms that the integration of blockchain, artificial intelligence, and multi-objective optimization technologies can provide a robust pathway to build an autonomous, resilient, and sustainable smart city energy ecosystem. The original framework proposed in this study can

comprehensively manage distributed renewable energy resources, support the operation of decentralized energy markets, and improve grid performance. It will play a key role in advancing the efficient integration of renewable energy, empowering prosumers, and supporting the implementation of next-generation smart cities.

References

1. H. Lund, P. A. Østergaard, D. Connolly, and B. V. Mathiesen, "Smart energy and smart energy systems," Energy, vol. 137, pp. 556-565, Oct. 2017, doi: 10.1016/j.energy.2017.05.123.

2. A. Bibri and J. Krogstie, "Smart sustainable cities of the future: An extensive interdisciplinary literature review," Sustainable Cities and Society, vol. 31, pp. 183-212, May 2017, doi: 10.1016/j.scs.2017.02.016.

3. M. Z. Jacobson, M. A. Delucchi, Z. A. F. Bauer, et al., "100% clean and renewable wind, water, and sunlight all-sector energy roadmaps for 139 countries of the world," Joule, vol. 1, no. 1, pp. 108-121, Sep. 2017, doi: 10.1016/j.joule.2017.07.005.

4. T. Adefarati and R. C. Bansal, "Integration of renewable distributed generators into the distribution system: A review," IET Renewable Power Generation, vol. 10, no. 7, pp. 873-884, Aug. 2016, doi: 10.1049/iet-rpg.2015.0378.

5. A. Ulbig and G. Andersson, "Analyzing operational flexibility of electric power systems," International Journal of Electrical Power & Energy Systems, vol. 72,

   pp. 155-164, Nov. 2015, doi: 10.1016/j.ijepes.2015.02.028.

   IEEE

1. T. Morstyn, A. Teytelboym, and M. D. McCulloch, "Bilateral contract networks for peer-to-peer energy trading," IEEE Transactions on Smart Grid, vol. 10, no. 2, pp. 2026-2035, Mar. 2019, doi: 10.1109/TSG.2017.2786668.

2. J. Kang, R. Yu, X. Huang, et al., "Enabling localized peer-to-peer electricity trading among plug-in hybrid electric vehicles using consortium blockchains,"

   Transactions on Industrial Informatics, vol. 13, no. 6, pp. 3154-3164, Dec. 2017, doi: 10.1109/TII.2017.709784.

3. Y. Du, F. Li, L. Zandi, and F. Xue, "Approximating Nash equilibrium in day-ahead electricity market bidding with multi-agent deep reinforcement learning,"

   Journal of Modern Power Systems and Clean Energy, vol. 9, no. 3, pp. 534-544, May 2021, doi: 10.35833/MPCE.2020.000502.

4. J. Foerster, I. A. Assael, N. de Freitas, and S. Whiteson, "Counterfactual multi-agent policy gradients," in Proc. AAAI Conference on Artificial Intelligence, San Francisco, CA, USA, Feb. 2018, pp. 2974-2982, doi: 10.1609/aaai.v32i1.11794.

5. R. T. Marler and J. S. Arora, "Survey of multi-objective optimization methods for engineering," Structural and Multidisciplinary Optimization, vol. 26, no. 6,

   pp. 369-395, Apr. 2004, doi: 10.1007/s00158-003-0368-6.

6. A. Soroudi, "Multi-objective optimization in power system operation," in Power System Optimization Modeling in GAMS, Cham, Switzerland: Springer, 2017,

   pp. 237-267, doi: 10.1007/978-3-319-62350-4_11.

7. S. Zhao, F. Li, H. Li, R. Lu, S. Ren, and H. Bao, "Smart contracts for energy systems: A review of opportunities and challenges," Applied Energy, vol. 309, Art. no. 118445, Mar. 2022, doi: 10.1016/j.apenergy.2021.118445.

8. C. Jung and D. Schindler, "Global comparison of the goodness-of-fit of wind speed distributions," Energy Conversion and Management, vol. 133, pp. 216-234, Feb. 2017, doi: 10.1016/j.enconman.2016.12.006.

9. H. Yang, W. Zhou, L. Lu, and Z. Fang, "Optimal sizing method for stand-alone hybrid solar-wind system with LPSP technology by using genetic algorithm,"

   Solar Energy, vol. 82, no. 4, pp. 354-367, Apr. 2008, doi: 10.1016/j.solener.2007.09.005.

10. M. Nijhuis, M. Gibescu, and J. F. G. Cobben, "Analysis of reflectance anisotropy of vegetation for the prediction of solar irradiance in urban environments,"

    IEEE Journal of Photovoltaics, vol. 4, no. 1, pp. 447-458, Jan. 2014, doi: 10.1109/JPHOTOV.2013.2294764.

    IEEE Transactions on Evolutionary

1. K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, "A fast and elitist multiobjective genetic algorithm: NSGA-II,"

   Computation, vol. 6, no. 2, pp. 182-197, Apr. 2002, doi: 10.1109/4235.996017.

   Proc. IEEE Power & Energy

1. L. N. An, T. Quoc-Tuan, "Optimal energy management for grid connected microgrid by using dynamic programming method," in

   Society General Meeting, Chicago, IL, USA, Jul. 2017, pp. 1-5, doi: 10.1109/PESGM.2017.8274142.

   IEEE

1. A. Paudel, K. Chaudhari, C. Long, and H. B. Gooi, "Peer-to-peer energy trading in a prosumer-based community microgrid: A game-theoretic model,"

   Transactions on Industrial Electronics, vol. 66, no. 8, pp. 6087-6097, Aug. 2019, doi: 10.1109/TIE.2018.2874578.

2. F. Lezama, J. Soares, P. Hernandez-Leal, M. Kaisers, T. Pinto, and Z. Vale, "Local energy markets: Paving the path toward fully transactive energy systems,"

   IEEE Transactions on Power Systems, vol. 34, no. 5, pp. 4081-4088, Sep. 2019, doi: 10.1109/TPWRS.2018.2833959.

   Renewable and

1. M. Andoni, V. Robu, D. Flynn, et al., "Blockchain technology in the energy sector: A systematic review of challenges and opportunities,"

   Sustainable Energy Reviews, vol. 100, pp. 143-174, Feb. 2019, doi: 10.1016/j.rser.2018.10.014.

2. C. Zhang, J. Wu, Y. Zhou, M. Cheng, and C. Long, "Peer-to-peer energy trading in a microgrid," Applied Energy, vol. 220, pp. 1-12, Jun. 2018, doi: 10.1016/j.apenergy.2018.03.010.

3. L. Mengelkamp, J. Gärttner, K. Rock, S. Kessler, L. Orsini, and C. Weinhardt, "Designing microgrid energy markets: A case study: The Brooklyn Microgrid,"

   Applied Energy, vol. 210, pp. 870-880, Jan. 2018, doi: 10.1016/j.apenergy.2017.06.054.

   IEEE

1. S. Wang, A. F. Taha, J. Wang, K. Kvaternik, and A. Hahn, "Energy crowdsourcing and peer-to-peer energy trading in blockchain-enabled smart grids,"

   Transactions on Systems, Man, and Cybernetics: Systems, vol. 49, no. 8, pp. 1612-1623, Aug. 2019, doi: 10.1109/TSMC.2019.2916565.

2. M. Mihaylov, S. Jurado, N. Avellana, K. Van Moffaert, I. M. de Abril, and A. Nowé, "NRGcoin: Virtual currency for trading of renewable energy in smart

   Technical Report

   Proc. 11th International Conference on the European Energy Market (EEM)

grids," in , Krakow, Poland, May 2014, pp. 1-6, doi: 10.1109/EEM.2014.6861194.

https://bitcoin.org/bitcoin.pdf

1. S. Nakamoto, "Bitcoin: A peer-to-peer electronic cash system," , 2008. [Online]. Available:

   https://ethereum.org/en/whitepaper/

   Technical Report

2. V. Buterin, "Ethereum: A next-generation smart contract and decentralized application platform," , 2014. [Online]. Available:

   IEEE Cloud

1. T. Salah, M. Zemerly, and Y. Salah, "The evolution of distributed ledger technologies: From Bitcoin mining pools to blockchain clouds,"

   Computing, vol. 5, no. 2, pp. 40-50, Mar. 2018, doi: 10.1109/MCC.2018.111951688.

2. S. Ren, D. Zhang, S. Ren, and J. Zhang, "A dynamic economic dispatch model incorporating wind power based on chance-constrained programming,"

   Renewable Energy, vol. 75, pp. 501-509, Mar. 2015, doi: 10.1016/j.renene.2014.10.033.

3. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.

   https://arxiv.org/abs/1312.5602

4. V. Mnih, K. Kavukcuoglu, D. Silver, et al., "Playing Atari with deep reinforcement learning," arXiv preprint arXiv:1312.5602, Dec. 2013. [Online]. Available:

   Applied

1. T. T. Nguyen, V. T. Phan, A. K. Srivastava, et al., "Reinforcement learning-based energy management for an electric vehicle integrated microgrid,"

   Energy, vol. 288, Art. no. 116619, Apr. 2021, doi: 10.1016/j.apenergy.2021.116619.

   IEEE Journal on Selected Areas in

1. Z. Zhao, W. C. Lee, Y. Sharma, and C. P. Goh, "Artificial intelligence aided next-generation networks relying on UAVs,"

   Communications, vol. 37, no. 12, pp. 2884-2902, Dec. 2019, doi: 10.1109/JSAC.2019.2927072.

   Electric

1. M. Shakeri, M. Pasandideh, E. Pouresmaeil, J. C. Vasquez, and J. M. Guerrero, "An overview of demand response programs in electrical markets,"

Power Systems Research, vol. 141, pp. 265-273, Dec. 2016, doi: 10.1016/j.epsr.2016.08.001.