Evaluating Socio-Economic and Geopolitical Events Impact on UK Electricity Demand using Explainable AI

Evaluating Socio-Economic and Geopolitical Events Impact on UK Electricity Demand using Explainable AI

1st Oluwakemisola Christiana Adewole Research and Doctoral College University of Greater Manchester Bolton, United Kingdom oa7res@bolton.ac.uk

ORCID: 0009-0005-2027-9496

2nd Olayinka Anthony Ojo Computing Department University of Greater Manchester Bolton, England a.ojo@greatermanchester.ac.uk ORCID: 0000-0003-3209-549X

3rdJohn Olalere Ogunlola Research and Doctoral College University of Greater Manchester Bolton, United Kingdom jo3res@bolton.ac.uk

ORCID: 0009-0001-1546-1526

4th Opeyemi Victor Omolade MRes Articial Intelligence University of Greater Manchester Manchester, United Kingdom ovo1res@bolton.ac.uk

ORCID: 0007-6034-0778

5th Mariam Omolola Oparinde Research and Doctoral College University of Greater Manchester

Bolton, United Kingdom omo1res@bolton.ac.uk

ORCID: 0009-0007-4167-633X

6th Bisola Kafayat Oyebamiji Research and Doctoral College University of Greater Manchester

Bolton, United Kingdom oyebamijikafayat8@gmail.com ORCID: 0009-0006-1658-1964

AbstractThis paper examines how socio-economic and geopo- litical events affect electricity consumption in the United King- dom from 2009 to 2024 using a multi-source, event-aware dataset. The study integrates electricity demand, climate, economic, COVID-19, holiday, renewable generation, and geopolitical indi- cators, and evaluates Random Forest, XGBoost, and LightGBM models with time-aware validation. LightGBM achieved the best performance, with an RMSE of 1,590 MW and an R2 of 0.920 on the holdout test set, outperforming XGBoost by 137 MW in RMSE. SHAP analysis further revealed that temporal, weather, and event-based features were key demand drivers. The framework improves forecasting accuracy while providing inter- pretable, policy-relevant insights for resilient energy planning.

Keywords: Electricity, Demand, Explainable, Forecasting, machine learning, LightGBM, Random Forest, SHAP.

1. Introduction

   Energy consumption is one of the important indicators of socio-economic development, technological advancement and quality of life of nations [1]. It is related to the size of industrial production, transport infrastructure, housing charac- teristics, and the general character of an economy. Global elec- tricity demand increased by nearly 2.2% in 2023, according to the International Energy Agency [2], driven by continued population growth, digitalisation, and the accelerating electri- cation of industries and mobility. However, energy demand trends over the last decade have been extremely volatile because a series of disruptive events has called into question long-held expectations of how demand will behave. [3] noted that while total electricity demand in the UK fell relatively steadily between 2009 and 2019 as a result of improvements in energy efciency, continued shift away from heavy industry,

   and there were anomalies during the pandemic (2020-2021) and as a result of the geopolitical tension affecting energy markets. Additionally, the continued policy reforms toward the promotion of renewable energy adoption and energy efciency improvement and structural eliciting have further changed the temporal and structural attributes of electricity consumption [4]. These dynamic patterns show that progress in pricing, technologies and behaviours, as well as larger socio-economic shocks and international forces, play a role in shaping the demand for electricity in the UK. Traditional time-series and econometric regression forecasting models have encountered difculties to account for these complex, nonlinear and event- driven variations [5][6]. Although recent models of machine learning have shown better predictive performance by being able to capture hidden dynamics and seasonality [7][8], they are a black box so we lack easy understanding of how certain features affect predictions. [9] points out that this lack of in- terpretability makes it difcult to make decisions about policy, where it is needed not only to predict consumption but also to see what drives variations. This study aims to ll these gaps by incorporating heterogeneous event information into explain- able machine learning models to analyse the pattern of the UK electricity consumption from 2009 to 2024. Also to analyse the effects of socio-economic and geopolitical occurrences on the UK electricity consumption trends and to prepare infer- ential and analysable data-driven models to facilitate proper energy prediction and policy consideration. The objectives include ranking important external events, which affect the UK electricity consumption patterns, to predict consumption in the presence of various possible external conditions through explainable machine learning models, gauge the magnitude

   and direction of effect of each category of events on electricity demand with feature attribution and counter-factual analysis, and offer practical strategic responses to policy-makers and energy operators in the light of the necessity of better energy planning and response to future disruptions. The research is limited to the UK and secondary data is used based on ofcial sources of information.

2. RELATED WORK

   Statistics and econometric models as linear regression, ARIMA, SARIMA, exponential smoothing, and vector au- toregression models have been traditionally used to predict electricity demand [10]. These techniques are appreciated due to their transparency and the capacity to model trends and seasonality, but do not consider the linearity and structural stability. Consequently, they do not do well in volatile environ- ments that are brought about by external shocks like economic crises, pandemics, geopolitical tensions, and sudden changes in policies [11].

   To address these shortcomings, recent research has used machine learning and hybrid modelling. Random Forest, XG- Boost, articial neural networks and long short term memory networks algorithms have been shown to have better predictive power as they are able to be used to model non-linear relation- ships and complex time dynamics in electricity consumption data [12]. Hybrid models incorporating both statistical and machine learning elements enhance the performance even further through the modelling of both linear and non-linear structures [13]. In spite of these benets, the majority of machine learning models act like black boxes and are not very transparent and policy-relevant.

   Issues of interpretability have resulted in the increasing popularity of explainable machine learning (XML). Methods like SHAP and LIME have been suggested to assign the signicance of features and also offer local and global expla- nations regarding the predictions of the models [14]. Previous research indicates that XML improves trust, accountability and decision-making in energy forecasting, especially when it comes to households or buildings [15]. Regardless, the literature reports a signicant gap on the national level. Not many studies combine the structured socio-economic and geopolitical events into forecast models, and even fewer use explainable machine learning models in the national electricity demand in the UK setting [9][16]. Such an inability to forecast events in an interpretable manner restricts the capability of policymakers and energy planners to learn and control demand volatility, which compels the development of explainable and event-sensitive modelling.

   available on UK Ofce for National Statistics [19], COVID-

   19 policy and health data available on Oxford COVID-19 Government Response Tracker [20], calendar and holiday data on WorldPop [21], and geopolitical events that were manuall coded: Brexit milestones, the Ukrainian war and the energy price crisis.

   1. Research Design

      The research is structured as a supervised regression prob- lem in which electricity demand is treated as the target variable and a set of external explanatory variables are used as predictors. Let yt denote the observed electricity demand at time t, and let Xt represent the vector of associated explanatory features, including weather, calendar, economic, renewable generation, COVID-19, and geopolitical variables. The modelling objective is to learn a function f (·) such that:

      yt = f (Xt)+ t (1)

      where t captures unobserved variation and random noise. In practical terms, the study does not aim only to forecast demand values; it also seeks to understand how specic classes of events and conditions shift consumption patterns over time. This makes the research both predictive and explanatory, with a strong emphasis on interpretability. The paper explicitly frames this as a secondary-data investigation using explainable machine learning to study UK electricity consumption over

      20092024.

   2. Data Preprocessing

      Data cleaning was done to remove missing values and in- consistencies, time synchronisation of heterogeneous datasets, and variable normalisation to make them compatible with electricity demand time series. Exploratory data analysis was performed to investigate trends, anomalies and correlation of electricity demand and external events.

   3. Model Development and Evaluation

      Random Forest, XGBoost, and LightGBM types of ensemble-tree-based machine learning models were used be- cause they can capture non-linear associations and use high- dimensional features. The records were divided into training (2009-2021) and testing (2022-2024) groups, cross-validation of the time-series was used. RMSE, MAE, MAPE and R² were used to determine model performance.

   4. TrainTest Split and Validation Strategy

      To preserve temporal ordering, the dataset was split sequen- tially (not randomly). Let

3. Methodology

   D = {(Xt, yt)}T

   , (2)

   t=1

   The methodology is based on predictive accuracy and explainable machine learning that is based on secondary data analysis. The datasets used consist of UK electricity demand and renewable generation data on Kaggle [17], climate vari- ables data on Copernicus ERA5 reanalysis [18], economic data

   where yt is electricity demand and Xt is the feature vector at time t. The training set contains the earliest observations and the test set contains the most recent observations. In this study, 80% of the data (20092021) was used for training, and 20222024 was held out for nal testing.

   Model selection used forward-chaining (expanding-window) time-series cross-validation. For fold k, the training and vali- dation sets are

   1. Temporal Patterns in Electricity Demand

      Fig. 1 illustrates the temporal pattern of UK electricity de- mand from 2009 to 2024, with some of the major geopolitical

      D(k) = {(Xt, yt)}tk

      , (3)

      events such as the Brexit referendum (June 2016), Covid-

      and

      train

      t=1

      19 lockdowns (March 2020) and the Ukraine war (February 2022). The monthly average trend shows the evolution in

      D(k) = {(Xt, yt)}tk +h

      , (4)

      val

      t=tk+1

      demand over a long period and a general declining trend over

      where h is the validation horizon. This strategy respects temporal dependence and prevents look-ahead bias.

      1. Explainability and Interpretability Analysis

         p

         To complement predictive accuracy with interpretability, the study employed SHAP (Shapley Additive exPlanations) and conventional feature-importance measures. SHAP explains an individual prediction by decomposing it into additive feature contributions:

         the last years despite population growth.

         y(x)= 0 + j(x), (5)

         j=1

         where 0 = E[y(X)] is the baseline prediction and j(x) is the contribution of feature j.

         The Shapley value for feature j is dened as

         j(x)=

         SF\{j}

         |S|! (p |S| 1)! f

         p!

         S{j}

         (x) fS

         (x) ,

         (6)

         Fig. 1. UK Electricity Demand Pattern (2009 to 2024)

         where F is the set of p features and fS denotes the model evaluated with only features in subset S.

         N i=1

         L

         Global importance was obtained by aggregating absolute attributions (e.g., 1 N |j(xi)|), while local explanations were inspected for disruption periods. Counterfactual analysis assessed sensitivity by perturbing selected drivers and re- estimating predictions:

         y = y(x ) y(x), (7)

         linking demand changes to specic socio-economic and geopolitical conditions.

      2. Ethical Considerations

      This study uses only secondary, publicly available, aggregate-level datasets and does not involve human subjects, clinical records, or personally identiable information. The electricity-demand dataset, climate reanalysis data, economic statistics, COVID-19 policy indicators, and holiday variables were obtained from public sources and used strictly for aca- demic analysis and forecasting purposes.

4. ANALYSIS

   This section presents a comprehensive analysis of electricity demand forecasting in the UK using machine learning tech- niques. The analysis aims to assess the predictive accuracy of the models, identify the key factors driving electricity demand, and quantify how major external shocks affect consumption patterns. The models were trained using practical, deployable input features and evaluated with a time-aware split, using earlier years for training and the most recent years for testing.

   This indicates seasonal decomposition, revealing a pro- nounced annual cycle in electricity consumption. The observed series was decomposed into trend, seasonal, and residual com- ponents, showing that seasonal effects account for uctuations of approximately ±4,000 MW around the underlying trend. Demand is consistently higher in winter than in summer, primarily due to increased heating requirements.

   The hourly and weekly patterns (Fig. 2) shows strong daily cycles. Average demand showed two peaks: in the morning at 09:00 (31,344 MW) and in the evening at a higher level of consumption at 18:00 (34,395 MW), reecting commercial and residential consumption patterns. The weekly assessment shows signicant differences between weekdays and week- ends, with average demand on weekdays 29,413 megawatts and average demand on weekends 25,748 MW, a reduction of 12.5% owing to reduced commercial and industrial activity.

   Ukraine war and energy crisis (Fig. 3) shows continued suppression of demand. Pre-war average (January 2021 to February 2022) was 26,408 MW decreasing to 23,667 MW during the conict and energy price crisis (February 2022 to December 2023), a 10.4% reduction. This signicant decline was due to both price-driven conservation behaviour and industrial reductions in response to high energy costs, building on the Previous reductions driven by the pandemic.

   Monthly and seasonal trends (Fig. 4) conrms winter- dominated demand proles. December had the highest average demand and August the lowest, indicative of heating and cooling demands between seasons. Year-on-year boxplot com- parisons showed a progressive decrease in demand between 2009-2024, with median demand decreasing despite economic

   Fig. 2. Hourly and Weekly Pattern Analysis.

   growth and population growth, possibly due to improvements in energy efciency and growth in renewable generation.

   1. Climate and Energy Relationships/p>

      Temperature-demand correlation (Fig. 5) shows a non-linear U-shaped correlation. Demand rose at both ends of the tem- perature spectrum, low temperatures spurring the heating and high temperatures spurring cooling. The polynomial regression curve showed this pattern, with medium temperatures (15-20 degrees Celsius) showing the lowest demand levels. Binned temperature analysis supported this relationship with high demand below 5 degrees Celsius and above 25 degrees Celsius.

      Fig. 3. Ukraine War Effect on Demand

      Fig. 4. Monthly and Seasonal Patterns.

      The polynomial regression curve shows this pattern, with medium temperatures (15-20 degrees Celsius) showing the lowest demand levels. Binned temperature analysis supported this relationship with high demand below 5 degrees Celsius and above 25 degrees Celsius.

   2. Economic and Calendar Effects

      GDP sector correlations at monthly aggregation level with electricity demand shows moderate positive relationships. Total GVA was poorly correlated, and the production and manufacturing sectors were more highly correlated in line with their electricity intensive nature. Scatter plots showed linear trends, but there was a large amount of variance suggesting that other factors apart from economic activity affected demand. Population trends (Fig. 7) displays paradoxical trends. While UK population rose from 2009-24, per capita demand for electricity has fallen, indicating greater energy efciency, technological progress and structural changes in the economy (energy intensive industries) counteracting the increased de- mand due to population growth.

   3. Model Performance Results

      This section introduces the performance of three tree-based machine learning models; Random Forest, XGBoost, and

      Fig. 5. Temperature-demand correlation.

      LightGBM, using time series cross-validation, and holdout test data. LightGBM coutperforms the other models across all metrics, achieving RMSE of 1,590 MW, MAE of 1,243 MW, MAPE of 5.33%, and R² of 0.920. XGBoost ranked second with RMSE of 1,727 MW and R² of 0.906, while Random Forest shows RMSE of 1,745 MW and R² of 0.904.

      Table 4 summarises the performance metrics on the holdout test set (20222024).

      TABLE I

      Model Performance on Test Set (20222024)

      Model RMSE MAE R² MAPE Training

      (MW) (MW) (%) Time

      Random

      Fig. 6. Population Trends and Electricity Demand

   4. Prediction Accuracy and Residual Behaviour

      The predicted-versus-actual comparison in Fig. 10 shows that LightGBM predictions cluster closely around the 45- degree reference line, indicating high agreement between predicted and observed demand values. The scatter patterns for the other models are also strong, but LightGBM is visually the most concentrated around the ideal prediction line. The numerical performance results are conrmed by this, and the model is shown to generalise well to unseen data.

      Slight underestimation at the extreme upper range, partic- ularly above 50,000 MW, is also identied across all three models. Very high peak events continue to be difcult to capture perfectly, even with exible ensemble methods. Such underestimation is not unusual in electricity forecasting, be- cause the highest-demand periods are often associated with

      Forest 1,745

      XGBoost 1,727

      1,331 0.904 5.75 563s

      1,372 0.906 5.92 6s

      rare combinations of weather, behaviour, and system stress.

      Overall, the residual behaviour indicates that the proposed

      LightGBM 1,590 1,243 0.920 5.33 22s

      LightGBMs performance proves the efcacy of gradient boosting combined with the leaf-wise growth of the tree for learning complex temporal relationships in electricity demand forecasting.

      modelling framework is stable and suitably accurate for national-level forecasting applications.

   5. Interpretability Insights

   The interpretability analysis is shown to produce broadly consistent feature-importance rankings from tree-based meth-

   Fig. 7. Model Performance Metrics Comparison

   ods and SHAP values, strengthening condence in the model conclusions. Temporal variables are brought forward as the most important predictors, providing conrmation of the dom- inant role of hourly, daily, weekly, and seasonal cycles in electricity demand. Additional explanatory power is provided by climate variables, while longer-term shifts and shock- related deviations are accounted for by economic and event- based features.

   Importantly, short-term lag features were deliberately ex- cluded by the study, even though they would have improved predictive performance in a retrospective setting, because they would not be available in a realistic operational forecasting environment. As a result, practical explainability is given priority over maximum back-tested accuracy. This makes the ndings more policy-relevant because the feature set reects information that decision-makers can plausibly know in ad- vance. As illustrated in Fig. 9, the SHAP results therefore support the central conclusion of the study: UK electricity demand is driven by a combination of predictable temporal structure and less predictable event-based shocks, and both must be represented in future forecasting frameworks.

5. CONCLUSION

   This study set out to examine how socio-economic and geopolitical events inuence electricity consumption in the

   Fig. 8. SHAP Summary Plot

   United Kingdom over the period 2009 – 2024, and to determine whether explainable machine learning can provide both accu- rate forecasts and interpretable insights for energy planning. Using a multi-source dataset that combined electricity demand, renewable generation, climate variables, economic indica- tors, COVID-19 measures, holiday information, and manually coded geopolitical events, the research developed a national- scale framework for analysing demand dynamics under routine and disruptive conditions. The results show that UK electricity demand is shaped by a combination of stable temporal struc- ture and irregular external shocks. The exploratory analysis revealed strong daily, weekly, monthly, and seasonal cycles, with higher demand in winter, clear morning and evening peaks, and lower consumption during weekends. The long- term series also indicated a gradual decline in electricity demand over time, despite population growth, suggesting that improvements in energy efciency, industrial restructuring, and renewable expansion have altered the traditional relation- ship between economic growth and electricity use . In addition, major events such as the COVID-19 pandemic, Brexit-related uncertainty, and the RussiaUkraine energy crisis produced visible changes in consumption behaviour, conrming that national electricity demand is highly responsive to external disruption. Tree-based ensemble learning methods were shown to be well suited to electricity-demand forecasting in a com- plex, nonlinear environment. Random Forest, XGBoost, and LightGBM were assessed using time-aware validation, and

   strong performance on unseen data was achieved by all three models. The best overall predictive accuracy was delivered by LightGBM, along with the lowest error and the highest explanatory power among the models considered. The inter- actions among climate, calendar, renewable, economic, and event-based features in the UK electricity system are therefore suggested to be captured especially effectively by gradient- boosting approaches. More importantly, transparent evidence of the variables shaping demand changes was provided by the SHAP-based interpretability analysis, moving the model beyond pure prediction and toward explanation relevant to policy.

   he main contribution of this research is threefold. First, an event-sensitive forecasting framework is introduced, and heterogeneous datasets are brought together to examine elec- tricity demand at the national scale. Second, evidence is provided that both the magnitude and direction of demand drivers can be represented by explainable machine learning, with transparency for decision-makers improved as a result. Third, support for energy policymakers and grid operators is generated through evidence that can inform planning for shocks, demand volatility, and long-term system transitions. Taken together, the ndings show that UK electricity demand should not be viewed as a smooth time series alone, but rather as a system inuenced by climate, behaviour, structural economic change, and geopolitical instability.

6. LIMITATION AND FUTURE WORK

Despite its contributions, several limitations are acknowl- edged in the study. Reliance is placed on secondary data from multiple public sources, creating possible differences in tem- poral granularity, measurement standards, and documentation quality. Although careful preprocessing and time alignment were carried out, some residual inconsistency may still remain. In addition, a degree of subjectivity is created by the manual coding of geopolitical events. The model also omits some potentially inuential variables, such as electricity prices, grid constraints, fuel costs, and real-time behavioural indicators, which could further enhance forecast performance if included in future work.

References

1. M. Baltruszewicz, J. K. Steinberger, J. Paavola, D. Ivanova, L. I. Brand-Correa, and A. Owen, Social outcomes of energy use in the United Kingdom: Household energy footprints and their links to well-being, Ecol. Econ., vol. 205, p. 107686, Mar. 2023, doi: 10.1016/j.ecolecon.2022.107686.

2. International Energy Agency, Electricity 2024: Analysis and Fore- cast to 2026. Paris, France: IEA, 2024. [Online]. Available: https://www.iea.org/reports/electricity-2024

3. P. Jiang, Y. V. Fan, and J. J. Klemes, Impacts of COVID-19 on energy demand and consumption: Challenges, lessons and emerging opportunities, Appl. Energy, vol. 285, p. 116441, Feb. 2021, doi: 10.1016/j.apenergy.2021.116441.

4. Q. Hassan et al., The renewable energy role in the global energy transformations, Renew. Energy Focus, vol. 48, p. 100545, Mar. 2024, doi: 10.1016/j.ref.2024.100545.

5. N. U. Moroff, E. Kurt, and J. Kamphues, Machine learning and statistics: A study for assessing innovative demand forecasting models, Procedia Comput. Sci., vol. 180, pp. 4049, 2021, doi: 10.1016/j.procs.2021.01.127.

6. R. Roma´n-Collado and M. Economidou, The role of energy efciency in assessing the progress towards the EU energy efciency targets of 2020: Evidence from the European productive sectors, Energy Policy, vol. 156, p. 112441, Sep. 2021, doi: 10.1016/j.enpol.2021.112441.

7. R. Banik and A. Biswas, Enhanced renewable power and load forecast- ing using RF-XGBoost stacked ensemble, Electr. Eng., vol. 106, no. 4,

   pp. 49474967, Aug. 2024, doi: 10.1007/s00202-024-02273-3.

8. M. M. Go¨kc¸e and E. Duman, Performance comparison of simple regression, random forest and XGBoost algorithms for forecasting electricity demand, in Proc. 3rd Int. Inform. Softw. Eng. Conf. (IISEC), 2022, pp. 16, doi: 10.1109/IISEC56263.2022.9998213.

9. T. M. Ghazal, Energy demand forecasting using fused machine learning approaches, Intell. Autom. Soft Comput., vol. 31, no. 1, pp. 539553, 2022, doi: 10.32604/iasc.2022.019664.

10. P. Perera and A. Ferno, Impact of energy prices and macroe- conomic variables on GDP prediction UK: Machine learning ap- proach, J. Bus. Manag. Stud., vol. 6, no. 5, p. 113, 2024, doi: 10.32996/jbms.2024.6.5.13.

11. M. F. I. Sumon et al., Environmental and socio-economic impact assessment of renewable energy using machine learning models, J. Econ. Finance Account. Stud., vol. 6, no. 5, p. 112, 2024.

12. S. Pokharel and P. Ghimire, Data-driven ML models for accurate prediction of energy consumption in a low-energy house: A comparative study of XGBoost, random forest, decision tree, and support vector machine, J. Innov. Eng. Educ., vol. 6, no. 1, pp. 1220, 2023, doi: 10.3126/jiee.v6i1.54965.

13. G. P. Zhang, Time series forecasting using a hybrid ARIMA and neural network model, Neurocomputing, vol. 50, pp. 159175, 2003, doi: 10.1016/S0925-2312(01)00702-0.

14. F. M. Talaat, A. E. Kabeel, and W. M. Shaban, Towards sustainable energy management: Leveraging explainable articial intelligence for transparent and efcient decision-making, Sustain. Energy Technol. Assess., vol. 78, p. 104348, 2025, doi: 10.1016/j.seta.2025.104348.

15. S. M. Malakouti, Explainable machine learning for mapping energy poverty: Identifying low-income, high-cost households across Europe, 2025.

16. D. A. Bala and M. Shuaibu, Forecasting United Kingdoms en- ergy consumption using machine learning and hybrid approaches, Energy Environ., vol. 35, no. 3, pp. 14931531, May 2024, doi: 10.1177/0958305X231206972.

17. J. Vidal Rodriguez, UK Electricity Consumption Dataset (20092024). Kaggle, 2022. [Online]. Available: https://www.kaggle.com/datasets/jvidalrod/uk-electricity-consumption- 2009-2024

18. Copernicus Climate Change Service (C3S), ERA5 Reanalysis Dataset. European Centre for Medium-Range Weather Forecasts (ECMWF). [Online]. Available: https://cds.climate.copernicus.eu/

19. Ofce for National Statistics (ONS), Gross Domestic Product and Population Statistics. UK Government. [Online]. Available: https://www.ons.gov.uk/

20. T. Hale et al., Oxford COVID-19 Government Response Tracker. Blavatnik School of Government, University of Oxford, 2020. [Online].

    Available: https://www.bsg.ox.ac.uk/covidtracker

21. WorldPop, Global Holiday Dataset. University of Southampton. [On- line]. Available: https://www.worldpop.org/