 Open Access
 Authors : Ayanangshu Das , Anirban Sengupta
 Paper ID : IJERTV8IS110152
 Volume & Issue : Volume 08, Issue 11 (November 2019)
 Published (First Online): 21112019
 ISSN (Online) : 22780181
 Publisher Name : IJERT
 License: This work is licensed under a Creative Commons Attribution 4.0 International License
Forecasting Electrical Energy Consumption using Artificial Neural Networks
Ayanangshu Das1
1Data Scientist
Cognizant Technologies Solutions United States
Anirban Sengupta2
2Asst. Professor
Sikkim Manipal Institute of Technology India
Abstract Knowing how much the demandfor energy would be for the coming months would be very useful for electricity producing companies. Electricity cant be stored and energy companies face a challenge where the demand is always more and the supply is less. Hence the companies look for factors that affect electricity consumption and accurately forecast its usage. In this paper Neural Network methodology is proposed as an effective methodology for load forecasting
Keywords: Neural Network, Random Forest,LSTM

INTRODUCTION
In the contemporary energy sector, the prediction of energy consumption is done by traditional forecasting methods
such as machine learning models. This paper aims at establishing the fact that Artificial Neural Networks play a better role in making more accurate predictions provided that there is no restriction on the time to make the predictions and resources in terms of hardware to be used
The electrical energy consumption depends on the following factors/ loads

Social Load
The consumption varies during the holidays, the weekdays and the weekends
Fig 1: Variation of social load with time

Weather Dependent Load
Factors of weather are temperature, global radiation, wind, cloud etc. The load consumption increases due to high
temperature in summer or may increase due to more heater consumption during the winters
Rain
Winter
Summer
Rain
Winter
Summer
200
180
160
140
120
100
80
60
40
20
0
200
180
160
140
120
100
80
60
40
20
0
1
2
3
4
5
6
7
8 9
Month
1
2
3
4
5
6
7
8 9
Month
Figure 2Variation of Weather Dependent Load

Stochastic Load
These types of loads are the ones that couldnt be accounted for e.g. a sudden storm has increased the load for a certain period of time. Now it would be necessary to understand that whether to take the load into consideration.
A hurricane might be a one of a phenomena in some place whereas it might be a regular phenomenon in some other place. As such a decision is necessary to be taken whether or not to account for these factors
Fig 3: Variation of stochastic load


DATA USED
The data is fed into the machine learning model and was trained to identify the pattern so that the model is able to predict the usage amount for new databy date. The target variable is USAGE and the predictors areDate, Time,
Global_active_power,Global_reactive_power, Voltage and Global_intensity.
If this energy consumption is analyzed over the days of the week, it can be seen that there is a difference in the pattern
of energy consumption on a weekday vs a weekend/holidays.
An additional column called IS_HOLIDAY is introduced to understand if the date is a holiday or not.
Now the usage values aggregated over one month, two months and so on till five months are introduced in the dataset
The consumption of energy over time is analyzed as follows
Fig 4: Consumption of energy over time

RANDOM FORESTSMETHODOLOGY
A random forest is an ensemble of decision trees. The final value of random forest is obtained by taking the average of each of the values of the decision tree.
The Random Forest algorithm adds extra randomness when growing trees. Instead of searching for the very best feature when splitting a node it searches for the best feature among
a random subset of features. As a resultit creates a greater tree diversity, which introduces higher bias and a lower variance, generally yielding an overall better model.
The features derived as explained above are used for making the predictions using the random forest algorithm.
With Random Forest Regressor, the following actual against predicted result is obtained
Fig 5: Result from Random Forest Prediction
It can be seen that the model predicted the values accurately to a certain extent. The grey line is the prediction value on new data. The validation score on the validation data is
RMSE= 0.18412616700088405

SOLVING WITH NEURAL NETWORK
APPROACH
Neural networks is another approach and has been widely used for time series forecasting. These are feedforward
networks which employ a sliding window over the input sequence. Common examples of this approach are stock market predictions and predicting telecommunication load [1, 2]. The neural network based prediction can be considered as a model in which function f is assumed to be a nonlinear combination of a fixed number of previously input values [3].
The neural network is constructed as a three layer architecture as depicted in the following diagram
Fig 6: Neural Network Model
Loss function: Mean Squared Error (MSE) Activation function: ReLU for hidden layers and linear for the output layer Number of epochs=1000
Batch size (for batch gradient descent) =250 Optimizer=adam
On training the network on the train data, the following validation error is obtained
Fig 7: Variation of train data
It isseen that the training loss decreased over time
The predicted data
The below graph shows the result on how the model predicted the train data and the grey line represents the prediction on new unseen data
The accuracy of the model is RMSE= 0.02003
Figure 8Result from Neural Network
Artificial Neural Networks: LSTM
Ho and Xie performed a comparative study between ARIMA and RNN showing that RNN has performed better than ARIMA[4] in modelling time series prediction. Connor and Martin also discussed the effectiveness of Neural Networks in time series forecasting [6]. The LongShortTermMemory (LSTM) cell proposed in 1997 by Sepp Hochreiter and JÃ¼rgen Schmidhuber [5] proved to be more effective than traditional recurrent neural networks in time series predictions. LSTM cell can be considered as a black box. It can be used very much like a basic cell, except it will perform much better; training will converge faster and it will detect longterm dependencies in the data. LSTM is a kind of Recurrent Neural Networks which is depicted by the following architecture:
Figure 9LSTM Architecture
In a Recurrent Neural Network, the output of the previous layer is fed into itself and also to the next layer. This type of architecture is very effective in timeseries forecasting. The equations for a RNN architecture is as follows:

Y (t)is anmÃ—nneuronsmatrix containing thelayers outputs at time step t for each instance in the minibatch (m is the number of instances in the minibatch and nneurons is the number of neurons).

X (t)is anmÃ—ninputsmatrix containing theinputs for all instances (ninputs is the number of input features).

Wxis aninputsÃ—nneuronsmatrix containing theconnection weights for the inputs of the current time step.

Wyis anneuronsÃ—nneuronsmatrix containingthe connection weights for the outputs of the previous time step.

bis a vector of sizenneuronscontaining eachneurons bias term.

The weight matices Wx and Wy are often concatenated vertically into a single weight matrixW of shape (ninputs + nneurons)
Ã— nneurons

The notation [X(t)Y(t1)] represents the horizontal concatenation of the matrices X(t) and Y(t1)
The basic LSTM architecture is depicted as follows:
Figure 10LSTM Gates
The LSTM computations are as follows:

Wxi,Wxf,Wxo,Wxgare the weightmatrices of each of the four layers for their connection to the input vector x(t).

Whi,Whf, Who, andWhgare the weightmatrices of each of the four layers for their connection to the previous shortterm state h (t1).

bi, bf, bo,andbgare the bias terms for eachof the four layers.
The training and validation losses using LSTM came down to as follows:
Figure 11Train and Test Loss
The performance of the model is shown as below. The grey line is an approximation of how the model predicted the value in the training data against the actual value in the green line.
MSE = 0.1586783446921463
CONCLUSION
Figure 2Result from LSTM
It can be concluded that the Artificial Neural Networks did a better job in terms of predicting the actual values than the random forest. If the emphasis is on predicting accurate value, the NN models to be used. However, if the data volume is huge and there is a constraint on how many servers can be used and how quickly the predictions can be made,a simpler machine learning algorithm such as lasso regression or Random Forest can be used.
REFERENCES

Bengio, S., Fessant, F., and Collobert, D.: A connectionist system for mediumterm horizon time series prediction, in: Proc. of Internat. Workshop Application Neural Networks to Telecoms, 1995, pp. 308315

Edwards, T., Tansley, D. S. W., Davey, N., and Frank, R. J.: Traffic trends analysis using neural networks, in: Proc. of the Internat. Workshop on Applications of Neural Networks to Telecommunications, Vol. 3, 1997, pp. 157164

Dorffner, G.: Neural networks for time series processing, Neural Network World 4/96 (1996), pp. 447468

Ho, Xie, Goh, April 11, Singapore, A comparative study of neural network and BoxJenkins ARIMA modeling in time series prediction, V42, pp. 371375

SeppHochreiter and JÃ¼rgen chmidhuber,Nov, 1997, MIT, Neural Computation V9, pp. 17351780

Connor, Martin, Atlas, Mar 94, NJ, Recurrent neural networks and robust time Seriesprediction,pp.240254