A SYSTEMATIC MACHINE LEARNING APPROACH TO SHORT TERM ELECTRICITY LOAD FORECASTING

Abstract

In the energy sector, for an efficient electricity load management which includes viable utilization and allocation of energy assets, Electricity Load Forecasting plays a critical role. Precise long term and short-term electricity demand forecast is significant as it enables complete utilization of produced electric power, preventing over-production and sometimes wastage of energy and resources.short term load forecasting is necessary, asit is used to maintain and regulate the optimal performance in the daily operation of electrical power systems. With enhancement in technology and automation, and the rise of artificial intelligence it becomes an eminent need to bring about revolutionizing changes for forecasting the future energy needs. In this regard researchers are trying to explore, study, innovate and improving the existing machine learning and deep learning methodologies for the purpose of accurate and efficient electricity load forecasting. Hence through this study, Ensemble based methods and neural network-based methods are explored and the results have been provided for comparative evaluation. This thesis presents a comparative proof of ensemble learning based algorithm Extreme Gradient Boosting Technique (XGBoost) with Deep Recurrent Neural Network (RNN) and Stacked Long Short-Term Memory Network (LSTM) for short term electricity demand forecast on the Dominion Energy Data taken from PJM energy market. The aim of this thesis is to investigate and prove that stacked LSTM performs better as compared to an ensemble machine learning model XGBoost and deep RNN algorithms on PJM energy data, by using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R2 score as evaluation metrics for performance validation. Here in this study, RNN, stacked LSTM model and XGBoost model is compared with a Hypertuned stacked LSTM model. The Hypertuned model which had increased number of hidden nodes from the initial LSTM network tend to give an improved performance, proof of which is shown through this study. This work sheds light on the internal architecture of the models and the different values of hyper-parameters used while training the models to justify the observed day-ahead predictions.