Forecasting Techniques for Estimation of Per Capita Energy Consumption in the Electric Grid

Abstract

This study presents a thorough analysis of per capita load forecasting techniques used to predict electrical energy consumption on a per capita basis. Further, examines the historical time series data on per capita energy consumption to develop models that predict future energy demand with considerable accuracy. Statistical forecasting techniques as well as the Machine Learning techniques were used to analyze the time series data of per capita electricity consumption and subsequently predict the future trends. The dataset used for the forecasting of per capita electricity consumption is obtained from the Website of Central Electricity Authority, Ministry of Power, Government of India. Statistical time series analysis techniques used in this research are Auto-Regressive Integrated Moving Average and HoltWinters Smoothing also known as Triple Exponential Smoothing. Machine Learning model that was used to examine the per capita electricity consumption is Long-Short Term Memory based Recurrent Neural Network. An additive regression model known as Facebook Prophet is implemented for forecasting the per capita load demand. Finally, the performance of each of these techniques is evaluated by factor MAPE i.e. mean absolute percentage error. Thus, analyses the performance of each of the four techniques on the per capita consumption data for various states. Based on the performance of the implemented models, two of the best models were utilized for the implementation regional and national energy forecasting. In view of the accuracy of the forecasts, the Facebook Prophet and Holt-Winters Smoothing techniques are quite suitable for estimation of per captia energy consumption.