Comparative Analysis of Short-Term Load Forecasting Methods

Abstract

One of the primary tasks of electric utilities is to accurately predict the load demand requirements of consumers, especially for short term prediction. In view of this, different methods have been proposed for load prediction. In this paper, three methods (i.e. Multiple Linear Regression (MLR), Seasonal Auto Regressive Integrated Moving Average with Exogenous Variables (SARIMAX) and Long Short Term Memory (LSTM)) are compared to forecast load consumption in a typical Nigerian University. The main objective is to determine which of the techniques best model the load consumption pattern of the University accurately. Load forecast was made for weekdays (Monday-Friday) and weekends (Saturday and Sunday). The result showed that the LSTM technique is the best performing model achieving the least errors. The technique returns the mean absolute error (MAE) that varies between 0.029-0.093, mean square error (MSE) ranging between 0.0014-0.014 and root mean square error that has values between 0.037-0.12.