Development and Comparison of Two Computational Intelligence Algorithms for Electrical Load Forecasts with Multiple Time Scales

Abstract

Electricity load forecasting provides the critical information required for power institutions and authorities to develop rational, effective, and economic dispatch plans. The load forecasting at the regional power system is important for optimal management and accommodating local renewable energy sources, which is a challenging task as the demand variations are more sensitive to local weather changes (such as temperature, humidity, precipitation, and wind speed) and consumers' activities and behaviours. The paper aims to develop a new prediction method using intelligent computational algorithms. Long Short-Term Memory (LSTM), a deep recurrent neural network, explores the long-term dependency of network memory sequence data to identify intrinsic variations in both horizontals (time series) and vertical (network depth) dimensions over a longer historical period. Support Vector Machine (SVM) is a typical learning method that has been successfully implemented to solve nonlinear regression and time series problems. This paper studies the two methods and adapts the two methods to become suitable algorithms for load prediction. The paper presents the algorithms, their applications and prediction results. The prediction performance is compared for using LSTM and SVM at ultra-short, short-term, medium-term, and long-term forecasting. The results show that LSTM has higher prediction accuracy than SVM in both ultra-short and short-term forecasts, but SVM is more capable of medium-term and long-term forecasting. Finally, the epoch time for LSTM and SVM is also calculated and compared.