Forecasting Short-Term Electric Load with a Hybrid of ARIMA Model and LSTM Network

Abstract

Smart Meters in the recent years have led to the generation of large consumer data sets which have enabled more energy forecasting algorithms to be designed. Two such algorithms are discussed in this paper with their minor deficiencies and a hybrid approach is proposed. First algorithm being Autoregressive Integrated Moving Average (ARIMA) model which turns out to be futile in determining nonlinear relationships that are involved. Secondly, Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) can not correctly model seasonal variations in energy consumption. This paper blends Seasonal ARIMA (SARIMAX) with LSTM network by integrating their benefits for a improved electric load forecast. The major contribution is the implementation of the combining algorithm to form the hybrid network. The proposed hybrid implementations provides an almost 13.08% decrease in the mean absolute error when compared with the two algorithms. The slight superior performance of the proposed method in the power load forecasting application is highlighted in the results section.