A granular deep learning approach for predicting energy consumption

Abstract

This paper proposes a granular deep learning approach consisting of maximal overlap discrete wavelet transformation (MODWT) and long short-term memory (LSTM) network for predicting the energy consumption of different sectors at macro levels. Input features are first evaluated using Boruta algorithm-based feature selection model. MODWT is then used to decompose the energy consumption time series to alienate the linear and nonlinear components. The LSTM network, a deep learning tool, is used to make predictions on individual sub-series at a granular level. The final prediction is obtained by aggregating the forecasts obtained on decomposed components. Statistical analyses rationalize the efficacy and superiority of the proposed hybrid framework over six other well-known prediction algorithms. Monthly data for residential, commercial, industrial and transportation sectors of the USA have been taken for analyses. It is observed that energy consumption in commercial and transportation sectors are easier to predict than residential and industrial sectors.