A novel hybrid deep correction approach for electrical load demand prediction

Abstract

Based on technical reports, the high demand of proper tool for load forecasting (LF) and precise planning in recent combative and challenging markets of electrical energy is highly uprising. Therefore, this paper intends to suggest a creative hybrid deep estimation model for short term LF (STLF) using Generative Adversarial Network (GAN), Auto-Regressive Integrated Moving Average (ARIMA) and wavelet package. To get the stationary behavior, the time series in non-stationary behavior case would be differenced in the required number of times. The appropriate order for the model of ARIMA is found using Akaike Information Criterion (AIC). When the linear part of the electrical demand time series is captured by ARIMA, the remaining nonlinear part would be hard to model. The discrete wavelet transform would be utilized to decompose the residual nonlinear component into its sub-frequencies. To estimate the future nonlinear samples, several GAN models are then applied to approximation and detail components of residual signal. Finally, the results of GAN and ARIMA models would be added together to construct the final signal. The observed experimental results indicate the proper improvement of the proposed accurate LF model.