The electricity consumption forecast: Adopting a hybrid approach by deep learning and ARIMAX-GARCH models

Modeling and forecasting electricity consumption (EC) help industry managers make better strategic decisions. In this study, a hybrid approach for predicting EC is proposed which first EC is decomposed into approximate and detail parts based on wavelet transfer (WT). Next, the modeling of wavelet components is accomplished based on the adaptive WT (AWT)-long short-term memory (LSTM) and autoregressive integrated moving average with explanatory variable (ARIMAX)-generalized autoregressive conditional heteroscedasticity (GARCH) type models. The key concept of AWT-LSTM is adding an optimal set of adaptive layers through which AWT-LSTM can automatically concentrate on various frequency components based on the dynamic evolution of the input sequence revealing the multi-frequency pattern of EC. Then, the components are predicted separately and eventually, the results are combined nonlinearly through random forest (RF) algorithm. Besides the lag calendar setting, the affecting exogenous variables are applied on the electricity consumption in the structure of hybrid models which include humidity, water precipitation, solar radiation, effective temperature, wind direction, and wind speed. To measure the predictive power, the proposed models were tested using a 24 h ahead forecasting horizon using a sample rate of 30 min in two different seasons (winter and summer). The empirical results indicate the hybrid model of AWT-LSTM-ARIMAX-exponential (EGARCH) in mean-RF with root mean square error (RMSE) = 1.536, mean absolute percentage error (MAPE) = 1.276, normalized mean absolute error (NMAE) = 0.0116, normalized RMSE (NRMSE) = 0.0143, relative RMSE (RRMSE) = 0.1906, Theil's inequality coefficient (TIC) = 0.01025, sum of squared errors (SSE) = 113.2 values in summer test and RMSE = 1.249, MAPE = 0.869, NMAE = 0.0078, NRMSE = 0.0092, RRMSE = 0.1223, TIC = 0.00533, SSE = 74.9 values in winter test outperforms the benchmark models (WT-feedforward neural network (FFNN)-RF, WT-LSTM-RF, WT-exponential smoothing (ETS)-RF, and WT-seasonal ARIMA with explanatory variable (SARIMAX)-GARCH types models) and alternative filtering methods (Curvelet, Ridgelet, and Contourlet). The results suggest that employing the WT for the decomposition of data and RF for the nonlinear aggregation of sub-signals (predicted values of detail and approximate parts) can significantly improve the prediction power of predictive models. (C) 2022 Published by Elsevier Ltd.