Deep learning modeling in electricity load forecasting: Improved accuracy by combining DWT and LSTM

Abstract

Forecasting electricity load plays a vital role in the planning and management of sustainable power systems, considering the multifaceted impacts of social, economic, technical, environmental, and cultural factors on electricity consumption. Addressing this complexity requires the development of robust models capable of handling high levels of nonlinearity. In this study, we used four machine learning-based methods for forecasting short to long-term electricity load. Electricity load Data were collected from the Iranian Grid Management Company (IGMC) online electricity data and German electricity market data platform. Environmental factors (ambient temperature, cloud cover, solar radiation, precipitation), social events (vacations, festivals), and time series features (Hour Lag, Day Lag, Week Lag, Year Lag) were considered as input variables. The methods include Long Short-Term Memory (LSTM), a combination of LSTM and Discrete Wavelet Transformation (DWT-LSTM), Nonlinear Auto-Regressive with eXogenous inputs (NARX), and Support Vector Machine (SVM) regressor. We apply these methods to forecast electricity load under normal conditions and during social events in both Iran and Germany, evaluating their performance using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Our results demonstrate that the DWT-LSTM method achieves the highest accuracy, with MAPE ranging from 0.59% to 4.2% for Iran and 0.29% to 3.02% for Germany, across hour-ahead to year-ahead forecasts. Moreover, during special events and festivals, DWT-LSTM exhibits precise forecasting capabilities, with MAPE ranging from 0.55% to 3.07% for Iran and 0.33% to 6.01% for Germany, spanning hour-ahead to week-ahead predictions. Comparative analysis of the implemented methods confirms the superior accuracy of DWT-LSTM, followed by LSTM, NARX, and SVM methods, respectively. Our proposed forecasting approach demonstrates high performance in anticipating electricity load under both standard conditions and during significant social events in diverse geographical contexts.