Short-term electricity consumption forecasting with deep learning

Abstract

Global electricity demand is surging due to population growth, industrialization, and technological advancements. While renewable energy sources are expanding, fossil fuels still remain the primary source of electricity generation, posing challenges due to resource limitations and environmental concerns. To address these challenges and optimize energy use, accurate prediction of electricity consumption is crucial. Therefore, this work introduces novel short-term (24-hour) electricity consumption forecasting models based on customized long short-term memory (LSTM) networks, convolutional neural networks (CNNs), and their ensemble. The models utilize time-series electricity consumption data and meteorological features, including temperature, relative humidity, and wind speed. Trained and evaluated on two geographically distinct datasets spanning 2.5 years, our models utilizing appropriate feature sets surpass the recent studies and achieve significantly high forecasting performance with normalized root mean square error (N-RMSE) reaching 0.16, normalized mean absolute error (N-MAE) reaching 0.13, and mean absolute percentage error reaching 4%. The inclusion of meteorological features contributed notably to prediction performance, demonstrating the benefit of integrating external features in electricity forecasting models. The results highlight the effectiveness of customized deep learning architectures in capturing complex temporal and contextual dependencies within electricity consumption data. Also, these findings offer valuable insights for future research and practical applications in energy management and grid optimization.