A Hybrid Deep Learning Approach for Multi-Output Short-Term Electricity Demand Forecasting

Abstract

This study proposes hybrid deep learning architectures that integrate convolutional and recurrent layers for short-term electricity demand forecasting. A multivariate half-hourly dataset from Great Britain's National Grid Electricity System Operator (ESO), covering January 2009 to early 2024 (279,264 records), was used for model development. Features include national demand (ND), transmission system demand (TSD), embedded wind and solar generation, interconnector flows, and calendar indicators. Models were evaluated using normalized root mean squared error (nRMSE), normalized mean absolute error (nMAE), and symmetric mean absolute percentage error (SMAPE). Across averaged test metrics, the standalone LSTM achieved the lowest errors (Loss 8.8 × 10−4, MSE 0.0018, and MAE 0.0320), while the hybrid CNN + LSTM + DNN and CNN + GRU + DNN attained comparable accuracy and demonstrated greater robustness during peak-load and holiday intervals. Statistical testing indicated that CNN + GRU + DNN significantly outperformed GRU (p = 0.035), but no significant difference was observed when compared with LSTM. These results highlight that while LSTM provides the most accurate overall performance, hybrid architectures offer enhanced stability under volatile demand conditions, ensuring a balanced trade-off between predictive accuracy and operational reliability.