An Application of Deep Learning for Short-Term Electricity Load Forecasting
TL;DR
The results show high accuracy with LSTM without temperature providing the best average performance; however, an analysis per regime shows that the errors are concentrated in the ramping conditions where GRU and Transformer–LSTM are more robust, pointing to the importance of regime-aware evaluation.
Abstract
Short-term electricity load forecasting is a challenging time-series problem due to nonlinear demand dynamics, strong temporal dependence, and recurring calendar effects. This paper evaluates deep learning models using open operational data for Albania (2024–2025). Recurrent architectures based on Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) are used as baseline models, while a hybrid Transformer–LSTM is assessed to examine the contribution of attention mechanisms. Calendar effects are incorporated using trainable embeddings, and temperature is analyzed through a controlled ablation study based on ERA5-Land data. All models were trained and evaluated under a unified framework using standard regression metrics. The results show high accuracy (MAPE < 2%, R2 > 0.99) with LSTM without temperature providing the best average performance; however, an analysis per regime shows that the errors are concentrated in the ramping conditions where GRU and Transformer–LSTM are more robust, pointing to the importance of regime-aware evaluation.