Two-stage transformer encoder and attention-guided GRU-based residual correction for short-term load forecasting

Accurate short-term load forecasting (STLF) plays a vital role in reliable power system planning and operations. The increasing electricity demand is driven by electrification, digitalization, population, and economic growth, which necessitates the need for accurate short-term load forecasting to ensure the real-time balance between supply and demand. Traditional forecasting approaches struggle to capture long-range temporal dependencies, and correcting the forecasting errors remains challenging due to increasing demand and dynamic load behavior. To address this limitation, this study proposes a two-stage forecasting architecture (transformer encoder and attention-guided GRU-based residual corrector [TE-AG-GRU-RC]) for short-term load forecasting. In the first stage, a transformer encoder (TE) processes historical load and exogenous features to generate the base forecast and capture long-range temporal dependencies from historical load sequences. In the second stage, an attention-guided GRU residual correction mechanism (AG-GRU-RC) captures short-term dynamics and refines the initial prediction with the temporal information derived from the transformer encoder. The final forecast is attained by combining the predicted residual with the base forecast, enabling the first stage to model dominant load patterns, while the second stage corrects the remaining errors. The proposed framework is evaluated using real-world multivariate hourly time series datasets from two geographically distinct regions. The forecasting performance of the model is evaluated against several baseline models, such as recurrent neural network (RNN), long short-term memory (LSTM), Gated Recurrent Unit (GRU), temporal fusion transformer (TFT), and ResidualLSTMPlus, to demonstrate its effectiveness. The proposed model attains mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R ²) values of 0.0297, 0.0014, 0.0376, 3.58, and 0.879, respectively, on the Tamil Nadu test data, and 0.0334, 0.0023, 0.0482, 4.99, and 0.790, respectively, on the Panama test data. These findings demonstrate the generalizability of the proposed model across both datasets with low prediction error and improved forecasting accuracy.