Hybrid CNN–BiGRU model with attention mechanism for enhanced short-term load forecasting

Reliable short-term load forecasting has emerged as one of the most urgent challenges in modern power systems, where stochastic weather variations and the nonlinear nature of high-dimensional data necessitate advanced predictive methodologies. Overcoming this challenge is essential to improving grid reliability, optimize dispatch plans, and increasing the integration of renewable energy sources. This study proposes a hybrid deep learning model integrating convolutional neural networks (CNN) and bidirectional gated recurrent units (Bi-GRU) with an attention mechanism to improve short-term load forecasting. The proposed framework integrates stacked Conv1D layers with L2 regularization for robust local feature extraction, a deeper GRU encoder with 20 units to capture long-range temporal dependencies, and a post-concatenation soft attention mechanism to emphasize the most informative temporal patterns adaptively. The model was evaluated using the publicly available AEP and ISONE datasets, and a comprehensive comparison was conducted against existing benchmark models. In single-step forecasting, the proposed model achieved outstanding precision, with RMSE values of 135.94 (AEP) and 134.26 (ISONE), MAE values of 101.96 (AEP) and 70.96 (ISONE), and remarkably low MAPE scores of 0.68 (AEP) and 0.51 (ISONE), indicating minimal prediction error. For multi-step forecasting, the CNN–BiGRU with attention sustained its effectiveness, producing RMSE values of 731.73 (AEP) and 720.21 (ISONE), MAE values of 564.54 (AEP) and 524.66 (ISONE), and MAPE scores of 3.85 (AEP) and 3.53 (ISONE). These results indicate that the proposed model effectively generates accurate forecasts and shows its potential applicability to real-world energy systems, where improved predictive performance can contribute to greater reliability, efficiency, and grid stability. • Hybrid CNN–BiGRU with attention boosts short-term load forecasting accuracy. • Model captures local & long-range patterns using Conv1D + GRU + soft attention. • Achieved low RMSE (135.94 AEP, 134.26 ISONE) in single-step forecasting. • Multi-step forecasts show sustained precision with low MAE and MAPE values. • Framework enhances grid reliability, renewable integration, and stability.