Trustworthy hybrid model with dual dilated attention and BiConvLSTM for multi-horizon load forecasting

Load forecasting has gained attention with the increasing integration of renewable energy into smart grid systems and urbanization, and it has become significantly important for enhancing grid stability and reliability. However, the complex black-box nature of many deep learning models compromises their trustworthiness and causes reluctance among system operators to use them to reduce operational and maintenance costs in real-time. To address this problem, an explainable AI-assisted hybrid deep learning model is proposed that integrates a dual dilated attention mechanism with bidirectional convolutional long short-term memory for efficient and robust multi-horizon short-term load forecasting. The proposed model captures complex long-range temporal and spatial dependencies and enhances feature extraction and correlations within the load data through explainability. Extensive simulation experiments are conducted, and the results are compared with benchmark models. The results demonstrate notable performance improvement, with mean absolute error reductions of 2.38% ∼ 74.42%, 2.76% ∼ 50.07%, 3.37% ∼ 70.30%, and 2.05% ∼ 86.60% in 1h, 6h, 12h, and 24h forecasting horizons, respectively, across the AEP, ComED, PJME, PJMW, Panama, Johor, London, Turkey, and ISONE datasets, respectively. These results confirm the effectiveness and robustness of the proposed model. • xAI-assisted DDA-BiConvLSTM improves multi-horizon STLF with adaptive learning. • DDA uses parallel dilated convolutions for long-range temporal and spatial features. • DDA and BiConvLSTM hybrid enhances multi-horizon STLF accuracy and robustness. • Overall improvement of up to 85.47% in MAE, 81.53% in RMSE, and 81.64% in MAPE. • MAE reduced by 2.38% ∼ 74.42%, 2.76% ∼ 50.07%, 3.37% ∼ 70.30%, and 2.05% ∼ 86.60% in all horizons (1h, 6h, 12h, 24h).