Explainable deep learning combined attention-based LSTM for building energy prediction: a framework from the perspective of supply side

Accurate prediction of dynamic building energy demand is essential for energy conservation and operational flexibility and in the premise of satisfied indoor environment. The development of data science and machine learning introduced powerful tools for predictive modeling in the built environment. This study presents a hybrid novel framework combining Attention based Long Short-Term Memory (ALSTM) with the Hilbert Haung Transform (HHT) for generalized accuracy predictions. Six hotel buildings’ operational data of two months (March and July) combined with meteorological data were taken to demonstrate the effects of this method. Critically, to ensure these forecasts translate into actionable operational strategies, Explainable AI (XAI) techniques (LIME and SHAP) were embedded to deconstruct model logic and highlight key drivers of energy behavior. The results show that ALSTM consistently outperformed benchmark models (LSTM, GRU, and ANN), achieving R² values up to ∼0.94 and RMSE values as low as ∼17–26, compared to baseline values as low as 0.57 for R² and above 190 for RMSE. These outcomes correspond to improvements of up to 54 % in R² and 22 % reduction in RMSE relative to the baseline models. XAI diagnostics revealed that the hybrid model produced transparent and physically consistent prediction with real-time consumption metrics, signal decomposition features, and temporal patterns emerging as dominant drivers of energy predictions. This integration of advanced hybrid modelling with XAI demonstrated significant potential for optimizing sustainable energy operations by enhancing buildings’ responsiveness to dynamic grid and environmental demands.