Bayesian-optimized CNN-LSTM for residential building energy consumption prediction with SHAP explainability

This study examines methods to improve predictive accuracy and interpretability of residential building energy consumption (RBEC) prediction, aiming to support energy retrofitting, demand response, and sustainable energy management that reduce environmental impacts. This study proposes a Bayesian-optimized hybrid Convolutional Neural Network and Long Short-Term Memory (BO-CNN-LSTM) that fuses static household/building attributes with temporal climate signals and is integrated with SHapley Additive exPlanations (SHAP) for feature-level interpretability. Results demonstrate that the proposed BO-CNN-LSTM achieves an overall R 2 of 0.884 on the test set, outperforming CNN (0.672), LSTM (0.738), and unoptimized CNN-LSTM (0.812). It also exhibits lower MAE and RMSE, with a slower R 2 degradation in both seasonal-window and cross-seasonal forecasting, confirming its short-term precision and long-term stability. Compared with tree-based baselines such as XGBoost and LightGBM, the BO-CNN-LSTM delivers superior predictive accuracy, whereas conventional models typically achieve R 2 between 0.6 and 0.8. Furthermore, the SHAP analysis identifies gas heating, building area, monthly heating degree days, heating equipment control systems, building age, and number of minors as key determinants of RBEC. Dependence and interaction plots reveal crucial conditional effects, providing actionable insights for scenario-oriented retrofit prioritization and operation strategies. Overall, the integration of Bayesian optimization, hybrid deep learning, and SHAP analyses enhances both the accuracy and interpretability of RBEC prediction. Furthermore, the findings contribute to explainable AI for energy systems, support urban policy, and supply transferable prediction workflows for digital-twin and Life Cycle Assessment applications.