An interpretable building air conditioning load forecasting framework using SSA-optimized Bi-LSTM and SHAP analysis

Short-term air conditioning load forecasting in industrial buildings plays a key role in energy management and carbon neutrality efforts. Yet internal disturbances, latent heat variations, and seasonal distribution shifts often hinder model generalization. Here we propose an interpretable framework that couples Sparrow Search Algorithm (SSA)-optimized Bidirectional Long Short-Term Memory (Bi-LSTM) networks with SHapley Additive exPlanations (SHAP). Working with real 5-min data from an industrial facility across all four seasons, we construct a 17-dimensional physics-informed feature vector comprising environmental states, historical load inertia, dynamic trend indicators, and cyclical time encodings, employing strict chronological splits to avoid data leakage. The pure Bi-LSTM model avoids the parameter redundancy of complex hybrid networks and clearly outperforms traditional machine learning, basic deep learning, and attention-based baselines. Furthermore, SSA proves significantly more robust and efficient for hyperparameter tuning than PSO, DBO, ISSA, and SCSSA. In winter testing, the SSA-optimized model reaches R 2 = 0.965 and RMSE = 1.20 kW. SHAP analysis highlights physically plausible feature contributions-indoor enthalpy and historical load dominate, demonstrating clear seasonal patterns. Direct cross-season transfer still yields R 2 0.93, and a parameter-based transfer learning strategy using only 10% of target-season data reduces prediction errors by another 30%–50% compared to training from scratch. Overall, the framework delivers accurate, automated, and explainable forecasting tailored to industrial HVAC systems, offering practical engineering guidance.