Hybrid model for robust and accurate forecasting building electricity demand combining physical and data-driven methods

The high-precision prediction of building electrical demand is paramount significance for optimal control of building systems and the reduction of operational energy consumption. The pure physical simulation software struggles to depict buildings' energy usage behavior accurately; while purely data-driven methods lack the interpretability in real-world physical scenarios. To address the challenges existing in current models and algorithms, this paper proposes a hybrid model that combines physical mechanism simulation and machine learning to complement each other. Empirical Mode Decomposition (EMD) is utilized to decompose complex data into simple components in pre-processing (RMSE, MAE, and SMAPE all decreased by around 0.2–0.3). And K-means clustering algorithm is used to extract features of building electrical data and determine employee working states (R2 increased from 0.3 to 0.85). Shapley Additive explanations (SHAP) values and Pearson correlation coefficients are used for model interpretability analysis. Combining weather forecasts, the Monte Carlo method is employed to obtain prediction intervals. Furthermore, this study explores the use of transfer learning to save computational costs and reduce the amount of data needed for predicting, especially for high-rise buildings or floors with insufficient historical data. The hybrid model results indicate that the R2 can exceed 0.85, up to 0.9, while pure physical simulation can only achieve 0.3. MAE, MSE, and RMSE remain below 0.07, while SMAPE stays around 0.2. The R2 value of the transfer learning model can surpass 0.8 (MSE, MAE. MAE, MSE, and RMSE remain below 0.2, while SMAPE stays around 0.4). The model's residuals obey normal distribution, indicating the model has strong generalization capabilities.