Novel energy consumption forecasting method employing weighted occupant behavior probabilities and physics-informed network with thermodynamic constrain in office buildings

Addressing the global energy crisis and excessive emissions has heightened the critical importance of reducing energy consumption and carbon emissions in the building sector, making accurate building energy forecasting a fundamental research focus. While existing methods predominantly prioritize forecasting accuracy by advanced algorithms, considerations of computational efficiency and model interpretability remain scarce. To bridge this gap, this study proposes a novel forecasting method that simultaneously optimizes for accuracy, efficiency, and interpretability. The method integrates three strategies: (a) incorporating weighted occupant behavior probabilities as novel inputs; (b) incorporating physics-informed loss function calculated by thermal resistance-capacitance (R-C) models; and (c) developing a hybrid CNN-LSTM-Attention algorithm that integrates convolutional neural networks and an attention mechanism with a long short-term memory network. Validation of 48 cases from four office buildings shows the proposed method significantly enhances performance. These three strategies reduce the mean absolute percentage error (MAPE) by 25.78 % and the coefficient of variation of the root mean square error (CV-RMSE) by 21.31 %, and average contributions are 40 %, 15 % and 45 % for Strategies (a)–(c), respectively. Strategy (c) is the primary contributor to efficiency gains, which can reduce time consumption by 7343.69s and 146.81s compared to Transformer-LSTM-Adaboost and LSTM-SSA, respectively. Strategies (a) and (b) improve interpretability by embedding occupant behavior patterns and thermal constraints. Moreover, the priority of these strategies for buildings with varying behavioral and functional complexities is analyzed. In summary, based on theoretical considerations and practical validation, the proposed method can improve the accuracy, efficiency, and interpretability simultaneously.