Physics-informed long short-term memory networks for price-incentive heating, ventilation and air conditioning loads prediction

Accurate prediction of price-incentive heating, ventilation and air conditioning (HVAC) loads is an important prerequisite for their participation in demand response (DR). However, traditional physics-based prediction models exhibit poor scenario adaptability that leads to modeling bias, while data-driven models struggle to characterize the physical relationships between inputs and output and thus suffer from poor interpretability. To address these limitations, this paper proposes a physics-informed long short-term memory network (PILSTM) for price-incentive HVAC loads prediction. The model dynamically adjusts feature weights to capture the time-varying effects of temperature and electricity prices on load, and optimizes the loss function of the model by incorporating physical monotonic consistency and load boundary constraints, thereby concurrently enhancing both the model prediction accuracy and interpretability. Experimental results across different scenarios show that: in real-time pricing scenario, compared to LSTM, the proposed model reduces the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) by 13.57 %, 12.45 %, and 12.42 %, respectively, while its predictions are more consistent with the actual response characteristics of HVAC; in time-of-use pricing scenario, the R² of PILSTM is 10.18 % higher than that of LSTM, demonstrating superior prediction accuracy; seasonal load forecasts reveal that the performance indicators of PILSTM exhibit better values and smaller fluctuations, indicating its adaptability to diverse HVAC operating conditions and better prediction stability. By dynamically adjusting feature weights and incorporating physical constraints, the proposed model accurately predicts HVAC loads under diverse operating conditions, thereby providing reliable data support for energy system optimization.