Physics-guided LSTM model for heat load prediction of buildings

Abstract

The thermal process of heating buildings is characterized by the multivariable, nonlinear, and large time lag. Accurately predicting building heat loads is crucial for regulating building heating systems. However, the current black-box method of heat load prediction has challenges such as extensive data usage and weak integration with the building's heat transfer process. To address these challenges, a physics-guided long short-term memory (LSTM) model for dynamic heat load prediction in the next seven days based on data-driven and mechanism analysis was proposed. The data for the model were collected from in-situ measurement and numerical simulation. Specifically, the thermal mechanism was introduced to the training loss function of LSTM and the selection of input variables. The influence of the adaptive data time step on prediction accuracy was investigated. The main conclusions obtained from the current data in this paper are drawn as follows. 1) Refining solar radiation intensity calculation for each orientation reduced mean absolute percentage error (MAPE) of heat load prediction results by 0.02%–1.8% compared to horizontal solar radiation intensity. 2) Improving the training loss function decreased MAPE of heat load prediction results by 0.04%–0.43% compared with the original loss function. 3) Adaptive data time step lowered the MAPE of heat load prediction results by 0.08%–1.8% compared with fixed time step. 4) Mean absolute errors corresponding to those three aspects decreased by 25 kJ–558 kJ, 6 kJ–157 kJ, and 41 kJ–1881 kJ. In the future, wider promotion and application of the prediction model will lead to better performance and be more beneficial to energy saving and carbon reduction of the heating system.