Real-time indoor thermal comfort prediction in campus buildings driven by deep learning algorithms

Abstract

Thermal comfort prediction is vital in achieving a good indoor environment and efficient energy management. However, thermal comfort is strongly nonlinear and dynamically changing over time, making it difficult to predict thermal comfort accurately. Two real-time thermal comfort prediction models on multi-time scales are proposed based on deep learning algorithms. The indirect prediction model for thermal comfort integrates the bidirectional long and short-term memory network to predict real-time environmental parameters. It subsequently combines the thermal comfort calculation method with the predicted environmental parameters to obtain the predicted thermal comfort value. The indirect prediction model for thermal comfort combines the bidirectional long and short-term memory network to realize real-time prediction of environmental parameters and then calculates thermal comfort by the thermal comfort calculation method to realize indirect thermal comfort prediction. The other is the direct prediction model of thermal comfort, which is realized using the bidirectional long short-term memory network for real-time thermal comfort prediction. The two real-time prediction models are compared and analyzed for 10-min, 30-min, and 60-min scales. Experimental results show that the prediction accuracy of the indirect prediction model is better than the direct prediction model, but the robustness of the indirect prediction model is lower than the direct prediction model. In addition, the accuracy of the thermal comfort prediction model on the small time scale is higher than that on the large time scale, in which the RMSE and MAE reached 0.0551 and 0.0543 (10-min scale), respectively. Therefore, the proposed thermal comfort prediction models can be integrated into the control model of heating, ventilation and air conditioning systems, providing a reference for optimizing indoor thermal comfort.