A two-stage deep learning framework for building cooling load prediction based on occupancy information

Abstract The proportion of HVAC energy consumption in buildings is the highest and has a rising trend. Building cooling load prediction is crucial for reducing energy waste, where occupancy information plays a vital role. While existing occupancy prediction methods have progressed, their application is limited by deployment and reliability challenges. This study proposes a two-stage deep learning framework to improve cooling load prediction accuracy in buildings without real-time occupancy monitoring systems. First stage applies a BiLSTM model to predict building occupancy levels, then integrates this occupancy information with meteorological and historical load data for cooling load prediction. Results show that BiLSTM outperforms other models (GRU, LSTM) in occupancy prediction, achieving a test accuracy of 93.11%. For second stage of load prediction, incorporating occupancy information significantly improved prediction accuracy: the CNN-LSTM model using real occupancy data achieved an MAPE of 8.39%, a notable improvement from 11.13% without occupancy information. More importantly, when using BiLSTM-predicted occupancy information, the CNN-LSTM model maintained good performance with an MAPE of 9.12%, approaching the prediction accuracy achieved with actual occupancy data.