A novel deep-learning framework for short-term prediction of cooling load in public buildings

Abstract

Optimal control of heating, ventilation, and air conditioning (HVAC) systems, along with demand-side management, are both cost-effective methods in the process of energy conservation and carbon reduction. The successful implementation of these initiatives largely hinges on accurate cooling load predictions. Due to the complex nonlinear and dynamic time-varying nature of demand loads, however, it is a formidable challenge to accurately predict the cooling load. To address these issues, a novel deep learning-based prediction framework, aTCN-LSTM, is proposed. First, a gate-controlled multi-head temporal convolutional network is designed to capture the inherent nonlinear and local temporal features from the time series of cooling loads. Second, a sparse probabilistic self-attention mechanism is integrated with a bidirectional long short-term memory (BiLSTM) network to extract the long-term dependencies within the cooling load sequences. Finally, through integration with the proposed two components, the framework is developed and validated through a 14-month real cooling load forecasting problem for a 51-story hotel building in Guangzhou, China. Experiments and comparison studies demonstrate the effectiveness and superiority of the proposed method. The mean absolute percentage error of the proposed method's 1-step, 6-step, and 12-step prediction results is reduced by 27.48%, 14.05%, and 13.38%, respectively, compared with the state-of-the-art baseline model. Consequently, it stands poised to serve as an effective guide for HVAC chiller scheduling and demand management initiatives.