Energy Consumption Prediction of Building Air-Conditioning Systems Based on a 1DCNN-BiLSTM-AM Framework

Abstract

: Accurate energy consumption prediction for building air-conditioning systems is essential for improving building energy efficiency, reducing operational costs and supporting low-carbon operation and intelligent energy management. Traditional statistical models and shallow machine learning methods have limitations in feature representation and time-series modeling, which restricts their prediction accuracy under complex operating conditions. To address these limitations, this study proposes a hybrid energy consumption prediction model based on a one-dimensional convolutional neural network (1DCNN), bidirectional long short-term memory network (BiLSTM), and attention mechanism (AM), namely the 1DCNN-BiLSTM-AM framework. In this framework, the 1DCNN is first employed to extract feature information from chilled water parameters, cooling water parameters, and outdoor meteorological variables. The extracted features are then fed into the BiLSTM to model temporal dependencies in both forward and backward directions, thereby capturing the dynamic evolution of air-conditioning system energy consumption. Finally, the AM is introduced to emphasize key temporal features and reduce the interference of redundant information. Cooling capacity and total system power consumption are selected as prediction targets to evaluate the model performance. Experimental results show that the proposed model achieves R 2 values of 0.932 and 0.976 for cooling capacity and total system power consumption prediction, respectively, outperforming least squares support vector machine (LSSVM), 1DCNN and 1DCNN-BiLSTM models. The results indicate that the proposed method can effectively characterize the nonlinear and time-varying energy consumption behavior of building air-conditioning systems and provide data-driven support for energy consumption prediction, operational optimization and intelligent building energy management.