An innovative model integrating dual-channel graph attention network and multi-task learning for building energy consumption anomaly detection

Building energy consumption data are subject to the combined influence of multidimensional factors such as meteorological conditions, equipment status, and human factors, posing significant challenges to the effective identification of anomaly patterns within these data. To address these challenges, a multi-task learning model based on a dual-channel graph attention network (GAT) is proposed. The model processes building energy consumption data with strong temporal periodicity and multivariate coupling, and realizes efficient and accurate anomaly detection. The proposed method adopts a feature graph attention layer to model the coupling relationships between the energy consumption of circuits (such as air conditioning, lighting, and power) and environmental parameters, while simultaneously using a temporal graph attention layer to capture periodic patterns like workdays and non-workdays. Based on the prediction error of the temporal convolutional network (TCN) model and the reconstruction error of the variational autoencoder (VAE) model, a dynamic balancing coefficient is designed to achieve adaptive adjustment of the anomaly score to changes in data distribution. The feasibility and efficiency of the proposed novel multi-task model were validated using an energy consumption dataset collected from an office building in Jinan, China. Experimental results show that for anomaly detection of air conditioning, lighting, power, and other circuit energy consumption, the proposed model achieves F1-scores of 0.936, 0.951, 0.905, and 0.892, respectively. Compared with comparative models, the F1-scores are improved by at least 6.1 %, 4.4 %, 3.2 %, and 4.7 %. This method provides an efficient and reliable technical solution for refined management and control of building energy consumption.