Exécution de notre outil de synthèse sur un nouvel article

Vérifiez l’article ci-dessous, puis lancez sa synthèse.


Short term energy consumption prediction of regional building group based on spatiotemporal graph convolutional network considering spatiotemporal correlation between building nodes

Yamei Ma (Tongji University), Yingjun Ruan (Tongji University), Tingting Xu (Tongji University), Yuting Yao (Tongji University), Hua Meng (Tongji University)

Journal of Physics Conference Series
Licence : CC BY

Résumé : Abstract With the rapid development of high-performance computing technology, accurate regional building energy consumption prediction based on data-driven is the basis for the refined management of regional energy consumption and energy saving and emission reduction. Owing to climatic conditions, human movement, and other factors, the energy consumption of buildings is influenced not only by their inherent characteristics but also by surrounding buildings. However, there is a lack of a systematic approach to modeling the spatial dependence of regional buildings. Therefore, this study proposed a spatiotemporal graph convolution model combining different graph construction methods for short-term regional building energy consumption prediction. First, Spearman correlation coefficient was used to determine the key influencing factors of building power consumption. Then, the graph convolution is based on an adaptive adjacency matrix trained end-to-end with the model to capture potential spatial dependencies between buildings and a correlation coefficient adjacency matrix, and WaveNet learns time dependencies. This study was conducted on a publicly available dataset from the Arizona State University online system, results show that compared with the SVR, LSTM, CNN-LSTM models, the proposed model has higher prediction accuracy, with the mean absolute percentage error(MAPE) of 6.07% for the whole region and the minimum MAPE of single building prediction only 2.11%, which means the proposed method can fully capture the spatial dependence of different buildings in the region to improve the prediction accuracy of regional building energy consumption. Comparing the prediction results of different graph construction methods, we can see that the node correlation captured by combining adaptive adjacency matrix and DTW adjacency matrix is the most effective.

Publié le 01 Avr. 2025