A knowledge embedded graph neural network-based cooling load prediction method using dynamic data association

Abstract

Accurate cooling load prediction has continually attracted interest due to its important role in building energy saving. Much attention to existing methods has been paid to the temporal information of the features that affect the load prediction. However, facing the time-varying interactions between various types of features, existing approaches have limited ability to extract the implicit information of the complex associations among them. This limitation leads to underutilization of features, especially those that indirectly affect the load prediction through such interactions, and further constrain prediction accuracy. To this end, this paper proposed a graph neural network-based cooling load prediction method to integrate the associations and temporal information of the features. Using the weighted mask mechanism, a knowledge embedded dynamic association graph is constructed to handle the time-varying feature associations rationally. By introducing empirical knowledge and adjusting the graph structure automatically according to the real-time data, the proposed method can provide accurate cooling load prediction. Comprehensive experiments are carried out on the onsite data, and the experimental results show the practicability and effectiveness of the proposed method for cooling load prediction.