Temporal graph attention network for building thermal load prediction

Machine learning models have seen widespread application in predicting building thermal loads. Yet, these existing models generally predict the thermal load for a single thermal zone or an entire building, overlooking buildings with multiple thermal zones. Moreover, interactions between thermal zones are seldom accounted for these methods. Addressing this gap, our study introduces a graph-based approach that employs a Temporal Graph Attention Network (TGAT) model for building thermal load prediction. This model accommodates both the spatial and temporal dependence across different thermal zones. Specifically, the TGAT model initially utilizes graph attention neural networks (GATs) to identify the spatial relationships between thermal zones. The spatial representation thus obtained is subsequently processed via Gated Recurrent Units (GRUs) to ascertain temporal dependencies. Lastly, a fully connected layer is engaged to decode the representation for the final output. Our model incorporates 11 features as inputs to depict the characteristics of outdoor environment and thermal zones. Testing the proposed TGAT model in a small office virtual environment for heating load prediction revealed that a configuration of four GAT attention heads achieved optimal result for predicting the ensuing hour's heating load with a six-hour lookback. Additionally, we scrutinized and discussed the efficacy of feature selection and the GAT module. This research thus establishes a novel groundwork for forthcoming studies on zone-based building energy management optimization.