Spatiotemporal Electric Energy Efficiency Evaluation via Hybrid Graph Neural Network and Transformer Architecture

Abstract

Traditional power efficiency assessment mechanisms face limitations in processing spatiotemporal data. Therefore, we combined deep learning models to construct a power resource efficiency evaluation method that integrates spatiotemporal information. Regarding the model, we combined graph neural networks to process spatial information. Transformers were used to process the features implicit in continuous information, and a hybrid GNN-Transformer approach was employed to achieve dual modeling of temporal and spatial features. The model employed an attention mechanism to capture the characteristics of key time nodes and critical loads in both temporal and spatial data. This paper evaluates the model on two real-world power system datasets: (1) hourly load and day-ahead price data from the US PJM grid (2018–2023; ~43,800 samples across 13 control areas); and (2) 15-minute wind, solar, and load data from a Chinese provincial grid (2020–2022; ~105,000 samples). Both datasets span diverse weather and seasonal conditions, enabling robust assessment of the model’s generalization. Experimental results showed that our proposed model and several baseline models were evaluated on real-world power grid datasets. The proposed hybrid GNN-Transformer model aggregates spatial dependencies in the power grid topology through graph convolutional networks (GCNs) and utilizes a multi-head self-attention mechanism to model long-term temporal dynamics. On the PJM dataset, this method reduces the mean squared error (MSE) from 0.1089 to 0.0823 on the LSTM baseline, a relative reduction of 24.4%, which is significantly better than existing methods.