Short-term load forecasting using spatial-temporal embedding graph neural network

Accurate short-term load forecasting is of great significance to the efficient operation of power grids. Nowadays, graph neural network (GNN) is widely used to capture the implicit spatial dependencies among electricity users for spatial-temporal load forecasting. However, most existing methods have the following shortcomings: (1) Lack of spatial dependencies represented by directed dynamic graphs. (2) Being limited to plain GNNs, which usually suffer from over-smoothing when the network goes deeper, thus harming the prediction accuracy. (3) Failing to capture the periodicity in load data. To address these problems, we propose a novel Spatial-Temporal Embedding Graph Neural Network (STEGNN). We first construct the directed static graph and directed dynamic graphs. Then we utilize a novel exponential moving average graph convolutional network (EMA-GCN) to capture the spatial dependencies. The periodicity is captured by two sets of trainable temporal embeddings. Finally, we make predictions based on these spatial and temporal features. To evaluate the effectiveness of our method, we conduct comparative experiments on two real-world datasets. Results demonstrate that our model outperforms the state-of-the-art baselines and considerably improves the prediction accuracy.