LoadSeer: Exploiting Tensor Graph Convolutional Network for Power Load Forecasting With Spatio-Temporal Characteristics

Power load forecasting plays a crucial role in ensuring the stable operation of the power system and avoiding system collapse or resource waste caused by power shortages or surpluses. However, the complex spatio-temporal property of power load makes it difficult to predict, which poses a great challenge to the power system. Existing spatio-temporal prediction methods can only handle one factor in each dimension of time and space. In reality, power load is influenced by various factors. Especially in terms of time dimension, crowd flow, weather, and historical load all have significant impacts on load forecasting. Inspired by tensor time series, considering the structure of spatial geographic location, we propose Tensor Graph Convolutional Network for Power Load Forecasting, LoadSeer. A distance adjacency matrix is designed to represent the geographical location relationship and land use nature. A spatio-temporal processing layer integrating graph convolution module (GCN) and T-Transformer is mapped out to extract the spatio-temporal features, which are then sent to the fully connected layer to provide the refined expression. The experimental results on three public datasets, PeMSD4, PeMSD7, and PeMSD8 show that our proposed method outperforms baseline models on all indicators. To further validate the effectiveness of our proposed approach, we apply LoadSeer to real load data during the Asian Games in a certain city, and the results also demonstrate the superiority of our method.