Given the increasing demand for high-accuracy power load forecasting, traditional load forecasting methods can capture long-term dependencies in time series, but cannot fully capture the complex relationships between multi-dimensional features. This paper proposes an innovative method to convert time series data into graph features. By constructing a graph structure based on time nodes, the time series forecasting problem is transformed into a graph-based load forecasting problem. On this basis, the Edge Graph Attention Network (EGAT) is used to combine the feature information of nodes and edges to further enhance the ability to represent feature interactions and improve the accuracy of load forecasting. This paper compares the EGAT model with common load forecasting methods, including gated recurrent units (GRU), multi-layer perceptron networks (MLP) and long short-term memory (LSTM). The results show that EGAT is effective at finding important features and understanding complex time patterns, which means it shows strong potential in predicting energy demand. A limitation of the proposed approach is its increased computational cost introduced by graph construction and attention-based aggregation, which may raise training time and memory usage for large-scale graphs. In addition, the forecasting performance can be influenced by the design of the time-series graph (e.g., connectivity patterns) and the availability/quality of edge features.
Gu et al. (Thu,) studied this question.