Spatial–temporal short-term load forecasting framework via K-shape time series clustering method and graph convolutional networks

Abstract

Load forecasting, especially short-term load forecasting, is of great significance to the safe operation of power grids and power system optimization. Historical data is commonly applied in the most of existing methods of load forecasting to predict the future data in one time step, which is considered to have a positive effect on the forecasting performance. However, a serious problem is ignored, which is the assumption that the interactions between end-users are often not considered. Such interaction could be considered as a paired non-Euclidean relationship, which could enhance the accuracy of forecasting effects. In this paper, we propose a graph convolutional network prediction framework based on K-shape time series clustering method for short-term load forecasting. User clusters with similar electricity consumption habits could be separated from a large number of users by the proposed framework. Simultaneously, the paired non-Euclidean relationship between users can also be captured by this framework. The proposed framework is evaluated by the dataset of business users in the southeastern United States. The results show that compared with baselines, the framework significantly improves the prediction accuracy in any scale of user groups, as well as future data in one time step and multi-time steps.