Dynamic spatio-temporal correlation and hierarchical directed graph structure based ultra-short-term wind farm cluster power forecasting method

• The spatio-temporal correlation of adjacent wind farms is dynamic and time-varying. • Dynamic spatio-temporal correlation can alleviate lagging problem of forecast values. • Hierarchical directed graph can clarify the causal relationship among input variables. • Reduce error sources by hierarchical directed graph can improve forecast accuracy. Accurate wind farm cluster power forecasting is of great significance for the safe operation of the power system with high wind power penetration. However, most of the current neural network methods used for wind farm cluster power forecasting have the following three problems: (1) lack of consideration of dynamic spatio-temporal correlation among adjacent wind farms; (2) simultaneously forecasting all wind farms’ power to obtain the total power will produce numerous error sources; (3) ignoring the causal relationship among input variables. Therefore, to solve the above problems, this paper proposes an ultra-short-term wind farm cluster power forecasting method based on dynamic spatio-temporal correlation and hierarchical directed graph structure. Firstly, three different types of nodes (wind speed nodes, wind power nodes, and target node) and input samples are defined, and then the spatio-temporal correlation matrices that can describe the correlation of adjacent wind farms are also calculated. Secondly, directed edges are defined to connect different nodes in order to obtain the hierarchical directed graph structure. Finally, this graph structure with dynamic spatio-temporal correlation information is used to train the forecasting model. In case study, compared with other benchmark methods, the proposed method shows excellent performance in improving accuracy of power forecasting.