Wind farm cluster power prediction based on graph deviation attention network with learnable graph structure and dynamic error correction during load peak and valley periods

Abstract

The power prediction accuracy of wind farm cluster (WFC) seriously affects its consumption and the safe and stable operation of power system. The fluctuation of power between wind farms (WFs) significantly affects the wind power ultra-short-term prediction (WPUP) accuracy of WFC. In this regard, this paper proposes a graph deviation attention network (GDAN) considering improved clustering distance and learnable graph structure (LGS) for predicting and correcting the wind power of WFC. And used a weighted distance function combining sequence convergence smoothing effect and correlation to dynamically divide the WFC, and to learn and construct the graph structure. Proposed the GDAN with LGS to mine the convergence correlation of WF sub-clusters and establish power prediction model. Considering the characteristics of load peak and valley periods (LPVP), introduced a power correction coefficient to reduce the error, and used the successive variational mode decomposition (SVMD) to extract its key components to achieve power prediction and correction. The proposed method is applied to the WFC in Western Inner Mongolia, China. Compared with the comparison model before correction, the RMSE, MAE and MAPE are reduced by 4.27 %, 3.55 % and 17.92 % respectively, and the R2 and Pr are increased by 11.87 % and 9.88 % respectively.