Ultra-short-term wind power forecasting jointly driven by anomaly detection, clustering and graph convolutional recurrent neural networks

Abstract

Accurate ultra-short-term regional wind power forecasting is crucial for real-time power grid dispatching and frequency regulation. However, recent works have primarily focused on designing complex model structures, often overlooking computational efficiency or the balance between efficiency and prediction accuracy, which limits practical applications. This paper aims to improve data quality, reduce computational cost, and enhance prediction accuracy by proposing a spatiotemporal prediction method for large-scale wind farms. To address the parameter sensitivity issue of the density-based spatial clustering of applications with noise (DBSCAN) and to enhance its anomaly detection capability and robustness, we propose an improved version of DBSCAN, applied to wind turbine data anomaly detection. Simultaneously, leveraging the high-performance advantage of the lightweight gradient boosting machine, abnormal data are quickly corrected. Using spectral clustering based on graph theory, we optimally partition the data graph to form wind farm clusters. Subsequently, a two-layer adaptive graph convolutional recurrent neural network (AGCRN) is employed to capture complex spatiotemporal correlations between wind turbines in each cluster. Finally, the regional total power forecast is obtained by summing the forecast outputs of all clusters. Through numerical simulations using measured data from Dataset 1 (134 wind turbines) and Dataset 2 (200 wind turbines), the results indicate that the proposed data preprocessing scheme can achieve at least a 50 % improvement in the model. By forecasting in clusters, the mean absolute error (MAE) can be reduced by 45.84 %, training time shortened by 70.84 %, and GPU memory saved by 94.04 %. Compared with advanced models such as Transformer variants and TimeNet, the multi-layer AGCRN achieves the highest prediction accuracy, exceeding 85 %.