基于层次图神经网络与极值理论的短期区域风电功率预测方法

Abstract The total potential of wind power generation in the jurisdiction attracts more attention compared to an individual wind farm from the prospect of power system operators. Exploiting the spatio-temporal dependencies at multiple wind farm sites within the target region can significantly enhance the forecasting performance. However, multitudes of wind farms pose ignorable modelling challenges due to the intricate correlations arising from the continuity of different spatial-scale weather systems, and traditional loss function based on mean square error manifests natural weaknesses in the face of extreme events. To tackle these issues and further boost the forecasting accuracy, this paper constructs a hierarchical spatio-temporal graph neural network with a correction module and an improved loss function grounded in extreme value theory. Sub-regions are first divided based on agglomerative hierarchical clustering considering geographic distance information and long-term climate characteristics. Then, the hierarchical spatio-temporal forecasting framework consisting of graph neural network and redesigned bidirectional long-short term memory network is established to address the multifaceted interconnections across the wind farms and sub-regions. In addition, the correction module and corresponding revised loss function are encapsulated to enhance the forecasting performance at certain moments especially with extreme values. Finally, the proposed method is performed on the real-world regional wind power forecasting, and case studies confirm the superiority over the benchmark methods.