基于三维卷积与时空注意力机制的风光荷典型场景生成方法OA

To address the complex and varied source-load scenarios in the new power system,the scenario clustering method improves optimization efficiency by selecting typical scenarios.However,traditional scenario generation methods fail to adequately capture source-load correlations,making it difficult to generate representative typical scenarios.This paper proposes a wind-photovoltaic(PV)-load typical scenario generation method based on three-dimensional(3D)convolution and spatio-temporal attention mechanism.First,historical wind-PV-load data are preprocessed using Z-score normalization and cubic spline interpolation.Then,a deeply embedded autoencoder is constructed by integrating 3D convolution and spatio-temporal attention mechanism.The 3D convolution introduces the temporal dimension to extract spatio-temporal features of source-load scenarios,while the spatio-temporal attention mechanism enhances the extraction of key spatio-temporal features.Furthermore,a three-stage training strategy is proposed to train the reconstruction ability and the clustering performance of the model in stages,avoiding performance degradation caused by single-task training.Finally,a case study based on real wind-PV-load datasets is carried out.The results show that,compared with both traditional clustering methods and deep learning based clustering approaches,the proposed method demonstrates significant advantages in terms of clustering metrics,and can effectively meet the requirements of economic operation optimization of power system in terms of both accuracy and timeliness.