基于时空特征自适应提取与状态转移融合的表后光伏分解方法OA

With the advancement of the"Thousands of Households Embracing Sunlight Initiative",an increasing number of distributed photovoltaic(PV)installations are being placed behind smart meters.Their unobservable nature poses significant challenges to distribution network planning and operation,highlighting an urgent need for accurate and reliable methods for behind-the-meter PV disaggregation.However,existing disaggregation methods remain inadequate in extracting and deeply integrating periodic temporal features and spatial correlation features.To address this issue,this paper proposes a behind-the-meter PV disaggregation method based on adaptive extraction and state transition fusion of spatiotemporal features.First,an adaptive adjacency matrix is constructed to dynamically characterize the spatial correlations among users,mapping net load data into a dynamic spatiotemporal graph structure.Second,the net load data is periodically decoupled by integrating time-domain and frequency-domain information,enabling the extraction and aggregation of multi-scale periodic features.Subsequently,graph diffusion convolutional neural networks are employed in the spatial dimension to extract wide-area spatial features.Finally,a selective spatiotemporal state model based on dynamic spatiotemporal graph embedding is used to deeply integrate the spatiotemporal features and generate behind-the-meter PV disaggregation results.Experimental results on a real-world dataset comprising 268 users demonstrate that the proposed method reduces the mean absolute error by 58.8%compared to existing disaggregation methods.Even under a metering data missing rate of up to 15%,the error is reduced by at least 61.7%.