Spatial-Temporal Attention Mechanism and Graph Convolutional Networks Based Distributed PV Ultra-Short-Term Probabilistic Power Forecasting

Abstract

To ensure the safe and stable operation of the distribution network with high penetration of distributed photovoltaics (PVs), reliable power forecasting of the distributed PV cluster is significant. However, the lack dynamic spatial-temporal correlation modeling ability and representative spatial-temporal feature extraction ability of the existing methods limits the further improvement of forecasting accuracy. Moreover, the existing power forecasting results are mostly presented as deterministic form, which is difficult to describe the forecasting credibility. Therefore, a spatial-temporal attention mechanism and graph convolutional networks based distributed PV ultra-short-term probabilistic power forecasting method is proposed in this paper. Firstly, the target distributed PV cluster is divided into sub-regions and the original distributed PV graph-structured time series is generated by abstracting each sub-region as a graph node and regarding the aggregated power of each sub-region as the original node feature. Then, the spatial-temporal attention mechanism is used to adaptively capture dynamic spatial-temporal correlations from the distributed PV graph-structured times series. Supported by the dynamic spatial-temporal correlations, spatial-temporal convolution is applied to extract representative spatial-temporal features from the distributed PV graph-structured times series. Finally, a non-crossing joint quantile loss function is introduced to the training process to construct a complete probabilistic power forecasting model of a distributed PV cluster that can provide power forecasting intervals at multiple confidence levels at once for the next four hours. The validity and progressiveness of the proposed model have been verified in real-world datasets.