基于相似云况匹配的分布式光伏集群功率概率预测方法OA

Distributed photovoltaic(PV)systems lack on-site meteorological measurements,while the cost of acquiring large-scale and high-resolution numerical weather prediction(NWP)data is very high,which makes it difficult for existing power forecasting methods for centralized photovoltaics to be applicable.Satellite cloud images,offering wide-area,high-resolution cloud observations at a low cost,which is expected to become an important source of meteorological information to solve this problem.Therefore,this paper proposes a probabilistic power forecasting method for distributed PV clusters based on similar cloud condition matching.First,an optical-flow-enhanced 3-dimensional convolutional neural network(3D-CNN)is constructed to forecast future satellite cloud images.By incorporating a physical consistency loss,the model is guided to focus on the dynamic evolution of cloud structures,thereby improving the physical plausibility of the forecasting results.Secondly,a cloud condition matching method is introduced,by searching for the cloud conditions that are most similar to the forecasting period in history,the corresponding photovoltaic power is extracted as input features for subsequent power forecasting.Finally,the common information such as similar historical power and measured PV power at adjacent times will be combined as inputs for the spline quantile regression model to achieve accurate probabilistic forecasting of PV power.Case studies are conducted on a large-scale real dataset containing over 600 distributed PV sites,and the results show that the proposed method shows significant improvement compared to classical kernel density estimation methods in terms of the continuous ranked probability score index.