Uncertainty Analysis And Short-term Prediction Of Photovoltaic Output Considering Temporal And Spatial Dependence

With the proposal of "carbon peak" and "carbon neutrality",the development of large-scale photovoltaic（PV）power stations represented by centralized development has reached its peak.In the process of PV power station operation,PV output has obvious fluctuation in a short period of time,which increases the difficulty of the power system to effectively absorb PV output and make scheduling plans.Therefore,high-precision PV output probability prediction technology is particularly important.Due to the influence of climatic conditions,quality of unit fittings,maintenance time and other factors,PV power stations are prone to abnormal output data,which hinds the analysis of PV output data characteristics.Therefore,this paper firstly analyzes the characteristics of normal and abnormal PV output data,summarizes the causes and classification of abnormal data,and puts forward a PV output abnormal data identification model based on least square filters-shoville criterion,which realizes the function of locating abnormal output data and makes preparation for PV output prediction.According to eliminate the abnormal data after the irradiance,lack of the efforts of both,could be divided into partial missing and complete missing two types,through the principle of maximum correlation minimum redundancy of output data and multi-source meteorological data correlation is analyzed,based on hybrid network matching method-long-term and short-term memory（LSTM）of missing data filling model,the accuracy of output missing data is more than 90% in cloudy weather.The characteristics of historical output data of single-field,multi-field and large-scale cluster PV power stations are described from the perspective of multi-space-time scale,and the changes of multiple indicators are analyzed.At the same time,the prediction error is analyzed,and the conclusion is that the prediction error of different PV power stations in the same area is spatiotemporal dependent,which lays a theoretical foundation for improving the accuracy of single value prediction of PV output.Various uncertain factors existing in the process of PV output prediction are analyzed and summarized.From the perspective of spatio-temporal dependence of prediction errors among multiple stations,the appearance similarity updating（ASU）is used to quantitatively evaluate the leader-lag relationship between prediction errors,and optimize the short-term single-value prediction results of the target power station.Combined with kernel density estimation（KDE）,a short-term uncertainty analysis and prediction model based on appearance similarity updateLSTM was proposed.The results of two examples show that the model can obtain relatively reliable short-term probability prediction results,which can provide a reference for power grid dispatching and energy storage.