Short-term Probabilistic Forecasting Of Photovoltaic Power Based On Outlier Identification And Reconstruction Technology

Probabilistic forecasting of photovoltaic(PV)power uses interval,quantile or probability distribution to characterize the uncertainty of photovoltaic output,and provides comprehensive forecast information for grid dispatching,thereby improving the reliability of dispatching operation decision-making.However,at present,it is still difficult to achieve high-reliability photovoltaic power probability prediction.On the one hand,there is a lack of true and credible historical photovoltaic power data due to artificial power curtailment,measurement and communication equipment failures,etc.On the other hand,existing prediction models have“quantile crossing” and high training costs when realizing multi-quantile predictions.In this regard,considering the characteristics of photovoltaic power generation,this paper conducts research in two aspects,data quality improvement and prediction model improvement.A shortterm photovoltaic power prbabilisitc forecasting method is proposed based on outlier identification and reconstruction technology.The specific research content includes:1)To facilitate the preprocessing of photovoltaic power data and the construction of the input features of the prediction model,the characteristics of photovoltaic output power and its influencing factors are analyzed.The volatility of the PV power is quantified in four aspects:the average value,the degree of dispersion,the degree of asymmetry,and the degree of steepness.With the help of clustering and dimensionality reduction visualization technology,the relationship between photovoltaic power and weather types is analyzed.The self-correlation of photovoltaic power time series and mutal correlation between PV power and meteorological factors such as irradiance,temperature and humidity are analyzed from two aspects,linear correlation and rank correlation.2)A preprocessing technology for PV power data with high proportion of outlier is proposed.The outlier types are divided based on the distribution of the outlier data in the irradiance-photovoltaic power scatter diagram,and the causes of each type of outlier are analyzed.Based on the LSCP framework,a variety of basic anomaly identification models are integrated to improve generalization performance,which effectively improves the recognition accuracy of high-proportion abnormal data.To repair the PV power outliers in a low cost and high precision way,a reconstruction model based on a Light GBM(Light Gradient Boosting Machine)is established for discontinuous and continuous anomalies respectively,and the hyperparameters of the reconstruction model are optimized by random search.3)A vine copula based quantile regression model for day-ahead probabilistic forecasting of PV power is proposed.The vine copula is used to express the dependent structure between the PV power and its conditions analytically,and the vine copula structure and the parameters are optimized with the optimization algorithm.On this basis,the PV power conditional quantile regression model is established.The point prediction value of the PV power is added into the conditions,and the minimum continuous rank probability score(CRPS)is used to select the optimal combination of conditions,which can weigh the reliability and sharpness.Simulation results show that the proposed method overcomes the shortcomings of the existing quantile regression methods and further improves the performance of the PV power probabilistic forecasting.