Method Of Photovoltaic Power Generation Forecast Based On Improved Ensemble Learning Algorithm

Under the background of the global energy transition and the energy production and consumption revolution in China,PV power generation has received widespread attention as an important clean alternative,and the rapid growth of global PV installed capacity has been shown.However,the uncertainty of PV power generation is inevitable.As the PV penetration rate continues to increase,the problem of energy consumption and grid connection security has become increasingly prominent.Therefore,further research on PV power generation forecast to improve the accuracy and reliability of the forecast system is of great significance to ensuring the power system safety and economic operation.There are many classification methods for PV power forecast.According to the different forms of forecast results,it can be divided into the spot forecast and the probabilistic forecast.The expectation of PV power can be obtained by the spot forecast,while probabilistic forecast can provide more effective information such as quantile,prediction interval and probability density function.At present,the research on the former is quite mature,while on the latter is relatively rare.With the deepening of research,the advantages of probability forecast are becoming obvious,which can not only improve the risk assessment ability of PV plants,but also provide more references for grid regulators.So far there are several following issues in PV power forecast.On the one hand,the existing PV power forecast methods only use a single forecast model with limited generalization performance;or only combining models through averaging them,which lack sufficient theoretical support.On the other hand,most researches on the probability forecast of PV power are based on parameter methods,and the selection of the parameters has a great influence on the forecast results.Moreover,the prior distribution in actual situations is uncertain and irregular,which leads to poor universality of the model.This thesis successively proposes spot and probabilistic forecast methods for PV power based on improved ensemble learning algorithm.Firstly,a spot forecast method of PV power based on Ensemble Adaptive Boosting Random Forests is proposed.This method integrates the random forest and the AdaBoost to mine the nonlinear relationship between PV power and meteorological elements with numerical weather prediction and historical power data of PV plants to obtain accurate forecast results.First,the initial weights are set for training samples,and sub-training sets with weights are generated by bootstrap sampling to train the decision trees.Then,the results of the random forest can be obtained by averaging the decision trees.Furthermore,by calculating the error rate and assigning a corresponding weight to each random forest,the result of the weighted random forest is obtained.Finally,a high-precision spot forecast method of PV power is achieved through iterative optimization.In the case studies,the actual data of four PV plants in Ningxia are used to verify the accuracy of the proposed method compared with other single-model methods.Secondly,a probabilistic forecast method of PV power based on Multi-Model Stacking Quantile Regression is proposed.This method uses the Stacking ensemble model,combined with the quantile regression and the kernel density estimation(KDE)to obtain reliable probabilistic forecast results.First,under the Stacking framework,sub-models such as the random forest,the support vector machine,and the XGBoost are combined to reduce the forecast error of a single model.Then,the quantile regression is integrated on the basis of Multi-Model Stacking to obtain the quantile results and the prediction interval of PV power.Furthermore,using the KDE,the obtained quantile results are converted into probability density functions.Finally,through iterative optimization,a reliable probabilistic forecast method of PV power is achieved.In the case studies,the reliability and sensitivity of the proposed probabilistic forecast method is verified with the data of a PV plant in Shandong compared with other forecast methods.