Research On Short-term Power Prediction Methods For Photovoltaic Power Plants

The instability of weather conditions make the output of photovoltaic(PV)power have strong fluctuations and intermittence.If the PV power can be predicted ahead of time,it can effectively reduce its impact on the grid and provide a reference for the dispatching of the grid.Based on the analysis of PV power prediction theory and meteorological influencing factors,the short-term prediction based on historical power data and meteorological data and the ultra-short-term prediction with lack of historical meteorological parameters are studied in this paper,respectively.Then the short-term and ultra-short-term prediction models are established combined with machine learning algorithms.The main tasks are as follows:Firstly,for short-term power prediction,a method of selecting similarity days and optimal similarity days based on clustering and grey relational analysis(GRA)is proposed.The K-means++ approach is applied to cluster the historical datasets,the meteorological characteristic vectors are established by using the historical meteorological parameters of the category of the forecasting day,the similarity days and the optimal similarity day are determined by the GRA algorithm.Then a short-term power prediction model is developed combined with the Elman neural network.The generation power of the forecasting day for the selected four seasons is predicted by modeling the historical datasets of a PV power plant provided by the Desert Knowledge Australia Solar Center(DAKSC)website.The prediction results show that the prediction effect of the model is better than that of the 8 selected comparative prediction methods.Secondly,for ultra-short-term power prediction,a SAGA-FCM-LSSVM-based ultra-short-term forecasting model is proposed.The SAGA-FCM algorithm,which optimizes the initial clustering centers of the fuzzy c-means algorithm(FCM)by the simulated annealing genetic algorithm(SAGA),is adopted to cluster the historical datasets.The least square-support vector machine(LS-SVM)technique is applied to map the nonlinear relationship between multivariate meteorological factors and power.Compared with BP neural network and LS-SVM model,this method has higher prediction accuracy.Furthermore,a chaos-BP neural network(C-BPNN)based ultra-short-term forecasting model is proposed for ultra-short-term power prediction in the absence of historical meteorological parameters.The prediction accuracy is higher than chaosgeneralized regression neural network(C-GRNN)method and chaos-LSSVM(CLSSVM)method.Finally,in order to combine the prediction models with the actual system and observe the results of power prediction intuitively,the upper monitor with the function of prediction and result query and printing is developed with MATLAB/GUIDE tool in the study.The comprehensive test results show that the prediction models presented in this paper have better prediction performance under different working conditions,which have certain reference value for the design of short-term power prediction technology for PV power plants.