Research On Short-Term Power Prediction Method For Photovoltaic Grid-Connected System

Driven by the energy and environment problems,solar energy has attracted more and more attention all over the world.As one of main forms of solar energy development,photovoltaic power(PV)generation installing capacity has increased rapidly.Compared with traditional thermal and hydroelectric power generation,PV power generation is affected by many factors,such as solar radiation intensity,temperature,humidity,sunshine duration and so on.PV output power has obvious intermittence,randomness and uncertainty.Therefore,in order to ensure the stability of grid-connected operation and improve the absorption ability of PV power generation,it is necessary to propose an accurate PV power prediction method,which provides a scientific reference for the planning and scheduling of the power sector.The main contents of this thesis are as follows:(1)The principle and model building process of the commonly short-term power forecasting methods for PV power generation were introduced,including extreme learning machine(ELM),least squares support vector machine(LSSVM)and BP neural network.Person coefficient method was used to analyze the main factors affecting photovoltaic power generation,and the principal component analysis method was used to reduce the input of the model,improve the training efficiency of the model,and establish the prediction model of photovoltaic power generation.(2)The weights of input layer and threshold of hidden layer of traditional ELM were chosen randomly.The accuracy of the model still needs to be improved.Therefore,a short-term PV power prediction method based on improved ELM was proposed in this thesis.Levenberg-Marquardt(LM)algorithm was used to modify the parameters of the ELM model.The optimal network of the ELM prediction model was obtained.The short-term power prediction was carried out based on the improved ELM model.The effectiveness of the improved model was validated by the simulation.(3)The single prediction model has limited ability to analyze data.In order to improve the prediction accuracy of rainy weather with large fluctuation,two other non-linear models based on improved ELM was introduced in this thesis.A combined forecasting method of PV power generation based on similar samples and improved entropy weight method was proposed.Firstly,the samples of PV power and meteorological data were clustered to form similar day samples under different weather types.The weighted Euclidean distance,cosine similarity and grey correlation coefficient were used to define the comprehensive selection index of similar samples,so as to select model training samples.Then GA-BP,PSO-LSSVM and LM-ELM were used to predict the generation power on the forecasted day.Finally,the improved entropy weight method was used to dynamically set the weight values of the single prediction model in different prediction periods.The single model was reasonably combined to obtain the prediction results of photovoltaic power generation.The results of simulation show that the combined prediction model based on similar samples and improved entropy weight method has better prediction accuracy in weather with large fluctuation.