Photovoltaic Power Prediction In Non-stationary Period Based On Time Series And Bayesian Neural Network

Due to the randomness and intermittency of photovoltaic power generation,large-scale photovoltaic power stations will have impact on the power grid when they are incorporated into the power system.Especially in the non-stationary period when photovoltaic power generation fluctuates violentiy,the impact will lead to more serious consequences,which will affect the safety of the power grid.Therefore,the prediction of photovoltaic power in non-stationary period has very imporrant research significance.Aiming at the problem of photovoltaic power prediction in the non-stationary period of photovoltaic output,this paper proposes a hybrid prediction model based on multi-objective optimization,which improves the prediction accuracy of photovoltaic power in the non-stationary period.The main research contents are as follows:(1)A non-stationary period discrimination method based on radial power ratio difference is proposed.It has been verified that this method can accurately identify the non-stationary output period,which is more accurate and efficient than the traditional method of dividing photovoltaic stationary output period and non-stationary output period by weather type.(2)The main influencing factors of photovoltaic power are qualitatively analyzed,and nonlinear correlation analysis is carried out.Several features with strong correlation are selected as the input of the model,which lays a foundation for subsequent power prediction.(3)Based on the prediction model of LSTM(Long Short-Term Memory),a Stack-LSTM prediction model is proposed.By constructing new feature factors to mine the feature information hidden in the data,the experiment shows that the prediction accuracy of this model can be improved by 9%-17%.(4)A hybrid prediction model based on multi-objective optimization was proposed,which combined the point prediction and interval prediction models to make full use of the advantages between the two models,and the prediction results were respectively optimized in two forms.The results show that the prediction accuracy of the optimization results based on point prediction is improved by 5%-8%.Compared with ordinary interval prediction,the optimization results based on interval prediction have narrower interval width and improved prediction accuracy by 15%-20%on the premise that the interval coverage rate remains unchanged.