Forecasting Photovoltaic Power Using The Wavelet Transformation And Convolutional Neural Network

With the global energy crisis intensifying and the environmental pollution caused by fossil fuels increasing,lots of countries pay more attention to the strategy of integrating new energy into power grid for power generation.Because solar energy is inexhaustible and environmental,it has rapidly penetrated into modern power networks and energy systems in recent years.However,the solar power generation technology is affected by many other factors,and there are strong uncertainty and intermittent in photovoltaic(PV)power sequences,which can strongly affect the stability and reliability of power grid operation if they cannot be predicted accurately or the reliable basis cannot be provided for power grid dispatching.In order to reduce the negative impacts of photovoltaic power generation on power system,we need to forecast the photovoltaic power as accurately as possible to know how to schedule the operation of power grid.Therefore,it is necessary to find a prediction method with high precision.A new deterministic forecasting method of PV power based on the wavelet decomposition and convolutional neural network is proposed.At first,wavelet decomposition is used to decompose the original PV power sequence into a series of different frequency sub-sequences.Then,convolution neural network sub-models are built based on those sequences.The input of each sub-model corresponds to different sub-sequence,and the output corresponds to the predicted value of the input.At last,the prediction series obtained from each sub-model is reconstructed and the complete PV power forecast result is acquired.A probabilistic prediction model which combines the above deterministic forecasting model with quantile regression analysis is proposed.First,the error sequences between true and forecasting values are calculated.Then,for a given confidence level,quantile regression at different quantiles is calculated to determine the limits of prediction interval,so as to forecast value corresponds to a prediction interval.This interval indicates the range that the forecasting value could vary.Last,the probability of the true value falls in the prediction interval and the interval sharpness of the prediction interval can be calculated.Experiments on deterministic and probabilistic prediction model are carried out.The experiment datasets are collected from two photovoltaic power plants in Europe,including all power values in 2015.Platform MATLAB2014 a is for deterministic forecasting and R language platform is for probabilistic prediction.In this paper,BP neural network(BPNN),support vector machine(SVM)and support vector machine(SVM)based on wavelet decomposition(WD)are selected for comparison.Analysis and comparison based on predictive evaluation index prove the effectiveness and reliability of the proposed method.The PV power forecasting method based on wavelet decomposition and convolutional neural network can effectively extract the internal features when dealing with time series with strong non-linearity.In deterministic prediction,the proposed model performs better than other methods with higher accuracy and lower volatility of evaluation index.In probabilistic prediction,it has high range coverage and more precise prediction interval.