Short-term Photovoltaic Power Forecasting And Uncertainty Analysis Based On CIWOA-LSTM Network

In recent years,the installed capacity of photovoltaic(PV)power generation has increased rapidly,and the proportion of PV power generation is increasing in the total annual power generation.However,due to the complex and changeable meteorological factors,the PV power has the characteristics of randomness,volatility and intermittent.With the continuous increase of photovoltaic gridconnected capacity,it brings great challenges to the stable operation and dispatching of power grid.Accurate PV power prediction can provide reliable decision-making basis for power dispatching departments,and improve the power grid’s capacity to absorb renewable energy such as solar energy,so as to ensure the stable operation of power system.It is an effective way to address the aforementioned issues.At present,the accuracy of PV power prediction remains to be improved and most of the studies do not provide interval prediction result.This paper explores a PV power forecasting model based on the improved grey wolf algorithm(IGWO)optimized support vector machine(IGWO-SVM).And a deterministic short-term PV power prediction method based on Cauchy variation improved whale optimization algorithm(CIWOA)optimized long and short-term memory(LSTM)network(CIWOA-LSTM)is proposed.The prediction error of CIWOA-LSTM model is analyzed by nonparametric kernel density estimation(KDE)method,and the PV power prediction interval under given confidence is estimated.The main research contents are as follows :(1)Analysis the key affecting factors of PV power output.By analyzing the mechanism and physical model of PV power generation system,the main influencing factors of PV power output are selected.Through the grey correlation analysis of historical data of PV power station,the solar irradiance,temperature,wind speed and wind direction are selected as the input variables of PV power prediction model.(2)Research on short-term PV power prediction based on IGWO-SVM model.A PV power forecasting model based on support vector machine(SVM)is proposed.Then,aiming at the key parameters(C and gamma)are difficult to determine,IGWO is used to optimize the parameters of SVM.Compared with SVM,LSTM,PSO-SVM,GWO-SVM and WOA-SVM,the effectiveness of the proposed model is verified.(3)A deterministic short-term PV power prediction method based on CIWOA-LSTM model.Although the SVM-based model can achieve PV power prediction,SVM is insufficient to analyze the correlation characteristics of PV power station time series data.LSTM network has good timing data mining ability and it’s more suitable for PV power prediction.Therefore,a short-term PV power forecasting method based on LSTM network is proposed.Aiming at selecting the hyper-parameters of LSTM network(batch size and the neurons of hidden layer)depends on experiments and takes a lot of time.Firstly,an improved whale optimization algorithm(CIWOA)is proposed,then CIWOA is used to automatically optimize the hyper-parameters of LSTM.And superiority of proposed model is verified by compared with LSTM,PSO-LSTM and WOALSTM models.(4)PV power interval prediction based on KDE method.The mutation of input variables and the robustness of neural model will bring uncertainty to PV power prediction result.It is necessary to analyze the uncertainty of PV power prediction result to determine its fluctuation range.Therefore,the KDE method is proposed to analyze the prediction error distribution of CIWOA-LSTM model,so as to solve the problem that it is difficult to assume the error distribution parameters of PV output power in advance in practical engineering and the parameter method has low accuracy of interval estimation.And the prediction interval ranges under the given confidence levels of 80%,90% and 95% are calculated.Then,the Gaussian distribution method and Gamma distribution method are selected for comparison.The results prove that:(1)The accuracy of SVM model is improved by IGWO optimizing the parameters of C and gamma.(2)By Cauchy mutation strategy,CIWOA overcomes the WOA’s problems of premature development and local optimization,and has better ability of global search and local development.(3)The LSTM model optimized by CIWOA has better prediction accuracy in four seasons.(4)KDE method precisely describes the distribution of PV power at a given confidence level,and avoids the problem that the parameter method needs to assume the error distribution parameters of PV output power in advance.