Ultra Short Term Prediction Of Photovoltaic Power Based On Machine Learning

At present,China is facing the dual challenges of increasing energy demand and energy structure transformation.The traditional thermal power generation has gradually been unable to meet the social requirements of energy conservation and emission reduction.The development of clean energy with wind energy and solar energy as the main components is the inevitable trend of energy development in the future.In recent years,the proportion of grid connected photovoltaic power generation has increased year by year.However,after grid connected photovoltaic power generation,affected by irradiance,temperature,humidity and other conditions,the electric energy emitted by photovoltaic power generation system has certain volatility and periodicity,which will affect the stable operation of power grid.In order to accurately evaluate the impact of photovoltaic access on the power grid,this paper mainly uses three methods:error feedback(BP)neural network,particle swarm optimization(PSO)BP neural network and long-term memory(LSTM)depth neural network to predict the ultra short-term photovoltaic power under different weather and season types,so as to assist the power grid dispatching decision-making.This research is of great significance to accurately predict the ultra short-term photovoltaic power.(1)This paper expounds the principle of photovoltaic power generation,the composition and structure of photovoltaic power generation system,the source of data set and the method of data preprocessing,analyzes the meteorological conditions affecting photovoltaic power prediction,and uses the Pearson correlation coefficient method to make it clear that the main factors affecting photovoltaic power are irradiance and temperature.(2)Taking the historical power,temperature and irradiance of photovoltaic power station as the input characteristics,a BP neural network prediction model is established to predict the ultra short-term photovoltaic power.The prediction results show that the prediction effect of BP model in different seasons is quite different,and the prediction effect is the worst in winter.The mean square error(MSE)and mean absolute error(MAE)are 0.0368 and 0.1277respectively.Aiming at the disadvantage that BP neural network is easy to fall into local extremum,PSO algorithm is used to optimize the initial weight and threshold of BP neural network,so as to improve the prediction accuracy of BP neural network,and the prediction error of BP model after PSO optimization is significantly reduced.The MSE and Mae predicted in winter are 0.0176 and 0.0784,respectively,which are reduced by 0.0192 and0.0493 respectively.(3)Considering the time series characteristics of photovoltaic power data and the characteristics that the long-term and short-term memory network is good at processing time series data,build the LSTM model,compare and analyze the influence of the parameters of the LSTM model time step,the number of neural network layers,dropout value and optimization algorithm on the prediction results,so as to gradually determine the selection of model parameters.Among them,the time step and the number of neural network layers mainly affect the prediction speed,and the setting and optimization algorithm of dropout mainly affect the prediction accuracy.The prediction results show that under different weather and seasonal conditions,the MSE and Mae predicted by LSTM model in the three models are the smallest,but when the photovoltaic output power fluctuates greatly,the goodness of fit(R~2)of LSTM model will decrease.Under cloudy conditions,MSE predicted by LSTM model decreased by 0.0322 and 0.0202 respectively compared with BP model and PSO-BP model,but R~2decreased by 0.028 and 0.038 respectively.Whether using PSO algorithm to optimize the initial parameters of BP model,or using LSTM structure to deepen the depth of neural network and further mine data features,the prediction accuracy can be improved.Compared with LSTM model,the prediction effect of LSTM model is better and can be better applied to the field of ultra short-term prediction of photovoltaic power.