Research On Prediction Technology Of Photovoltaic Power Based On Deep Learning

Energy consumption and environment change are two mutual restricted problem human facing in the process of social development.Solar energy is outstanding among clean energies because of its characterstics of large reserves,permanent,clean and pollution-free,which makes it is energetically developed by many countries in the world.However,photovoltaic power generation is stochastic and intermittent.If a large number of photovoltaic power generation systems are connected directly to the power grid,they will bring great challenges to the stability and safety of the existing grid.Prediction of photovoltaic power is conducive to solve this problem.This thesis combines the deep learning which is popular in recent years and the historical data of Baoding Yingli photovoltaic power plant to study the prediction technology of photovoltaic power.The research contents include the following parts:Firstly,the first part is data preprocessing,data analysis and data clustering.The historical data of photovoltaic power are processed with time window unification,missing value complementation,time interval transformation and data normalization.Then,the power historical data under different weather types and same weather types are analyzed separately according to their numerical values.A clustering algorithm based on the average power value of each data set under specific weather types and K-means is proposed.Based on the clustering results,the historical power data for a whole year are divided into four parts.Secondly,a photovoltaic power prediction model based on Deep Belief Network(DBN)is built.In order to solve the input problem of the forecasting model,a similarity day selection algorithm based on temperature is proposed.The power data of the similar day corresponding to the forecasting day and the meteorological data of the forecasting day are combined as the input of the prediction model.Then,based on the input and output of the forecasting model,the training set,test set and verification set are established for the divided histrical power data,and four prediction models based on DBN are thus established.The prediction examples results of four DBN forecating models show that this kind of forecasting model achieves a good prediction performence on four verification sets.Finally,a prediction model is also constructed based on LSTM which is a neural network specially used to process sequence data in deep learning.Because the network can not receive daily meteorological data,its inputs are only the power data of similar day of forecasting day.And,the prediction effect of this model is slightly worse than that of DBN prediction model because of this reason.In addition,the two deep learning prediction models are also compared with the traditional statistical ARIMA model.The comparison results show that the two prediction model based on deep learning is better than that of ARIMA model.