Application In Solar Irradiance Prediction Based On Deep Learning

With the development and progress of society,people’s demand for energy is increasing day by day,which also forces the transformation and upgrading of energy structure relying on traditional fossil fuels to develop in the direction of new energy.Solar is a kind of inexhaustible clean energy,but with the increase of photovoltaic installed capacity year by year,it brings great challenges to the grid connection of photovoltaic power stations.The accurate prediction of global horizontal irradiance(GHI)can effectively monitor the grid connection process,deal with the grid fluctuation caused by sudden weather changes,improve the grid connection efficiency and enhance the competitiveness of enterprises.In recent years,modeling methods such as statistical learning and neural network have been successfully applied to the prediction of solar irradiance.As a rapidly developing intelligent computing method,deep learning is introduced into the prediction and modeling example of solar irradiance in this thesis because of its strong nonlinear modeling ability.Deep echo state network(DESN),as a method with the characteristics of reserve pool calculation,is widely used because of its fast training and efficient calculation.Recurrent neural networks(RNN)has great advantages in processing time series data.Long short-term memory(LSTM)and gated recurrent unit(GRU)are its improved algorithms,which can not only effectively improve the network performance,but also solve the problems in gradient calculation.The research contents of this paper include the following aspects:(1)This thesis briefly introduces the process of photovoltaic power generation and the influence of solar irradiance on it,also summarizes the background and significance of predicting solar irradiance.The prediction process of solar irradiance is analyzed,and the evaluation indexes to measure the performance of the model are introduced.(2)The echo state network and deep echo state network with different training methods are studied.In addition to generalized inverses(GI)algorithm,the output training algorithm of deep echo state is improved by different regression algorithms.Among them,kernel ridge regression(KRR)and support vector regression(SVR)have greatly improved the nonlinear modeling ability of the models because of the use of kernel function,and these models are applied to the prediction and modeling of solar irradiance.In order to measure the effectiveness of the model,it is compared with persistence algorithm,BP algorithm and SVM algorithm under the same conditions.The experimental results show that the prediction effect of deep echo state network is better than persistence algorithm,BP algorithm,SVM algorithm and simple echo state network.While DESN-KRR and DESN-v-SVR methods using kernel function have better modeling effect.(3)The modeling method of recurrent neural network(RNN)in solar irradiance prediction is studied.For the problem that RNN is prone to gradient disappearance and gradient explosion in the training process,the improved algorithm is further studied,that is,the gated recurrent neural network,including LSTM and GRU and their bidirectional structure,which not only solves the problem of gradient calculation,but also solves the problem of long-term dependence.Back propagation through time(BPTT)algorithm and gradient descent Adam optimization algorithm are introduced.Based on the above methods,the clear sky index prediction model is established.In order to verify the effectiveness of the model,it is applied to the solar irradiance prediction modeling 1-hour ahead,and compared with the deep echo state network.The experimental results show that under the same conditions,the modeling effect of recurrent neural network is better than that deep echo state network.In the recurrent neural network,the prediction effect of LSTM and GRU is better than RNN.The corresponding bidirectional structures Bi LSTM and Bi GRU further improve the prediction accuracy of the network.Moreover,compared with Bi LSTM,Bi GRU has better tracking effect where the target changes rapidly,and the overall error is relatively more stable.Because Bi GRU has fewer network parameters and faster training speed than Bi LSTM,it saves a lot of time and computational cost.The recurrent neural network model is also applied to the multi-step prediction modeling of solar irradiance.The experimental results show that the multi-step irradiance prediction model has higher prediction results than the single step,which verifies the effectiveness of information sharing in the multi-step prediction model.