Research On Intra-hour Solar Irradiance Prediction Model Based On Deep Learning

With the increasingly global energy crisis and environmental issues,governments have focused attention on renewable energy sources.As a kind of clean energy with a broad application prospect,solar power generation has developed rapidly in recent years and is becoming crucial renewable energy.The integration of solar power generation into an electrical network intensifies the complexity of the grid management and the continuity of the production/consumption balance due to their intermittent and unpredictable nature.Solar irradiance variations are the most critical factor that leads to the instability of solar power generation,so the accurate prediction of solar irradiance plays an increasingly significant role in electric energy planning and management and has important application value.According to the characteristics of irradiance time series,five kinds of network structures and methods are proposed and designed to realize the accurate prediction of irradiance based on the LSTM(long short-term memory),GRU(gated recurrent unit),and Elman networks in RNN(recurrent neural network).The main work is as follows:(1)The correlation between weather forecast parameters and irradiance is analyzed,and the temporal characteristics of irradiance are studied.Six input features are selected,which contain three weather forecast parameters,temperature,humidity,and weather type,plus the time parameters of the month,day,and hour.On this basis,a solar irradiance forecasting method of the GRU model based on the weather forecast is proposed.The experimental results show that the GRU network with six input features can achieve accurate 24-hour irradiance prediction;Compared with BPNN(back propagation neural network,BPNN)and RNN,the root mean squared error(RMSE)of GRU is reduced by 23.3% and 11.9%;compared with the LSTM,the training time is reduced by 36.6%;the forecast skill(FS)is 0.4201.(2)The irradiance prediction model based on wavelet transform and Elman network is proposed.In this model,firstly,the irradiance time series is decomposed into different frequency sub-series by wavelet transform.Then,each sub-series is input into the Elman network to train,test,and predict.Finally,the predicted irradiance value is obtained by wavelet reconstruction of the wavelet coefficients predicted by each subseries.The periodic low-frequency information generated by the earth motion and the high-frequency information caused by cloud motion in the irradiance sequence can be separated by wavelet transform.Elman can be used to predict the information of different frequencies separately,and its data correlation is more potent,so the accuracy is higher.The simulation results based on the measured data show that the proposed model has good performance in hourly irradiance prediction,and the forecast skill(FS)reaches 0.7590.(3)The irradiance prediction model based on CEEMDAN(complete ensemble empirical mode decomposition with adaptive noise)data decomposition method and CNN-LSTM model is proposed.After CEEMDAN decomposition of irradiance data,IMFs(intrinsic mode functions)and R(residual)components with different frequency components were obtained.According to the structure characteristics of IMFs and R data in a period of time,five kinds of input structures of CNN-LSTM are constructed and discussed in detail.The results show that the performance of CNN-LSTM-V is the best.Meanwhile,the interpretability of the model is also the strongest.The CEEMDAN-CNN-LSTM model is verified and compared on four different climate data sets.The result shows that the prediction accuracy of the model is high,and the robustness is strong.(4)The LSTM network for irradiance time series prediction is constructed,and two input structures with different input features are designed.The LSTM-MLP model with a dual branch input is proposed.Two different main inputs and two different auxiliary inputs are designed for the network.The prediction performance of the above six structures with different lagging time is discussed.The experimental results show that the performance of the model is the best when the historical irradiance and meteorological data are the main input,and the weather forecast parameters at the next time are as the auxiliary input.This indicates that the next time weather forecast parameters play an important role in improving the prediction accuracy of the model.(5)A WPD-CNN-LSTM-MLP hybrid deep learning irradiance prediction model is proposed.Based on LSTM-MLP double branch,the new model is further improved to multi-branch and multi-input structure,and WPD(wavelet packet decomposition)and CNN(convolutional neural network)are added to extract feature information.The irradiance sequence is decomposed into four channels by WPD,and each channel is further extracted by CNN,then which are input into the LSTM network.Historical irradiance and meteorological parameters are input into the LSTM network as the second input.The weather forecast parameters,which are the third input,are input into the fully connected layers,and the final prediction results are obtained from the output.In this model,the frequency domain information and the time domain information are effectively fused through different inputs to achieve the purpose of learning frequency and temporal correlations jointly.The experimental results show that the proposed WPD-CNN-LSTM-MLP model with multi-branch and multi-input improves prediction accuracy.