Research On Solar Irradiance Prediction Algorithm Based On Deep Learning

As the global population grows,energy consumption is increasing.In developing countries,traditional thermal power generation still accounts for a large proportion,but combustion will produce a large number of harmful gases and pollute the atmospheric environment.In order to cope with climate problems and energy consumption problems,countries around the world have put forward the goal of "carbon neutrality,carbon peak" to promote the development of green energy.As a new type of pollution-free clean energy,solar energy is an effective way to achieve this goal.Accurately predicting the changing trend of photovoltaic power generation can provide data guidance for energy storage and energy supply in the power system,which is conducive to the adjustment of energy structure.Solar irradiance is closely related to photovoltaic power generation output,so it is widely used in photovoltaic power generation forecasting research,but the irradiance is affected by many external factors,resulting in complex data characteristics,and it is difficult for traditional neural networks to accurately mine their internal characteristics.Based on the above background and problems,this paper analyzes the specific characteristics of irradiance time series data,improves the current deep learning prediction model,and achieves more accurate irradiance prediction.The main research contents of this article include:(1)This paper proposes a series of data processing optimization methods for the problems existing in the irradiance prediction process,including data decomposition,weighted selection of input features,etc.to improve the learning ability of neural networks.At the same time,the bidirectional LSTM model is selected for irradiance prediction.The experimental results show that these processing methods can effectively improve the prediction effect.(2)Based on the problem of poor prediction effect of traditional neural networks,on the basis of wavelet transform,the short-term irradiance prediction method of Long Short-Term Memory Neural Network(LSTM)combining Convolutional Neural Network(CNN)and Attention Mechanism is creatively proposed.Referred to as WCNN-ALSTM.This model uses the wavelet transform method to decompose the irradiance to obtain the high and low frequency sub-sequences of the original data,and each sub-sequence is extracted by CNN separately and then learned by ALSTM.After wavelet decomposition noise reduction and feature extraction of CNN,ALSTM can adaptively focus on more important input features from different dimensions.Experimental results show that the model proposed in this paper has better predictive power on multiple time steps,and the mean absolute percentage error(MAPE)is reduced by 22% compared with LSTM.(3)Based on the research of WCNN-ALSTM,a multi-channel(MC)hybrid prediction method(MC-WT-CBi LSTM)combining convolutional neural network and bidirectional Long Short-Term Memory Neural Network(Bi LSTM)is proposed for the shortcomings of the prediction process.On the basis of the wavelet transform,the model selects the feature data with the greatest correlation with the irradiance data for multi-channel parallel learning,and at the same time,for the before-after correlation of the irradiance sequence,the two-way long and short memory neural network is used to obtain the overall change characteristics of the sequence,thus further improving the prediction results.Compared with the predictions of the LSTM model,MAPE was reduced by 34%.