Research On Photovoltaic Power Prediction Method Based On Deep Learning Hybrid Model

In order to achieve the goal of carbon neutrality and carbon peaking,the development of new energy has received unprecedented Attention and research at home and abroad.In the context of China’s carbon neutrality and carbon peaking goals,photovoltaic energy has great potential and is bound to become one of the main energy sources of the future.Despite the rapid development of solar energy,there are still several challenges in grid-connected PV.The capacity of PV power generation systems is subject to uncertainty and uncontrollability due to environmental conditions.Accurate prediction of PV power can effectively reduce the uncertainty of energy connected to the grid and improve the stability and utilization of the system.Based on this,this paper focuses on a deep learning model approach and firstly designs a two-channel SSA-CNN-Bi LSTM model for PV generation power prediction.Then,an improved Att-GN-SCB model is proposed for the problem of low prediction accuracy of specific weather patterns.Finally,a transfer learning-based PV power prediction method is proposed for the problem of low historical data of new PV systems.The main research contents are as follows.Firstly,for the problem of low prediction accuracy of traditional hybrid models,a dual-channel SSA-CNN-Bi LSTM model prediction method based on deep learning is proposed in this paper.First,the pre-processed historical sequences are decomposed using singular spectrum analysis to obtain high-frequency component and low-frequency component feature maps.Then,two CNN-Bi LSTM units are used as feature extraction networks for the high-frequency feature maps and low-frequency feature maps,respectively.Finally,the high-frequency features and low-frequency features are integrated by a fully connected layer to predict the future generation power values in advance.Then,to address the problem of degraded model performance for highly fluctuating weather types,an improved Attention network is proposed in this paper without introducing a other models.The structure takes into consideration the different influence of different weather types on the prediction day,and introduces an activation unit to weight the activation of the historical series and similar day series.This improves the model’s feature representation of extreme weather types and obtains more accurate power generation prediction results.In addition,this paper combines the advantages of Res Net and GLU networks and proposes a GN network.On the one hand,the input vector is further globally calibrated through the gating mechanism to carve the input sequence in a more refined way.On the other hand,the introduction of GN network can mitigate the model gradient disappearance/explosion problem.For the problem of weak model generalization ability due to too little historical data at the early stage of PV plant operation,this paper proposes three transfer scenario learning methods based on deep learning models: global parameter fine-tuning model,local parameter freezing model,and global parameter freezing model on the foundation of Att-GN-SCB model.The effectiveness and applicability of the different transfer models are verified by comparing them with the benchmark model in different scenarios.