Ultra-Short-Term Solar Irradiance Prediction Based On Multi-model Fusion

Photovoltaic power generation is a renewable energy generation technology that converts solar energy into electricity,which not only provides a high-quality,clean and sustainable solution to the energy crisis,but also helps to reduce dependence on fossil fuels and reduce environmental impact.However,due to the instability of photovoltaic power generation,it brings serious safety and efficiency problems to power grids and photovoltaic power generation institutions.To this end,it is necessary to accurately predict the solar irradiance on the ground.This is not only a key step to improve the prediction accuracy of photovoltaic power generation,but also to ensure stable power supply,optimize grid integration,improve market operation efficiency,and improve the utilization of light resources and photovoltaic facilities.important means of efficiency.On the basis of summarizing the current research status in the field of photovoltaic prediction at home and abroad,this paper studies the ultrashort-term prediction of ground solar irradiance,innovatively combines satellite data with meteorological data and historical irradiance data,based on different principles Three integrated ultra-short-term solar irradiance prediction models were constructed,and the model effects were mutually verified at two observation sites.The main research results are as follows:1.With reference to the relevant knowledge of physics,feature engineering is carried out on the original data,and multiple features that are helpful to improve the performance of the model are generated,and the clear sky index is used as an intermediate variable to convert the predicted ground solar irradiance into the predicted clear sky index,and The superiority of the method is proved by comparative experiments.2.Optimize the parameters of a single machine learning model,select the base model for stacking according to the prediction performance,and improve the traditional stacking model fusion method by adding original features and using pseudo-labels during the training process,and through ablation Experiments prove the effectiveness of the innovation.3.Construct LSTM_CNN₁D,a fusion model based on feature-level long-term short-term memory network and convolutional neural network.LSTM is used to extract temporal features in data,CNN₁D is used to extract spatial features in data,and the features are stitched together to output the final prediction results.And the regularization and Early Stopping mechanisms are used during the training process to prevent the model from overfitting,and the performance of the final model is better than the fusion model of machine learning.4.Considering the characteristics of machine learning and deep learning model extraction features,use adaptive weight methods on different sites to carry out decision-level weighted fusion of machine learning model prediction results and deep learning prediction results,for some unreasonable Forecasting,use the improved K-nearest neighbor algorithm to further correct the weighted results.And a variety of measures were taken to avoid over-fitting during the research process.Finally,the prediction results of the three fusion models are compared,and the results show that the weighted fusion model modified based on the improved KNN algorithm has the highest prediction accuracy,indicating that the model has strong adaptability in solar irradiance prediction.5.This study proposes three integrated solar irradiance prediction models,and the performance of the models is better than that of a single model,which illustrates the necessity and effectiveness of model fusion.And through more accurate prediction of ground solar irradiance,it plays an important role in improving the stability of the grid,increasing the photovoltaic consumption capacity of the grid,reducing the economic losses caused by photovoltaic power rationing,and improving the operation and management efficiency of photovoltaic power generation institutions.