Short Term Photovoltaic Power Generation Prediction Method Based On Machine Learning

In recent years,with the rapid development of energy technology and the increasing scarcity of fossil fuels,renewable energy represented by photovoltaic(PV)power generation has developed rapidly.The output power of PV power generation is influenced by various factors and exhibits characteristics such as instability and strong randomness,which can negatively impact the efficiency of PV power generation and the stability of the power grid.Timely and accurate prediction of PV power generation is crucial for ensuring the stable operation of power systems and promoting the widespread application of PV technology.Traditional PV power generation prediction methods cannot effectively mine and extract features from historical PV power generation data to meet the requirements of current practical applications.Moreover,traditional prediction methods do not yield satisfactory results in small-sample scenarios.This study aims to address the shortcomings of existing PV power generation prediction methods and focuses on the following research works:(1)Data mining of the PV power generation dataset: Firstly,this study conducts data preprocessing on the PV power generation dataset,including error data identification,missing data filling,and data normalization.Secondly,two auxiliary features are constructed to improve the training effectiveness of the model.Then,correlation analysis and other methods are employed to investigate the features that influence PV power generation.Quantitative analysis of the correlations between various features is conducted,and multiple features with strong correlations are selected as input variables for the prediction model.(2)Development of a PV power generation prediction model based on an improved embedding encoding method and convolutional self-attention Transformer-CNN.The model uses local convolution and residual connection as the embedding encoding method,and employs convolutional sparse self-attention.Meanwhile,a CNN model based on multi-channel one-dimensional convolution is introduced to learn features among relevant variable trends from multiple dimensions.The improved Transformer model is then connected with the CNN network to enhance the overall performance of the model.The prediction performance of the model on the PV dataset is improved through data preprocessing and feature construction.The experimental results demonstrate that the proposed improved model has good fitting and prediction performance.(3)Development of a small-sample PV power generation prediction method based on a fusion model.Firstly,a PV power generation prediction model based on Stacking is proposed,and multiple comparative experiments are conducted to select base learners and meta-models.Furthermore,an improved grey wolf optimization algorithm is proposed to optimize the parameters of the model,achieving significant performance improvement.Experimental results show that the improved grey wolf algorithm has obvious effects on optimizing model parameters,and the proposed fusion model has significantly better performance in small-sample scenarios compared to comparative models.