Research On Short-Term Power Forecasting Method Of Photovoltaic Power Station Based On Machine Learning

As the proportion of photovoltaics in the power grid continues to increase,photovoltaic power with strong randomness and volatility is bound to have a serious impact on the power system.Prediction of photovoltaic power can effectively solve the problem of efficient photovoltaic power generation access to the grid.In this thesis,the machine learning model is used to analyze the short-term prediction of photovoltaic power generation.The main tasks are as follows:First,based on the principle of photovoltaic power generation,the influencing factors of photovoltaic power generation are analyzed.From the perspective of data visualization,the factors that affect the photovoltaic output are analyzed.The specific analysis is based on stable weather and sudden weather,and several important influencing characteristics are selected.Through feature analysis,temperature difference,The distance characteristic of the distance peak,the preprocessing of the original data of the power station.In order to ensure the reliability of the data,the outliers are identified,deleted and filled,and the data is standardized.Secondly,the tree model and neural network model in machine learning are analyzed.And The mathematical principles of three algorithms commonly used in machine learning e Xtreme Gradient Boosting(XGBoost),LightGradient Boosting Machine(LightGBM),and Long Short Term Memory(LSTM)networks are introduced in detail.This thesis studies and analyzes the combined model prediction method to enhance the photovoltaic power generation.Introduce the process of indirect and direct forecasting of photovoltaic power.Then aiming at the irradiation intensity of photovoltaic power station,an indirect prediction method of photovoltaic power generation based on LSTM and Lightg GBM was proposed.In the first step,according to the data characteristics of radiation intensity,LSTM and LightGBM prediction models are constructed.The model parameter is then determined by the experiment.In the end,the prediction is made based on the actual photovoltaic power station data.Take the result of LightGBM model as an input of LSTM,and then the reciprocal error method is used to compare the above two Time series data results weighted prediction.The prediction results are analyzed through evaluation indicators,and the results show that the photovoltaic power plant radiation intensity prediction method based on the combined model of LSTM and LightGBM has higher prediction accuracy.Finally,in the part of direct photovoltaic power prediction method,the photovoltaic power prediction model of multi-layer fusion method is adopted.First test the XGBoost model with photovoltaic data.It is found that the tree model XGBoost has lower prediction results under stable weather,but better prediction results under abrupt weather,indicating the limitations of a single model prediction,and then a multi-layer combination of three models Improve the prediction accuracy of photovoltaic power.Then,Pearson’s method is used to analyze the degree of correlation between each characteristic factor.Second,it is proposed to use the LightGBM model to verify the impact of different data combinations on the prediction effect.It is found that the combined effect of several data features has a greater impact on the photovoltaic output,and then focuses on the photovoltaic data.Feature construction feature engineering,find out new features with higher correlation and add them to the PV power prediction model.In the end,an example is used to verify the validity of the multi-layer fusion solar power prediction model proposed in this thesis.At the same time,it shows that adding feature engineering can better improve the prediction accuracy.