Data-driven Photovoltaic Power Generation Forecasting

Photovoltaic(PV)power output is easily influenced by external factors such as solar radiation,atmospheric temperature,wind speed,rainfall,wind direction,and humidity.Large-scale integration of PV power into the power grid can affect the power balance within the grid.Accurate prediction of PV power generation is beneficial for precise scheduling and arrangement of the power system.It can support decisions related to peak-shaving and frequency regulation of the power grid,and is of significant importance for maintaining the safety,stability,and efficient operation of the power system.Based on this,this paper predicts PV power generation using semi-data-driven and data-driven models.In the semi-data-driven model,the jellyfish search optimization algorithm is used to optimize the single diode model parameters of the PV module,and the parameter correction formula is used to correct the parameters,achieving the prediction of the working state of the module under different conditions.In the data-driven method,the mapping relationship between meteorological and time data and power data is studied using artificial neural networks,and the accuracy of power prediction is improved through data classification.In the convolutional neural network data-driven method,the power time series and meteorological time series data are organized into a two-dimensional format,and the features of the two-dimensional data are extracted using convolutional neural networks,achieving higher accuracy power prediction.The research conclusions of this paper are as follows:(1)The semi-data-driven method utilizes the intelligent optimization algorithm of jellyfish search to avoid direct solving of nonlinear model unknown parameters and to optimize model parameters accurately,ensuring the accuracy of the model parameters.(2)In the semi-data-driven method,the single-diode model,through the parameter correction formula,can achieve relatively accurate prediction of the working conditions of photovoltaic modules under laboratory conditions.However,the accuracy of the model decreases under high-temperature and low-light conditions because the parameter correction formula is not accurate for parameter correction under low-light and high-temperature conditions.(3)In the data-driven model,using K-means clustering algorithm to classify data based on power data,it was found that data classification had a positive effect on network training,and power prediction accuracy was significantly improved.(4)In the prediction results of the data-driven artificial neural network model,it was found that the meteorological data and power data have different mapping methods under different weather conditions.This conclusion indirectly confirms the effectiveness of classification.In addition,it was also found that there are cases of large fluctuations in power output that cannot be reflected by meteorological data,indicating that there are other factors that can have a huge impact on photovoltaic power generation.Furthermore,there is a certain relationship between the large fluctuations in power output and the rapid fluctuations in wind direction and the large deviation of the wind direction angle.(5)From the perspective of network training and network structure of the datadriven artificial neural network power prediction model,it cannot learn the characteristics of the variation patterns between time series data and the influence of data such as wind direction on power prediction.Therefore,time series prediction methods are needed to achieve accurate power prediction.(6)The convolutional neural network model that takes power and meteorological time series data as input has improved the prediction accuracy of photovoltaic power.At the same time,classifying the data of the convolutional network model also improved the accuracy of power prediction,indicating that classification can improve the prediction accuracy of data-driven models.