Ultra-short-term Prediction Of Photovoltaic Cluster Power Based On Graph Neural Network

With the rapid development of the world economy,environmental and energy issues have become increasingly prominent.As a clean and renewable energy,solar energy has received extensive attention,developing new energy sources such as wind and solar has become the first choice to solve energy and environmental problems.China has vast territory,abundant resources and a lot of light resources.The efficient use of photovoltaic power can solve many sustainable development problems.Large scale and clustering will become the trend and main characteristics of photovoltaic power stations in China.However,The rapid development of photovoltaic power stations and the large-scale photovoltaic grid connection are bound to bring serious challenges to the safe operation and stability of traditional power systems.Electric power department puts forward higher requirements for the reliability and accuracy of photovoltaic power prediction,Meanwhile,it pays more attention to the overall forecasting results of photovoltaic output power in an area.Therefore,improving the power prediction accuracy of photovoltaic clusters is very important and necessary.The existing researches on photovoltaic power forecasting mostly focus on the power prediction of a single station,and relatively few researches have been carried out on regional clusters.The traditional neural network prediction method makes less use of the spatiotemporal characteristics between different stations when predicting the cluster power,and the prediction accuracy is limited when the output power fluctuates greatly.Therefore,this paper proposes an ultra-short-term power prediction method for photovoltaic cluster that considers spatiotemporal correlations.The main work is as follows:1.The essence of predicting photovoltaic clusters power is actually a non-linear fitting problem based on existing data.Therefore,the quality of historical data and its correct use directly determine the prediction results.This paper firstly conducts preprocessing research on cluster historical data,including normalization operation,vacancy value interpolation,abnormal data repair and other processes,and then analysis the historical data of different power stations in the regional cluster from the spatial dimension and the time dimension.2.A prediction model based on GCN network and GRU network is proposed.Establish a prediction model with the historical output data of different stations in the region as the model input and the overall power as the output for a period of time in the future.Among them,the GCN network is used for spatial feature extraction,and the GRU network is used for temporal feature extraction.Combining the two models can take into account both the time and space dimensions interfering factors,prediction accuracy of overall output of photovoltaic cluster is significantly improved3.In order to further optimize and improve the GCN-GRU model performance,a photovoltaic cluster prediction model based on attention mechanism is proposed.This method introduces an attention mechanism in the graph convolution layer,and reweights the influence of the output power states from different power stations,that is,the importance and weight of the input data features are adaptively allocated,In this way,the global trend of the overall power output state is captured,so that the important features in the original data can be more effectively applied.At the same time,based on the provided data set,the BP neural network is continuously used to optimize the prediction results.The simulation results of the model show that compared with the traditional model,the proposed model has higher prediction accuracy,especially in complex weather conditions.