Research On Multi-task Learning Based Method For Power Forecasting In Wind And Solar Power Station Cluster

Large-scale and clustering are one of the main characteristics of the distribution of wind farms and solar power stations in China.With the increase of the grid-connected scale of wind and solar power,the dispatch department puts forward higher requirements for the accuracy and reliability of wind power and solar power forecasting,and pay more attention to the power forecasting results of wind farms and solar power stations in the region.The existing wind and solar power forecasting methods mostly focus on single station,and there is less research on the power forecasting in a region.Therefore,focusing on improving the power prediction accuracy of wind and solar power station cluster,the power forecasting models of wind/solar power stations based on the improved fully connected neural network and the joint forecasting models based on multi-task learning are proposed.The main work and conclusions are as follows:1.Explor the volatility and relevance of the output of wind and solar power stations in the regionFirst,the cross-correlation coefficient and autocorrelation coefficient are used to analyze the correlation the power between single station to the wind farm/solar stations cluster.Then,the volatility is take as the evaluation index to analyze the probability distribution and time effect of the power fluctuation.Finally,The smoothing coefficient is takes as the evaluation index to analyze the spatial effect of the fluctuation of wind and solar power.It is found that as the scale of wind and solar power stations increases,the autocorrelation of the output of wind and solar power stations increases,and the volatility decreases.2.Establish a wind or solar stations power forecasting model based on improved fully connected neural network.power forecasting model of wind-solar stations is established based improved fully connected neural network which use the stacked autoencoders to initialize the neural network parameters layer by layer.The input of the model is multi-stations numerical weather forecast and the output is the multi-stations power.Take the operation data of eight wind farms and seven photovotalic stations distributed in the southern of China as example to verified the validity of the proposed model.The results show that the proposed model has strong learning ability,can better extract the input data characteristics,and effectively improve the prediction accuracy of each station in the area.3.Establish a joint forecasting of wind and solar stations power model based deep multi-task learningEstablish a wind-solar stations joint power forecasting model base on convolutional neural network and multi-task learning.This model take numerical weather forecast data of multiple wind farms and photovoltaic stations as input,and the output are total power of wind farms and photovoltaic stations.The results show that the propose model can not only improve the forecasting accuracy but also realize the joint forecast of wind and solar power.