| Weather prediction is related to national economy and people's livelihood.From traditional forecast to numerical forecast,with the progress of science and technology,especially the improvement of computing ability,the accuracy of prediction has been greatly improved.However,it is still limited by the degree of understanding the law of atmospheric change,especially the prediction of precipitation,which is still a worldwide problem.Data-driven weather forecasting methods emerge in endlessly in order to help people find more weather patterns,explore different forecasting approaches and improve the accuracy of forecasting.With its concise and intuitive features and extensive learning and reasoning ability,probabilistic graphical model has been widely used in various fields,and has unique advantages in weather prediction.In order to explore the temporal-spatial correlation of meteorological elements,a multi-scale spatial correlation graphical model is proposed by using graphical signal method.According to different time scales,the optimal spatial correlation scale is calculated,and the spatial correlation among stations is studied.The calculation efficiency is improved,and the prediction accuracy is also improved.First,the conditional Gaussian graphical model is used to learn the explanable spatial correlation of precipitation and temperature among 21 stations in East China.On this basis,the spatial range of correlation exploration is expanded.The proposed multi-scale spatial correlation graphical model is used to predict the elements of global meteorological observatories.Finally,the joint conditional Gaussian graphical model is used to predict for multi-site and multi-factor jointly.Compared with other linear forecasting models which do not consider spatial correlation and output variable correlation,The prediction accuracy is improved comparing with other linear prediction models without considering spatial correlation and output variable correlation. |
PDF Full Text Request