Research On The Key Technology Of Typhoon Prediction Based On Deep Learning

Typhoon is an extreme weather event that can cause heavy damages to urban economy in coastal areas.It is very important to predict the formation and intensity of typhoon to give early warnings for the typhoon disasters.Traditional numerical forecast models based on thermodynamic equations and statistical models based on empirical relationships are still hard to predict intensity of typhoon accurately.Some researches have tried using to use machine leaning methods to predict the formation and intensity of typhoon,but they did not consider the spatial and temporal relationships among typhoon related variables.Here we propose a deep hybrid spatio-temporal model to make up for the shortcomings of existing methods.Our model introduces 3D convolutional neural networks(3DCNN)and 2D convolutional neural networks(2DCNN)to learn the spatio-temporal correlation of atmospheric and ocean variables.We utilize LSTM to learn the temporal sequence relations in the path of typhoon.Extensive experiments based on the Western Pacific(WP),East Pacific(EP),and North Atlantic(NA)show that our model is better than existing methods,including numerical forecast models used by many official organizations,statistical forecast methods and machine leaning based methods.In the experiments of 24h typhoon formation,the accuracy of our model can reach 85.2%,the Auc is 92.2%,and the intensity prediction error is 7.4kt.At the same time,it is found that the super-parameters such as the learning rate during model training are preferably set to the order of 10-4,and the training can be stopped after 25?35 epochs.The best latitude and longitude range around the typhoon center of input data set is found in range of g°×g°??13°×13°,and the best predictive effect is obtained when the time step(LSTM time step)is 2.