Determination And Prediction Of Tropical Cyclone Intensity In The Northwest Pacific Based On Deep Learning

Tropical cyclones occur frequently in the coastal areas of China,and after upgrading to typhoons,they are easy to cause serious threats to people's personal and property safety.Therefore,the research on tropical cyclones has always been an important direction for meteorological researchers,and the research on the intensity of tropical cyclones is one of the important topics.Tropical cyclone intensity research generally has two directions: intensity determination and intensity prediction.Among them,the former is the real-time determination of tropical cyclone intensity,and the latter is the intensity forecast for the future time.For intensity determination,the traditional methods generally have related methods based on Dvorak technology and its variants,but this kind of method has strict requirements on the knowledge reserve and professional skills of researchers.Moreover,the classification accuracy of machine learning method needs to be improved.At present,the bottleneck of tropical cyclone intensity determination mainly lies in :(1)the traditional method has strict requirements on relevant personnel's professional knowledge and is relatively subjective;(2)The machine learning method fails to take into account the multi-channel cyclone cloud images and mine the connections between channels;And the imbalance of sample data has some influence on the final result.For intensity prediction,there are generally numerical prediction,statistical analysis and machine learning methods.Numerical prediction method requires researchers to have a deep understanding of tropical cyclone operation mechanism and atmospheric dynamics,and experts to calculate the corresponding fluid dynamics equations.The statistical analysis method needs to find the important factors influencing the intensity change from the historical data.Machine learning methods construct different models to mine historical features and patterns to predict future intensities.The main bottlenecks of tropical cyclone intensity prediction are as follows :(1)traditional methods require complex derivation and calculation by professionals,and have low accuracy in long-term prediction;(2)The machine learning method fails to make full use of cyclone characteristics and related meteorological features to mine correlations,and the accumulation of long-term prediction errors of general serial-to-series models will affect the prediction accuracy.With the continuous development of deep learning technology,its advantage of automatically extracting data feature association is very effective for solving complex intensity correlation research.Therefore,deep learning-based tropical cyclone intensity determination model and intensity prediction model were proposed in this paper to study the Northwest Pacific Ocean,and to some extent solve the problems of poor accuracy in estimating tropical cyclone intensity and difficulty in long-term intensity prediction.The main research contents of this paper are as follows:(1)For the determination of tropical cyclone intensity,multi-channel infrared satellite data was adopted to avoid insufficient information of single channel;The generative adversarial network is used to enhance the data of a few types of samples and under sample the majority of the samples to balance the data classes.Finally,channel attention mechanism was added into the classification model to extract the connections between channels and improve the classification accuracy.(2)For the prediction of tropical cyclone intensity,the deep correlation between cyclone characteristics and meteorological characteristics was explored;The stack denoising autoencoder and convolution operation were used to extract the spatial and temporal features of cyclone and meteorological grid,respectively.Finally,the one-step prediction structure in Transformer model can directly predict the intensity in the future period of time,avoiding the error accumulation problem of traditional sequence-tosequence model long-term prediction.