Research On Multi-model Algorithm For Typhoon Prediction Based On Deep Learning

In the past few decades,typhoons have hit our country frequently,causing immeasurable losses to people's lives and property in coastal areas.How to accurately predict the typhoon path and reduce the loss of people in coastal areas has become a hot research topic now.Due to the many factors affecting typhoon trajectories and the difficulty of feature extraction,traditional methods require more prior knowledge in related fields,which is not only time-consuming and labor-intensive,but also has low prediction accuracy.In recent years,with the development of deep learning technology,more and more researchers have begun to introduce deep learning technology into typhoon trajectory prediction,and have also achieved many good results.Based on this,this article uses deep learning to predict the typhoon trajectory,and uses multiple model fusion technology to fuse the prediction results of sequence data and image data to improve the prediction accuracy of the current method:1)For the typhoon trajectory sequence data,the Seq2 seq model based on the attention mechanism is used to improve the prediction accuracy.First of all,for the defect that the results of the Kalman filter algorithm are greatly affected by abnormal observations,an improved Kalman filter algorithm is proposed to effectively remove outliers and improve the quality of trajectory data.Then,for the problem of trajectory sampling too frequently,the minimum fan-shaped simplified algorithm is used.Simplify the trajectory and improve the training speed.Finally,build a model to predict the typhoon trajectory,use encoders and decoders to solve the problem of unequal sequence input and output,use attention mechanism to optimize the model's timing dependency,and improve the prediction accuracy of existing methods.The experimental results show that the method can make good use of data features and time series to predict the direction of the trajectory.2)Aiming at the typhoon satellite image data,a GAN model based on time series is proposed to improve the prediction accuracy.First use digital coding algorithms to preprocess image data in order to better focus on local features of the image;then the temporality is introduced into the GAN model and applied to the prediction of typhoon trajectories.In traditional CNN networks,the image generation is just the average of various possible situations,and the generation accuracy is not high.The GAN directly generates the distribution of data.The generator is responsible for generating new images from the noise,and the discriminator is responsible for identifying the "authenticity" of the generated images.The two form a dynamic game until the Nash equilibrium.As the input of the GAN,the pictures at the previous sequence time can predict the typhoon trajectory at the later sequence time,which can effectively improve the prediction accuracy of the existing method on the typhoon image data.Experimental results show that the method can learn the shape of the trajectory well and cope with unexpected situations.3)Multi-model fusion of the prediction results of sequence data and the prediction results of image data is innovatively used to further improve the prediction accuracy.Different algorithms perform differently on different datasets,and each algorithm has its own area of expertise.Multi-model fusion can better "boiler people." The genetic algorithm GASEN is used for multi-model fusion.The weights between the results of different models are weighted and averaged based on the weights between the evolutionary prediction results of the genetic algorithm.Experimental results show that the prediction accuracy after fusion is higher than all existing methods.