Research On Sea Surface Temperature Prediction Model Based On Deep Learning

With the rapid development of science and technology,deep learning has been widely used in all walks of life,from automation to data mining,from image recognition to pattern recognition,from data analysis to model building,and remarkable achievements have been made.In the field of deep learning,common neural network models are convolutional neural networks,autoencoders,recurrent neural networks,and generative adversarial networks,all of which can be used to solve complex problems.To effectively apply these models and obtain optimal experimental outcomes,it is crucial to analyze the weight feature values,attractors,and stability of deep neural networks.Such analysis not only facilitates a better understanding of the network structure,but also guides the design of network architectures to attain the desired dynamic behavior and characteristics,resulting in more advanced network structures.This thesis focuses on the study of deep neural network models and their application to predicting sea surface temperature.The study is divided into two parts:The first part examines the continuous attractors of a class of neural networks with a transfer function of ReLU and their applications in the reconstruction of MNIST datasets and the prediction of sea surface temperature.Theoretically,a continuous attractor is a set of connected stable equilibrium points,and in practice,the information of continuous mode can be regarded as continuous attractors.Therefore,according to the correlation theories of attractors and Lyapunov’s stability theorem,the existence of continuous attractors of the recurrent neural networks satisfying certain conditions with and without external inputs was proved,and the mathematical expression was given.Then the theories were verified by simulation experiments.Secondly,the recurrent neural network with visible layer and hidden layer,which was equivalent to the repeated iteration of feedforward autoencoder,was analyzed,and the network model was applied to the reconstruction of MNIST dataset and the prediction of sea surface temperature,which had good results in both aspects.The second part of this thesis explored the use of generative adversarial networks for the prediction of sea surface temperatures.A generative adversarial network was designed to generate images of future sea surface temperatures to provide a visualization method for predicting sea surface temperatures.The generative adversarial network consisted of a generator and a discriminator.The generator used multiple composite layers to capture changes in sea surface temperature and produced a clear image of future sea surface temperatures.The discriminator used the structure of the PatchGAN to obtain more characteristics of the sea surface temperatures,allowing the discriminator to more accurately identify whether the picture was real or generated.In addition,this thesis improved the loss function of the generative adversarial network and analyzed the convergence,and proved that minimizing loss function was equivalent to minimizing Pearson χ2 divergence.Finally,by training the generative adversarial network,the generative adversarial network can produce predictions closer to the real sea surface temperature maps.