Research On Integrated Climate Prediction Method Based On Echo State Network And Wavelet Transform

Climate prediction is a challenging task that is crucial to the development of human society and the protection of the ecological environment.To better understand and predict climate change,scientists have been exploring various climate prediction methods,which include physical modelling methods,statistical methods and machine learning methods.In recent years,machine learning methods have been increasingly used in climate prediction,with integrated methods based on Echo State Networks and Wavelet Transforms attracting much attention.Echo State Network（ESN）is a machine learning method based on recurrent neural networks,which has the advantages of fast training and good generalisation ability.Wavelet Transform（WT）is a mathematical transformation method that can decompose a signal into multiple sub-signals of different frequencies and has good localisation properties in the time-frequency domain.In this thesis,we combine Echo State Network and Wavelet Transform,making full use of their respective advantages,and propose an integrated method of climate prediction based on Echo State Network and Wavelet Transform to improve the accuracy and stability of climate prediction.The innovation of this thesis is reflected in the following two aspects.First,an integrated method of climate prediction based on echo state network and wavelet transform is proposed for predicting climate change in the future period.The raw meteorological observation data are then wavelet transformed to decompose the signals at different scales,and each sub-signal is used as input for training and prediction by an echo state network model.Finally,the prediction results at different scales are weighted and integrated to obtain the final prediction result.After comparing the experimental results with other comparative models（e.g.PCA-ESN model,ARMA model with integrated wavelet transform,LSTM model with integrated wavelet transform）,it is found that our proposed model has better prediction performance.Secondly,after the model prediction proposed in this thesis is conducted,we are faced with the problem of how to improve the efficiency of perturbations to the model.We would use the L₁-loss function as the main method to evaluate the adversarial attack,analyzed the validity,applicability and transferability of the perturbation terms,proposed and used the adversarial attack method based on the importance measure,and according to the experiments plotted the relative absolute error RAE,relative squared error RSE,and after plotting the correlation coefficient CORR,we found that the adversarial perturbation based on the top 5%importance measure has almost the same effect as the 100%perturbation of the original time series.Thus,an adversarial attack algorithm based on importance measures can significantly reduce the cost of perturbation and only a small number of perturbations are sufficient for an adversarial attack.