ENSOMIM:a novel spatiotemporal model for ENSO forecasts
In order to improve the accuracy of El Niño-southern oscillation(ENSO)prediction and solve the problem related to the difficulty in capturing long-range precursors of ENSO of convolution kernels,a deep learning model,called ENSOMIM,for ENSO unsteady spatiotemporal prediction based on attention mechanisms and recurrent neural networks was proposed via considering the ENSO prediction as a spatiotemporal series prediction problem.This model was used to learn space features of local and global inter-action via new attention mechanism BGAM,while long-term time series features was encoded by high-order nonlinear spatiotempo-ral networks.Due to the small number of samples in the ENSO observation data set,transfer learning method was adopted to train the model more fully,in which the historical model simulation data was used for pre training,and the observation data was used to correct the model.The experimental results show that ENSOMIM is more suitable for large-scale and long-term prediction.During the validation period from 1984 to 2014,the seasonal correlation technique of ENSOMIM's Niño3.4 index increased by 16% com-pared to the classical convolutional neural network,and the mean square error decreased by 17% .It can provide effective predictions for a lead time of up to 18 months,and can achieve relevant skills of 0.45 within a lead time of 23 months.Therefore,ENSOMIM can be a powerful tool for predicting ENSO events.
ENSOclimate disastersspatiotemporal series predictiondeep learningneural network
NETL
NSTL
万方数据
