Simulation And Prediction Of North Atlantic Oscillation Based On Numerical Model And Deep Learning

NAO(North Atlantic Oscillation)is a prominent seesaw oscillation of the atmospheric circulation between subtropical high and subpolar low.It is one of the most dominant circulation modes in the Northern Hemisphere.In recent decades,the NAO has been characterized by anomalies and frequent changes,which are particularly pronounced in winter.Studies have shown that extreme weather phenomena such as European blizzards are closely associated with NAO events.The low-frequency variability and seasonal characteristics of the NAO also profoundly affect climate change in the Northern Hemisphere and even that of the globe.Improving the forecasting skills of the NAO is of great practical importance for disaster prevention on a global scale.NAO prediction is mainly conducted through dynamical numerical models,but there are still large forecast errors.Therefore,it is necessary to explore the causes of forecast errors and to further reveal and understand the fundamental rules of NAO dynamics mechanisms.Furthermore,as a phenomenon with complicated contributing factors,the developmental mechanism of NAO has not been fully understood by humans yet.The numerical models have been simplified to a certain extent for efficiency reasons.Under the circumstances,the data-driven approach provides an alternative way for NAO forecasting.In this paper,the CNOP(Conditional Nonlinear Optimal Perturbation)method and CESM(Community Earth System Model),which has the capability to simulate NAO,are adopted to explore reasons for forecast errors in terms of the model initial field and model parameters.In addition,predictors that are directly and indirectly related to NAO are selected according to the coherence analysis of ENSO(El Ni?o-Southern Oscillation)and NAO variations,then forecast models for NAO based on the deep neural network are constructed.Finally,this paper adopts the CNOP method to optimize the model forecasts of CESM,and conduct the ensemble forecast together with the NAO forecasts based on deep learning.The main research of this paper is as follows.1.Optimal Precursor(OPR)Identification of NAO based on CNOP-I(CNOP related to Initial perturbation)First,the parallel PGAPSO(PCA-based Genetic Algorithm and Particle Swarm Optimization),which is based on a dynamic inertia weight strategy and explorationexploitation balance strategy,is proposed to solve the CNOP-I of CESM.Simultaneously,the atmosphere module of CESM is accelerated,and the speedup ratio reaches 40.2.The two cases,whose reference states would not become NAO events,are selected to solve CNOP-Is corresponding to positive-phase NAO(!)and negative-phase NAO(").The experimental results show that the CNOP-I type perturbation can trigger two types of NAO events and has the largest prediction uncertainty.Thus,they can be described as OPRs of NAO.For the physical variables that make up CNOP-Is,wind directions of OPRs corresponding to ! and "have opposite distribution around Iceland,and temperature perturbations over Greenland,Iceland,and eastern Europe have significant effects on triggering NAO events.Specific humidity and surface pressure show similar spatial structures in two types of NAO in both cases,while surface potential height is distinctive in the region from the North Pole to the Chukchi Sea.The CNOP-I perturbation is capable of forming typical dipole modes in both ! and " cases,and nonlinearity is also demonstrated to play a role mainly in the later stages of the simulation.2.Model Parameter Sensitivity Analysis of CESM based on CNOP-P(CNOP related to Parameter)First,28 sensitive parameters distributed among the atmosphere,sea ice,and river modules are filtered from 237 parameters in each module of the F component set in CESM.The PSO(Particle Swarm Optimization)algorithm is applied to solve CNOPP of CESM with single parameters and two-parameter combinations,while CMA-ES(Covariance Matrix Adaptation Evolution Strategy)is adopted to solve CNOP-P of CESM for three-parameter combinations.Then 10 single parameters,10 two-parameter combinations,and 10 three-parameter combinations,which can cause the maximum prediction uncertainty of ! and ",are chosen and their values are also determined.The experimental results show that the interaction between the parameters may contribute to the growth of the prediction error,that is to say,the prediction errors triggered by parameter combinations are significantly larger than that of the single parameters,specifically,three-parameter combinations > two-parameter combinations >> single parameters.For single parameters,the minimum relative humidity of low stable clouds(cldfrc＿rhminh)and the characteristic adjustment time scale(hkconv＿cmftau)are the most sensitive parameters for ! and ",respectively;For two-parameter combinations,cldfrc＿rhminh and the rainwater autoconversion coefficient(hkconv＿c0)are the most sensitive parameters for !,while the factor applied to the icefall velocity(cldsed＿ice＿stokes＿fac)and the total solar irradiance(solar＿const)are the most sensitive parameters for ".As for threeparameter combinations,the parameter for deep convection cloud fraction(cldfrc＿dp1),cldfrc＿rhminh,and solar＿const is the most sensitive combination for !,and the tunable constant for evaporation of precip(cldwat＿conke),hkconv＿cmftau and solar＿const is the most sensitive combination of ".The sensitive parameters obtained in this paper provide a theoretical basis for later correction of model parameters to reduce model prediction errors.3.NAO Prediction based on Artificial IntelligenceWe select predictors directly related to NAO events,such as NAOI(NAO Index),SLP,and GH(Geopotential Height),meridional and zonal winds,and indirectly related predictors,such as SST(Sea Surface Temperature),and we propose the DWT-LSTM model to forecast NAOI;the RF-Var model is adopted to predict NAOI variations;the DWT-Conv LSTM model and the Acc Net model are applied to forecast SLP.The experimental results show that the DWT-LSTM model has higher forecast accuracy at the peak point;the RF-Var model outperforms several other classifiers with an accuracy of 68%;the DWT-Conv LSTM has higher multi-step forecasting reliability than the NAO forecasting products GFS[1] and ENSM[2];the Acc Net model has greater ability to capture spatial and temporal correlations and has higher flexibility in forecasting NAO events with different seasons and types.In addition,combining the features of numerical models and deep learning,this paper attempts to conduct ensemble forecasting using CESM forecasts optimized by CNOP method along with the deep learning models for SLP.4.NAO Ensemble Forecasting based on Numerical Models and Deep LearningFirst,we conduct the data assimilation for the potential temperature field of the ocean module named pop2 in CESM using the Nudging method.Then,the initial field and the model parameters are optimized separately using the CNOP-I and CNOP-P method.The ensemble forecasts members of CESM compose of the above forecasts,and the NAO ensemble forecast members of deep learning are also formed by SLP forecast models DWT-Conv LSTM and Acc Net.Finally,the ensemble forecasts of SLP and NAOI are performed by these two kinds of ensemble members.