Research On Data-driven Data Assimilation Methods Incorporating Machine Learning Algorithms

Data assimilation(DA)is an important scientific computing method in the field of physics and earth science.Its goal is to fuse model output and observation information to obtain the optimal state estimation of the system.With the advent of the era of big data,using data to solve scientific problems has become a new development direction,and it also brings new opportunities for the development of data assimilation methods.At present,data assimilation methods rely on known physical models,but using machine learning algorithms to improve assimilation performance and deal with problems such as uncertainty in the model still has great potential.In this thesis,the machine learning algorithm and the classical sequential assimilation method are combined to explore the spatial and temporal evolution relationship of state variables from the data to design an alternative model.The data-driven model is used to replace the traditional(physics-based)model to optimize the assimilation method of fusion observation information and model simulation.The research work of this paper mainly includes the following two aspects:(1)A data-driven(DD)data assimilation method combining support vector regression(SVR)and ensemble Kalman filter(En KF)is proposed.Gaussian sampling is used to generate a training sample set,and then SVR is used to train and optimize a simulation operator for prediction.In the assimilation process,the operator will replace the known physical model to evolve the next predicted data value,and use the ensemble Kalman filter to complete the data assimilation process with the observed value.In this thesis,the data assimilation performance is compared by changing the sensitivity parameters such as sample set size,noise variance and observation step size.The results show that the proposed SVR-DD-ENKF hybrid data assimilation method is superior to the traditional data assimilation method in the case of large data sets,which proves the effectiveness of the new method.(2)A data-driven data assimilation method based on echo state network(ESN),differential evolution(DE)and Kalman filter(KF)is proposed.Firstly,the differential variation optimization algorithm is used to optimize the parameters of the echo state network,and the ENS model is used to predict each subsequence of the Lorenz system to obtain the predicted value.Furthermore,the predicted values are substituted into the next assimilation to obtain data-driven data assimilation analysis values.Experiments show that the newly proposed ENS-DD-KF hybrid data assimilation method is better than the classical Kalman filter assimilation method in 10 Lyapunov times.This method provides new development ideas in integrating machine learning and deep learning prediction alternative models.The data-driven data assimilation method proposed in this thesis can integrate machine learning algorithms to design alternative models and apply them to the data assimilation method system.Especially in the case of model understanding defects and incomplete data,it is possible to achieve complementarity and integration between models and data,which has certain theoretical significance for the study of hybrid data assimilation methods under the background of big data.