Data Assimilation Procedure by Recurrent Neural Network

Abstract

Data assimilation is a process to combine a model prediction of a state variable at a given time with a set of measurements available at this particular time in order to obtain a suitable set of data for model initialization. The state of the art in data assimilation techniques are based on Extended Kalman Filter (EKF) and Four-Dimensional Variational Analysis (4D-Var), but this methodology has high computational complexity. In this paper, the authors propose emulating a Kalman filter using a neural network as a proposal to reduce the computational complexity of the problem. This work applies a recurrent neural network paradigm, named Elman Neural Network (E-NN), to the data assimilation problem of a non-linear shallow water model. The performance of E-NN on emulating the Kalman filter (KF) and the evaluation of application of the technique at high dimension problems of operational numerical weather forecasting are analyzed. The results with the shallow water 1D dynamics show that E-NN converges faster than standard Multilayer Perceptron Neural Network (MLP-NN) in the training phase, and its computational complexity is less than that of extended Kalman filter. However, there is a loss of accuracy in the results when comparing E-NN to MLP-NN and KF.