Numerical weather prediction (NWP) uses atmospheric general circulation models (AGCMs) to predict weather based on current weather conditions. The process of entering observation data into mathematical model to generate the accurate initial conditions is called data assimilation (DA). It combines observations, forecasting, and filtering step. This paper presents an approach for employing artificial neural networks (NNs) to emulate the local ensemble transform Kalman filter (LETKF) as a method of data assimilation. This assimilation experiment tests the Simplified Parameterizations PrimitivE-Equation Dynamics (SPEEDY) model, an atmospheric general circulation model (AGCM), using synthetic observational data simulating localizations of meteorological balloons. For the data assimilation scheme, the supervised NN, the multilayer perceptrons (MLPs) networks are applied. After the training process, the method, forehead-calling MLP-DA, is seen as a function of data assimilation. The NNs were trained with data from first 3 months of 1982, 1983, and 1984. The experiment is performed for January 1985, one data assimilation cycle using MLP-DA with synthetic observations. The numerical results demonstrate the effectiveness of the NN technique for atmospheric data assimilation. The results of the NN analyses are very close to the results from the LETKF analyses, the differences of the monthly average of absolute temperature analyses are of order 10-2. The simulations show that the major advantage of using the MLP-DA is better computational performance, since the analyses have similar quality. The CPU-time cycle assimilation with MLP-DA analyses is 90 times faster than LETKF cycle assimilation with the mean analyses used to run the forecast experiment.

No abstract

No abstract

Predicting the weather is dependent on the initial states specified in the computer model used to make the prediction. The data assimilation (DA) schemes are state-estimation techniques to generate an appropriated initial states for numerical models. DA deals with observations and data from the nonlinear dynamical models, both data set are very large in use on operational weather centers. The output from the DA procedure is called analysis. Some DA techniques become computationally intensive. The artificial neural networks (NN) can be employed to improve the computational performance. Two DA schemes are analized here: the Local Ensemble Transform Kalman Filter [17], and a version of a variational assimilation method [2] named the representer method. The EnKF was applied to a 3D atmospheric global spectral model (SPEEDY model), while the representer scheme was applied to the 2D shallow-water modelfor simulating the ocean circulation. These DA techniques were emulated by multilayer perpectron neural network (MLP-NN). The goal of this paper is to show the speed up for the DA computer performance in comparison to the methods emulated. The data assimilation process by NN preserves the analysis quality of the former DA techniques. In our experiments, the NN applied to DA on the SPEEDY model was 75 times faster than EnKF.

Predicting the weather is dependent on the initial states specified in the computer model used to make the prediction. The data assimilation (DA) schemes are state-estimation techniques to generate an appropriated initial states for numerical models. DA deals with observations and data from the nonlinear dynamical models, both data set are very large in use on operational weather centers. The output from the DA procedure is called analysis. Some DA techniques become computationally intensive. The artificial neural networks (NN) can be employed to improve the computational performance. Two DA schemes are analized here: the Local Ensemble Transform Kalman Filter [17], and a version of a variational assimilation method [2] named the representer method. The EnKF was applied to a 3D atmospheric global spectral model (SPEEDY model), while the representer scheme was applied to the 2D shallow-water modelfor simulating the ocean circulation. These DA techniques were emulated by multilayer perpectron neural network (MLP-NN). The goal of this paper is to show the speed up for the DA computer performance in comparison to the methods emulated. The data assimilation process by NN preserves the analysis quality of the former DA techniques. In our experiments, the NN applied to DA on the SPEEDY model was 75 times faster than EnKF.