Global Data Assimilation Using Artificial Neural Networks in Speedy Model

Weather forecasting systems require a model for the time evolution and an estimate of the current state of the system. Data assimilation provides such an initial estimate of the atmosphere where it combines information from observations and from a prior short-term forecast producing an current state estimate. An Artificial Neural Network (ANN) is designed for data assimilation. The use of observations from the earth-orbiting satellites in operational numerical prediction models is performed for improving weather forecasts. The data related to atmospheric, oceanic, and land surface state from satellites provides increasingly large volumes. However, the use of this amount of data increases the computational effort. The goal here is to simulate the process for assimilating temperature data computed from satellite radiances. The numerical experiment is carried out with the global model Simplified parameterizations, primitive-Equation Dynamics (SPEEDY ) with simplified physical processes of an atmospheric general circulation in tri-dimensional coordinates. For the data assimilation scheme was applied an ANN: a Multilayer Perceptron(MLP) with supervised training. The MLP-ANN is able to emulate the analysis from the Local Ensemble Transform Kalman Filter(LETKF). LETKF is a version of Kalman Filter with Monte-Carlo ensembles of short-term forecasts. In this experiment, the MLP-ANN was trained with supervision from first six months considering the years 1982, 1983, and 1984. A hindcasting experiment for data assimilation performed a cycle for january of 1985 with MLP-NN, LETKF and SPEEDY model. The synthetic temperature observations were used. The numerical results demonstrate the effectiveness of this ANN technique on atmospheric data assimilation. The results for analysis with ANN are very close with the results from LETKF data assimilation. The simulations show that the major advantage of using MLP-NN is the better computational performance, with similar quality of analysis. The CPU-time assimilation with MLP-NN is 75% less than LETKF with the same observations. Actually, considering the supervised ANN for data assimilation, the most relevant issue is the computational speed-up for computing the analyzed initial condition for state model that accelerates the whole process of numerical weather prediction.