Multilayer Perceptron applied to Data Assimilation for the Global FSU Atmospheric Model

Abstract:

The better quality of forecasts is given the more accurate of the initial conditions. Data assimilation (DA) is the process by which short-forecast and observations are combined to obtain an accurate representation of the state of the modeled system, e.g. is a technique to generate an initial condition to a weather forecasts. This paper shows the results of a DA technique using artificial neural networks (NN) to obtain the analysis to the atmospheric model for the Florida State University. The Local Ensemble Transform Kalman filter (LETKF) is implemented with Florida State University Global Spectral Model (FSUGSM). The ANN data assimilation is made to emulate the initial condition from LETKF to run the FSUGSM. LETKF is a version of Kalman filter with Monte-Carlo ensembles of short-term forecasts to solve the data assimilation problem. The model FSUGSM is a multilevel spectral primitive equation model with vertical sigma coordinates, at resolution T63L27. The data assimilation experiments are based in simulated observations data and FSUGSM 6-hours forecasts. For the NN data assimilation, we use Multilayer Perceptron (MLP) with supervised training algorithm where NN receives input vectors with their corresponding response from LETKF data assimilation. The surface pressure, absolute temperature, zonal component wind, meridional component wind and humidity results are presented. A self-configuration method finds the optimal NN and configures a set of 52 MLPs to DA experiment, referred as MLP-DA. A methodology developed with self-configuration using a meta-heuristic called the Multiple Particle Collision Algorithm to compute the optimal topology for NN. The MLP presents four input nodes, two nodes coordinates vector, one for the 6-hours forecast vector and one node for observation vector; one output node for the analysis vector results. The vector represents the values for one grid model point. The ANNs were trained with data from each month of 2001, 2002, and 2003. The results demonstrate the effectiveness of the MLP-DA for atmospheric data assimilation, with similar quality to LETKF analyses. An experiment for data assimilation cycle using MLP-DA was performed with simulated observations for January of 2004. The major advantage of using MLP-DA is the computational performance, which is faster than LETKF.