Multilayer Perceptron on data assimilation system applied to FSU global model
Rosangela Cintra1; Haroldo Campos Velho1; Steven Cocke2
1 Brazilian National Space Research, INPE; 2 Center of Ocean-Atmospheric Prediction Studies, Florida State University
doi:10.20906/CPS/USM-2016-0018
Resumo
Numerical weather prediction (NWP) uses atmospheric general circulation models (AGCMs) to predict weather based on current weather conditions. The atmosphere could not be completely described due to inherent uncertainty. These uncertainties limit forecast model accuracy to about five or six days into the future. The process of entering observation data into mathematical model to generate the initial conditions is called data assimilation (DA). DA techniques seek to obtain an accurate representation of the state of the modeled system. This paper shows the results of a data assimilation technique using artificial neural networks (ANN) to obtain the initial condition to an AGCM used in Florida State University (FSU) in USA. The Local Ensemble Transform Kalman filter (LETKF) is implemented applied to Florida State University Global Spectral Model (FSUGSM). LETKF is a version of Kalman filter with Monte-Carlo ensembles of short-term forecasts to represent the model uncertatanty. The ANN data assimilation is made to emulate the initial condition from LETKF applied to the FSUGSM. The FSUGSM is a multilevel spectral primitive equation model with a vertical sigma coordinates. All variables are expanded horizontally in a truncated series of spherical harmonic functions at resolution T63L27. The data assimilation experiment is based in synthetic observations (surface pressure, absolute temperature, zonal component wind, meridional component wind and humidity). We use Multilayer Perceptron data assimilation (MLP-DA) with supervised training algorithm where ANN receives input vectors with their corresponding response or target output from LETKF initial conditions. An automatic tool that finds the optimal representation to these ANNs configures the MLP-DA in this experiment. After the training process, the MLP-DA is a function of data assimilation where the inputs are observations and a short-range forecast to each FSUGSM grid point. The MLP_DA were trained with data from each month of 2001, 2002, 2003, and 2004. The data ass
Palavras-chave: data assimilation; artificial neural networks; numerical weather prediction; inverse problem