NEURAL NETWORKS FOR DATA ASSIMILATION IN THE METROPOLITAN AREA OF RIO DE JANEIRO

Resumo

The use of neural networks for data assimilation in the terminal area of Rio de Janeiro - emulating the
3D-Var method as implemented in the Weather Research and Forecasting Data Assimilation module - is explored
here. Surface and upper-air data (air temperature, relative humidity and wind speed and direction) from airport
stations and 6-hour forecast from WRF are used as input for the model and the 3D-Var analysis for each grid
point is used as target variable. Periods of 168h from 2014 and 2015 are used with 6h and 12h assimilation cycles
for surface and upper-air data, respectively. The neural network model was built using two different approaches:
multilayer perceptron with static topology in Weka and a metaheuristic called Multi-Particle Collision Algorithm,

where different topologies are tested until the optimum solution is found. Results show that Weka-NN and MPCA-
NN are able to emulate the 3D-Var method with negligible differences in comparison of the magnitude of the

assimilated data. Also, the neural networks are able to speed up the data assimilation process. In this study, neural
network models were able to run from 70 to 100 times faster than the conventional method (3D-Var analysis) under
the same hardware configurations. This time reduction enables the execution of data assimilation methods using
less CPU time and using personal computers.