...
Data assimilation. Local ensemble transform Kalman filter. Artificial Neural Networks. Recurrent Neural Network. Multilayer perceptron. Long short-term memory.
This project presents a proposal to optimize the data assimilation process for numerical
weather forecasting by replacing the Local Ensemble Transform Kalman Filter (LETKF)
using sets of Multilayer Perceptron Artificial Neural Network (MLP ANN) and Long Short-
Term Memory Recurring Neural Networks (LSTM RNN). The assimilation methods will be
tested using the Simplified Parametrization, primitivE-Equation DYnamics – SPEEDY –
which is a General Circulation Model (GCM). The forecasts will be generated in an area of
25 km² over the city of Santarém, Pará, for the period from January 1, 2011 to December 31,
2013. The forecast variables for this model are: absolute temperature (T ), surface pressure
(ps), zonal wind component (u), southern wind component (v) and specific substance (q). The
assimilation results generated by the neural networks will be compared with the data
assimilation from the classic Local Ensemble Transform Kalman Filter method through the
Root Mean Square Error and execution times.