A Fast Algorithm for Training Dynamical Neural Networks Using Steady- State Prior Information of Offshore Oil Platform

Resumo

In systems identification, the use of auxiliary information configures a grey-box approach. This paper
describes a methodology to estimate model parameters including auxiliary information about the static behavior
of the system in a bi-objective approach and discusses a decision maker based on the auxiliary information. The
procedure can be applied to many model structures, such as polynomial models or multilayer perceptron (MLP)
neural networks, without the need of computing the model fixed points. The grey-box modeling procedure was
applied to design a soft-sensor for the downhole pressure of a real gas-lifted deep-water offshore oil well. To this
end, steady-state values of the downhole pressure were estimated from historical data from (almost) stationary
conditions. The available training and validation (dynamical) data had information over a limited operating range,
while test data had operating ranges not present in the training and validation data. The identified dynamic models
used only platform variables with a fixed MLP structure. The results indicate that the procedure yields suitable
models with good static and dynamic performance. Besides, the use of auxiliary information helped to find models
with better dynamical performance on operating regimes not originally represented in the dynamical data. Whereas
an identified black-box MLP model obtained a root mean squared error (RMSE) of 6.7 kgf/cm2

in a free-run

simulation over test data, the proposed approach achieved an RMSE of 3.7 kgf/cm2

. This is very relevant for many
practical situations where the available dynamical data does not cover all operating regimes of the system. The
procedure described in this work can be applied with different model classes with greatly reduced computing time.