Including steady-state information in nonlinear models: an application to the development of soft-sensors

TL;DR

This paper introduces a novel method for training nonlinear models, including neural networks, using both dynamical data and steady-state information, significantly reducing computation time and improving model performance across wider operating ranges.

Contribution

The proposed bi-objective training procedure incorporates steady-state data into nonlinear models without fixed point computations, applicable to complex structures like neural networks.

Findings

Effective soft-sensors for downhole pressure estimation developed

Method reduces computation time by about three orders of magnitude

Models show good dynamical and static performance

Abstract

When the dynamical data of a system only convey dynamic information over a limited operating range, the identification of models with good performance over a wider operating range is very unlikely. Nevertheless, models with such characteristic are desirable to implement modern control systems. To overcome such a shortcoming, this paper describes a methodology to train models from dynamical data and steady-state information, which is assumed available. The novelty is that the procedure can be applied to models with rather complex structures such as multilayer perceptron neural networks in a bi-objective fashion without the need to compute fixed points neither analytically nor numerically. As a consequence, the required computing time is greatly reduced. The capabilities of the proposed method are explored in numerical examples and the development of soft-sensors for downhole pressure estimation for a real deep-water offshore oil well. The results indicate that the procedure yields suitable soft-sensors with good dynamical and static performance and, in the case of models that are nonlinear in the parameters, the gain in computation time is about three orders of magnitude considering existing approaches.