Identifiability and physical interpretability of hybrid, gray-box models -- a case study

TL;DR

This study explores the balance between model identifiability and physical interpretability in hybrid models, demonstrating that non-identifiable models can still be physically meaningful and effective in industrial applications.

Contribution

It provides a case study showing that hybrid models can maintain physical interpretability despite non-identifiability, and highlights the benefits of hybrid models over purely mechanistic ones.

Findings

Hybrid models achieved 35% lower median prediction error than mechanistic models.

Physical interpretability can be preserved even when models are non-identifiable.

Data quality and quantity significantly influence model performance and interpretability.

Abstract

Model identifiability concerns the uniqueness of uncertain model parameters to be estimated from available process data and is often thought of as a prerequisite for the physical interpretability of a model. Nevertheless, model identifiability may be challenging to obtain in practice due to both stochastic and deterministic uncertainties, e.g. low data variability, noisy measurements, erroneous model structure, and stochasticity and locality of the optimization algorithm. For gray-box, hybrid models, model identifiability is rarely obtainable due to a high number of parameters. We illustrate through an industrial case study - modeling of a production choke valve in a petroleum well - that physical interpretability may be preserved even for non-identifiable models with adequate parameter regularization in the estimation problem. To this end, in a real industrial scenario, it may be beneficial for the model's predictive performance to develop hybrid over mechanistic models, as the model flexibility is higher. Modeling of six petroleum wells on the asset Edvard Grieg using historical production data show a 35\% reduction in the median prediction error across the wells comparing a hybrid to a mechanistic model. On the other hand, both the predictive performance and physical interpretability of the developed models are influenced by the available data. The findings encourage research into online learning and other hybrid model variants to improve the results.