Interpretable Scientific Machine Learning Approach for Correcting Phenomenological Models: Methodology Validation on an ESP Prototype

The electric submersible pump (ESP) is widely used in
oil extraction
processes and is recognized for its effectiveness as an artificial
lift technique in the petroleum sector. Developing and improving predictive
models for these systems can contribute to optimizing operational
efficiency and maximizing oil production. These improvements facilitate
ESP performance control and monitoring, potentially making the extraction
process more economically and environmentally sustainable. This paper
aims to develop a framework for creating an interpretable corrective
model for the ESP phenomenological model using experimental data.
Neural networks were employed to analyze the complexities and nonlinearities
of the processes, followed by symbolic regression to generate a simplified
and interpretable equation to improve the model’s predictive
capacity. An uncertainty assessment of model parameters was performed
using Markov chain Monte Carlo (MCMC) and propagated to the model
prediction. Additionally, a regression model was created directly
from the process data for comparison purposes. The comparative analysis
indicated that the approach incorporating neural networks to generate
synthetic data, followed by symbolic regression, improved the model’s
ability to predict key variables such as intake pressure, choke pressure,
and production flow.