Data Driven Virtual Flow Meter: A Use Case of Managing AI Models at Scale

This paper introduces a data-driven virtual flow meter that complements physical well tests and well models to estimate daily well production rates in the oil and gas industry. Traditional methods are time-consuming and subject to uncertainties. The proposed approach leverages data science methods and historical production data, along with additional data sources such as pressure measurements and fluid properties, to train machine learning models. The models establish correlations between the available data and actual well production rates, providing accurate estimations.

The virtual flow meter offers advantages such as real-time estimations, enabling continuous monitoring and rapid decision-making. It also reduces costs by gradually reducing the physical well tests frequency while utilizing existing data sources.

To ensure the reliability of the rate estimation, models are built exclusively for individual well. By scaling the models to hundreds of wells across Malaysia Asset, a sustainable model management workflow is required. A Machine Learning Operations, MLOps framework is established to automate and optimize the overall model management process.

In conclusion, with the proper MLOps framework, the full potential of data-driven virtual flow meter models can be realized by having reliable model churning out accurate rate estimation on daily basis across the entire well life cycle.