Abstract
Traditionally, physics-driven simulation models have primarily been employed for design and long-term forecasting purposes. However, maintaining the accuracy of the model's boundary conditions and achieving convergence for operational decisions have posed significant challenges. Consequently, operators do not commonly use these physics-based models for optimizing daily operations, relying instead on personal judgment and experience. In this work, we introduce a novel and agile approach for optimizing gas delivery using hybrid physics–machine learning (ML) modeling. It is applied in a complex onshore pipeline system connected to multiple midstream sales points. Given the fluctuating gas prices and daily operational constraints with the need to meet the daily production quotas, there is a significant demand for a robust, agile optimization solution. This solution must manage day-to-day decision-making processes, considering the ever-changing field and market conditions.
The system in this project comprises 80 pads and four delivery sites, each with its unique set of flow and pressure constraints. These include minimum and maximum flow rates at delivery points and a cap on the maximum allowable pressure within the network. The flow into delivery points is regulated through per-site choke adjustments, adding another layer of complexity to the optimization process. The core objective of the proposed solution is to maximize total revenue. This is achieved by optimizing gas production across the four delivery sites while adhering to all flow and pressure constraints. This balance is crucial for operational efficiency and profitability. The solution is built on a novel hybrid of physics-ML models, employing surrogate modeling techniques. This approach combines the accuracy of physics-based models with the efficiency and adaptability of ML. The model is trained on thousands of scenarios generated from high-fidelity physics-based flow simulations.
These simulations incorporate detailed aspects of multiphase flow, pressure losses, heat transfer, and rigorous pressure-volume-temperature analysis. Additionally, the model's flexible application programming interface facilitates integration across various platforms, enhancing its applicability. The hybrid ML model demonstrates remarkable efficiency, capable of solving complex scenarios in mere seconds. This efficiency translates into approximately 2,000 solves per optimization, completed within 2 to 10 seconds—a stark contrast to the 33 hours required by a flow simulator. Despite its incredible speed, the surrogate model's predictions have an impressive accuracy rate of 99% against physics-based models.
This solution marks a significant transformation from traditional methods, which heavily relied on personal judgment and experience. By integrating data-driven insights integrated with the detailed physics of multiphase flow and gas production, the proposed approach allows for rapid, informed decision making. This maximizes our gas production revenues and ensures the integrity of our pipelines.
