Software based multiphase flow simulation enables production optimization by monitoring and planning oil and gas flow in wells and pipelines. Optimization can focus on carbon-effective targets and maximizing production outputs from cost-effective operations. The outcome is reduced energy consumption in reaching production targets with lower carbon emissions from the production process.
Virtual Flow Metering (VFM) provides real-time well rates, crucial for optimal production monitoring & optimization. This software technology empowers operators to react promptly to optimize operations proactively and avoid deferred or lost production. A self-adjusting, hybrid VFM, combining physics-based and Machine Learning (ML) approaches, ensures accurate real-time well rates. This article explains the process and significance of this approach.
Physics-based vs ML VFMs
Physics-based VFM relies on multiphase flow simulations, utilizing well-established concepts in thermodynamics, fluid dynamics, fluid modelling, and optimization. While accurate physics models offer a robust solution, they demand profound domain expertise and can be computationally expensive.
In contrast, ML VFMs employ learning algorithms to discern correlations between sensor data and target variables in historical datasets. This approach doesn’t rely on in-depth physics knowledge but requires a good grasp of learning algorithms and statistics. However, limited datasets or changes in operational conditions can restrict its widespread application.
Leveraging ML to Enhance Physics-Based Modeling
ML methods harness large datasets, simplify complexity, and reveal new data features. One approach is to use ML models to generate synthetic data for physics-based models. Unlike physics-based simulations that rely only on specific sensor data, ML can incorporate data from various sensors, enhancing the robustness and accuracy of physics-based models.
Another application involves creating proxy models through ML, mimicking the behavior of physics-based models. Real-life data and data from physics-based simulators are combined to train machine learning models, resulting in faster and more scalable deployment.
Enhancing Machine Learning with Physics Integration
The typical method to incorporate physics into ML involves Feature Engineering. This approach combines various measurements to create meaningful features, necessitating expertise in both domain knowledge and ML.
To further refine Feature Engineering, measurements can be combined with simple physics models that include static data like fluid properties, geometry data, and equipment information. These high-level features not only enhance the ML model’s accuracy but also increase its transparency and comprehensibility.
Turbulent Flux’s Hybrid Modeling Approach
At Turbulent Flux, we view physics-based modeling and ML as complementary, not competing, strategies. Our philosophy involves using ML to enhance specific components of physics-based models. For instance, our FLUX VFM employs ML models with high-level Feature Engineering to augment physics-based simulations, resulting in average full-scale errors below 5% for all phase flow rates.
In a concrete case study, our hybrid FLUX VFM delivers real-time well rates for gas, oil, and water in a mature oil field. While a pure physics-based solution provides accurate estimates for gas and total liquid flow rates, ML models with Feature Engineering predict the oil/water ratio precisely. Combining physics-based simulations with ML models achieves the best accuracy for three-phase well rates, offering a unified solution independent of changing operating conditions.
