Artificial Lift

Oilfield Optimization

Oilfield optimization using AI-driven methods for artificial lift with rod pumping and electrical submersible pumps is becoming increasingly popular in the oil and gas industry. These methods incorporate advanced predictive maintenance and reservoir modeling to optimize production and improve operational efficiency. AI-based predictive maintenance algorithms can predict equipment failure before it happens, thereby reducing downtime and maximizing equipment uptime. Reservoir modeling enables engineers to optimize well placement and production rates, leading to increased oil recovery. Scheduling maintenance at optimal times is another benefit of using AI in oilfield optimization. AI-driven methods are a valuable tool in the oil and gas industry, helping operators to optimize production, minimize downtime, and increase profitability.

Artificial Lift Optimization

The PRISM group has created a mechanical digital twin of a sucker-rod pump and dynamic optimization methods to increase oil recovery with artificial lift. This work applies the methods to depleted oilfields with high water-cuts to increase reservoir recovery and maximize pump fillage. Hybrid AI technologies are also deployed for fault diagnosis and in key parameter predictions for oil well optimization and energy efficiency improvements. Combinations of physics-based and data-driven approaches are developed to capture value with Hybrid AI methods.

The project involves several stages with field testing and online application.

Data cleansing, parsing, and assessment
Simulator building and application of oil well
Build physics-based digital twin of the oil well
Soft sensor for reservoir fluid
Soft sensor for production rate
Diagnose oil well based on Hybrid AI using indicator card and subsurface data
Optimize the simulator and digital twin with parameter updates
Field testing and online application for real-time optimization

Deliverables include a simulator of the mechanical sucker rod system to estimate pump fillage, a soft sensor of the fluid flow, and soft sensor of the reservoir fluid level. The software contains a method to adapt the soft sensors and mechanical simulator to each well. Optimized adjustable pumping rate is an outcome of the updated models to suggest the pump-off rates for each well.

Reference

Hansen, B., Tolbert, B., Vernon, C., Hedengren, J.D., Model Predictive Automatic Control of Sucker Rod Pump System with Simulation Case Study, Computers & Chemical Engineering, 121, pp. 263-284, 2019.

Proposal: Electrical Submersible Pump (ESP) Predictive Maintenance

We propose to create multi-physics simulation libraries for decision support with Electrical Submersible Pumps (ESPs). This is in support of the target to increase average ESP run life. The planned multi-physics simulation libraries include hybrid machine learning (ML) to adapt to new conditions and perform rapid root cause analysis for asset performance management (APM). The ESP monitoring solution will receive real-time or offline data and produce alerts and recommended actions. The predictions include a probability of failure over a prediction time horizon and remedial actions that can either be automatically implemented or else recommended to the surveillance engineer. The analysis seeks to identify root cause of historical failures or predicted future events. The proposed work includes 4 Milestones.

Milestone 1: Curate data sets and develop models for training and validation
- Phase 1A: Collect ESP failure cases for training and validation
- Phase 1B: Characterize control performance
- Phase 1C: Characterize sensor performance
- Phase 1D: Create ESP digital twin with multi-physics simulation
Milestone 2: Hybrid ML for early failure detection
- Phase 2A: Self-adapting hybrid ML model for real-time surveillance
- Phase 2B: Hybrid ML for early failure condition classification
- Phase 2C: Uncertainty quantification of remaining run life
Milestone 3: Real-time advisory models
- Phase 3A: Generate actions for 1A scenarios
- Phase 3B: Probability assessment and visualization
- Phase 3C: Validation of advisor performance
Milestone 4: Model library integration
- Phase 4A: Integrate hybrid ML with training and validation tool
- Phase 4B: Stream data to surveillance tool
- Phase 4C: Training and documentation
- Phase 4D: Monitoring and support

The proposed methods build upon well-established techniques for empirical and fundamental modeling for fault detection and parameter estimation. This is intended to supplement the existing fault detection and prediction surveillance tools that are already in use such as Principal Component Analysis (PCA). Modern and extensible programming tools such as Python are proposed for this development to better incorporate the library package with other methods. The proposal also includes adaptation of methods with cutting-edge techniques in nonlinear optimization, large-scale dynamic simulation, and ML that extend the base capabilities of Python. The software libraries will be incorporated into the existing surveillance application for ESP monitoring and maintenance planning. An on-site (member location, 5 days) course on the ESP advanced predictive model library is proposed to transfer knowledge to engineers and data scientists. Software documentation and support will be provided to encourage model library adoption in the ESP surveillance workflow.