Abstract
Failure Prediction in Oil and Gas Artificial Lift Systems is materializing through the implementation of advanced analytics driven by physics-based models. During the Phase I of this project, two early failure prediction machine learning models were trained offline with historical data and evaluated through a blind test. The next challenge, Phase II, is to operationalize these models on Real-Time and re-assess their accuracy, precision and early prediction (in days) while having the assets focusing on either extending the runtime through optimization, chemical injection, etc. or proactive pump replacement (PPR) for high producers wells with triggered early prediction alarms. The paper details Phase II of live prediction for two assets consisting of 740 wells to enable data-driven insights in engineers’ daily workflow.
In Phase I, a collaboration between SMEs and Data Scientists was established to build two failure prediction models for Electrical Submersible Pumps (ESP) using historical data that could identify failure prone wells along with the component at risk with high precision. Phase II entails the development of a Real-Time scoring pipeline to avail daily insights from this model for live wells. To achieve this, PDO leveraged its Digital Infrastructure for extraction of high-resolution measured data for 750 wells daily. A Well Management System (WMS) automatically sustains physics-based ESP models to calculate engineering variables from nodal analysis. Measured and engineered data are sampled, and referencing learnt patterns, the machine learning algorithm (MLA) estimates the probability of failure based on a daily rolling data window. An Exception Based Surveillance (EBS) system tracks well failure probability and highlights affected wells based on business logic. A visualization is developed to facilitate EBS interpretation. All the above steps are automated and synchronized among data historian, WMS and EBS System to operate on a daily schedule. From the Asset, at each highlighted exception, a focus team of well owners and SME initiate a review to correlate the failure probability with ESP signatures to validate the alarm. Aided by physics-based well models, action is directed either towards a) optimization, b) troubleshooting or c) proactive pump replacement in case of inevitable failure conditions. This workflow enables IT infrastructure and Asset readiness to benefit from various modeling initiatives in subsequent phases.
Live Implementation of Exceptions from Predictive Analytics is an effective complement to well owners for prioritization of well reviews. Based on alarm validity, risk of failure and underperformance – optimizations, PPRs or workover scheduling are performed with reliability. This methodology would enable a Phase III of scaling up in Real-Time with growing assets wherethe system would be periodically retrained on True Negatives and maintained automatically with minimum manual intervention.
It is experienced that a high precision model alone is not enough to reap the benefits of Predictive Analytics. The ability to operate in a production mode and embedding insights into decisions and actions, determines ROI on Data Science initiatives. Digital Infrastructure, a Real Time Well Modeling Platform and Cognitive adaptation of analytics by Well Owners are key for this operationalization that demands reliable data quality, computational efficiency, and data-driven decisions philosophy.
DATA-DRIVEN FAILURE PREDICTION IN BRITTLE MATERIALS: A PHASE-FIELD BASED MACHINE LEARNING FRAMEWORK