Production Optimization

Upstream Operator — CIS

Well Stock Optimization via Virtual Metering and Predictive Maintenance

Deployed two integrated ML tools across a large ESP well fleet — restoring data reliability and eliminating unplanned downtime — without new wells or capital investment.

84%

ESP failure prediction accuracy

+1.6%

Average oil rate per well across the portfolio

<1.8%

Average virtual flow meter error

The Situation

A mature upstream operator with a large ESP well fleet faced two compounding problems that were eroding production performance and driving up maintenance costs.

First, production data was unreliable. Without accurate per-well flow measurement, regime decisions were based on estimates — creating a gap between planned and actual production that widened over time.

Second, ESP failures were reactive. Pump failures were detected after the fact, triggering unplanned workovers, lost production days, and emergency repair cycles that were difficult to plan or budget for.

What Was Not Working

No accurate per-well flow measurement — production accounting based on estimates

ESP failures detected reactively — no early warning system in place

Well regimes not optimized due to data quality gaps upstream

Maintenance planned on schedule, not on actual equipment condition

The Approach

Tool 01

Virtual Flow Meter

A hybrid physical-ML model was built and deployed to calculate per-well oil and fluid production rates in real time — without additional hardware. The physical model computes derived parameters (pump discharge pressure, motor cooling efficiency, fluid dynamics) which feed as engineered features into the machine learning layer.

Model inputs

  • 69 technological parameters per well
  • Well construction and ESP passport data
  • Historical failure and downtime records
  • Integration with OSISoft PI data infrastructure

<1.8%

Average error across full well stock

86%

Of wells below 5% error threshold

72%

Of wells below 1% error threshold

Tool 02

Predictive ESP Failure Detection

A hybrid model combining physics-based computation with statistical machine learning was trained on the operator's full failure history to predict both the occurrence and type of failure before it happens — enabling planned maintenance instead of emergency workovers.

Training data

  • 233 documented failures across 4 failure types
  • Physics-based features: motor cooling efficiency, pump discharge pressure, TMS data recovery
  • Integral (time-aggregated) signal features to capture degradation trends
  • Outputs: failure flag, failure type classification, equipment health index

84%

Overall failure prediction accuracy

4 types

Failure classes identified and classified

Verified Results

+1.6%

Average oil production rate per well

Across the full well stock, achieved through accurate flow measurement enabling better regime management.

84%

ESP failure prediction accuracy

Failures predicted before occurrence, enabling planned maintenance scheduling and elimination of emergency workovers.

<1.8%

Average virtual flow meter error

72% of wells achieved sub-1% measurement error. Full integration with existing enterprise data platforms.

Integration

Both systems were fully integrated with the operator's existing data infrastructure (OSISoft PI) and deployed within the client's IT environment. No additional field hardware was required. The anomaly detection and health monitoring interface consolidated all platform data into a single operational view accessible to production teams.

Similar challenges in your well stock?

We start with a four-week diagnostic. No obligation beyond that.