Improving Compressor Station Uptime
A multinational energy company transports over 23 billion cubic feet (4 million barrels of oil equivalent) of natural gas per day across North America. The company operates 1,100 compressor units to transport natural gas across a 92,000 km (57,000 mi) pipeline network. When a compressor unit shuts down, natural gas throughput of the pipelines network is reduced. Reduction in throughput can impact the company’s profitability and the quality of service to its customers. Unplanned compressor unit outages cause between 20 and 30 days of downtime per unit annually. This translates to a profitability loss of $82 million per year.
Before this project, the company would address compressor unit outages by rapidly mobilizing resources, resulting in high-cost reactive maintenance. Addressing this concern, the company completed a trial of BHC3 Predictive Asset Maintenance™ to predict unplanned gas generator outages at compressor stations. Upon successful completion of the trial, the company can now effectively identify high-risk compressor units and dispatch resources up to 48 hours before a shutdown, thereby reducing maintenance costs and downtime.
About the Midstream Company
- $10+ billion annual revenue
- 90,000+ km of natural gas pipelines
- 20+ billion cubic feet per day of natural gas
transported across North America
"In about 12 weeks a team of 4 individuals built a predictive maintenance application.”
Forward Deployed Solutions Leader
Building and deploying an AI-based application
The team worked with the company to ingest, clean, and integrate five years of operational data from six source systems, totaling 4.6 billion records. The unified data image, including work orders, telemetry, outage events, gas composition measurements, and external weather, was used to predict unplanned gas generator outages. The team used both supervised and unsupervised (anomaly detection) machine learning model to optimize performance. The team trained and tested more than 4,000 supervised machine learning model permutations using more than 10,000 machine learning features, including over 250 SME-designed metrics to codify unit health. In parallel, the team configured 300 unsupervised machine learning model permutations to detect anomalous behavior on units with inconsistent or poor data quality. The team configured three user interface screens to visualize all data received and identify and prioritize high-risk compressor units.
- 12 weeks from kick-off to pre-production application
- Ingested 5 years of historical data from 6 source systems and external weather, comprised of over 4.6 billion records
- Created a unified data image across all systems
- Constructed 10,000+ time-based expressions to serve as feature inputs to the machine learning pipeline
- Built and tested 4,000+ machine learning model permutations to predict unplanned gas generator outages and detect anomalous behavior
- Configured BHC3 Predictive Asset Maintenance application user interface