Anomaly detection: Small deviations can mean big bucks – Combined Cycle Journal

Anomaly detection: Small deviations can mean big bucks

Primex, the latest entrant into the field of operational anomaly detection and machine-learning (ML) services for gas-turbine-based plants, introduced its technology to the wider combined-cycle (CC) community through a series of four webinars held in June and July 2023. The legacy application of the technology has been for SO2 scrubber and BOP performance at coal-fired plants. The firm lists several big owner/operators as customers, several of which apply the service to CC units, and considers itself a services provider, not a software supplier.

ML, in its basic configuration, harnesses computing power to recognize patterns in large data sets, in this case the input and outputs of an operating system, either the entire plant, or its subsystems. ML first “trains” on the system to develop the baseline, or normal, operating patterns among the variables important to the system. Real-time data going forward is compared to the baseline to detect anomalies.

The technology is system agnostic—that is, it doesn’t care what the system is, just its streams of input and output data. However, deep domain expertise with the operating system is necessary to convert the ML results into actionable insights. For this and other reasons, the Primex team stresses that ML is not a substitute for human expertise. The technology pulls the data from the plant-wide data network, such as PI.

The webinars make a clear distinction between machine learning (ML) and artificial intelligence (AI) and they answer other questions commonly asked by powerplant managers. Many more questions remain and interested parties should contact Primex for additional information. This is typical for virtually all ML technology firms. The one requirement in ML pattern recognition is a stable baseline for comparisons, which may be difficult for CC units which start and stop frequently and cycle up and down in load.

Stewart Nicholson, founder, believes that the software offers more granularity and a deeper level of precision and confidence over traditional methods of analyzing performance–such as vendor performance curves. One example cited: An additional megawatt of output could have been worth nearly half a million dollars over 12 months (2021-2022) in the PJM market.

Use cases presented in the webinars include generation monitoring and diagnostics (for example, fault detection), performance optimization, predictive maintenance (for example, early warning of unusual degradation), performance comparison (such as before and after a major outage, or unit event like a GT water wash or HRSG chemical clean), generation and resource demand forecasting (improving bidding strategies and fuel procurement), and regional supply and demand forecasting.

“The service helps you make better decisions and make them sooner,” the experts conclude. Visit for details.

Scroll to Top