Primex is an established provider of mission-critical information for power plant managers. Stewart Nicholson, the founder, believes Primex’s machine learning (ML) technology represents a step-change improvement in timeliness, precision, and confidence over traditional performance information systems. “The service helps managers make better decisions and make them sooner,” he concludes.
In their webinar series, viewable here, Primex shares how plant and operations managers use their ML technology to improve decision-making in power generation use cases. Using real-world examples, benefits are discussed, along with lessons learned and challenges remaining. The firm lists several big owners and operators as customers using their information systems at CCGT facilities.
The first session introduces an ML framework tailored to plant time-series data. Presenters define ML as a set of algorithms that map inputs, such as ambient conditions and system parameters, to outcomes like output, heat rate, and efficiency. They emphasize that these systems support, not replace, human decisions. Primex organizes use cases into three actions: detect, diagnose, and optimize. Detect identifies persistent deviations from a trained “normal,” diagnose traces which predictors drive the deviation, and optimize quantifies how changing controllable setpoints or configurations influences outcomes.
Case studies
In the Generation Monitoring and Diagnostics session, Primex demonstrates process-level and component-level models. Process models track KPIs like plant net output, generator set efficiency, boiler efficiency, and SCR performance, while component models focus on pumps, compressors, motors, or individual instruments for predictive maintenance.
A CCGT case study modeled steam turbine-generator output using weather, CT exhaust conditions, duct-firing energy, and pressures. After training on a normal post-outage period, the model flagged months with 5-10 MW underperformance. Root cause was a hot-reheat bypass leak. The estimated lost revenue before detection was about 250,000, illustrating earlier detection and faster value capture compared to conventional trending.
Hydropower modeling extends the approach using detailed basin weather and topography. Primex integrates hundreds to thousands of weather predictors to compare expected versus actual production. One run-of-river facility produced over 98% of expected MWh in 2022, with peak months within 0–1% of model expectations, confirming a strong fit and practical value for daily operations.
Diagnostics use model internals to correlate anomalous outcomes with likely drivers. In an HRSG efficiency example, ML reduced a large parameter set to a shorter list of the most correlated signals, accelerating on-site investigation. Primex stresses that domain expertise and walkdowns remain essential and that more than half of apparent issues trace to instrumentation faults.
Optimization examples include identifying the single most influential controllable factor on HRSG thermal efficiency, such as reheater outlet steam temperature, and quantifying the gain and decay from offline CT water washes to set optimal wash intervals.
Lessons learned include reducing false anomalies by addressing character drift, extrapolation error, and model complexity. Primex highlights the need to simplify models, ensure data quality, avoid information overload, and tailor outputs to what is actionable for operators. They also note that ML systems have certain operating costs and that realized benefits depend on anomaly frequency and magnitude.
The third webinar, Capacity and Fuel Demand Forecasting, applies ML to day-ahead capacity declarations and fuel planning for merchant markets. Primex notes that plants can reduce declaration errors by 50% or more by replacing static correction curves with continuously updated ML models. Because models are retrained daily, routine degradation, maintenance effects, and abnormal changes are quickly incorporated.
Market context matters. Using PJM examples, Primex shows monthly average day-ahead prices of roughly 20–50 per MWh and persistent day-ahead vs real-time spreads averaging 5–26, with monthly maxima often above 100 and sometimes exceeding 500. Accurate declarations mitigate risk when actual delivery differs from day-ahead commitments.
A 2×1 CCGT case illustrates the shift from spreadsheets to ML. With conventional methods, annual declaration error totaled about 23,000 MWh. ML, using more ambient predictors and daily updates, cut error to roughly 10,700 MWh on the same data set. The approach also optimizes weather inputs by blending multiple forecasts to reduce temperature-forecast error, a primary driver of capacity and heat-rate variance.
Fuel demand forecasting improves in parallel. Instead of a static base-load heat rate, ML predicts heat-rate variation across ambient conditions, typically within about 0.5%, supporting more precise gas nominations and lower imbalance costs.
How to watch. Primex has posted recordings for Introduction to Machine Learning and Power Plant Use Cases, Generation Monitoring and Diagnostics, and Capacity and Fuel Demand Forecasting. Visit www.primexprocess.com for access and additional materials. CCJ
