Reliability-Oriented Predictive Emission Monitoring of Gas Turbines Under Multiyear Distribution Shift and High-Emission Tail Risk

Reliable deployment of predictive emission monitoring systems (PEMSs) requires evaluation under future operating conditions and explicit attention to high-emission events. This study presents a pollutant-specific, risk-constrained framework for predicting Gas Turbine CO and NOx emissions under multiyear distribution shift. A dataset containing 36,733 observations was divided chronologically: 2011–2013 were used for model development, 2014 for configuration selection and calibration, and 2015 as a locked final test. Matched random within-prefix and forward temporal validation were compared, and Extra Trees was selected for both pollutants. Conventional, tail-weighted, covariate-shift-weighted, and combined configurations were evaluated under global-RMSE budgets of 2%, 5%, and 10%. Random within-prefix validation produced RMSE estimates 25.0% lower for CO and 34.0% lower for NOx than matched temporal validation. In 2015, CO tail weighting produced modest changes, reducing global RMSE by 0.9% and Q99 mean underprediction loss by 2.0%, without statistically resolved Q99 improvement. For NOx, tail weighting reduced Q99 RMSE by up to 5.5% while increasing global RMSE by 5.8%. Linear calibration reduced NOx global RMSE from 13.246 to 6.333 mg/m3 but worsened Q99 RMSE. Overall, temporal generalization, global calibration, and high-emission performance were distinct criteria, and no adaptation strategy was uniformly preferable.