Process inputs
One scenario, shared by both predictors
Loading the training-data ranges…
THE EVIDENCE BEHIND THE ESTIMATE
Model analytics
Saved predictions, measured outcomes and a closer look at error.
Charts follow each predictor’s model selection.
CO · Actual vs predicted
NOx · Actual vs predicted
Development plots use earlier fold fits, not the main fitted artifact. Pooled metrics use individual saved observations; R² can be negative. Later years were previously inspected and are retrospective evidence.
UNDERSTAND THE INVESTIGATION
Data & methods
A historical virtual sensor: estimate concurrent emissions from nine turbine measurements.
36,733hourly aggregate observations
9 → 2operating inputs → pollutants
5 yearschronological source order
Learn from earlier years
Main models fit all 2011–2013 observations. Development folds fit 2011 to assess 2012, then fit 2011–2012 to assess 2013. Configurations are selected on those folds.
Review later performance
2014–2015 stay outside model fitting. Since those years have been inspected, subsequent comparisons are retrospective. Strong training fit does not establish useful later performance.
Make uncertainty visible
Separate models fit 2011 and calibrate on 2012. Their nominal 90% or 95% intervals can under-cover later observations, especially NOx. They are not intervals for the main 2011–2013 models.
Measurements, units and data-quality notes
No fabricated dates, unit conversions, sensor corrections or outlier removal. Source row order does not prove uninterrupted hourly coverage. This is concurrent regression, not a forecast of future emissions.
Explore distributions & operating conditions
Distribution by source year
Turbine energy yield & inlet temperature
All source observations are retained. Associations are descriptive and do not establish causation.
Public UCI Gas Turbine CO and NOx Emission Data Set · DOI 10.24432/C5WC95 · CC BY 4.0. An engineering learning project and interactive proof of concept.
CHECK THE CALCULATIONS
Verification
Independent checks of calculations, data handling, model pipelines and reporting.
Evidence loads when this section is viewed.
Search individual check records
Verification checks software and reporting. Predictive validation examines behaviour on observations outside fitting. Neither certifies sensor accuracy or operational suitability.
FOLLOW THE SOURCE
Evidence library
Methods, experiment comparisons, original charts and downloadable records.