Definition
Percentage difference between actual measured emissions during transient conditions (startups, shutdowns, grade transitions) and the AI's predicted emissions for the same transients: detecting whether the compliance system underreports during the highest-emission periods of plant operation.
The Measurement Gap
Predictive emissions monitoring (PEMS) and AI-augmented emission prediction systems train on steady-state operation because that is where most of the data lives. The transient periods (startup, shutdown, grade transition, feed change) are short, complex, and underrepresented in training data. They are also the periods when emission rates are highest. A reformer fired to bring catalyst up to operating temperature emits more NOx in two hours than the next twelve hours of steady operation. A flare with feed-quality variation during a transition can spike VOC emissions in a way that no steady-state model has learned to predict. CEMS data and PEMS-reported emissions both report numbers; only the gap between them during transients reveals whether the prediction system is operationally trustworthy. Annual emission inventories average across operating modes, hiding transient-specific underprediction. Reportable event counts capture exceedances after they happen, not the unmeasured emissions that fell below the legal threshold because the prediction system told the plant they did.
Decision Supported:
Is the AI emissions prediction system accurate during the transient conditions that produce the highest emission rates, and if not, is the facility unknowingly exceeding permit limits during startups, shutdowns, and transitions?
Formula
Plain language: For each transient event in the measurement period, take the actual measured emissions and the AI-predicted emissions. Sum the differences across events, divide by the sum of predictions, multiply by one hundred. Positive means the AI underestimated; negative means it overestimated.
Notation:
TEAG = ( Σ ( E_actual_transient_k − E_predicted_transient_k ) ) / ( Σ E_predicted_transient_k ) × 100
Variables:
| Variable | Definition | Unit |
|---|---|---|
| E_actual_transient_k | Actual measured emissions during transient event k (CEMS or reference method) | lb of pollutant |
| E_predicted_transient_k | AI-predicted emissions for the same transient event k | lb of pollutant |
| k | Index of transient events (startups, shutdowns, grade transitions, feed changes) | n/a |
| Transient definition | Non-steady-state operating periods pre-specified in the prediction system's scope | n/a |
Track TEAG per transient type; startup emissions, shutdown emissions, and grade transition emissions often have different signatures.
Proposed formula, pending validation pilots.