NOx Emission Soft Sensor for Thermal Power Plants
A just-in-time-learning random forest soft sensor predicts NOx concentration from thermal power plant process data, reaching R² = 0.93 on the test set and outperforming the six comparison models of the study.
Soft Sensor Solution
Approach
Just-in-time learning (JITL) with a random forest: for every prediction, a local training set is selected from historical process data using a similarity index that combines temporal and spatial measures, and a random forest is fitted on it. An improved permutation-importance algorithm selects the relevant input variables, and moving-window process monitoring detects changes in working conditions and triggers online model updates.
Input Variables
- Process variables selected by improved permutation importance from the plant's operating data
Output Variables
- NOx concentration (mg/m³)
Model Type
- Random forest
- Just-in-time learning
Update Strategy
- Local model per query (just-in-time learning)
- Moving-window monitoring of working conditions
Key Performance Indicators
Results
-
On the test set the proposed method reached R² = 0.9319, RMSE = 3.696 mg/m³ and MAE = 2.772 mg/m³, with 99.71 % of samples within 15 mg/m³ absolute error — better than all six comparison models in the study (PLS, JIT-PLS, MW-PLS, random forest, moving-window random forest and plain JIT random forest; the closest was JIT random forest with R² = 0.9252).
Why It Matters
- Real-time, accurate NOx prediction is the basis for combustion optimization in thermal power plants: it gives the combustion control a fast signal to act on, alongside the regulatory emissions measurement.
- Working conditions change with load and fuel, so a single global model degrades. Local, just-in-time models with monitored updates keep the prediction aligned with the current operating regime.
Have a control challenge? Let's talk.
Sources
He, K., Ding, H. — open access. JITL random forest with improved permutation importance, temporal-spatial similarity and moving-window monitoring; results on the test set of plant operating data.
Pattern Overview
A soft sensor estimates NOx from process variables the plant control system already records, so combustion adjustments can be made before the emissions measurement reacts. It complements, and does not replace, the regulatory continuous emissions monitoring.
When to Use This Pattern
- NOx varies strongly with load, fuel and air distribution, and a faster signal than the emissions analyser would help combustion control.
- Several years of historized process data are available.
- The operating regime changes often enough that a single global model drifts.
Why just-in-time learning
A global model averages over operating regimes that behave differently. Just-in-time learning builds a local model around the current operating point from the most similar historical samples, and moving-window monitoring decides when the model must be updated. The cited study reports that this combination outperformed all six comparison models it evaluated on the same test data.
What to check before building one
- Whether the historized data covers the load and fuel range the sensor will see.
- How the prediction will be used (operator display, combustion optimizer, cross-check of the analyser) — this sets the accuracy and latency requirements.
- How drifts in the underlying instruments are detected, so the soft sensor does not silently learn a faulty measurement.
Related Patterns
Contact
Send a message
Direct contact
Dr. Rafał Noga
Stay Updated
Get insights on Industrial AI, APC, and process optimization delivered to your inbox.