NOx Emission Soft Sensor for Coal-Fired Power Plant
A Just-In-Time Random Forest soft sensor predicts NOx emissions from coal-fired boiler process variables, achieving R²=0.93 and outperforming six benchmark models across variable load conditions.
Soft Sensor Solution
Approach
A Just-In-Time (JIT) learning framework selects historically relevant operating samples based on similarity to the current operating point, then fits a local Random Forest model on the selected subset. A Gaussian Process model provides uncertainty estimates alongside point predictions. A sliding window mechanism ensures the model reflects recent boiler behaviour, which changes with coal quality variation and load scheduling patterns.
Input Variables
- Coal feed rate (t/h)
- Primary air flow rate (Nm³/h)
- Secondary air flow rate (Nm³/h)
- Furnace exit temperature (°C)
- O2 content at furnace exit (%)
- Unit load (MW)
- Flue gas recirculation rate (%)
- Over-fire air damper position (%)
Output Variables
- NOx concentration at stack (mg/Nm³)
Model Type
- Just-In-Time Random Forest
- Gaussian Process
Update Strategy
- Just-in-time learning (query-based local fitting)
- Sliding window retraining
Technology Stack
- Python
- scikit-learn
- DCS integration
Key Performance Indicators
Results
-
The JIT-Random Forest model achieved R²=0.93 on a held-out test set from a 300 MW coal-fired unit, demonstrating robust performance across varying load conditions (150–300 MW range) and seasonal coal quality changes.
-
The JIT learning mechanism reduced prediction error by approximately 18% compared to a global (non-local) Random Forest baseline, confirming the value of locally adaptive modelling for processes with nonstationary operating regimes.
Why It Matters
- NOx emissions from coal-fired power plants are subject to strict regulatory limits (EU IED, US EPA MATS). Continuous Emissions Monitoring Systems (CEMS) are mandatory but expensive to install, calibrate, and maintain. A soft sensor provides a real-time predictive signal for combustion optimisation before CEMS limits are approached.
- Coal quality (calorific value, moisture, ash content) varies between shipments and within a stockpile, causing the NOx-emission relationship to shift over time. Just-in-time learning adapts automatically to these drifts without manual model retuning.
- By predicting NOx in real-time from existing DCS variables, the soft sensor enables closed-loop combustion optimization: air staging, over-fire air damper positioning, and load distribution can be adjusted proactively to minimize NOx while maintaining efficiency, rather than reacting to post-hoc CEMS readings.
Have a control challenge? Let's talk.
Sources
He et al. 2024 — JIT-RF soft sensor for NOx at 300 MW coal-fired unit, R²=0.93, outperforms 6 benchmarks. Field-validated results on industrial DCS data.
Reference for JIT-learning methodology in industrial soft sensors; contextualizes the local modelling approach used in power plant applications.
Pattern Overview
This pattern applies to coal-fired power plant boilers (typically 100–600 MW units) where NOx emissions must be continuously monitored for regulatory compliance. While a CEMS provides the legally required measurement, a soft sensor adds a predictive layer: by estimating NOx from existing DCS variables, the combustion control system can take corrective action before the CEMS reading rises toward the regulatory limit.
When to Use This Pattern
- The plant operates under regulatory NOx limits with mandatory CEMS measurement obligations.
- NOx varies significantly with load changes, coal quality shifts, and air distribution patterns.
- Combustion engineers want a fast-response NOx signal for manual or automated air staging optimization.
- A CEMS is already installed but a soft sensor is needed as a cross-validation or backup layer.
JIT Learning Rationale
A coal-fired boiler operates across a wide range of loads (turndown ratio typically 40–100%) and with varying coal properties. A single global model trained on all historical data averages across operating regimes that behave differently — resulting in poor predictions at extremes. Just-in-time learning solves this by querying the historical database for the K most similar operating points at prediction time and building a local model on the fly. This is computationally efficient because the query-and-fit cycle takes milliseconds on modern hardware, well within the 1-minute prediction interval required for combustion control.
Integration with DCS
The soft sensor receives inputs from the DCS historian via OPC-UA. Predictions are written back as DCS tags and displayed on the operator interface alongside the CEMS reading. Discrepancies between the soft sensor and CEMS are flagged as potential CEMS drift or sensor fouling events.
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.