Dynamic Prediction Model for NOx Emission of SCR System Based on Hybrid Data-driven Algorithms

Zhenhao Tang, Shikui Wang, Shengxian Cao, Yang Li, Tao ShenarXiv 2021 · arXiv

Abstract

Accurate real-time prediction of nitrogen oxide (NOx) emissions is critical for the efficient operation of Selective Catalytic Reduction (SCR) systems in coal-fired power plants. This paper presents a novel hybrid data-driven framework, designated as EC-VMD-ELM. By integrating non-linear correlation analysis for time-delay compensation, Variational Modal Decomposition (VMD) for signal denoising, and an Extreme Learning Machine (ELM) with an error-correction (EC) loop, the model achieves high precision. Experimental results from a 1000MW power unit demonstrate a Mean Absolute Percentage Error (MAPE) of 2.61% and a prediction speed of 0.011 seconds, making it suitable for real-time industrial control.

What Problem Does This Paper Address?

Selective Catalytic Reduction (SCR) systems serve as the primary technology for reducing NOx emissions in coal-fired power plants, converting pollutants into harmless nitrogen and water via ammonia injection. However, controlling these systems is inherently challenging due to high non-linearity, frequent disturbances, and significant time delays. The interval between ammonia injection and the subsequent sensor reading at the exhaust can span several minutes, complicating real-time optimization.

Schematic of an SCR reactor showing inlet and outlet monitoring parameters.

Traditional physical models based on computational fluid dynamics are often too computationally intensive for real-time applications. Meanwhile, standard machine learning models frequently struggle with the non-stationary, "noisy" nature of industrial operational data and the inherent system lag. This research aims to provide a robust, dynamic prediction framework that overcomes these limitations.

Methods

The authors propose a multi-stage hybrid pipeline termed EC-VMD-ELM. The architecture is designed to handle data quality issues and system dynamics through several specialized steps:

  1. Time-Delay Alignment (MIC): The Maximum Information Coefficient (MIC) is utilized to quantify non-linear correlations between 15 input variables and the output NOx. This allows for the "re-alignment" of historical data by shifting input timestamps to match their actual impact at the exhaust.
  2. Combined Feature Selection: To prevent model overfitting and reduce redundancy, a "voting" ensemble of three tree-based models (CART, Random Forest, and XGBoost) identifies the most influential operational variables.
  3. Signal Denoising (VMD): Industrial signals, particularly ammonia injection rates, are often non-stationary. Variational Modal Decomposition (VMD) decomposes these complex signals into simpler, cleaner sub-signals (modes), facilitating more effective learning by the neural network.
  4. Core Prediction (ELM): The primary predictor is an Extreme Learning Machine (ELM), a feedforward neural network favored for its rapid training speed and generalization capabilities.
  5. Error Correction (EC): A secondary ELM model is implemented to predict the residuals (errors) of the initial model based on previous time steps. The final output is the sum of the initial prediction and the predicted correction.
The proposed hybrid prediction framework combining an initial model with an error correction loop.

Datasets

The model was validated using a high-fidelity dataset from a 1000MW ultra-supercritical coal-fired power unit in China. The dataset comprised 9,210 sets of historical operating data sampled at 10-second intervals. For the study, 7,500 sets were used for training and 1,710 sets for testing. The 15 initial candidate variables included flue gas temperature, pressure, oxygen concentration, and ammonia flow rates.

Key Insights and Results

The study highlights that accounting for system lag is paramount; using MIC to align data significantly improved accuracy across multiple model architectures (ELM, DNN, SVR). Sensitivity analysis further revealed that while ammonia injection is the primary control variable, inlet oxygen concentration and flue gas temperature also exert substantial influence on NOx emissions, suggesting that SCR control strategies should be integrated with upstream boiler combustion parameters.

The proposed EC-VMD-ELM model demonstrated superior performance:

  • Accuracy: It achieved a MAPE of 2.61%, outperforming benchmarks such as Deep Neural Networks (DNN) and Support Vector Regression (SVR).
  • Efficiency: With an average prediction time of 0.011 seconds, the model is well-suited for integration into Distributed Control Systems (DCS) where sensors update every 10 seconds.
  • Robustness: The VMD component successfully mitigated the impact of signal noise, maintaining accuracy during load fluctuations.

Conclusion and Future Work

The EC-VMD-ELM framework provides a high-precision, low-latency solution for NOx emission prediction. While the current results are promising, future research will focus on exploring more advanced hybrid architectures for enhanced stability and testing the model's "transferability" to different plant configurations. The ultimate goal is the integration of this predictive model into a fully automated, closed-loop control system for real-time ammonia injection optimization.

Summary updated

BibTeX
@misc{tang2021dynamic,
  author = {Zhenhao Tang and Shikui Wang and Shengxian Cao and Yang Li and Tao Shen},
  title = {{Dynamic Prediction Model for NOx Emission of SCR System Based on Hybrid Data-driven Algorithms}},
  year = {2021},
  eprint = {2108.01240},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.01240},
}
Figure