Immunological AI Optimizer Deployment in a 330 MW Lignite-Fired Unit for NO_x Abatement

Abstract

This study presents an advanced NO_x reduction strategy for a 330 MW lignite-fired boiler using an immunological AI system: the SILO (Stochastic Immune Layer Optimizer) combustion optimizer inspired by artificial immune systems. The immunological AI optimizer adaptively models multi-variable interactions and fireball shape in real time, optimizing fuel–air mixing to reduce NO_x formation at the source. Unlike reactive secondary methods, the combustion optimizer reshapes the combustion process to reduce emissions while improving efficiency. Real-time temperature data from the AGAM acoustic system inform the combustion optimizer’s fireball modeling, ensuring combustion uniformity. A urea-based SNCR system serves as a secondary layer, controlled based on local furnace conditions to target thermal zones. Field results confirmed that SILO reduced NO_x emissions below 200 mg/Nm³, decreased urea consumption by up to 34%, and improved boiler efficiency by 0.29%. The architecture offers a scalable, DCS-integrated solution for aligning fossil-fueled operations with tightening emission standards.

1. Introduction

Despite the global shift toward renewable energy, fossil-fueled power generation remains a critical backbone of electricity supply in many countries of Europe, particularly in Central and Eastern Europe. In 2022, nearly 40% of electricity was generated from coal, oil, and gas [1]. However, the environmental costs of coal combustion, particularly the emission of nitrogen oxides (NO_x), have become a central regulatory and operational challenge. NO_x contributes significantly to acid rain, photochemical smog, and tropospheric ozone formation [2,3,4], making it a high-priority pollutant under both the EU Industrial Emissions Directive (IED) and Best Available Techniques Reference Document (EU BREF).

Traditionally, NO_x control in fossil fuel-fired boilers has relied heavily on Selective Non-Catalytic Reduction (SNCR) [5] and Selective Catalytic Reduction (SCR) [6] systems. While these secondary measures offer high removal efficiencies, their performance is often constrained by the following:

• Narrow temperature operating windows (especially for SNCR);
• High reagent costs and ammonia slip risks;
• Limited adaptability to fuel quality variations or boiler aging effects.

In response, the focus of modern combustion engineering has shifted toward primary NO_x reduction strategies—namely, optimizing the combustion process itself to minimize NO_x formation at the source [7]. This approach involves precisely regulating flame geometry, air–fuel ratios, and thermal distribution to suppress the formation of thermal and fuel-bound NO_x species.

Within this evolving framework, the immunological AI combustion optimizer emerges as a powerful AI-inspired tool that manages combustion conditions through real-time adaptive control. Built on immune system principles, the optimization system continuously models the process, identifies optimal operating points, and adjusts combustion variables to maintain both efficiency and low emissions. When coupled with AGAM acoustic temperature measurement, the system can infer and respond to temperature profiles in the furnace, allowing finer control over the flame shape, thermal zoning, and ultimately the effectiveness of NO_x mitigation strategies like SNCR [8,9].

This study documents the comprehensive modernization of a lignite-fired 330 MWe boiler, focusing on the following:

• Deploying the SILO combustion optimization system;
• Integrating AGAM acoustic thermometry for real-time fireball profiling;
• Enhancing SNCR performance by ensuring ideal temperature zones;
• Achieving strict compliance with NO_x emission limits (<200 mg/Nm³) across all operating loads using only native lignite fuel.

This study introduces an innovative temperature control strategy by integrating AGAM acoustic thermometry directly into the SILO optimization framework. Unlike conventional systems, the SILO uses real-time spatial temperature data to infer and adjust the fireball geometry inside the furnace, ensuring thermal uniformity. This temperature-aware optimization enables precise control of combustion zones, which enhances NO_x reduction and improves SNCR reagent efficiency. By presenting the technical framework, test methodology, and performance outcomes of this deployment, the paper aims to demonstrate how AI-driven combustion control can serve as the primary measure for NO_x reduction, with SNCR as a secondary measure, in modern coal-fired power generation.

2. Background and Literature Review

2.1. Advanced Combustion Control

The global market for combustion control optimization solutions is expected to grow from USD 4.3 billion in 2024 to USD 7.4 billion by 2031 [10]. The market for industrial combustion control components and systems is projected to increase from USD 26.4 billion in 2024 to USD 35.2 billion by 2030, with a CAGR of 4.9% [11]. China is forecasted to reach USD 7.8 billion by 2030 with a CAGR of 7.5%, while the U.S. market was estimated at USD 6.9 billion in 2024.

Traditional combustion control in coal-fired boilers relies on PID-based closed-loop or open-loop schemes, typically reactive and manually tuned [12]. The growing complexity of emissions constraints and operational variability has led to wider use of advanced methods like Model Predictive Control (MPC) and steady-state optimization. Evolutionary algorithms (e.g., GA, PSO) and AI models (e.g., neural networks, SVM) are increasingly adopted for real-time boiler optimization [13].

Recent advances in combustion optimization have focused on the integration of adaptive, self-learning systems to handle process nonlinearities and time-dependent variability in power generation environments. These intelligent control architectures use real-time plant data to continuously update control policies, enabling dynamic response to fluctuating loads and varying fuel properties. For example, Ref. [14] discusses a nonlinear Model Predictive Control (MPC) implementation in MATLAB for boiler optimization, demonstrating enhanced adaptability. Similarly, the DeepThermal system applied offline reinforcement learning (RL) techniques to large-scale thermal units with promising results [15]. Notably, a MIMO NARX neural network-based nonlinear MPC reduced NO_x emissions to 23.84 mg/m³ and cut temperature imbalances by 64.2% in a 660 MW plant [9]. A “W”-type flame furnace study used orthogonal experimental design to evaluate burner configuration, wind distribution, and oxygen content on NO_x and efficiency across various loads [16]. AI and genetic algorithms were applied in another case to a coal-fired plant, achieving 15% NO_x and 12% ammonia reduction, as well as a 0.57% improvement in net heat rate [17]. The Improved Constrained Fuzzy Association Rule (ICFAR) framework combined with multi-objective particle swarm optimization (MOPSO) facilitated real-time adaptive tuning, increasing boiler efficiency by 0.082% and lowering NO_x by 6.49 mg/m³ [18]. The PDOBLPSO algorithm outperformed conventional methods for a 600 MW boiler, cutting NO_x by up to 37.72 mg/Nm³ with high stability [19]. A twin-delayed deep deterministic policy gradient (TD3) approach enhanced thermal efficiency by 0.411% and reduced NO_x by 17.7 mg/m³ in a 600 MW down-fired unit while preserving wall temperature integrity [20]. Hybrid optimization frameworks using Kernel Extreme Learning Machine (KELM) and MOEA/D regulated airflow under static loads, whereas ARMAX-PSO and TG-MOEA/D optimized dynamic control parameters, achieving a 12.18% NO_x reduction without efficiency loss [21]. Furthermore, an online framework combining an improved cultural algorithm (CIPSODM), LSTM neural networks, and mutation-enhanced case-based reasoning achieved a 0.588% thermal efficiency increase and a 14.18 mg/m³ NO_x reduction [22]. Another comprehensive architecture with IBBO-LSTM, fuzzy safety indices, and INSGA-II delivered 0.653% higher efficiency and reduced NO_x by 22.89 mg/m³, improving operational safety and emissions [23].

These findings underline the transformative potential of intelligent, real-time adaptive combustion control systems. By integrating AI, evolutionary algorithms, and robust data-driven models, modern power plants can achieve significant environmental, economic, and performance gains. Compared to traditional methods like PID and Model Predictive Control (MPC), the SILO offers greater adaptivity and robustness by continuously updating its internal static model using immune-inspired mechanisms. Unlike MPC, which requires periodic re-identification and manual tuning, the SILO autonomously adapts to load variations, fuel changes, and system aging using real-time plant data. It eliminates the need for costly offline model development and instead builds process knowledge online, enabling faster deployment and lower operational costs. While MPC performs better in transients due to its dynamic modeling, the SILO’s self-learning and stochastic layers ensure long-term performance without manual intervention.

2.2. SILO—Immunological AI Combustion Optimizer

The SILO, as an advanced control (AC) [24]-class solution, is specifically designed for the automatic, real-time optimization of industrial processes, with a particular emphasis on combustion processes in power boilers. Its design concept draws inspiration from the biological immune systems of living organisms, a methodology detailed in [25].

The SILO is an immunologically inspired combustion optimization system composed of two independent modules: knowledge gathering and optimization. The term “Immunological” reflects the system’s biological foundation, drawing on principles such as memory, adaptation, and affinity-based learning—similar to the human immune system—to manage complex industrial processes. The knowledge gathering module monitors process signals and identifies static relationships between manipulated variables (MVs) and controlled variables (CVs) under constant disturbance variables (DVs). These learned relations, stored as “lymphocytes”, enable the SILO to predict process behavior across varying operating conditions. The optimization module uses this knowledge to continuously calculate and adjust MV setpoints, transmitting corrections to the boiler’s control layer. A single internal parameter—the quality indicator—guides optimization by integrating all control objectives and their relative priorities.

$$\begin{array}{cc}\mathrm{J}=& {\displaystyle \sum _{\mathrm{k}=1}^{{\mathrm{n}}_{\mathrm{m}}}}\left[{\mathsf{\alpha }}_{\mathrm{k}}{\left(\left|{\check{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}-{\tilde{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{l}\mathrm{m}}\right)}_{+}+{\mathsf{\beta }}_{\mathrm{k}}{\left({\left(\left|{\check{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}-{\tilde{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{s}\mathrm{m}}\right)}_{+}\right)}^2\right]+\hfill \\ & +{\displaystyle \sum _{\mathrm{k}=1}^{{\mathrm{n}}_{\mathrm{y}}}}\left[{\mathsf{\gamma }}_{\mathrm{k}}{\left(\left|{\check{\mathrm{y}}}_{\mathrm{k}}-{\tilde{\mathrm{y}}}_{\mathrm{k}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{l}\mathrm{y}}\right)}_{+}+{\mathsf{\delta }}_{\mathrm{k}}{\left({\left(\left|{\check{\mathrm{y}}}_{\mathrm{k}}-{\tilde{\mathrm{y}}}_{\mathrm{k}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{s}\mathrm{y}}\right)}_{+}\right)}^2\right]\hfill \end{array}$$

(1)

The goal function of the optimization module to minimize the value of the quality indicator is presented below:

$$\begin{gathered} \begin{array}{c}\min \\ {∆\mathrm{m}}^{\mathrm{d}}\end{array}\left\{\sum _{\mathrm{k}=1}^{{\mathrm{n}}_{\mathrm{m}}}\left[\begin{array}{c}{\mathsf{\alpha }}_{\mathrm{k}}{\left(\left|{\check{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}+{∆\mathrm{m}}^{\mathrm{d}}-{\tilde{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{l}\mathrm{m}}\right)}_{+}\\ +{\mathsf{\beta }}_{\mathrm{k}}{\left({\left(\left|{\check{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}+{∆\mathrm{m}}^{\mathrm{d}}-{\tilde{\mathrm{m}}}_{\mathrm{k}}^{\mathrm{c}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{s}\mathrm{m}}\right)}_{+}\right)}^2\end{array}\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}+\sum _{\mathrm{k}=1}^{{\mathrm{n}}_{\mathrm{y}}}\left[\begin{array}{c}{\mathsf{\gamma }}_{\mathrm{k}}{\left(\left|{\check{\mathrm{y}}}_{\mathrm{k}}+{∆\mathrm{m}}^{\mathrm{d}}{\mathrm{K}}_{\mathrm{k}}-{\tilde{\mathrm{y}}}_{\mathrm{k}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{l}\mathrm{y}}\right)}_{+}+\\ {\mathsf{\delta }}_{\mathrm{k}}{\left({\left(\left|{\check{\mathrm{y}}}_{\mathrm{k}}+{∆\mathrm{m}}^{\mathrm{d}}{\mathrm{K}}_{\mathrm{k}}-{\tilde{\mathrm{y}}}_{\mathrm{k}}\right|-{\mathsf{\tau }}_{\mathrm{k}}^{\mathrm{s}\mathrm{y}}\right)}_{+}\right)}^2\end{array}\right]\right\} \end{gathered}$$

(2)

The K_k matrix represents static gains in the mathematical model, which are automatically identified at each optimization period [26]. In the artificial immune system framework, the process model is constructed based on a set of selected lymphocytes that correspond to the current operating point of the process. This model is assumed to be static and linear. Throughout operation, the artificial immune algorithm continually gathers new lymphocytes, which are then incorporated into subsequent optimization cycles to ensure that the model adapts to evolving process conditions.

One of the core advantages of this approach is the SILO’s inherent ability to track long-term drift in system dynamics without requiring manual intervention or periodic re-identification. Unlike conventional MPC frameworks that often necessitate comprehensive re-modeling every 1–2 years to correct for gradual system wear or fuel variability, the SILO’s immune-inspired memory and continuous update mechanism maintain model relevance. In cases where previously stored lymphocytes no longer match new behavior patterns, the system can switch to stochastic exploration, enabling it to re-learn the updated optimum autonomously. This capacity for lifelong adaptation ensures robust performance in the face of long-term changes.

2.3. Acoustic Gas Temperature Measurement (AGAM)

AGAM is an acoustic measurement system designed to monitor in-furnace temperature distribution by capturing the thermal profile of combustion gases across a horizontal cross-section of the boiler. AGAM technology has been comprehensively detailed in studies [27,28,29,30,31], with [30,31] specifically highlighting system characteristics critical for combustion optimization.

The underlying physical principle of AGAM relies on the relationship between the speed of sound in a gas and the temperature and composition of that gas.

In industrial settings, the AGAM system is composed of multiple transmitters and receivers installed at the same elevation within the furnace. Each measuring path is defined as the direct line between a transmitter and a receiver. Temperature measurement along each path is based on the travel time of an acoustic impulse using the following equation.

$$\mathrm{T}={\displaystyle \frac{{\mathrm{l}}^2}{\mathsf{\kappa }·\mathrm{R}·{\mathsf{\tau }}^2}}$$

(3)

A network of such paths forms a measurement mesh, enabling the system to capture a detailed two-dimensional temperature distribution across the furnace cross-section. Figure 1 illustrates a sample configuration of the AGAM system installed at a coal-fired power plant, showing the arrangement of transmitters and receivers along with the resulting measurement mesh (24 paths).

As analyzed in [30], the precision of temperature measurements obtained using AGAM technology is highly robust, with deviations due to changes in combustion gas composition remaining below 1.4% and the overall measurement error not exceeding 2%.

2.4. SNCR: Urea-Based Mechanisms, Limitations, and Temperature Sensitivity

In fossil fuel-fired pulverized coal boilers, nitrogen oxide (NO_x) emissions arise from three key mechanisms: thermal NO_x, prompt NO_x, and fuel NO_x [31].

Thermal NO_x is produced at flame temperatures exceeding 1300 °C through the Zeldovich mechanism, governed by the following reactions:

N₂ + O → NO + N

(4)

$$\mathrm{N}+{\mathrm{O}}_2 \to \mathrm{NO}+\mathrm{O}$$

(5)

Prompt NO_x results from the interaction of hydrocarbon radicals with molecular nitrogen, forming intermediates such as HCN via reactions like

C_mH_n + N₂ → C_mH_nN + N

(6)

and

CH + N₂ → HCN + N

(7)

followed by oxidation to NO.

Fuel NO_x, produced from nitrogen bound in coal, is the dominant NO_x source during combustion. It forms through devolatilization—where volatile nitrogen species like NH₃ or HCN are released—and from the oxidation of char-bound nitrogen. Combustion conditions significantly influence these pathways; high excess air favors NO formation, while low O₂ promotes CO formation, thereby suppressing NO. Primary combustion measures thus aim to create conditions that favor nitrogen reduction, with optimal boiler outlet O₂ typically maintained between 2.6% and 3.1% (dry, vol.).

Selective Non-Catalytic Reduction (SNCR) is a post-combustion technology that reduces NO_x by injecting ammonia or urea into the flue gas, where it reacts with NO and NO₂ to form N₂ and H₂O. This reaction is highly temperature-dependent, operating efficiently within an 850–1100 °C window [32,33].

In SNCR systems, ammonia or urea solution is atomized using water, which rapidly evaporates to enhance reagent dispersion across the furnace cross-section. Nozzle types include mechanical atomizers, retractable multi-nozzle lances, dual-fluid injectors, and air-aspirated nozzles. In situ nozzle adjustments are used to optimize droplet size and velocity according to furnace conditions.

A major challenge for SNCR in large boilers is achieving adequate reagent penetration into the furnace center. To address this, pre-installation CFD modeling is performed, using detailed temperature profiles to optimize injection strategies. Although SNCR effectively targets high-temperature, high-NO_x zones, larger boiler applications face issues like reagent distribution and ammonia slip. Direct urea injection has been adopted to improve performance for larger units. The simplified chemical reactions for NO_x reduction using urea are

2 NO + NH₂CONH₂ + 0.5 O₂ → 2 N₂ + CO₂ + 2 H₂O

(8)

2 NO + 2 NH₃ + 0.5 O₂ → 2 N₂ + 3 H₂O

(9)

These reactions are exothermic and most effective within a temperature window of 800 °C to 1300 °C. As flue gases cool through the superheater sections, the reaction rate declines, increasing the risk of unreacted ammonia (“ammonia slip”) downstream. Furnace temperature imbalances are further influenced by boiler loading conditions and mill configurations, which affect the spatial distribution of combustion heat release. Variations in mill performance or load introduce furnace temperature fluctuations, undermining NO_x reduction efficiency. Thus, continuous combustion monitoring and real-time control are essential to maintain optimal conditions for SNCR performance across all loads.

Numerical modeling of SNCR-based NO_x reduction in an OP 650 boiler showed that optimized urea injection at proper nozzle heights and within the effective temperature window can lower NO_x emissions below 200 mg/m³ (dry, 6% O₂), supporting cost-effective compliance with EU standards [33]. In [34], it was also demonstrated that multi-hole nozzles improve reagent distribution and reduction efficiency and that a 12% urea solution yields better NO_x reduction than a 32.5% solution due to improved droplet penetration and interaction with the flue gases.

CFD simulations using detailed chemical kinetics in a biomass boiler indicated that SNCR using ammonia could achieve up to 63% NO_x reduction, with performance highly dependent on injection height and ammonia flow rate [35]. An experimental study in a CFB lignite-fired boiler showed that spatial control of urea distribution and furnace flow dynamics enables effective NO_x reduction under both steady-state and transient conditions [36]. A comprehensive review of the SNCR process emphasizes the complex interplay of reagent evaporation, gas-phase chemistry, and flow physics, highlighting the importance of computational fluid dynamics (CFD) and advanced techniques such as large eddy simulation (LES) for system design and optimization [37].

Modern SNCR systems attempt to improve reliability by integrating real-time feedback from in-furnace temperature systems like AGAM and using load-adaptive injection strategies. Nonetheless, without simultaneous optimization of combustion conditions, SNCR performance can still fluctuate significantly, especially during load transitions or fuel quality changes.

2.5. Gaps in Current Research and Contributions of This Work

Despite extensive research into combustion optimization and nitrogen oxide (NO_x) control, key gaps persist in current literature and industrial practice. Traditional NO_x formation models, including thermal NO_x (Zeldovich mechanism), prompt NO_x (radical interactions), and fuel NO_x from devolatilization, have been well studied, but their real-time application in large-scale dynamic plant optimization remains limited.

Notable gaps include the following:

• Real-time application of NO_x formation models under varying load and coal quality;
• Dynamic combustion optimization integrating spatial furnace temperature profiling (e.g., acoustic thermometry) with intelligent setpoints;
• Seamless coordination of primary NO_x control (combustion tuning) and secondary SNCR strategies under transient conditions;
• Limited integration of AI-driven systems (MPC, RL) with direct in-furnace temperature measurements for enhanced control responsiveness.

This study addresses these challenges by implementing an integrated optimization platform combining the SILO combustion optimizer with AGAM acoustic temperature mapping to dynamically guide urea-based SNCR operation.

The research advances the field as follows:

• Prioritizing primary NO_x reduction via intelligent combustion and air staging;
• Creating a self-adaptive, data-driven optimization framework reacting to furnace asymmetries and NO_x generation zones;
• Validating the approach on a 330 MW lignite-fired boiler, achieving NO_x emissions consistently below 200 mg/Nm³ across varying loads.

3. Project Objectives

The primary objective of this project is to achieve stable and reliable operation of the boiler across the full load range (50–100%), utilizing exclusively lignite coal from domestic sources, while ensuring compliance with the legally mandated NO_x emission limits (ref.

Table 1. Operational and environmental objectives of the project.

Category	Indicator	Target/Requirement
Operational Range	Load range	50–100%
	Fuel type	Domestic lignite only
Efficiency	Boiler thermal efficiency	≥88.5% (rated conditions)
	Specific fuel consumption	≤design value
Emissions	NOx emissions	≤200 mg/Nm3 (6% O2) across all load conditions
	CO emissions	≤100 mg/Nm3 (Best Available Techniques (BAT) Reference Document for Large Combustion Plants)
	Ammonia slip	≤5 mg/Nm3 (Best Available Techniques (BAT) Reference Document for Large Combustion Plants)
Ash Quality	Unburned fuel in bottom ash	<35%
	Unburned fuel in fly ash	<3.5%
Reagent Use	Solid urea consumption for SNCR	≤382.3 kg/h (equivalent to 955.75 kg/h of 40% urea solution) (SNCR design parameter for optimized combustion)
	Water consumption for urea solution preparation	≤86.4 hectoliters/hour (SNCR design parameter for optimized combustion)
Cost Optimization	SNCR operation and maintenance costs	Minimized through adaptive control

). The target NO_x emission concentration is set at ≤200 mg/Nm³, corrected to 6% oxygen content (EU Industrial Emissions Directive IED 2010/75/EU for existing plants > 300 MWth).

Achieving these operational and environmental objectives is a fundamental requirement, forming the basis for evaluating the success of the combustion optimization and NO_x reduction measures.

4. Boiler Configuration and Research Platform

The boiler is a tower-type boiler with a design steam generation capacity of 1035 t/h and a rated thermal input of 878 MW_t.

The combustion system is equipped with six coal mills (each rated at 100 t/h), operating in a direct injection configuration. Coal is fed to the mills via an ERKO-type scraper feeder, where it undergoes drying and pulverization using recirculated flue gas drawn from the upper furnace. Pulverized coal, combined with primary air, is delivered through differentiated dust pipes to six slotted burners positioned along the furnace walls. Each burner facilitates the injection of a primary mixture—comprising coal dust, primary air, flue gas, and evaporated moisture—while secondary air is independently introduced through ports around the burners. The interaction of primary and secondary air jets promotes flue gas recirculation, supporting thermal decomposition, ignition, and sustained combustion of the fuel.

The boiler operates under a Distributed Control System (DCS), maintaining key design parameters at 100% load, including a high-pressure steam flow of 1035 t/h at 192 bar and 540 °C, as well as an intermediate steam outlet at 540 °C and 48.2 bar. This ensures efficient and stable operation across a broad load range (50–100%) while utilizing low-grade lignite. The fuel characteristics of the used lignite include a lower calorific value between 1600 and 1900 kcal/kg, with an elemental composition of carbon (20–23%), hydrogen (1.7–2.1%), sulfur (0.6–1.1%), oxygen (8.5–9.89%), nitrogen (0.5–0.8%); moisture content ranging from 41.21% to 45%; and ash content between 22% and 25%. These properties present notable combustion challenges, which are addressed through advanced control and optimization strategies.

Air distribution is managed via general air dampers supplying air to front-left mills (1, 2, and 6) and rear-right mills (3, 4, and 5), designated as “FL-1,2,6” and “RR-3,4,5”, respectively. Each mill is equipped with dedicated secondary air (SA) dampers. Additionally, four Over Fire Air (OFA) dampers distribute air to specific furnace zones: upper-front-left (“OFA-U-FL”), upper-rear-right (“OFA-U-RR”), lower-front-right (“OFA-L-FR”), and lower-rear-left (“OFA-L-RL”). The system also includes oil dampers used mainly during start-up and for air distribution optimization. Two general oil dampers supply air to the front and rear furnace zones (“OIL-FRONT” and “OIL-REAR”), each further divided into left (“OIL-F-L” and “OIL-R-L”) and right (“OIL-F-R” and “OIL-R-R”) sections. The overall fuel and air distribution configuration is illustrated in Figure 2.

At the furnace bottom, an afterburner grate ensures further slag combustion, with resultant slag transported via a scraper conveyor to slag crushers, after which slag is hydraulically removed.

4.1. Immunological AI Combustion Optimization System—SILO

The combustion optimization system is deployed as the primary NO_x reduction technology on the boiler. It is fully integrated with DCS, with dedicated modifications applied to existing DCS control logics to enable seamless interaction between SILO-generated setpoints and the native control loops.

The key control priorities for the SILO include maintaining NO_x emissions below 200 mg/Nm³ and CO concentrations below 100 mg/Nm³, minimizing solid urea consumption for the SNCR system, reducing excess air levels, and minimizing boiler efficiency losses related to unburned carbon in both bottom ash and fly ash, as well as excess O₂ in the flue gases.

A critical innovation of this study lies in the integration of AGAM acoustic thermometry directly into the SILO optimization framework to enhance combustion control and SNCR efficiency. Unlike conventional systems, the SILO leverages real-time, spatially distributed temperature data from AGAM sensors to infer and adjust fireball geometry, ensuring thermal uniformity across left–right and front–rear furnace axes. This uniformity is crucial, as the effectiveness of urea-based SNCR depends on maintaining the flue gas within the optimal temperature window (approximately 850–1100 °C) and in regions with high NO_x concentration. By continuously learning the relationships between air distribution settings and AGAM-measured temperature patterns, the SILO adjusts damper positions—including secondary air (SA), oil, and over fire air (OFA) dampers—on both the front–rear and left–right sides. For example, increasing airflow to the front or left and reducing it at the rear or right can shift the fireball toward the center of the furnace, aligning it with the SNCR injection zone. This dynamic and data-driven strategy enables the SILO to maintain optimal fireball placement across varying lignite mill configurations, each of which influences furnace heat distribution. As a result, the system ensures that SNCR injectors always target the most reactive zones, maximizing NO_x reduction while minimizing ammonia slip and reagent consumption.

The SILO configuration was as follows:

• Twenty-five MVs, including 6 × coal feeders, 6 × secondary air dampers, 2 × general air dampers, 4 × OFA dampers, 6 × oil dampers, and 1 × oxygen setpoint bias;
• Fifteen CVs, including NOX and CO emission, urea flow, O₂ content in flue gases—left and right flue gas duct, superheated and reheated steam temperatures, gas temperatures from AGAM system—8 temperatures for 8 zones;
• Three DVs: unit load, pulverizers’ configuration, and estimated fuel quality.

The SILO updates every 3 min, allowing timely adaptation to process dynamics. The learning rate is not fixed but adaptively determined based on model drift and performance changes. Identification experiments involve altering one manipulated variable and observing the response over a 15 min window, ensuring accurate capture of process relationships. This approach enables SILO to maintain model accuracy without manual intervention.

4.2. SNCR Installation

A Selective Non-Catalytic Reduction (SNCR) system using 40% urea solution was installed—see Figure 3. The SNCR system is composed of four independently operating mixing and distribution modules. Each module is equipped to supply urea, dilution water, and atomizing air, and it can independently dose two injection levels based on real-time furnace conditions. The active injection level—either at 37 m or 44.5 m elevation—is dynamically selected according to the AGAM-measured temperature field to ensure reagent spraying occurs within the optimal 850–1100 °C reaction window. Urea preparation begins with batch production (12,500 kg per batch) and storage in a 100 m³ insulated tank, supported by a booster pump station for dilution water. A total of 48 injectors, organized into 8 spray groups, are distributed across the two levels for efficient reagent delivery. The control and regulation layer is managed by an SNCR PLC interfaced with the DCS and SILO. It receives continuous input from NO_x analyzers, AGAM temperature sensors, and boiler load data to operate in closed-loop control mode. The system includes automatic ammonia slip correction and adaptive lance activation to maintain optimal performance under varying operating conditions.

The SNCR injectors at level 1 or level 2 are activated when SILO-optimized furnace conditions align with the urea reaction temperature window, ensuring high denitrification efficiency and minimal reagent loss.

5. Results and Discussion

The evaluation of the combustion optimization and NO_x reduction project was based on the analysis of historical process data collected from the Distributed Control System (DCS). The NO_x content in the flue gases was measured using a calibrated NO_x analyzer installed at the unit stack, which reports dry gas concentrations corrected to 6% oxygen, as required by the National Environmental Agency standards.

Boiler efficiency was calculated using a dedicated block performance evaluation system, incorporating an algorithm based on the European Standard EN 12952-15.

5.1. Performance Testing Methodology

Two primary performance tests were conducted:

• An initial test (baseline) prior to any system upgrades;
• A final test after commissioning the SNCR installation and the SILO combustion optimization system.

The final performance test was divided into two phases:

• Phase 1: SNCR operational, SILO deactivated;
• Phase 2: both SNCR and SILO operational.

The initial and final testing adhered to the following preconditions:

• Boiler load control maintained in AUTO mode with a stable load ± 5%;
• Collection of fly ash and bottom ash samples hourly;
• Automatic operation of other control loops (airflow, steam pressure, feedwater, etc.);
• No adjustments permitted to mill configurations;
• No sootblower operations allowed during the test;
• SILO combustion optimizer and SNCR systems active (only final performance test).

Each load condition (200 MW, 265 MW, 300 MW) was held for approximately three hours, with operator setpoints maintained constant.

5.2. Project Results

Below is a summary of 24 h averaged performance (final test, when both SNCR and the SILO are in operation):

• Average load: 246.5 MW;
• NO_x emissions (corrected to 6% O₂): 194.38 mg/Nm³;
• Unburned carbon in bottom ash: 30.85%;
• Unburned carbon in fly ash: 0.75%;
• Average urea flow: 788.5 kg/h (solid urea equivalent: 315.4 kg/h);
• Water consumption: 78.48 hl/h;
• CO emissions: 54.9 mg/Nm³;
• NH₃ slip: 1.05 mg/Nm³;

Detailed information on the optimization results is presented in the figures.

As illustrated in Figure 4 the NO_x emissions were significantly reduced after implementation of the SNCR system and SILO combustion optimization. Compared to the baseline values (ranging from 451.6 to 584.2 mg/Nm³), emissions were reduced below the regulatory limit of 200 mg/Nm³ at all tested loads (200 MW, 260 MW, 300 MW). Further improvement was observed when the SILO was activated, achieving the lowest NO_x levels, particularly at 260 MW and 300 MW.

Figure 5 clearly demonstrates that the activation of SILO combustion optimization significantly reduced urea consumption across all tested load conditions. Compared to operation with only SNCR (SILO OFF), urea flow was reduced by 12.7% to 21.8%, with the greatest improvement observed at 300 MW, where urea flow decreased from 869.6 kg/h to 680.3 kg/h. This demonstrates that SILO enhances SNCR efficiency by optimizing combustion and reducing reagent demand.

The SILO significantly reduced carbon in bottom ash at all loads, dropping from 47.62% to 28.98% at 200 MW—well below the 35% limit (ref. Figure 6). Improvements were also seen at 260 MW and 300 MW. Fly ash carbon remained under 3.5% in all cases, with a slight increase at lower loads when the SILO was active but still within limits (ref. Figure 7).

According to the data presented in Figure 8, SILO implementation improved boiler efficiency at all load levels. Increases of 0.28%, 0.55%, and 0.04% were observed at 200 MW, 260 MW, and 300 MW, respectively. Overall, efficiency was maintained or enhanced without adverse effects on plant performance.

As evidenced in Figure 9, CO emissions slightly increased with SILO activation but remained well below the regulatory limit of 100 mg/Nm³. At 260 MW, CO levels dropped significantly compared to the baseline, indicating improved combustion efficiency. The observed increase in CO at other load levels is a direct consequence of the optimizer targeting reduced excess air levels (lower flue gas O₂) to enhance thermal efficiency and suppress NO_x formation. Since CO and O₂ are inversely correlated in combustion systems, the marginal rise in CO reflects the SILO’s active search for an optimal excess air setting—balancing oxygen supply just above the threshold needed for complete combustion. Figure 10 clearly demonstrates the NH₃ slip also stayed below the 5 mg/Nm³ threshold across all loads. The elevated slip observed at 300 MW (2.01 mg/Nm³) is linked to a delayed SNCR response despite achieving low NO_x (186.5 mg/Nm³).

According to the data presented in Figure 11 and Figure 12, after SILO activation, SH and RH steam temperatures increased across all loads, with improved uniformity between left and right sides. The combined average temperature rise exceeded +1.6 °C, enhancing thermal efficiency and promoting more balanced heat distribution in the boiler.

As visualized in Figure 13, activation of the SILO led to reduced O₂ content in flue gases at all load levels, indicating more efficient combustion with less excess air. Notably low O₂ on the boiler’s right side is attributed to specific coal mill configurations affecting local combustion dynamics.

5.3. Post-Commissioning Tuning of SILO Combustion Optimizer

After initial commissioning, additional SILO tuning was performed to reduce high urea consumption at elevated loads by improving furnace temperature uniformity. Using AGAM data, the optimizer’s weighting for left–right and front–rear imbalances was increased to better align the combustion center with the SNCR injection zone. This improved reagent efficiency and reduced ammonia slip. The tuning results showed enhanced temperature balance, lower NO_x emissions, and significant urea savings across all tested load levels.

SILO tuning led to a significant reduction in urea consumption—up to 34% (ref. Figure 14)—while maintaining NO_x emissions below 200 mg/Nm³ (ref. Figure 15). Commissioning data reflect steady-state test conditions, whereas pre-tuning values represent full-range operation. Post-tuning results confirm that SILO adjustments enhanced reagent efficiency across variable loads without compromising NO_x control.

As shown in Figure 16, the refined SILO tuning notably reduced front–rear AGAM temperature differences, with left–right imbalances staying within new ±60 °C limits. Penalties for excess air were reduced, while penalties for AGAM asymmetries were increased and applied quadratically when thresholds were exceeded. This led to urea consumption reductions of 21%, 34%, and 30% at 200 MW, 260 MW, and 300 MW, respectively.

The improvements in AGAM temperature uniformity were achieved by the SILO through advanced control of the air distribution system. The SILO dynamically adjusted the positions of secondary air dampers, general air dampers, over fire air (OFA) dampers, and oil air dampers based on real-time furnace conditions. By identifying the relationships between specific damper positions and AGAM-measured left–right and front–rear temperature differences, the SILO continuously optimized airflows to minimize thermal asymmetries. This targeted adjustment of combustion air improved flame symmetry, enhanced NO_x reduction performance, and stabilized boiler operation.

6. Conclusions

The integration of the SILO optimizer with the SNCR system achieved all key performance targets. NO_x emissions remained consistently below 200 mg/Nm³ across all load conditions, while CO levels stayed within regulatory limits. AGAM-guided temperature balancing significantly improved combustion uniformity, enhancing SNCR effectiveness and boiler stability.

The SILO dynamically adjusted combustion parameters, reducing excess air and unburned carbon in both fly ash and bottom ash. Boiler efficiency improved by up to 0.55 percentage points. Post-commissioning tuning reduced urea consumption by up to 34%, without compromising performance or safety.

Overall, the project demonstrated that combining AI-based combustion control with SNCR delivers a reliable, cost-effective path to meet strict emission targets while improving efficiency. Real-time furnace monitoring and adaptive control—especially AGAM-based balancing—were critical for maintaining high performance under varying loads and fuel conditions.