Low-carbon and low-nitrogen collaborative control method for coal-fired boiler ammonia blending combustion based on machine learning and intelligent optimization

Abstract

This invention discloses a method for coordinated low-carbon and low-NOx control of ammonia-blended combustion in coal-fired boilers based on machine learning and intelligent optimization. First, historical operating data is acquired, and multi-dimensional parameters of fuel characteristics, combustion conditions, and air distribution strategies are extracted as input features to construct a training dataset. Then, multiple machine learning models are trained in parallel, and a baseline model is determined by combining cross-validation and multi-index evaluation. A heuristic optimization algorithm is then used to globally optimize the hyperparameters to obtain a high-fidelity NO emission concentration prediction model. Finally, using this model as the evaluation function, and with the optimization objective of maximizing the ammonia blending ratio under preset NO emission constraints, inverse optimization is performed by generating a virtual operating condition grid and eliminating solutions that exceed the limits, outputting the optimal combination of control parameters. This invention achieves coordinated optimization of low-carbon and low-NOx emissions and possesses closed-loop adaptive capability.

Description

Technical Field

[0001] This invention relates to the field of industrial process automatic control and clean combustion technology, specifically to a low-carbon and low-nitrogen synergistic control method for ammonia-blended combustion in coal-fired boilers based on machine learning and intelligent optimization. Background Technology

[0002] Combustion equipment using fossil fuels constitutes a significant proportion of my country's energy system. Against the backdrop of addressing global climate change and the energy crisis, emission standards for such equipment are becoming increasingly stringent. Utilizing zero-carbon fuels (NH3) in co-firing with coal is currently an effective technological approach to reduce carbon emissions from coal-fired boilers. However, because ammonia fuel contains nitrogen, the oxidation of ammonia during co-firing leads to a significant increase in the concentration of nitrogen oxides (primarily NO) in the flue gas. Therefore, industry has introduced technologies such as staged air combustion, flameless combustion, and oxygen-enriched combustion into the ammonia co-firing process of coal-fired boilers in an effort to achieve low NOx emissions.

[0003] In practical engineering applications, ammonia-blended combustion technology in coal-fired boilers faces the extremely challenging physical and aerodynamic problem of efficiently and uniformly blending ammonia fuel with pulverized coal. From a fuel characteristics perspective, coal is typically introduced into the furnace as a solid powder carried by primary air, while ammonia is injected in gaseous form. This fundamental difference in the gas-solid two-phase flow makes it difficult to achieve uniform mixing at the microscopic level within the extremely short residence time at the burner outlet. Simultaneously, the introduction of a large amount of ammonia gas drastically alters the original aerodynamic field distribution of the burner. If the ammonia blending method and location are not properly matched, local extreme concentrations of ammonia can easily form within the furnace. Uneven mixing of pulverized coal and ammonia not only causes flame instability and delayed ignition but also disrupts the local stoichiometric ratio within the furnace, leading to an exponential surge in the generation of nitrogen oxides (primarily NO) in oxygen-rich or high-temperature zones, or causing severe unburned ammonia escape. These combustion problems, caused by complex flow fields and limited gas-solid phase blending, make the pollutant generation mechanism during co-combustion even more difficult to control.

[0004] Currently, the combined application of staged combustion technology and ammonia-infused combustion in coal-fired boilers is relatively mature. However, in actual operation, the factors affecting nitrogen oxide emissions from boiler outlet flue gas are complex. These factors include not only operating parameters such as main combustion zone temperature, ammonia blending ratio, ammonia blending method, and residence time, but also coal quality characteristics such as volatile matter content and nitrogen content, as well as boundary conditions such as the stage ratio of staged combustion, the location of burnout air (OFA), and boiler structure. Existing low-NOx emission control methods are mostly limited to the independent adjustment of single factors, making it difficult to achieve coordinated control of multiple parameters to reach lower NOx emission levels in practical industrial applications. Furthermore, under the deep coupling of multiple factors, the specific influence mechanism and evolution trend of each parameter on NOx emissions are still unclear.

[0005] In recent years, machine learning methods have gradually gained attention in the emission prediction and operation optimization of combustion systems. Existing research attempts to establish predictive models of coal-fired boiler efficiency and pollutant emissions using methods such as support vector regression and extreme random trees, and then conduct optimization work based on these models. These methods have improved the combustion efficiency and pollution emissions of conventional coal-fired boilers to some extent, but there is still a gap in the emerging low-carbon emission technology field of ammonia blending in coal-fired boilers. For example, Chinese patent CN121576570A discloses a low-NOx combustion optimization control system and method for low-quality coal in large power plant boilers based on machine learning and multi-objective optimization, which mainly focuses on the NOx levels of low-quality coal during combustion in power plant boilers. x The emission issue does not address the control and optimization of other fuel combinations. Another patent, CN121631312A, discloses a machine learning-based multi-objective control optimization method based on coal quality characteristics and combustion state feedback. While this expands the application of coal quality factors in the emission control of coal-fired boilers, it remains limited to conventional combustion methods and fails to cover ammonia-coal co-firing technology that can achieve deep decarbonization.

[0006] In summary, existing technologies in the field of ammonia-coated combustion in coal-fired boilers still have significant technological gaps regarding how to control operating parameters to achieve synergistic optimization of low carbon and low nitrogen emissions. Therefore, there is an urgent need to develop a method that can deeply mine the multivariate coupling influence mechanisms of the ammonia-coal co-firing process based on machine learning, and accordingly perform intelligent system optimization. Summary of the Invention

[0007] To address the challenges of precise control of NO concentration at the outlet of ammonia-blended boilers, the inability to simultaneously optimize carbon reduction and pollutant emissions, and the unclear mechanisms of multiple parameter influences in existing technologies, this invention provides a collaborative control method based on machine learning and intelligent optimization. This method utilizes historical operating data to construct a high-fidelity NO emission prediction model. Combined with optimization algorithms, it seeks the optimal combustion control strategy that balances maximum carbon reduction (i.e., the highest ammonia blending ratio) with low NO emissions under different NO emission constraints. The model is continuously updated through real-time operational feedback, achieving adaptive collaborative optimization.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for coordinated control of low-carbon and low-nitrogen combustion in coal-fired boilers based on machine learning and intelligent optimization, comprising the following steps:

[0009] 1) Obtain the control parameters and corresponding NO emission concentration data of the coal-fired boiler under the historical operating conditions of ammonia co-firing, extract multi-dimensional control parameters including fuel characteristics, combustion conditions and air distribution strategy as input features, take the NO emission concentration at the boiler outlet as the output target, and form a standardized training dataset after preprocessing.

[0010] 2) Based on the training dataset, multiple machine learning regression prediction models are trained in parallel. Cross-validation is used in conjunction with multiple statistical evaluation indicators for multidimensional comparison. The algorithm with the smallest root mean square error or mean absolute error and the largest coefficient of determination under cross-validation is selected as the baseline prediction model. Then, a heuristic optimization algorithm is used to globally optimize the hyperparameters of the baseline prediction model to obtain a high-fidelity NO emission concentration prediction model.

[0011] 3) Using the high-fidelity NO emission concentration prediction model as the evaluation function, under the constraint of satisfying the preset NO emission limit, with the optimization objective of maximizing the ammonia blending ratio, the optimal combination of control parameters is determined and output by inverse optimization within the feasible value space of the control parameters through the generation of virtual operating condition grid and the elimination of solutions that exceed the limit.

[0012] Preferably, the multidimensional control parameters include at least: coal volatile matter content, coal nitrogen content, ammonia blending ratio, main combustion zone temperature, air stage ratio, burnout air location, and ammonia blending method.

[0013] Preferably, the multiple machine learning regression prediction models include at least two of random forest, extreme gradient boosting tree, and K-nearest neighbor algorithm; the multiple statistical evaluation indicators include root mean square error, mean absolute error, and coefficient of determination; and the heuristic optimization algorithm is particle swarm optimization algorithm.

[0014] Preferably, the method for generating the virtual operating condition mesh in step 3) is to generate a large number of virtual operating conditions in the control parameter space by using Monte Carlo sampling, and the method for eliminating solutions that exceed the limits is to introduce a penalty function to eliminate invalid solutions that violate the physical combustion law or exceed the equipment safety boundary.

[0015] Preferably, between step 2) and step 3), the method further includes: introducing an interpretable machine learning framework to perform feature attribution analysis on the high-fidelity NO emission concentration prediction model, quantifying the contribution and influence direction of each regulation parameter on NO emissions, extracting the nonlinear response curve of a single parameter, and analyzing the synergistic coupling relationship between parameters.

[0016] Preferably, the interpretable machine learning framework includes SHAP attribution analysis, partial dependency graphs, and bivariate interaction heatmaps.

[0017] Preferably, the method further includes the following steps: sending the optimal control parameter combination output in step 3) to the boiler control system, collecting the actual NO emission concentration after actual operation, and calculating the relative error between the model prediction value and the actual value; when the relative error exceeds a preset threshold, supplementing the current actual operating condition data into the training dataset formed in step 1), and re-executing step 2) to achieve iterative updating of the model.

[0018] Preferably, the preset threshold is 5%.

[0019] Preferably, step 1) further includes data augmentation processing: expanding the size of the training dataset by adding random noise to the continuous input features and applying physical boundary truncation.

[0020] The preprocessing in step 1) includes Z-score normalization of continuous features and one-hot encoding of discrete categorical features.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] (1) This invention maps core parameters such as ammonia blending ratio, main combustion zone temperature, air stage ratio, burnout air location, and coal quality characteristics to a feature space to construct a high-quality historical operating condition dataset. This breaks through the bottleneck of traditional mechanism models that are difficult to accurately describe multivariate strongly coupled processes, and realizes the fine quantification of NO emissions and carbon emission reduction characteristics of ammonia-coal co-firing.

[0023] (2) The present invention first determines the optimal baseline model through multi-model comparison and multi-index evaluation, and then uses the particle swarm optimization algorithm to globally optimize the hyperparameters of the model to improve the prediction accuracy. On this basis, combined with the Monte Carlo sampling and penalty function mechanism, the feasible solution is accurately locked in the huge virtual working space, which not only meets the low NO emission constraint, but also achieves the maximum ammonia doping ratio, truly achieving a synergistic win-win situation for carbon emission reduction and pollutant control.

[0024] (3) The present invention designs a model dynamic update mechanism based on actual operation error feedback. The system can automatically capture high-error edge condition data and feed it back to the training set, enabling the model to have continuous evolution and adaptive correction capabilities, which significantly improves the long-term reliability and optimization effect in complex industrial environments. Attached Figure Description

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0026] Figure 1 This is a schematic diagram of the overall process of the low-carbon and low-nitrogen synergistic control method for ammonia-blended combustion in coal-fired boilers based on machine learning and intelligent optimization in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only for explaining the invention and do not constitute a limitation on the scope of protection of this invention. Embodiments

[0028] like Figure 1 As shown, this embodiment takes an actual ammonia co-firing industrial project of a coal-fired boiler as an example to explain in detail the implementation process of the present invention.

[0029] Step 1. Historical data collection, feature engineering, and data augmentation:

[0030] This step aims to build a high-quality underlying dataset that couples fuel characteristics, thermal state, and wind distribution strategy, providing a physical basis for subsequent modeling.

[0031] First, relevant literature on ammonia-mixed combustion in coal-fired boilers was retrieved from academic databases such as Web of Science, extracting operating data of different boilers under steady-state conditions. After categorizing by boiler type, operating data with an excess air coefficient of 1.2 and an unburned coal rate not exceeding 5% were selected to eliminate secondary effects caused by differences in macroscopic flow fields. After cleaning and screening, a total of 80 sets of high-quality and effective one-dimensional pulverized coal boiler datasets were obtained.

[0032] To accurately characterize the physicochemical mechanism of ammonia-coal co-firing, this embodiment strictly selects seven key control parameters as input features: coal volatile matter content, coal nitrogen content, ammonia blending ratio, main combustion zone temperature, air stage ratio, burnout air location, and ammonia blending method. Simultaneously, the NO emission concentration at the flue gas outlet under the corresponding operating condition is used as the single output feature. All continuous input and output feature data are Z-score standardized to eliminate the influence of dimensions; for the discrete classification feature of "ammonia blending method," one-heat encoding is used to map it into orthogonal numerical variables to ensure data consistency and comparability for subsequent machine learning model training.

[0033] To overcome the risk of overfitting due to the scarcity of combustion experiment samples, a Gaussian perturbation method is introduced for data augmentation under physical constraints: injecting data with a mean of 0 and a variance of σ into the continuous feature space. 2 Random Gaussian noise was used to simulate actual feed fluctuations and sensor errors, expanding the dataset to 800 groups. Strict boundary constraints and feature isolation were implemented during the enhancement process: perturbations were applied only to continuous values, ammonia doping method classification features were isolated, and physical threshold truncation was performed to ensure that the synthesized operating conditions conformed to combustion kinetics logic.

[0034] Step 2. NO emission concentration prediction model construction and hyperparameter optimization

[0035] The 800 amplified samples were randomly divided into a training set (640 sets) and a test set (160 sets) at an 8:2 ratio. Three machine learning algorithms—random forest, extreme gradient boosting tree, and K-nearest neighbors—were used for parallel modeling and training. The three models were validated using the test set, and the root mean square error, mean absolute error, and coefficient of determination for each model were calculated and compared. In this embodiment, the algorithm with the best evaluation metrics (random forest) was selected as the baseline prediction model.

[0036] Building upon this foundation, a particle swarm optimization (PSO) algorithm is introduced, with a particle swarm size of 50 and a maximum number of iterations of 100. Minimizing the root mean square error (RMSE) of the model's predicted NO concentration is set as the fitness function for PSO. Within the defined hyperparameter boundaries (such as the maximum depth of the decision tree, learning rate, and minimum number of samples per leaf node), particles iteratively update their velocity and position until convergence is achieved. Finally, the globally optimal hyperparameter combination is output, and this combination is used to reconstruct the prediction model, resulting in a high-fidelity NO emission concentration prediction model.

[0037] The optimized particle swarm optimization-random forest model performs excellently on the test set: the coefficient of determination R0 2 The mean absolute error was reduced to 91.46 mg / m³, reaching 0.951. 3 The root mean square error was reduced to 149.85 mg / m 3 This provides a high-fidelity model foundation for subsequent mechanism analysis and virtual optimization.

[0038] Step 3. Analysis of the physical mechanism of the prediction model and feature attribution mapping

[0039] To break the inherent "black box" nature of machine learning models and enhance the engineering credibility and physical logic support of prediction results, interpretable machine learning methods are introduced to perform underlying dynamic analysis of the prediction model.

[0040] First, the SHAP framework is used to perform global and local attribution analysis on the prediction model optimized by particle swarm optimization. By drawing SHAP bar charts, the absolute importance of each regulatory parameter to NO emissions is accurately quantified; at the same time, by drawing SHAP swarm diagrams, the evolution direction of each parameter (i.e., the positive promotion or negative inhibition mechanism of NO generation) is intuitively revealed by using the distribution of color gradients and SHAP values.

[0041] Secondly, a partial dependency graph technique is introduced to extract the global average response curve of a single target variable to NO emissions. By marginalizing other features, threshold mutations or nonlinear inflection points of each control parameter within a specific range are accurately captured (e.g., to determine whether there is a "U-shaped" minimum window for NO emissions from volatile matter in coal).

[0042] Finally, considering the strong coupling characteristics of multiple variables in actual combustion systems, a bivariate characteristic interaction heatmap is further constructed. By explicitly characterizing the nonlinear mapping relationship under the synergistic effect of two core control parameters (such as the main combustion zone temperature and ammonia blending ratio, stage ratio and volatile matter, etc.) through two-dimensional color gradients, the synergistic influence mechanism between parameters in complex multiphase reactions is clearly revealed, providing rigorous physical logic support and theoretical boundaries for subsequent intelligent constraint optimization.

[0043] Step 4. Intelligent reverse optimization of boiler control parameters

[0044] Based on a high-fidelity NO emission concentration prediction model, we conduct reverse control optimization and decision generation for the synergistic goal of low carbon (high ammonia ratio) and low nitrogen (low NO emission).

[0045] First, based on the actual physical structure, safe operation specifications, and combustion dynamics limitations of the target coal-fired boiler, reasonable fluctuation ranges and operating boundaries are set for seven core control parameters (coal volatile matter content, coal nitrogen content, ammonia blending ratio, main combustion zone temperature, air stage ratio, burnout air location, and ammonia blending method). Then, using a high-density Monte Carlo sampling algorithm, high-throughput random sampling is performed within the given multi-dimensional parameter space to generate 50,000 sets of virtual operating condition grids that comprehensively cover potential operating states.

[0046] Secondly, the aforementioned 50,000 sets of virtual operating condition data are used as input vectors and imported in batches into the high-fidelity prediction model trained in step 2 to calculate the expected NO emission concentration at the boiler outlet for each set of virtual operating conditions. A penalty function mechanism is introduced in this process: data points where the predicted NO emission concentration exceeds the set environmental protection limit standard, or where the control parameter combination exceeds the physical safety operating boundary under this specific operating condition, are uniformly defined as "exceeding the limit solutions" and discarded, thereby selecting the "feasible solution set" that meets the low NO emission constraints.

[0047] Then, in the filtered "feasible solution set", the core optimization objective is established to maximize carbon emission reduction of the system. Since the carbon emission reduction benefit is directly positively correlated with the ammonia blending ratio, the system performs a global search within the feasible solution set to accurately extract specific optimal operating condition candidate points that can achieve the highest theoretical ammonia blending ratio (i.e., maximize carbon emission reduction benefit).

[0048] Finally, after locking in the optimal operating point, the system automatically extracts and outputs the best combination of underlying control parameters that matches the high ammonia blending ratio. These parameters include the optimal main combustion zone temperature, air stage ratio, optimal burnout air position, and optimal ammonia blending method. This parameter combination, while ensuring compliance with low NO emission standards, physically supports the achievement of the maximum ammonia blending ratio and ultimately translates into intelligent control commands for the distributed control system to execute.

[0049] Step 5. Model closed-loop feedback and retraining optimization

[0050] Establish a real-time combustion status feedback and model retraining mechanism to ensure the long-term reliability and adaptability of the control system when facing complex and variable operating conditions.

[0051] The optimal control parameter combination output from step 4 is sent to the control system to guide the actual ammonia-blended operation of the coal-fired boiler. The system continuously collects the actual control parameters and actual NO emission concentration data at the boiler outlet under the current operating conditions, and calculates the relative error between the model-predicted NO concentration and the actual NO concentration in real time.

[0052] A prediction error trigger threshold of 5% is set. If the calculated relative error exceeds this threshold, it is determined that there is a significant error between the model's predicted data and the actual data. This indicates that the current actual physical conditions (such as coal quality shifts, equipment characteristic drift due to aging, etc.) have deviated from the high-precision prediction generalization space of the original model.

[0053] Once the error exceeds the limit, the system automatically extracts the actual operating condition data that is currently generating a large error (i.e., the 7 actual input control parameters and the corresponding actual NO emission concentration at this time), and uses it as a high-value new sample to directly supplement and integrate it into the initial 80 sets of underlying original datasets to form a new dataset.

[0054] Based on the updated and expanded dataset, the system automatically re-triggers the machine learning baseline model training and particle swarm optimization hyperparameter global optimization in step 2. By retraining and optimizing the model, the predictive model quickly learns and absorbs the latest real-world physical characteristics, correcting prediction biases. After the model update is complete, step 4 is re-executed to generate new optimal control parameters and distribute them, thereby achieving closed-loop adaptive iteration of the entire predictive and optimization control system.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art can make various improvements and modifications without departing from the spirit and principles of the invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for coordinated control of low-carbon and low-nitrogen combustion in coal-fired boilers using ammonia blending based on machine learning and intelligent optimization, characterized in that, Includes the following steps: 1) Obtain the control parameters and corresponding NO emission concentration data of the coal-fired boiler under the historical operating conditions of ammonia co-firing, extract multi-dimensional control parameters including fuel characteristics, combustion conditions and air distribution strategy as input features, take the NO emission concentration at the boiler outlet as the output target, and form a standardized training dataset after preprocessing. 2) Based on the training dataset, multiple machine learning regression prediction models are trained in parallel. Cross-validation is used in conjunction with multiple statistical evaluation indicators for multidimensional comparison. The algorithm with the smallest root mean square error or mean absolute error and the largest coefficient of determination under cross-validation is selected as the baseline prediction model. Then, a heuristic optimization algorithm is used to globally optimize the hyperparameters of the baseline prediction model to obtain a high-fidelity NO emission concentration prediction model. 3) Using the high-fidelity NO emission concentration prediction model as the evaluation function, under the constraint of satisfying the preset NO emission limit, with the optimization objective of maximizing the ammonia blending ratio, the optimal combination of control parameters is determined and output by inverse optimization within the feasible value space of the control parameters through the generation of virtual operating condition grid and the elimination of solutions that exceed the limit.

2. The method according to claim 1, characterized in that, The multidimensional control parameters include at least: coal volatile matter content, coal nitrogen content, ammonia blending ratio, main combustion zone temperature, air stage ratio, burnout air location, and ammonia blending method.

3. The method according to claim 1, characterized in that, The various machine learning regression prediction models include at least two of random forest, extreme gradient boosting tree, and K-nearest neighbor algorithm; the various statistical evaluation indicators include root mean square error, mean absolute error, and coefficient of determination; the heuristic optimization algorithm is particle swarm optimization algorithm.

4. The method according to claim 1, characterized in that, The method for generating the virtual operating condition mesh in step 3) is to generate a large number of virtual operating conditions in the control parameter space using Monte Carlo sampling. The method for eliminating solutions that exceed the limits is to introduce a penalty function to eliminate invalid solutions that violate the physical combustion laws or exceed the equipment safety boundaries.

5. The method according to claim 1, characterized in that, Between step 2) and step 3), the following steps are also included: introducing an interpretable machine learning framework to perform feature attribution analysis on the high-fidelity NO emission concentration prediction model, quantifying the contribution and influence direction of each regulation parameter on NO emissions, extracting the nonlinear response curve of a single parameter, and analyzing the synergistic coupling relationship between parameters.

6. The method according to claim 5, characterized in that, The interpretable machine learning framework includes SHAP attribution analysis, partial dependency graphs, and bivariate interaction heatmaps.

7. The method according to claim 1, characterized in that, It also includes the following steps: The optimal control parameter combination output in step 3) is sent to the boiler control system. The actual NO emission concentration after actual operation is collected, and the relative error between the model prediction value and the actual value is calculated. When the relative error exceeds the preset threshold, the current actual operating condition data is added to the training dataset formed in step 1), and step 2) is executed again to achieve iterative update of the model.

8. The method according to claim 7, characterized in that, The preset threshold is 5%.

9. The method according to claim 1, characterized in that, Step 1) also includes data augmentation: the training dataset is expanded by adding random noise to the continuous input features and applying physical boundary truncation.

10. The method according to claim 1, characterized in that, The preprocessing in step 1) includes Z-score normalization of continuous features and one-hot encoding of discrete categorical features.

Images