Sintering flue gas co concentration optimization control method based on key influence factor identification

Abstract

The application discloses a sintering flue gas CO concentration optimization control method based on key influencing factor identification, and belongs to the technical field of flue gas emission reduction, and comprises the following steps: acquiring sintering production process data and corresponding sintering flue gas CO emission concentration measured values, and constructing a model sample set after pretreatment; constructing an improved grey wolf optimization algorithm, optimizing parameters of a regression algorithm, and building a sintering process CO concentration soft measurement fitting model based on the optimized regression algorithm; performing feature importance analysis by using a SHAP method, retraining the fitting model after key influencing factors are identified, and obtaining a simplified model; collecting key influencing factor data in real time, predicting CO concentration and gradient by using the simplified model; screening a regulation and control feature item that can make CO concentration decrease the most, fusing gradient descent and PPO reinforcement learning control instructions, and realizing multi-time scale optimization control of CO concentration. The application can accurately predict and efficiently reduce sintering flue gas CO concentration.

Description

Technical Field

[0001] This invention relates to the field of combustion flue gas emission reduction technology, specifically a method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors. Background Technology

[0002] Sintering is an important process in steel production. Its main purpose is to heat iron ore powder, flux, fuel and return ore and other iron-containing raw materials at high temperature (usually below their melting point), and through solid-phase reaction and partial liquid-phase generation, they are bonded together to form sinter with sufficient strength and metallurgical properties, providing high-quality raw materials for ironmaking.

[0003] However, during the sintering process, incomplete combustion of fuel and complex physicochemical reactions within the sintering bed generate a large amount of flue gas containing pollutants, among which carbon monoxide (CO) is one of the main harmful components. CO is not only highly toxic, and direct emissions severely pollute the environment and harm human health, but it is also an inefficiently utilized fuel, leading to energy waste. Therefore, accurate prediction and effective control of CO emissions from sintering flue gas are crucial for achieving green and low-carbon development in the steel industry.

[0004] Traditional CO emission prediction models mainly adopt two approaches: one is to directly use all process data that can be collected during sintering (such as material layer thickness, machine speed, wind box temperature, exhaust gas composition, etc.) as input features of the model to predict CO emissions; the other is to first use methods such as principal component analysis (PCA) to reduce the dimensionality of these high-dimensional process data, and then use the obtained principal components as model input for prediction.

[0005] While the first method incorporates all the original information, its high-dimensional input features, including numerous variables with weak or no relation to CO generation, make the model prone to learning noise and redundant information. This limits prediction accuracy and generalization ability. The second method, through PCA dimensionality reduction, addresses the curse of dimensionality to some extent, improving model stability and generalization. However, its generated "principal components" are linear combinations of the original variables, lacking clear physical meaning and failing to directly correspond to specific process parameters (such as fuel ratio and air volume) during sintering. This makes the model a "black box," significantly reducing interpretability. Consequently, operators struggle to identify key factors influencing CO emissions based on model results, hindering the provision of concrete and feasible guidance for optimizing sintering processes and reducing CO emissions at the source. Summary of the Invention

[0006] To address the problems of existing technologies, this invention provides a method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors, comprising: S1. Obtain the process parameter data of the sintering production process and the corresponding measured values ​​of CO emission concentration in sintering flue gas. After preprocessing including missing value filling, outlier removal and standardization, construct the model sample set. S2. Construct an improved gray wolf optimization algorithm that integrates random restart and stagnation escape mechanisms. Use the improved gray wolf optimization algorithm to optimize the parameters of the LightGBM regression algorithm. Based on the optimized LightGBM regression algorithm, build a soft measurement fitting model for CO concentration in the sintering process. S3. The SHAP method based on game theory is used to conduct feature importance analysis, quantitatively analyze the contribution of each process parameter in the sintering production process to the CO concentration prediction results, and identify the key influencing factors. S4. Based on the identified key influencing factors, the soft measurement fitting model of CO concentration in the sintering process was retrained to obtain a simplified model. S5. Real-time acquisition of key influencing factor data, prediction of CO concentration and gradient through simplified model; screening of the regulatory feature that reduces CO concentration the most, and using gradient descent regulation of the regulatory feature when the concentration exceeds the standard, and integrating gradient descent and PPO reinforcement learning control instructions to achieve multi-timescale optimization control of CO concentration.

[0007] Furthermore, S1 includes: S11. Collect historical data of the entire sintering process in the steel plant, including several process characteristic variables such as batching parameters, fan speed, wind box temperature, and oxygen content of the fume hood, as well as the corresponding measured values ​​of CO emission concentration. S12. Perform preprocessing on the collected data, including missing value imputation, outlier removal, and standardization. Preprocessing includes: identifying and removing outliers using box plots and the 3σ criterion; deleting entire rows of records with missing values ​​to ensure data consistency; aligning and resampling multi-source data based on a unified timestamp; and processing all numerical features using Z-Score normalization to eliminate the influence of units. S13. After preprocessing the collected data by filling missing values, removing outliers, and standardizing, construct the model sample set.

[0008] Furthermore, the improvements made to the standard gray wolf optimization algorithm in S2 include: Hybrid random exploration mechanism: Before updating each individual in each iteration, a probability determination is first performed: if the probability determination is triggered, the individual will ignore the current social level guidance and directly generate a new location coordinate randomly in the search space; otherwise, it will be updated according to the standard GWO formula. This is used to continuously jump to the unknown areas of the search space, increase population diversity, and prevent the algorithm from converging to a local optimum too early. Stagnation detection and population reset strategy: A stagnation counter is introduced into the algorithm to monitor the change in global optimal fitness in real time; a fitness stagnation criterion is defined. If stagnation is triggered, the top 3 wolves of the current best fitness are retained and the positions of all remaining wolves are forcibly reset. This is used to inject strong perturbation into the stagnant population and force the algorithm to jump out of the current local optimal region. Alpha Wolf-guided dynamic perturbation strategy: When updating the wolf pack population based on the location of the Alpha Wolf, a random perturbation term related to the search space range is superimposed.

[0009] Furthermore, the contribution of each process parameter in S3 to the CO concentration prediction result is represented by the SHAP value of each process parameter: ; in, For the first The SHAP value of each feature represents the impact of that feature value on the current prediction. The contribution of N is the total feature set {1,2,...,n}, S is any subset excluding feature j, and v(S) is the model prediction offset corresponding to feature subset S.

[0010] Furthermore, S5 includes: S51. Real-time acquisition of key characteristic data of a number of standard items to predict the current CO concentration value and its gradient; when the predicted value exceeds the standard, the gradient descent algorithm immediately fine-tunes parameters including damper opening and fuel supply to achieve real-time optimization of CO concentration. S52. At the same time, the PPO reinforcement learning algorithm is introduced, which uses the CO concentration gradient signal calculated by the LightGBM prediction model and the historical control action sequence as the state input. This enables the reinforcement learning agent to perceive the current operating conditions, identify the directional influence of each key process parameter on the CO concentration change, and comprehensively consider the cumulative effect of the preceding adjustment behavior on the current emission state. S53. By dynamically fusing gradient descent and PPO reinforcement learning control instructions, multi-timescale optimization control of CO concentration is achieved.

[0011] Furthermore, S5 also includes: combining the sintering process principle, equipment safety limits and historical stable operation data, defining the feasible control range of the key characteristic parameters of the number of standards, and limiting the gradient control within the feasible control range.

[0012] Furthermore, the control instructions for dynamically fusing gradient descent and PPO reinforcement learning in S53 are as follows: (3); Where a_gd is the gradient descent control output. Here, a_ppo represents the weights controlling the output of gradient descent, and a_ppo represents the output controlling PPO reinforcement learning. The weights for controlling the output of PPO reinforcement learning.

[0013] Furthermore, in the initial operation phase of the S5 system, the weight of the gradient descent control output is set relatively high to ensure the control response speed and safety. As the training rounds increase, the weight of the PPO reinforcement learning control output is gradually increased when the cumulative control effect continues to improve, so that the control strategy gradually transitions from short-term correction to long-term optimization.

[0014] Furthermore, in S5, when the CO concentration is detected to continuously exceed the set safety threshold, the system automatically reduces the PPO weight and switches to the gradient descent dominant mode to prioritize production safety.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an improved gray wolf optimization algorithm that integrates random restart and stagnation escape mechanisms. Instead of relying solely on the social hierarchy of the wolf pack for position updates, this algorithm introduces a forced balancing mechanism of "exploration-development." This mechanism monitors the evolutionary state of the population and actively resets the population when the algorithm stagnates, thus avoiding getting trapped in local optima. Furthermore, it applies micro-perturbations during the alpha wolf's guidance process, thereby efficiently finding the global optimum in the non-convex parameter space of LightGBM. This effectively solves the technical problems of local optima easily encountered when dealing with high-dimensional complex hyperparameter spaces, and insufficient convergence accuracy due to the loss of population diversity in the later stages.

[0016] 2. Based on the identified key influencing factors, this invention retrains and reconstructs a soft-sensor fitting model for CO concentration in the sintering process, resulting in a simplified model. Compared to the full-feature model, the simplified model reduces the number of input parameters from over 200 to 50 while maintaining essentially the same prediction accuracy. Compared to the full-feature model, the simplified model achieves a "lightweight" reconstruction with almost no loss of prediction accuracy, significantly reducing the system's computational load and making it more suitable for online deployment and real-time control.

[0017] 3. In this invention, the gradient descent algorithm, based on real-time gradient calculation, rapidly generates short-term process control commands aimed at immediately reducing CO concentration in the fusion model. The PPO reinforcement learning module learns long-term optimal operating strategies to continuously minimize CO concentration. After dynamic fusion, the two algorithms collaboratively adjust key parameters such as damper opening, forming an intelligent closed-loop optimization with CO emission reduction as the sole objective. This solves the problems of insufficient model accuracy and poor generalization ability caused by feature redundancy in traditional methods, as well as the weak interpretability and difficulty in guiding actual CO emission reduction control of PCA dimensionality reduction methods. It achieves accurate prediction and efficient emission reduction of CO concentration in sintering flue gas. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the CO concentration optimization control method for sintering flue gas based on the identification of key influencing factors according to the present invention. Figure 2 This is a residual plot of CO emissions fitted to a model according to an embodiment of the present invention; Figure 3 The SHAP value distribution of key influencing factors in one embodiment of the present invention; Figure 4 This is a graph showing the change of the SHAP value of the oxygen content in a smoke hood as a function of the oxygen content in a smoke hood according to an embodiment of the present invention. Figure 5 This is a comparison chart of key performance indicators between a simplified model and a full-feature model according to an embodiment of the present invention; Figure 6 This is a comparison chart showing the detailed running efficiency of a simplified model and a full-feature model according to an embodiment of the present invention; Figure 7 This is a comparison chart of the predicted values ​​vs. actual values ​​of a simplified model and a full-feature model according to an embodiment of the present invention. Figure 8 This is a distribution density map of the predicted residuals of a simplified model and a full-feature model according to an embodiment of the present invention; Figure 9 This is a simplified model relative performance gain diagram according to an embodiment of the present invention. Detailed Implementation To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0020] like Figure 1 As shown below, the present invention will be further described in detail with reference to embodiments. The present invention provides a method for optimizing and controlling the CO concentration of sintering flue gas based on the identification of key influencing factors, which includes the following steps: S1. Obtain the process parameter data of the sintering production process and the corresponding measured values ​​of CO emission concentration in the sintering flue gas. After preprocessing including missing value filling, outlier removal, and standardization, construct the model sample set, specifically including: S11. Collect historical data of the entire sintering process in the steel plant, including several process characteristic variables such as batching parameters, fan speed, wind box temperature, and oxygen content in the fume hood, as well as the corresponding measured values ​​of CO emission concentration.

[0021] The experimental data were obtained from an actual sintering production line in a factory, acquired by on-site online monitoring equipment. The data included process parameters for the entire sintering process and CO concentration data at the main exhaust fan outlet, covering normal continuous production conditions. Data was collected every 2 minutes over a uniform time span of 120 days. The experimental data was collected using a material tracking monitoring mode, tracking a single batch of sintered material. Monitoring was completed from primary mixing, secondary mixing, trolley movement, wind box operation, to the main exhaust fan outlet, encompassing 248 process characteristic variables, including batching parameters, fan speed, wind box temperature, and oxygen content in the fume hood, along with corresponding measured CO emission concentrations. Some data are shown in Table 1 below.

[0022] Table 1: Display of Partial Data Collected During the Sintering Process

[0023] S12. Perform preprocessing on the collected data, including missing value filling, outlier removal, and standardization.

[0024] Preprocessing of the raw data included: identifying and removing outliers using box plots and the 3σ criterion; deleting entire rows with missing values ​​to ensure data consistency; and aligning and resampling the multi-source data using a unified timestamp. This process aimed to improve the statistical quality of multi-source heterogeneous data. Outlier removal was based on the Laida criterion (3σ criterion) under the Gaussian distribution assumption. Its mathematical basis is that when data follows a normal distribution, samples falling within the interval... The probability of any of these exceptions is only about 0.27%, and these are considered outliers. ; To eliminate the influence of dimensions, the Z-Score standardization method is further used to process all numerical features. The transformation formula is as follows: z = (x - μ) / σ; Where μ is the feature mean, σ is the standard deviation, and x is the feature column in the original data.

[0025] For example, the CO concentration emitted from the sintering flue has a mean of 3240.32 and a standard deviation of 665.07. The CO concentration emitted from the sintering flue is calculated as (x - 3240.32) / 665.07. This step eliminates problems such as uneven gradient descent convergence speed caused by different sensor dimensions (e.g., temperature is on the order of thousands, pressure is on the order of units), providing a smooth hyperplane for subsequent optimization of the objective function.

[0026] S13. After preprocessing the collected data by filling missing values, removing outliers, and standardizing, construct the model sample set.

[0027] S2. Construct an improved gray wolf optimization algorithm that integrates random restart and stagnation escape mechanisms. Use the improved gray wolf optimization algorithm to optimize the parameters of the LightGBM regression algorithm, and build a soft measurement fitting model of CO concentration in the sintering process based on the optimized LightGBM regression algorithm.

[0028] A soft measurement model for CO concentration in the sintering process is constructed using the LightGBM regression algorithm. LightGBM is essentially an additive model based on gradient boosting decision trees (GBDT), commonly used for industrial time-series and regression-type prediction tasks. This invention applies it to predicting CO concentration at the main extraction outlet of the sintering process. The core idea of ​​this model is to train multiple decision trees in successive rounds, with each tree fitting the residuals generated by the preceding model to iteratively correct the prediction error. Finally, the outputs of all decision trees are weighted and summed to obtain the final prediction result for the sample.

[0029] The model uses a standard additive model, and the formula for predicting the CO concentration of the i-th sample is as follows: ; ; ; in, For the model to the first Predicted CO concentration at the main sampling outlet of each sample. For the first The 247-dimensional feature vector of each sample, x i,j This represents the j-th feature of the i-th sample, where j takes values ​​from 1, 2, ..., 247. This refers to the learning rate. The total number of decision trees (n_estimators). For the first The predicted output for each tree. The initial value is usually taken as the target variable in the training set. That is, the average value of the CO concentration actually measured at the main extraction outlet. After the model has been trained for the first time... After a certain number of trees, the difference between the predicted value and the actual value is called the residual. The goal of the tree is to fit this residual.

[0030] The core of this step lies in approximating the nonlinear mapping from the feature space to the target variable using the forward distribution algorithm. Its mathematical essence is minimizing the regularization objective function: ; in, The loss function is used to measure the difference between predicted and actual values. The regularization term is used to control the complexity of the decision tree. The improved Grey Wolf Optimization Algorithm (IGWO) effectively avoids hyperparameters from getting trapped in local minima during the optimization process by introducing an annealed "exploration-exploration" forced balancing mechanism, ensuring that the model converges to the global optimum.

[0031] The Grey Wolf Optimization (GWO) algorithm plays a crucial role in the LightGBM regression algorithm as an efficient, globally searched hyperparameter optimizer. By simulating the predatory behavior of a grey wolf, it automatically searches for the optimal parameter combination that maximizes the performance of the LightGBM model (such as minimizing prediction error) within a complex parameter space.

[0032] However, it is prone to local optima when dealing with high-dimensional complex hyperparameter spaces, and the loss of population diversity in the later stage leads to insufficient convergence accuracy. Based on this, the present invention proposes an improved GWO with Stochastic Restart & Dynamic Perturbation.

[0033] This optimization algorithm does not simply rely on the social hierarchy of the wolf pack for position updates, but instead introduces a forced balancing mechanism of "exploration-development". This mechanism monitors the evolutionary state of the population and actively resets the population when the algorithm stagnates, thus avoiding getting stuck in local optima. It also applies micro-perturbations during the alpha wolf's guidance process, thereby efficiently finding the global optimum in the non-convex parameter space of LightGBM.

[0034] This invention makes three core improvements to the standard GWO, as follows: First, Hybrid Stochastic Exploration: In the standard GWO algorithm, all individuals strictly follow the rules in each iteration. The wolves' guidance for movement caused all individuals in the algorithm to converge in the later stages, losing the ability to explore new areas. This improvement introduces a random restart probability during the position update phase to enhance the population's global exploration capabilities. . No. The generation The position of Sekiro The update rule has been revised to a piecewise function: ; in: The probability of random restart is set as follows in this invention. . for A random number between [a certain number of points]. The full-dimensional position vector is randomly generated within the search space, and its calculation formula is: ; in, These are the lower and upper bounds of the hyperparameters, respectively. ).

[0035] Before updating each individual in each iteration, a probability check is performed: if the check is triggered (SP < 0.25), the individual ignores the current social rank guidance and directly generates a new random position coordinate within the search space. Otherwise, it is updated according to the standard GWO formula. Through this mechanism, they continuously jump to unknown regions of the search space. This greatly increases the diversity of the population and prevents the algorithm from prematurely converging to local optima.

[0036] Second, Stagnation Detection and Population Resetting Strategies: The impact of LightGBM hyperparameters on model performance is often non-linear, leading to many local extremum traps. Standard GWO lacks a mechanism to escape these traps. Therefore, a stagnation counter is introduced into the algorithm to monitor the change in the global optimal fitness in real time.

[0037] Define the fitness stagnation criterion. Let... For the first The optimal fitness value (RMSE) of a generation is obtained if the following conditions are met: generation( ): ; Then the algorithm is determined to be trapped in a local optimum (where (This is the stagnation threshold). At this point, a reset mechanism is triggered: ; This formula indicates that, apart from the first three elite wolves ( Except for the individuals that are retained, all other individuals in the population are reinitialized to random positions, thereby forcing them to escape the local extreme value region.

[0038] If the improvement in the optimal RMSE is less than a threshold of 0.1 over five consecutive iterations, the algorithm is considered to have stalled. Once stalling is triggered, the top three wolves are retained, and the positions of all remaining wolves are forcibly reset. By retaining elite individuals to ensure that historical bests are not lost, and by introducing a large number of new individuals, a strong perturbation is injected into the stalled population, forcing the algorithm to jump out of the current local optimum.

[0039] Third, Alpha-Guided Dynamic Perturbation in AlphaWoW: In the standard GWO, The wolf represents the currently found optimal solution, and other wolves will directly gravitate towards it. However... The wolf's position may only be a rough approximation of the global optimum.

[0040] According to When the wolf updates the position of other individuals, it is no longer a simple linear approximation, but rather a random perturbation term related to the search space range is superimposed.

[0041] In calculating individual orientation When considering the components of the wolf's (optimal solution) movement, a dynamic perturbation term is introduced. To enhance the accuracy of local development. (Revised) Wolf Guiding Position The calculation formula is: ; Among them, the disturbance term Defined as: ; in: The disturbance intensity coefficient is set as follows in this invention. . for A random vector between [variables]. will with wolves and Wolf-led , The final position will be decided jointly: .

[0042] Finally, the LightGBM regression model was trained using the obtained optimal parameter combination, and its generalization performance was evaluated using time-series cross-validation, thereby constructing a high-precision CO concentration prediction model for the sintering endpoint, providing a reliable basis for the optimized control of the sintering process.

[0043] Let the hyperparameter combination of LightGBM be: The root mean square error on the validation set is the objective function. The optimal hyperparameters obtained by IGWO are: ; in, This represents the feasible interval for each hyperparameter. Each candidate hyperparameter... Encoded as the position vector of an individual gray wolf The position is iteratively updated through the random restart, stagnation reset, and dynamic perturbation mechanism introduced in this invention: ; Optimal position after convergence Decode to get Substituting this optimal hyperparameter into the LightGBM prediction formula: ; To achieve high-precision prediction of CO concentration in sintering flue gas in this invention, an improved Grey Wolf Optimization Algorithm (IGWO) combined with the LightGBM regression model is used on a real dataset.

[0044] LightGBM itself follows the standard additive model prediction formula. The improved gray wolf optimization algorithm does not change the mathematical structure of this prediction formula, but indirectly improves the generalization ability and accuracy of the prediction formula by globally optimizing the key hyperparameters of LightGBM. IGWO encodes hyperparameters as position vectors of individual gray wolves, uses the root mean square error (RMSE) on the validation set as the fitness function, and effectively avoids the defect of traditional GWO being prone to getting trapped in local optima by introducing random restart probability, stagnation detection and population reset, as well as alpha-guided dynamic perturbations, thus stably converging to the globally optimal hyperparameter combination.

[0045] Finally, the LightGBM model is retrained using these optimal hyperparameters to predict the formula. and The parameters were replaced with optimized values, which reduced the RMSE of the model on the actual sintering process dataset to approximately 150 ppm. 2 The accuracy reached 0.92. This indicates that IGWO has optimized the prediction process, ultimately achieving high-precision CO concentration prediction based on sintering data. The optimal parameter combination is as follows (based on the improved GWO optimization results): The parameters are: learning_rate: 0.05; n_estimators: 500; num_leaves: 31; max_depth: 6; min_child_samples: 20; subsample: 0.8; colsample_bytree: 0.8; reg_alpha: 0.1; reg_lambda: 0.1.

[0046] The essence of GWO optimization of LightGBM is to encode LightGBM hyperparameters as the positions of gray wolves and use the model's prediction errors (such as RMSE and MAE) on the validation set as the fitness function. Through the iterative mechanism of GWO, it finds the parameter combination that minimizes the error. The core role of the Gray Wolf Optimization (GWO) algorithm in the LightGBM regression algorithm is as a highly efficient, globally searched hyperparameter optimizer. By simulating the predatory behavior of gray wolves, it automatically searches for the optimal parameter combination that maximizes the performance of the LightGBM model (such as minimizing prediction error) in a complex parameter space. This combination is particularly suitable for solving high-dimensional, nonlinear hyperparameter optimization problems where finding a global optimum is difficult, and it performs exceptionally well in improving model accuracy, stability, and generalization ability.

[0047] The sample set was divided into training and test sets in a 7:3 ratio, and the optimal hyperparameters were determined through cross-validation. The model showed good fit on the test set, and its residual distribution is as follows. Figure 2 As shown, the residuals basically follow a normal distribution and have small absolute values, indicating that the model has good fitting accuracy and generalization ability.

[0048] S3. The SHAP method based on game theory is used to conduct feature importance analysis, quantitatively analyze the contribution of each process parameter in the sintering production process to the CO concentration prediction results, and identify key influencing factors.

[0049] To quantitatively assess the contribution of each production process parameter to the CO concentration prediction results, the SHAP (Shapley Additive exPlanations) method based on game theory was used for feature importance analysis.

[0050] Specifically, 20% of the samples are randomly selected from the dataset, the SHAP value of all features for each sample is calculated, and the global importance is determined by calculating the mean of the absolute values ​​of the SHAP values ​​of each feature, thereby identifying the key features that have the most significant impact on CO concentration.

[0051] ; ; ; in, For specific samples of the LightGBM model Predicted CO concentration at the main pump outlet The baseline value is the average prediction of the model across all training set samples (historical sintering data). For the first The SHAP value of each feature represents the impact of that feature value on the current prediction. The contribution of the feature is positive, indicating that the feature boosts the CO concentration prediction, while a negative value indicates that it lowers it. n is the total number of features, N is the set of all features {1,2,...,n}, S is any subset excluding feature j, and v(S) is the model prediction offset corresponding to feature subset S.

[0052] Based on the light-GBM model trained using the above steps, the contribution of each feature to the model output is analyzed using SHAP (Shapley Additive Explanations) values ​​to quantify the importance of features. SHAP values The marginal contribution of the j-th process feature to the CO concentration prediction result is represented by the weighted summation of the prediction increments under all feature subsets, which can quantitatively characterize the influence weight of each production parameter on the CO concentration.

[0053] Figure 2 The distribution of the top 20 most influential features (SHAP values) is shown, with parameters such as oxygen content in the fume hood, sintering endpoint temperature, and wind box negative pressure having a significant impact. The top 50 features in terms of SHAP importance are selected as the key feature variable set for subsequent modeling and control.

[0054] To reduce model complexity and enhance its robustness, this study selects the 50 features with the highest average absolute SHAP value based on SHAP feature importance ranking, forming a key feature subset. Subsequently, this key feature set is used as input variables to reconstruct a novel LightGBM prediction model. This method aims to eliminate redundant information and focus on core process parameters, thereby effectively improving the model's accuracy and generalization ability while ensuring prediction performance. The 50 selected features are shown in Table 2 below: Table 2: Detailed list of features after screening

[0055] Based on the trained light-GBM model, the contribution of each feature to the model output is analyzed using SHAP (Shapley Additive Explanations) values ​​to quantify the importance of features.

[0056] Figure 3The distribution of the top 20 most influential features (SHAP values) is shown, with parameters such as oxygen content in the fume hood, sintering endpoint temperature, and wind box negative pressure having a significant impact. The top 50 features in terms of SHAP importance are selected as the key feature variable set for subsequent modeling and control.

[0057] S4. Based on the identified key influencing factors, the soft measurement fitting model of CO concentration in the sintering process was retrained to obtain a simplified model.

[0058] Using the 50 selected key feature variables, a simplified model was retrained. Compared to the full-feature model, the simplified model, while maintaining essentially the same prediction accuracy, reduced the number of input parameters from over 200 to 50, significantly reducing model complexity and computational overhead, making it more suitable for online deployment and real-time control.

[0059] A simplified LightGBM model was constructed using 50 selected key feature variables, and its performance was comprehensively benchmarked against the original full-feature model.

[0060] Figure 4 This is a graph showing the change of SHAP value of the oxygen content in a smoke hood as a function of the oxygen content in the smoke hood, according to an embodiment of the present invention. The oxygen content data in the smoke hood is standardized data. Figures 5-9 The comparison between the simplified model and the full-feature model shows that the simplified model achieves the optimal balance between prediction accuracy and computational efficiency while significantly reducing the input dimensionality. It maintains stable accuracy, with a low coefficient of determination (R²). 2 The value remained stable above 0.85 (a decrease of approximately 0.99%), and the root mean square error (RMSE) fluctuation was controlled within an acceptable range of 3%, confirming that the key feature set effectively preserved the core information about CO concentration changes in the original data.

[0061] Furthermore, efficiency is significantly improved: There is a clear advantage in computational overhead, with the total model execution time reduced sharply from 152.4s to 48.7s, a decrease of 68%; the single-sample inference response time is shortened to 14.0ms, an improvement of 32%. Detailed comparison of evaluation metrics is shown in Table 3 below: Table 3: Comparison of Detailed Evaluation Indicators

[0062] By significantly reducing the input dimensionality through feature filtering, the model's input space shrinks considerably, effectively reducing the structural complexity and feature computation cost of LightGBM. This dimensionality reduction directly alleviates the computational overhead during training and inference, improving overall computational efficiency while maintaining predictive performance.

[0063] In summary, the simplified model achieves "lightweight" reconstruction with almost no loss of prediction accuracy, significantly reducing the computational load of the system and fully meeting the requirements of industrial sites for real-time response and online closed-loop control.

[0064] Compared with the full feature model, the simplified model significantly reduces model complexity and computational overhead while maintaining the same prediction accuracy, making it more suitable for online deployment and real-time control.

[0065] S5. Real-time acquisition of key influencing factor data; prediction of CO concentration and gradient through simplified model; screening of regulatory feature standards that reduce CO concentration the most; gradient descent adjustment of regulatory feature standards when concentration exceeds the limit; and fusion of gradient descent and PPO reinforcement learning control instructions to achieve multi-timescale optimal control of CO concentration, specifically including: S51. Real-time collection of 50 key feature data points to predict the current CO concentration value and its gradient; when the predicted value exceeds the standard, the gradient descent algorithm immediately fine-tunes parameters including damper opening and fuel supply to achieve real-time optimization of CO concentration.

[0066] This step performs gradient-based feature sensitivity analysis: p1. Baseline Prediction: Extract 50 key features from real-time data and input them into the newly trained LightGBM model to obtain the predicted CO concentration at the main extraction outlet under the current operating conditions (C_current) and its gradient with respect to each feature. The gradient is the partial derivative of the predicted CO concentration with respect to each input process feature, representing the magnitude of CO concentration change caused by a single unit change in a single feature, and is used to screen the process parameters with the best control effect.

[0067] p2. Single-factor perturbation analysis: For each key feature, a small operation adjustment is simulated using the trained lightGBM: its value is reduced by 10% (i.e., multiplied by 90%), while keeping the other 49 features unchanged, forming a new feature vector.

[0068] p3. Performance Evaluation: The perturbed feature vector is used as input to retrain the LightGBM model, resulting in a new CO concentration prediction value (C_new). The concentration change caused by this feature adjustment is calculated as: ΔC = C_new - C_current.

[0069] Taking data from June 23 to July 22, 2025 as an example, comparing the ΔC of all 50 features, we selected the 5 features that caused the largest decrease in CO concentration (i.e., ΔC was negative and had the largest absolute value). These are the most sensitive and effective control levers for reducing CO concentration under the current operating conditions. The gradient reflects the direction and sensitivity of the influence of each operating parameter on the current target indicator. A positive gradient indicates that increasing the parameter helps to improve the target value, while a negative gradient indicates that increasing the parameter will lead to a decrease in the target value; the larger the absolute value of the gradient, the more significant the influence of the parameter on the target indicator, and the more critical the control lever under the current operating conditions, as shown in Table 4 below: Table 4: Recommended Parameter Adjustments for the Current Environment

[0070] Furthermore, by combining the sintering process principles, equipment safety limits, and historical stable operation data, feasible control ranges for the aforementioned 50 key characteristic parameters are defined. Gradient control is limited to within these feasible control ranges.

[0071] S52. Simultaneously, the PPO reinforcement learning algorithm is introduced, which uses the CO concentration gradient signal calculated by the LightGBM prediction model and the historical control action sequence as the state input. This enables the reinforcement learning agent to perceive the current operating conditions, identify the directional influence of each key process parameter on the CO concentration change, and comprehensively consider the cumulative effect of the preceding adjustment behavior on the current emission state.

[0072] Proximal Policy Optimization (PPO) is a reinforcement learning algorithm based on policy gradients. Through continuous interaction within a sintering process simulation environment, it adaptively modifies the control strategy based on changes in CO concentration, thereby forming an optimal control strategy for long-term operating conditions. This addresses the problem that traditional instantaneous gradient-based methods struggle to maintain long-term stability. It includes: W1, State and Action Construction; In reinforcement learning control, the sintering process is abstracted as a Markov decision process.

[0073] The agent's state consists of two parts: the first part is the gradient signal of 50 key sintering process parameters with respect to predicted CO concentration, used to characterize the immediate sensitivity of each parameter to CO emissions under the current operating conditions; the second part is the historical control action sequence over the past five consecutive control cycles, used to characterize the cumulative effect of control behavior and system inertia. By introducing historical action information, the strategy can identify the relationship between short-term adjustments and long-term effects, avoiding frequent reverse adjustments.

[0074] The action space is set as a continuous vector, and adjustment values ​​are generated only for the five most sensitive process parameters currently identified. Strict physical and process constraints are set for each dimension of the action, such as limiting the adjustment range of the sintering machine speed to -0.2 to 0.2 m / min, to ensure that the control commands output by reinforcement learning can be directly used in the actual system.

[0075] W2. Reward Design and Long-Term Goal Guidance; At each control moment, the agent outputs a control action based on the current state. After the environment executes the action, it returns a new operating state and a corresponding immediate reward.

[0076] The reward function is used to comprehensively reflect the CO control effect, energy consumption level, and operational safety, and its form is shown below: (1); Here, ΔCO represents the change in predicted CO concentration before and after the current control action, and its absolute value is negative to guide the strategy to prioritize reducing CO concentration; the energy consumption ratio is the ratio of current energy consumption per unit time to baseline energy consumption, used to constrain the control strategy to avoid significantly increasing energy consumption; and the safety penalty is used to punish control behaviors that violate equipment operating limits or process constraints. Through this reward function, the agent gradually forms a long-term control objective that balances emission reduction, energy saving, and safety during the training process.

[0077] W3, PPO strategy updates and stability constraints; The reinforcement learning model employs an Actor-Critic structure, where the policy network outputs control actions, and the value network evaluates the value of the current state in terms of long-term reward. During training, the policy is iteratively updated based on sampled state, action, and reward information. The policy update uses a PPO (Pruning and Optimizing Objective) as its form, as follows: (2); Among them, A t This represents the advantage value of the control action in terms of long-term reward, the ratio is the probability ratio of the old and new policies for the same action, and ε is the pruning parameter used to limit the magnitude of policy change in a single update. This pruning mechanism avoids drastic oscillations in the policy during training, thereby improving the usability and stability of the control policy in industrial processes.

[0078] S53. By dynamically fusing gradient descent and PPO reinforcement learning control commands, CO concentration can be optimized and controlled at multiple time scales.

[0079] Integration and Deployment Strategy with Gradient Control: In the actual deployment phase, to avoid introducing excessive control fluctuations during the initial exploration stage of reinforcement learning, the long-term optimized control action output by PPO is fused with the short-term fast adjustment action calculated based on gradient descent to generate the final control command, as shown below: (3); Where a_gd represents the gradient descent control output, and a_ppo represents the PPO reinforcement learning control output.

[0080] During the initial operation phase, the gradient descent weight is set relatively high to ensure control response speed and safety. As the training rounds increase and the cumulative control effect continues to improve, the PPO weight is gradually increased, allowing the control strategy to gradually transition from short-term correction to long-term optimization. When the CO concentration is detected to continuously exceed the set safety threshold, the system automatically reduces the PPO weight and switches to gradient descent-dominated mode to prioritize production safety.

[0081] Training and Evaluation Mechanism: Reinforcement learning training employs a multi-round simulation approach, with each round corresponding to a complete sintering cycle. During training, the average reduction effect and fluctuation level of CO concentration are periodically evaluated on an independent test dataset. When the control stability or emission reduction effect fails to meet the preset requirements, the reward weights and strategy parameters are adjusted and retrained to ensure that the obtained control strategy has good robustness under different operating conditions.

[0082] The Replay Buffer serves as the core for dynamically storing training samples in the PPO reinforcement learning module, caching experience entries generated in real time through the interaction between the agent and the sintering simulation environment. Each experience entry consists of a quadruple ( Composition: State Includes real-time gradient signals of 50 key process parameters and historical action sequences from the past 5 control cycles; actions The continuous adjustment amount of the most sensitive process parameter at present; reward The three indicators are comprehensively quantified: CO concentration pressure reduction effect, energy consumption constraints, and safety penalties. This refers to the state at the next moment after the action is performed.

[0083] During the network update phase, the system randomly samples batches of historical experience from the Replay Buffer to iterate the parameters of the Actor-Critic network. Its core function is to break the strong correlation of industrial time-series data and endow the model with the ability to perceive system inertia and cumulative effects through historical action sequences. An example table of the Replay Buffer is shown below: Table 5: Example of Replay Buffer

[0084] In terms of motion space design, a dynamic weighted hybrid control mode combining gradient descent and PPO strategy output is adopted. Specifically: The gradient descent algorithm quickly generates short-term adjustment commands based on gradient information under the current operating conditions, which aim to adjust process parameters along the direction of CO concentration decrease, ensuring that the system has timely and effective response capabilities when CO concentration fluctuates or exceeds the standard.

[0085] The PPO module learns the optimal adjustment strategy that balances emission control effectiveness and operational stability across multiple time scales by continuously interacting with the environment.

[0086] In the early stages of operation, the system primarily uses gradient descent control to shorten the exploration phase of reinforcement learning and improve initial control stability. As policy training deepens and convergence increases, the weight of the PPO output action is gradually increased, allowing the control to smoothly transition from local rapid optimization to global long-term optimum.

[0087] The reward function is designed to comprehensively consider multiple objectives, including CO concentration deviation penalties, energy consumption rewards, and process safety constraints. Specifically, the CO concentration deviation penalty term guides the control strategy to continuously reduce emission levels, the energy consumption reward term suppresses increased energy consumption due to over-adjustment, and the safety constraint term limits the adjustment range of key parameters such as damper opening and machine speed, ensuring that the control strategy always operates within the safe range allowed by the equipment and process. Through this reward mechanism, a CO concentration control scheme that conforms to the actual sintering process is generated, achieving proactive and optimized control of CO emissions from sintering flue gas.

[0088] A further improvement of the technical solution of the present invention lies in that: the negative gradient direction adjustment specifically involves modifying the value of key features along the negative gradient direction based on the gradient direction of the model at the real-time data, so as to achieve effective control of CO emission concentration.

[0089] This invention solves the problems of low prediction accuracy and poor generalization performance caused by redundant input features in traditional modeling methods, as well as the lack of interpretability and difficulty in guiding actual process control of dimensionality reduction methods such as principal component analysis. It achieves accurate prediction and effective emission reduction of CO concentration in the sintering process.

[0090] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors, characterized in that, It includes: S1. Obtain the process parameter data of the sintering production process and the corresponding measured values ​​of CO emission concentration in sintering flue gas. After preprocessing including missing value filling, outlier removal and standardization, construct the model sample set. S2. Construct an improved gray wolf optimization algorithm that integrates random restart and stagnation escape mechanisms. Use the improved gray wolf optimization algorithm to optimize the parameters of the LightGBM regression algorithm. Based on the optimized LightGBM regression algorithm, build a soft measurement fitting model for CO concentration in the sintering process. S3. The SHAP method based on game theory is used to conduct feature importance analysis, quantitatively analyze the contribution of each process parameter in the sintering production process to the CO concentration prediction results, and identify the key influencing factors. S4. Based on the identified key influencing factors, the soft measurement fitting model of CO concentration in the sintering process was retrained to obtain a simplified model. S5. Real-time acquisition of key influencing factor data, prediction of CO concentration and gradient through simplified model; screening of the regulatory feature that reduces CO concentration the most, and using gradient descent regulation of the regulatory feature when the concentration exceeds the standard, and integrating gradient descent and PPO reinforcement learning control instructions to achieve multi-timescale optimization control of CO concentration.

2. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 1, characterized in that, S1 includes: S11. Collect historical data of the entire sintering process in the steel plant, including several process characteristic variables such as batching parameters, fan speed, wind box temperature, and oxygen content of the fume hood, as well as the corresponding measured values ​​of CO emission concentration. S12. Perform preprocessing on the collected data, including missing value imputation, outlier removal, and standardization. Preprocessing includes: identifying and removing outliers using box plots and the 3σ criterion; deleting entire rows of records with missing values ​​to ensure data consistency; aligning and resampling multi-source data based on a unified timestamp; and processing all numerical features using Z-Score normalization to eliminate the influence of units. S13. After preprocessing the collected data by filling missing values, removing outliers, and standardizing, construct the model sample set.

3. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 1, characterized in that, The improvements made to the standard Grey Wolf Optimization Algorithm in S2 include: Hybrid random exploration mechanism: Before updating each individual in each iteration, a probability determination is first performed: if the probability determination is triggered, the individual will ignore the current social level guidance and directly generate a new location coordinate randomly in the search space; otherwise, it will be updated according to the standard GWO formula. This is used to continuously jump to the unknown areas of the search space, increase population diversity, and prevent the algorithm from converging to a local optimum too early. Stagnation detection and population reset strategy: A stagnation counter is introduced into the algorithm to monitor the change in global optimal fitness in real time; a fitness stagnation criterion is defined. If stagnation is triggered, the top 3 wolves of the current best fitness are retained and the positions of all remaining wolves are forcibly reset. This is used to inject strong perturbation into the stagnant population and force the algorithm to jump out of the current local optimal region. Alpha Wolf-guided dynamic perturbation strategy: When updating the wolf pack population based on the location of the Alpha Wolf, a random perturbation term related to the search space range is superimposed.

4. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 1, characterized in that, The contribution of each process parameter in S3 to the CO concentration prediction result is represented by the SHAP value of each process parameter: ； in, For the first The SHAP value of each feature represents the impact of that feature value on the current prediction. The contribution of N is the total feature set {1,2,...,n}, S is any subset excluding feature j, and v(S) is the model prediction offset corresponding to feature subset S.

5. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 1, characterized in that, S5 includes: S51. Real-time acquisition of key characteristic data of a number of standard items to predict the current CO concentration value and its gradient; when the predicted value exceeds the standard, the gradient descent algorithm immediately fine-tunes parameters including damper opening and fuel supply to achieve real-time optimization of CO concentration. S52. At the same time, the PPO reinforcement learning algorithm is introduced, which uses the CO concentration gradient signal calculated by the LightGBM prediction model and the historical control action sequence as the state input. This enables the reinforcement learning agent to perceive the current operating conditions, identify the directional influence of each key process parameter on the CO concentration change, and comprehensively consider the cumulative effect of the preceding adjustment behavior on the current emission state. S53. By dynamically fusing gradient descent and PPO reinforcement learning control instructions, multi-timescale optimization control of CO concentration is achieved.

6. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 5, characterized in that, S5 also includes: combining the sintering process principle, equipment safety limits and historical stable operation data, defining the feasible control range of the number of key characteristic parameters of the standard, and limiting the gradient control within the feasible control range.

7. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 6, characterized in that, The control instructions for dynamically fusing gradient descent and PPO reinforcement learning in S53 are as follows: （3）； Where a_gd is the gradient descent control output. Here, a_ppo represents the weights controlling the output of gradient descent, and a_ppo represents the output controlling PPO reinforcement learning. The weights for controlling the output of PPO reinforcement learning.

8. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 7, characterized in that, In the initial stage of the S5 system operation, the weight of the gradient descent control output is set relatively high to ensure the control response speed and safety. As the training rounds increase, the weight of the PPO reinforcement learning control output is gradually increased when the cumulative control effect continues to improve, so that the control strategy gradually transitions from short-term correction to long-term optimization.

9. The method for optimizing and controlling CO concentration in sintering flue gas based on the identification of key influencing factors according to claim 8, characterized in that, When the CO concentration in S5 continuously exceeds the set safety threshold, the system automatically reduces the PPO weight and switches to the gradient descent dominant mode to prioritize production safety.

Images