Slagging and corrosion online prediction and prevention method for biomass blending combustion boiler

Abstract

The application discloses a biomass mixed combustion boiler slagging and corrosion online prediction prevention and control method, and belongs to the technical field of operation optimization of thermal power boilers.The application constructs a four-layer closed-loop intelligent system of a data sensing layer, a model prediction layer, a decision execution layer and a feedback optimization layer.The data sensing layer calculates mixed fuel ash components and ash melting points in real time through fuel ratio weighting, and establishes a heating surface partition monitoring network.The model prediction layer fuses mechanism slagging / corrosion indexes and an XGBoost multi-task model, and outputs partition slagging risk grades and corrosion rates.The decision execution layer dynamically optimizes soot blowing demand and soot blowing parameters, and is coupled with combustion air distribution for collaborative adjustment.The feedback optimization layer realizes model online updating through incremental learning.The application realizes real-time sensing, quantitative prediction and active collaborative prevention and control of slagging and corrosion risks, reduces soot blowing consumption, and prolongs the continuous operation cycle of a unit.

Description

Technical Field

[0001] This invention relates to the field of thermal power boiler operation optimization technology, specifically to an online prediction and intelligent collaborative prevention and control method for slagging and high-temperature corrosion of boiler heating surfaces under high proportion biomass co-firing conditions. Background Technology

[0002] Biomass fuels (such as tree bark, waste household scraps, and straw) have ash characteristics that are drastically different from conventional coal. First, the alkali metal (K, Na) content in biomass ash can be 5-10 times that of coal, the chlorine content can reach 1-3%, and the sulfur content varies greatly. Second, its low ash melting point (typically 200-400℃ lower than coal) leads to easy slagging in the furnace, and the slagging layer is characterized by rapid growth, strong adhesion, and difficulty in removal. Furthermore, high-temperature chlorine corrosion and sulfur corrosion coexist, with corrosion rates reaching 3-5 times that of pure coal combustion.

[0003] To address the aforementioned issues, existing technologies suffer from the following core deficiencies. In fuel monitoring, current offline sampling and testing methods result in a 4-8 hour lag, failing to reflect real-time changes in the characteristics of the fuel entering the boiler. In slagging assessment, reliance on operator experience is paramount, leading to strong subjectivity and hindering quantitative assessment of each heating surface of the boiler. Corrosion monitoring relies on periodic shutdowns for inspection or finned tests, a passive, post-event detection method that cannot provide real-time early warning. Regarding prevention and control measures, the use of timed and quantitative soot blowing and fixed air distribution strategies results in inefficient management, with ineffective soot blowing accounting for up to 40%, and failing to mitigate risks at their source.

[0004] In summary, existing technologies cannot achieve closed-loop intelligent management and control of "real-time perception-quantitative prediction-proactive prevention and control", resulting in frequent accidents of damage to the heating surface and seriously restricting the safe and economical operation of high-proportion biomass co-firing. Summary of the Invention

[0005] The present invention aims to solve the core technical problems in the prior art, such as the lack of real-time fuel characteristic perception, insufficient regional quantitative prediction, and the inability to coordinate and optimize the extensive prevention and control measures.

[0006] To address the aforementioned technical problems, this invention constructs a four-layer closed-loop intelligent system consisting of a "data perception layer, model prediction layer, decision execution layer, and feedback optimization layer," with the specific technical solution as follows.

[0007] (a) Data perception layer The data sensing layer is responsible for collecting parameters such as the characteristics of the fuel entering the furnace, the wall temperature of key heating surfaces and the near-wall flue gas temperature, and the composition of the flue gas in real time.

[0008] (1) Real-time acquisition method of ash composition of multi-component biomass fuel Establish a weighted calculation model based on fuel ratio and the characteristics of individual fuels.

[0009] Step 1: Create a system containing A database of individual fuel characteristics for each type of fuel. Key parameters include: ash content ; Gray components , , , , , , , , , (If data is missing, it can be approximated as) ); Ash melting point characteristic temperature: deformation temperature Softening temperature Hemispherical temperature Flow temperature .

[0010] Step 2: Obtain fuel through the fuel metering system Mass flow rate of each fuel at any time Calculate the mass ratio: Step 3: Weighted calculation of ash content of blended fuels: by For example, the calculation of the content of each ash component in the blended fuel: Step 4: Predict the ash melting point (softening temperature) of the blended fuel using the ash composition weighted index method. ):

[0011] (2) Real-time monitoring network for key heated surface areas The boiler heating surface is divided into 6 key risk zones: furnace water-cooled wall burner area (Z1), screen superheater (Z2), high temperature superheater (Z3), high temperature reheater (Z4), medium temperature superheater (Z5), and turning chamber and tail flue (Z6).

[0012] Monitoring parameter configuration: Collect wall temperature at a frequency of 1Hz at representative tube panels in each zone (no less than 3 points per zone). and near-wall smoke temperature (At a distance of 100-200 mm from the pipe wall); oxygen content in flue gas was collected at a frequency of 1 Hz at the furnace outlet and flue cross-section. and smoke Concentration; flue gas was collected at the furnace outlet at a frequency of 0.1 Hz. and concentration.

[0013] (ii) Model Prediction Layer The model prediction layer integrates the mechanistic model and the data-driven model to achieve online quantitative prediction of slagging and corrosion risks.

[0014] (1) Slagging Index Mechanism Calculation Module Define the following symbols (all are mass percentages of the corresponding components of a blended fuel): :Silica content, Aluminum oxide content, Iron oxide content, Calcium oxide content Magnesium oxide content, Sodium oxide content, Potassium oxide content, Sulfur trioxide content, Chlorine content, Titanium dioxide content (set to 0 if no data is available) : Softening temperature (°C).

[0015] 1. Silicon-to-aluminum ratio index : Criterion: It has a low tendency to slag; It has a moderate tendency to slagging; It exhibits a strong tendency to slag.

[0016] 2. Alkali-to-acid ratio index : Criterion: It has low slagging properties; It exhibits moderate slagging. High slagging; It exhibits severe slagging.

[0017] 3. Silicon Ratio Index : Criterion: It has low slagging properties; It exhibits moderate slagging. It has a high slagging rate.

[0018] 4. Comprehensive Slagging Index : ,in The reference softening temperature.

[0019] 5. Corrosion Risk Index: Chlorine corrosion index : Sulfur corrosion index : Comprehensive corrosion index : (2) Slagging / Corrosion Risk Prediction Model Based on XGBoost A multi-output ensemble learning model integrating mechanistic indices and operating parameters is constructed to achieve regional risk level prediction and quantitative calculation of corrosion rate.

[0020] 1. Model input feature vector 15 dimensions in total: 5-dimensional mechanism index characteristics: 6-dimensional operating parameter characteristics: wall temperature Near-wall smoke temperature Temperature difference Oxygen content in flue gas Smoke Concentration, unit load rate ; 2D time-series characteristics: wall temperature change rate The continuous running time of this partition 2D historical characteristics: frequency of historical slag formation in this area. Depth of historical corrosion .

[0021] 2. Model Output: Slagging Risk Level (Level 0 indicates no slag, Level 4 indicates severe slag blockage); Corrosion risk rate (Unit: mm / 1000h).

[0022] 3. Multi-tasking XGBoost architecture: For the task of classifying the risk level of slagging, the Softmax multi-class objective function is adopted: ,in For the first Tree integration model of classes: , Sample Mapped to the The leaf node index of the tree. The weights are those of the leaf nodes.

[0023] The overall objective function (with regularization) is: ,in The total number of training samples, For real labels, These are predicted values.

[0024] Multi-class cross-entropy loss function: ,in Indicator variable, sample Category The value is 1 if it is true, and 0 otherwise. Model prediction samples Category The probability of.

[0025] Regularization term: ,in The number of leaf nodes in a single decision tree. : No. The weight of each leaf node, , : Regularization hyperparameter.

[0026] For the corrosion rate regression task, the squared loss function is used: .

[0027] Model training strategy: Data preprocessing: Time-align historical operating data with shutdown inspection records and establish "operating parameters - actual slagging / corrosion results" tag pairs.

[0028] Feature engineering: 5-minute moving average smoothing was performed on continuous variables such as wall temperature and flue gas temperature; Z-score normalization was performed on various slagging indices. ; Constructing interactive features (Temperature-corrosion synergistic effect).

[0029] Hyperparameter optimization: Bayesian optimization is used, and the search range includes the maximum depth of the tree. Learning rate The number of trees .

[0030] Model validation: Time series cross-validation was used, with a target slag classification accuracy greater than 85% and a target mean absolute percentage error (MAPE) of corrosion rate prediction less than 20%.

[0031] (III) Decision-making and execution level The decision-making and execution level dynamically optimizes the intelligent soot blowing strategy and the coordinated adjustment of combustion air distribution based on the prediction results.

[0032] (1) Intelligent soot blowing strategy optimization module Step 1: Each risk zone Demand index for blowing soot The calculation is as follows: ,in: Slagging risk level (0-4). : Wall temperature rise rate (°C / min) The degree to which the wall temperature exceeds the design value : Time since last soot blowing (h), weighting coefficient: , , , .

[0033] Step 2: Apply the following to all partitions: Generate a priority queue for blowing off dust by sorting the dust in descending order.

[0034] Step 3: Adaptively adjust the soot blowing medium parameters according to the risk level.

[0035] Risk level 0: Steam pressure 1.0MPa, soot blowing time 15s, soot blowing frequency 1 time / shift, using conventional soot blowers; Risk level 1: Steam pressure 1.2MPa, soot blowing time 20s, soot blowing frequency 2 times / shift, conventional soot blower; Risk level 2: Steam pressure 1.5MPa, soot blowing time 30s, soot blowing frequency 3 times / shift, enhanced soot blower; Risk level 3: Steam pressure 1.8MPa, soot blowing time 45s, soot blowing frequency 4 times / shift, high frequency soot blower; Risk level 4: Steam pressure 2.0MPa, soot blowing time 60s, soot blowing frequency 6 times / shift or more for emergency soot blowing, supplemented by high-frequency soot blowing and sonic soot blowing.

[0036] Step 4: Energy-saving optimization strategy. For areas with a risk level of 0-1 for three consecutive prediction cycles, automatically extend the soot blowing interval: ,in Number of consecutive low-risk periods Basic soot blowing interval (h) : Maximum permissible soot blowing interval (h).

[0037] (2) Combustion and air distribution coordinated optimization module Optimization variable: Burner tilt angle (range of values) Secondary air distribution ratio ,satisfy ; Burnout wind rate (range of values) Excess air coefficient (range of values) ) Optimize the objective function: ,in , These are the normalized average slagging risk index and average corrosion index. For thermal efficiency, For nitrogen oxide emission concentrations, the weighting coefficients are as follows: All are positive numbers.

[0038] Constraints: Upper limit of wall temperature: Incomplete combustion: Furnace negative pressure: .

[0039] The solution is obtained using an improved Particle Swarm Optimization (PSO) algorithm: Particle position: Speed ​​updates: Location update: Inertia weights adaptively decrease linearly: (3) Cooperative control logic: Within each forecast period, the slagging risk level of each heated surface zone is statistically analyzed, and the risk level is assigned accordingly. The area is classified as a "high-risk area for slagging". If the current number of high-risk areas for slagging is... If this is not the case, the combustion optimization algorithm will be triggered, prioritizing reducing the furnace outlet flue gas temperature, increasing the burnout air rate, and increasing the burner tilt angle. After the adjustments are implemented, a reassessment will be conducted; if the number of high-slagging risk areas remains... If it decreases to..., then continue iterating; if it decreases to... If the initial number of high slagging risk areas is [not specified], then the current parameters will be maintained. If combustion optimization is not triggered, local control will be achieved solely through a soot blowing strategy.

[0040] (iv) Feedback Optimization Layer (1) Online model update mechanism An incremental learning strategy is used to continuously optimize the model. Triggering condition: Periodic trigger: Every 1000 hours of accumulated runtime data Event trigger: After an unplanned shutdown or a failure of the heated surface. Performance trigger: Model prediction accuracy is below the threshold for 7 consecutive days. Update process: Merge and clean the new data with historical data, retrain the XGBoost model using a warm-start approach, verify the performance of the new model through A / B testing, and replace the online model when the performance improvement exceeds 5%.

[0041] (2) Evaluation of prevention and control effectiveness and knowledge accumulation Establish an indicator system for evaluating the effectiveness of prevention and control: Safety indicators: Number of unplanned shutdowns, number of times heating surface tubes rupture. Economic indicators: Soot blowing steam consumption (tons / month), heat exchange efficiency decay rate Prediction indicators: Slagging prediction accuracy (TP / (TP+FP)), average absolute percentage error of corrosion rate prediction ( ).

[0042] Beneficial effects This invention overcomes the shortcomings of traditional offline sampling and testing, which lags by 4-8 hours, by establishing a real-time weighted calculation model for the ash composition of multi-element biomass fuels. It can obtain the ash content, ash composition and ash melting point data of the mixed fuels fed into the furnace in real time, providing highly timely input features for subsequent slagging and corrosion prediction.

[0043] This invention divides the boiler heating surface into 6 key risk zones and integrates the mechanism slagging / corrosion index with the XGBoost multi-task model to achieve online quantitative prediction of the slagging risk level (0-4) and corrosion rate (mm / 1000h) of each zone. This solves the problem of existing technologies relying on the experience judgment of operators and lacking quantitative assessment of zones.

[0044] Based on the prediction results, this invention dynamically generates the soot blowing demand index and adaptively adjusts the soot blowing medium parameters. At the same time, it couples combustion air distribution optimization (burner swing angle, secondary air ratio, burnout air rate, excess air coefficient), which overcomes the extensive management defects of traditional timed and quantitative soot blowing and fixed air distribution strategies, and realizes the transformation from "passive post-event processing" to "active source suppression".

[0045] The verification of the embodiments shows that the present invention can effectively reduce the consumption of soot blowing steam (by up to 31%), extend the continuous operation cycle of the unit (from an average of 4 months to 7.5 months), and at the same time maintain the accuracy of slagging prediction above 85% and the average absolute percentage error of corrosion rate prediction below 20%, providing real-time early warning and active prevention and control measures for unplanned shutdowns and accidents such as tube rupture of heated surfaces. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the overall architecture of the method of this invention.

[0047] Figure 2 Flowchart for real-time calculation of ash composition of multi-component biomass fuel.

[0048] Figure 3 This is a structural diagram of the slagging / corrosion risk fusion prediction model.

[0049] Figure 4 Flowchart of the intelligent dust removal decision-making logic.

[0050] Figure 5 Flowchart of the combustion-air distribution coordinated optimization algorithm. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of this invention.

[0052] Example 1

[0053] This embodiment takes a 600MW supercritical parameter variable pressure once-through boiler as an example. The boiler has a Π-type layout and is designed to use bituminous coal blended with 30% biomass (bark + straw) as fuel. The heating surface adopts a front and rear wall opposed combustion method and is equipped with a partition screen superheater, a rear screen superheater, a final stage superheater, and a high-temperature reheater. It is deployed and implemented according to the four-layer closed-loop intelligent system described in this invention.

[0054] I. Data Perception Layer (1) Real-time acquisition method of ash composition of multi-component biomass fuel Step 1: Establish a database of single fuel characteristics (bituminous coal, bark, straw):

[0055] Step 2: Real-time fuel ratio acquisition. Feed mass flow rate at a certain moment: , , Total flow rate: 75 t / h. Calculate the mass ratio: .

[0056] Step 3: Ash content of blended fuel:

[0057] by For example: Similarly, we can conclude that: , , , , , , , .

[0058] Step 4: Ash melting point prediction: Predicting the softening temperature of blended fuels using the ash composition weighted index method. :

[0059]

[0060] (2) Real-time monitoring network for key heated surface areas The boiler heating surface was divided into 6 key risk zones (Z1~Z6), and the wall temperature was collected at a frequency of 1Hz at representative tube screens in each zone (no less than 3 points in each zone). and near-wall smoke temperature Data were collected at a frequency of 1Hz at the furnace outlet and flue section. and Concentration, collected at a frequency of 0.1 Hz. and concentration.

[0061] II. Model Prediction Layer (1) Calculation of slagging index mechanism Silicon-to-aluminum ratio: ( (Severe tendency to slag) Alkali-acid ratio: ( (Medium slagging tendency) Silicon ratio: ( (Medium slagging tendency) Comprehensive Slagging Index: Chlorine corrosion index: Sulfur corrosion index: Overall Corrosion Index: (2) Slagging / Corrosion Risk Prediction Model Based on XGBoost 1. Data preprocessing and label creation Operating data from the past six months (approximately 4300 hours of valid data) was collected for the unit. Parameters such as feed rate, wall temperature, flue gas temperature, and flue gas composition were obtained from the plant-level monitoring information system (SIS) and uniformly resampled to 1Hz. Simultaneously, slagging photos and slagging thickness measurements for each heating surface zone were extracted from three shutdown maintenance records. Slagging levels were categorized according to industry standards, ranging from level 0 (no slagging) to level 4 (severe slagging), creating classification labels. Corrosion rates (unit: mm / 1000h) for each zone were calculated based on tube wall thinning measurement data from the high-temperature superheater and high-temperature reheater (using an ultrasonic thickness gauge), generating regression labels. Historical operating data and shutdown inspection records were time-aligned to establish "operating parameters - actual slagging / corrosion results" label pairs.

[0062] 2. Feature Engineering According to the present invention, a 15-dimensional input feature vector is constructed. Specifically, it includes: 5-dimensional mechanism index characteristics: , , , , (See the preceding content in this section for calculation examples). 6-dimensional operating parameter characteristics: wall temperature Near-wall smoke temperature Temperature difference Oxygen content in flue gas Smoke Concentration, unit load rate ; 2D time-series characteristics: wall temperature change rate The continuous running time of this partition ; 2D historical characteristics: frequency of historical slag formation in this area Depth of historical corrosion .

[0063] Continuous variables such as wall temperature and flue gas temperature were smoothed using a 5-minute moving average (300 sampling points in the window). Various slagging indices ( , , , , Perform Z-score normalization: ,in The mean, The standard deviation is used to construct interaction features. (Temperature-corrosion synergistic effect), added to the feature vector. This ultimately generates approximately 250,000 time-aligned samples (one sample every 5 minutes).

[0064] 3. Model Structure and Loss Function The model adopts a multi-task XGBoost architecture and outputs both the slagging risk level (0-4) and the corrosion risk rate (mm / 1000h).

[0065] For the task of classifying the risk level of slagging, a Softmax multi-class objective function is adopted. The model output is the first... kind( The probability of () is: in For the first Tree integration model of classes: ,here For the number of trees, Map the sample to the first The leaf node index of the tree. Leaf weights. The final predicted slagging risk level is: The overall objective function (with regularization) is: in The total number of training samples, For the first The true label of each sample These are the predicted values. The multi-class cross-entropy loss function is: in For indicator variables (samples) Category (1 if it is 1, 0 otherwise) Predict samples for the model Category The probability of.

[0066] Regularization term Used to control model complexity, defined as: in The number of leaf nodes in the decision tree. For the first The weight of each leaf node, and This is the regularization hyperparameter.

[0067] For the corrosion rate regression task, the squared loss function is used: 4. Hyperparameter optimization Bayesian optimization is used to search for the optimal hyperparameters. The search range for the key parameters is as follows: maximum tree depth. Learning rate Number of trees Regularization parameters , .

[0068] After Bayesian optimization (50 iterations), the optimal hyperparameters were determined to be: tree depth = 6, learning rate = 0.08, and number of trees = 350. , .

[0069] 5. Model Validation Time-series cross-validation is used to avoid data leakage: the data is divided chronologically, with the first 80% used as the training set and the last 20% as the validation set, ensuring that the validation set is later than the training set. The target accuracy for slag classification is greater than 85%, and the target mean absolute percentage error (MAPE) for corrosion rate prediction is less than 20%. In this embodiment, the model achieves a slag classification accuracy of 87.3% and a corrosion rate prediction MAPE of 18.5% on the validation set.

[0070] III. Decision-making and Execution Level This implements intelligent soot blowing strategy and combustion air distribution optimization.

[0071] (1) Optimization of intelligent soot blowing strategy Taking the high-temperature superheater region (Z3) as an example, the model predicts the slagging risk level in this region. (Level 3). Real-time monitoring data: wall temperature rise rate Current wall temperature Design wall temperature Time since the last dust blowing .

[0072] Calculate the soot blowing demand index (weighting coefficient) ):

[0073] This value ranks first among all zones, triggering a soot blowing command. The soot blowing is executed according to the parameters corresponding to risk level 3: steam pressure 1.8 MPa, soot blowing duration 45 seconds, and high-frequency soot blower operation 4 times per shift.

[0074] For regions with a risk level of 0-1 for three consecutive prediction cycles (such as the Z5 medium-temperature superheater), the soot blowing interval is automatically extended according to the energy-saving optimization strategy of this invention. Let the base interval be... Number of consecutive low-risk cycles The new interval is:

[0075] (2) Combustion and air distribution synergistic optimization When the model predicts the slagging risk level in the high-temperature superheater or high-temperature reheater area And the comprehensive corrosion index At that time, combustion air distribution optimization is automatically triggered. In this embodiment, the slagging level in region Z3 is 3. This triggers optimization.

[0076] An improved particle swarm optimization (PSO) algorithm was used to solve for the optimization variable: burner tilt angle. Secondary air ratio Burnout wind rate Excess air coefficient The objective function to be optimized is:

[0077] Optimized output suggestion: Burner tilt angle from upward tilt Adjusted to (i.e., downward tilt) ), burnout wind rate from Down to The excess air coefficient is from Down to At the same time, the opening of the secondary air central damper will be increased. The recommended values ​​should be displayed on the monitoring system screen and confirmed by the operators before implementation.

[0078] IV. Feedback Optimization Layer After accumulating 1000 hours of operational data, the system automatically triggers incremental learning, retraining the XGBoost model using a warm-start approach. If the new model shows a performance improvement of more than 5% on the validation set, it automatically replaces the online model. The system automatically calculates the model's prediction accuracy daily and generates a prevention and control effectiveness report weekly. Statistical results from 12 consecutive months of operation show that the slagging prediction accuracy remained between 86% and 89%, the corrosion rate prediction MAPE was between 17% and 20%, the monthly soot blowing steam consumption decreased from 1200 tons to 830 tons (a decrease of 31%), and the unit's continuous operation period increased from an average of 4 months before the modification to 7.5 months.

[0079] Example 2

[0080] This embodiment is basically the same as Embodiment 1, except that the unit was originally designed to burn only bituminous coal, but has now been modified to burn 25% straw and garden waste, and no online wall temperature monitoring system has been installed. At the data sensing layer, infrared temperature probes (2 points per zone, sampling frequency 0.2Hz) are used to replace the original thermocouples. At the model prediction layer, due to the lack of accurately labeled data on slagging and corrosion, a transfer learning strategy is adopted: using the XGBoost model trained in Embodiment 1 as a starting point, fine-tuning is performed using only two months of actual operating data from this unit, improving the slagging classification accuracy from the initial 72% to 81%. The remaining implementation steps are the same as in Embodiment 1, ultimately reducing soot blowing steam consumption by approximately 18% and extending the continuous operating cycle from 3 months to 5 months.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can make various modifications, equivalent substitutions, and improvements without departing from the spirit and principles of the present invention, and all such modifications and substitutions should be included within the scope of protection of the present invention.

Claims

1. A method for online prediction and prevention of slagging and corrosion in biomass co-firing boilers, characterized in that, Includes the following steps: Data sensing: Based on the feed mass flow rate of each fuel and the pre-stored database of single fuel characteristics, the ash content, mass percentage of each ash component and ash melting point softening temperature of the mixed fuel are obtained in real time through weighted calculation; and the boiler heating surface is divided into multiple risk zones, and the wall temperature, near-wall flue gas temperature and flue gas composition of each zone are collected in real time. Model prediction: The slagging mechanism index and corrosion mechanism index are calculated based on the ash composition of the mixed fuel. The mechanism index, along with operating parameters, time series characteristics and historical characteristics, are input into the XGBoost multi-task model to output the slagging risk level and corrosion rate of each zone. Decision execution: Calculate the soot blowing demand index based on the slagging risk level, wall temperature rise rate, degree of wall temperature exceeding the design value, and time since the last soot blowing in each zone, and dynamically adjust the soot blowing parameters according to the index; and trigger combustion air distribution synergistic optimization based on the slagging risk level and corrosion index, and use particle swarm optimization algorithm to optimize burner sway angle, secondary air ratio, burnout air rate and excess air coefficient. Feedback optimization: An incremental learning strategy is adopted to retrain the XGBoost model periodically or triggered by events, and the model is replaced online based on the evaluation index of prevention and control effectiveness.

2. The method according to claim 1, characterized in that, In the data sensing step, the ash content and ash percentage of the blended fuel are calculated as follows: Ash content of the blended fuel ,in For the first The mass ratio of the fuels Its ash content; the mass percentage of a certain ash component in a blended fuel. .

3. The method according to claim 1, characterized in that, In the data sensing step, the ash melting point softening temperature of the mixed fuel is... The gray component weighted index method is used to predict using the following formula: 。 4. The method according to claim 1, characterized in that, The slagging mechanism index includes the silicon-aluminum ratio. Alkali-acid ratio Silicon ratio and comprehensive slagging index The corrosion mechanism index includes the chlorine corrosion index. sulfur corrosion index and comprehensive corrosion index The formulas for calculating each index are as follows: 。 5. The method according to claim 1, characterized in that, The input feature vector of the XGBoost multi-task model is 15-dimensional, including: 5-dimensional mechanistic index features. 6-dimensional operating parameter characteristics: wall temperature Near-wall smoke temperature Temperature difference Oxygen content in flue gas Smoke Concentration, unit load rate ; 2D time-series characteristics: wall temperature change rate Continuous running time of partitions 2D historical characteristics: frequency of historical slagging in this area Depth of historical corrosion The model outputs a slagging risk level. and corrosion rate , unit mm / 1000h.

6. The method according to claim 1, characterized in that, In the decision-making and execution steps, the soot blowing demand index The calculation formula is: ,in , , , ; The parameters of the soot blowing medium are set according to the slagging risk level. The higher the risk level, the greater the steam pressure, the longer the soot blowing time, and the higher the soot blowing frequency.

7. The method according to claim 1, characterized in that, In the decision execution steps, for partitions that are low-risk for multiple consecutive prediction periods, the dust-blowing interval is automatically extended: ,in The number of consecutive low-risk cycles.

8. The method according to claim 1, characterized in that, The objective function for the combustion-air distribution coordinated optimization is: The constraints include: upper limit of wall temperature, upper limit of CO emission concentration, and range of furnace negative pressure; the particle swarm algorithm adopts adaptive linearly decreasing inertial weights.

9. The method according to claim 1, characterized in that, In the feedback optimization step, the triggering conditions for incremental learning include: every 1000 hours of accumulated running data, unplanned downtime or heat-affected surface failure, and model prediction accuracy being below a threshold for 7 consecutive days; retraining is performed using a warm-start method, and the performance of the new model is verified through A / B testing. When the performance improvement exceeds 5%, the online model is replaced.

Images