A multi-coal blending dynamic optimization system

Abstract

The application discloses a multi-coal blending combustion dynamic optimization system, comprising an execution layer, a perception layer and a decision layer; the execution layer executes specific operations of mixed blending combustion, receives control instructions generated by the decision layer, and feeds back blending operation data after execution according to the control instructions to the perception layer; the perception layer collects data in real time through a combustion field perception network, an emission monitoring module and a coal quality online detection array, and uploads the real-time data stream to the decision layer; a dynamic coal blending model of the decision layer generates optimal coal blending proportions based on real-time coal quality data and operation feedback; and a multi-objective optimization engine of the decision layer can output adjustment coal blending and combustion optimization strategies. Through the synergistic effect of the dynamic coal blending model, the multi-objective optimization engine and the self-adaptive compensation system, the intelligent level of coal blending combustion is significantly improved; based on real-time data of the perception layer and a DRL intelligent agent, dynamic real-time optimization adjustment of coal blending is realized, and coal quality fluctuation and working condition changes are adapted.

Description

Technical Field

[0001] This invention relates to the field of coal blending technology, specifically to a dynamic optimization system for blending multiple coal types. Background Technology

[0002] Coal blending refers to the technology of burning a mixture of different types of coal (such as lignite, bituminous coal, and anthracite) in a specific ratio. Its core lies in utilizing the complementary characteristics of each coal type (such as calorific value, sulfur content, ash fusion point, and volatile matter) to optimize combustion performance and achieve goals such as increased efficiency, reduced costs, or reduced pollutant emissions. Coal blending not only solves the problem of insufficient performance of a single coal type (such as high-sulfur coal causing significant pollution and low-calorific-value coal causing low efficiency), but also allows for flexible responses to market coal price fluctuations and environmental policy requirements. It is a key technology for achieving a balance between economic, environmental, and safety objectives in the industrial sector.

[0003] Existing coal blending and combustion optimization systems are unable to respond in real time to coal quality fluctuations, load changes, and equipment status anomalies, resulting in a disconnect between the coal blending scheme and actual operating conditions, affecting combustion stability and efficiency. The existing ability to collect and analyze real-time data such as coal quality parameters and combustion status is insufficient, failing to effectively drive closed-loop optimization, resulting in low strategy update frequency and poor accuracy. Summary of the Invention

[0004] Therefore, the present invention provides a dynamic optimization system for blending multiple coal types to solve the problems in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A dynamic optimization system for multi-coal blending includes an execution layer, a perception layer, and a decision-making layer;

[0007] The execution layer performs the specific operations of mixing and co-firing. The execution layer receives control commands generated by the decision layer and feeds back the co-firing operation data after execution according to the control commands to the perception layer.

[0008] The perception layer collects data in real time through the combustion field perception network, emission monitoring module, and online coal quality detection array, and uploads it to the decision-making layer through real-time data stream.

[0009] The decision-making layer includes a dynamic coal blending model, a multi-objective optimization engine, and an adaptive compensation system;

[0010] The dynamic coal blending model generates the optimal coal blending ratio based on real-time coal quality data and operational feedback; the multi-objective optimization engine balances combustion efficiency, emission control, and cost targets, and outputs strategies for adjusting coal blending and combustion optimization; the adaptive compensation system dynamically adjusts compensation parameters according to real-time operating conditions to correct the strategies in real time.

[0011] The execution results are returned to the perception layer through execution feedback, triggering a new round of optimization iterations.

[0012] Furthermore, the execution layer includes an intelligent coal feeder unit, a three-dimensional coal blending device, and a burner co-controller. The intelligent coal feeder unit adjusts the coal feed rate according to control commands to ensure that fuel supply matches demand. The three-dimensional coal blending device accurately mixes different types of coal according to the instructions of the dynamic coal blending model. The burner co-controller adjusts the burner parameters to achieve efficient combustion and emission control.

[0013] Furthermore, the sensing layer includes a combustion field sensing network, an emission monitoring module, and an online coal quality detection array; wherein, the combustion field sensing network can monitor the combustion status in real time; the emission monitoring module can detect pollutant emission data; and the online coal quality detection array can analyze key parameters of coal, including composition and calorific value, in real time.

[0014] Furthermore, the specific implementation of the dynamic coal blending model is as follows:

[0015] (1) Feature extraction: Real-time raw coal quality data is obtained from the perception layer and input into the dynamic coal blending model. Then, physical features are extracted, and temporal features are extracted using a deep feature encoder to output a 32-dimensional normalized feature vector.

[0016] ;

[0017] Where v1, v2, ..., v32 refer to each feature value in a 32-dimensional feature vector;

[0018] (2) DRL agent decision-making: Input the 32-dimensional normalized feature vector into the DRL agent to make preliminary decisions;

[0019] (3) Physical constraint layer verification: The extracted physical features and the preliminary decision of DRL agent production are input into the physical constraint layer for verification;

[0020] Feasibility adjustments were made using the modified Lagrange multiplier method:

[0021] ;

[0022] in, The adjustment factor is xi, where xi represents the equipment constraint capability.

[0023] (4) Hybrid decision generation: calculate the conflict degree δ between DRL decision and physical constraints, then dynamically adjust the weight w allocation, and output the final coal blending scheme.

[0024] The formula for calculating the conflict degree δ is as follows:

[0025] ;

[0026] Among them, a DRL Let a be the decision vector given to the DRL agent. phy The decision vector to satisfy physical constraints;

[0027] The formula for calculating the dynamic weight allocation w is as follows:

[0028] ;

[0029] Where k is a constant used to adjust the sensitivity of weight allocation, that is, the degree of influence of conflict degree δ on weight w;

[0030] The final coal blending scheme calculation formula is as follows: .

[0031] Furthermore: The DRL agent employs the Proximal Policy Optimization (PPO) algorithm, and its state space, action space, and reward function are calculated using the following formulas:

[0032] State space:

[0033] ;

[0034] Where V represents a 32-dimensional normalized eigenvector; Q real Indicates the real-time heat load of the boiler (MW); η curr ENOx represents the current thermal efficiency (%); ENOx represents E NOx Emission concentration (mg / m³);

[0035] Action space:

[0036] ;

[0037] Let be the adjustment amount for the proportion of coal type i, representing the amount by which the proportion of coal type i needs to be increased or decreased relative to the current coal blending ratio; n is the number of coal types.

[0038] Reward function:

[0039] ;

[0040] Where η represents thermal efficiency, C coal Q represents the cost of coal. tar Q represents the target heat load. real The real-time heat load is represented by the reward function R, which encourages the DRL agent to improve thermal efficiency, reduce coal costs, reduce NOx emissions, and bring the real-time heat load close to the target heat load.

[0041] Furthermore: the DRL strategy network parameters are updated every 24 hours; the constraint boundaries are automatically updated when equipment is under maintenance or coal yard inventory changes.

[0042] Furthermore: the multi-objective optimization engine includes an input layer, a core algorithm layer, a decision layer, and an output layer;

[0043] The input layer includes an objective function, real-time operating data, and constraints. The objective function is a dual objective function of economic efficiency and environmental protection. The constraints are of three types: boiler safety, equipment capacity, and environmental regulations. The real-time data includes coal quality parameters, load demand, and emission monitoring values.

[0044] The objective function is:

[0045] ;

[0046] Among them, c i x is the unit price of the i-th type of coal (yuan / ton); i The blending ratio of the i-th type of coal (%) ej η is the conversion factor (mg / m³ / %) for the j-th emission type; act / tar Actual / target thermal efficiency (%); α, β, γ are weighting coefficients (α+β+γ=1);

[0047] The constraints are:

[0048] ;

[0049] Where, x i Let q be the blending ratio of the i-th type of coal. i The lower heating value (MJ / kg) of the i-th type of coal; s i The sulfur content (%) of the i-th type of coal; P min / max S represents the boiler's permissible heat load range (MW); max The upper limit of sulfur content (%) is set according to environmental standards.

[0050] The core algorithm layer includes improved NSGA-III, adaptive reference point generation, and dynamic population management. Improved NSGA-III is the main framework for multi-objective optimization. Adaptive reference points are used to dynamically adjust the distribution density of reference points. Dynamic population management can retain historical elite solutions and accelerate convergence.

[0051] The formula for adaptive generation of reference points is:

[0052] ;

[0053] Where Δfk is the rate of change of the kth objective function;

[0054] The decision-making layer includes Pareto front extraction and fuzzy decision-making modules. The Pareto front is used to visualize the non-dominated solution set, and the fuzzy decision-making is used to comprehensively allocate weights and screen the solution set. The output layer includes the optimal solution set and control instructions. The optimal solution set is the set of feasible solutions that satisfy all constraints, and the control instructions refer to the specific coal feed ratio and combustion parameters.

[0055] Furthermore, the implementation steps of the multi-objective optimization engine are as follows:

[0056] (1) Initialize the population; generate N random solutions, each solution is an m-dimensional vector (m = number of coal types); the encoding method adopts real number encoding to represent the blending ratio of each coal type;

[0057] (2) Non-dominated sorting; calculate the dominance relationship for each solution, and the hierarchical sorting formula is as follows:

[0058] ;

[0059] Where I=1, y dominates x, otherwise 0;

[0060] (3) Generate adaptive reference points;

[0061] The formula for adjusting the density of the reference point is:

[0062] ;

[0063] Where k=5, nobj=2, and t is the number of iterations;

[0064] (4) Crossover variation;

[0065] Using SBX crossover and polynomial mutation, the calculation formula is as follows:

[0066] ;

[0067] Where, η c =20, η c The distribution index;

[0068] (5) Dynamic population update; an elite retention strategy is adopted, retaining the top 20% of historical optimal solutions; the calculation formula for the elite retention strategy is: P t+1 =Select(P t ∪Q t (N);

[0069] (6) Fuzzy decision-making: The weights are determined by the entropy weight method, and then the optimal solution is selected from the Pareto solution set by TOPSIS sorting.

[0070] Furthermore: the formula for calculating the weights using the entropy weight method is as follows: ;

[0071] ;

[0072] The TPOSIS sort is calculated as follows:

[0073] ;

[0074] The ideal solution distance is:

[0075] ;

[0076] ;

[0077] Among them, w j E represents the target weighting coefficient. j p is the information entropy of the j-th index; ij is the proportion of the standardized value of the i-th sample under the j-th indicator; m is the total number of samples, and n is the total number of indicators; Let be the Euclidean distance from the i-th solution to the positive ideal solution (optimal solution); z is the Euclidean distance from the i-th solution to the negative ideal solution (worst solution); ij These are the standardized indicator values; Let j be the index value of the positive ideal solution; The j-th index value is the negative ideal solution.

[0078] Furthermore, the adaptive compensation system operates as follows:

[0079] (1) Data acquisition, real-time reception of coal quality, combustion and equipment status parameters;

[0080] (2) Anomaly detection: Multi-dimensional anomaly detection is performed every n seconds;

[0081] (3) Compensation decision: Match the best compensation type from the policy library and generate specific parameters through the transfer learning model; The transfer learning model uses MMD to minimize the distribution difference between domains to realize feature pairs; The model parameters are updated once every n new feedback data are received;

[0082] Its incremental learning formula is:

[0083] ;

[0084] in, To compensate for the loss of effectiveness; The domain adaptation loss is represented by λ = 0.5, which is the tradeoff coefficient.

[0085] (4) Security verification: if the verification passes, an instruction is issued; if it fails, a downgrade strategy is triggered.

[0086] (5) Effect feedback: Record the boiler efficiency and emission data after compensation for online updates of the transfer learning model.

[0087] The present invention has the following advantages: through the synergistic effect of dynamic coal blending model, multi-objective optimization engine and adaptive compensation system, the level of intelligence of coal blending and combustion is significantly improved; based on real-time data of perception layer and DRL intelligent agent, dynamic real-time optimization and adjustment of coal blending is realized to adapt to coal quality fluctuations and operating condition changes; through multi-objective optimization engine, thermal efficiency, emission control and coal cost are comprehensively optimized to improve the economic and environmental performance of the system.

[0088] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0089] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0090] Figure 1 This application provides an architecture diagram of a dynamic optimization system for multi-coal blending.

[0091] Figure 2 This is a flowchart of a dynamic coal blending model in a multi-coal blending dynamic optimization system according to the present invention.

[0092] Figure 3 This is an architecture diagram of a multi-objective optimization engine in a dynamic optimization system for blending multiple coal types, as described in this invention.

[0093] Figure 4 This is a flowchart illustrating the operation of the multi-objective optimization engine in a dynamic optimization system for multi-coal blending according to the present invention. Detailed Implementation

[0094] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0095] Please see Figures 1-4 A dynamic optimization system for multi-coal blending includes an execution layer, a perception layer, and a decision-making layer.

[0096] The execution layer includes an intelligent coal feeder unit, a three-dimensional coal blending device, and a burner co-controller. The execution layer can perform specific operations of blending and co-firing. The execution layer receives control commands generated by the decision layer and feeds back the blending operation data after execution according to the control commands to the perception layer.

[0097] Among them, the intelligent coal feeder unit adjusts the coal feed rate according to control commands to ensure that fuel supply matches demand;

[0098] The three-dimensional coal blending device precisely blends different types of coal according to the instructions of the dynamic coal blending model.

[0099] The burner co-controller adjusts burner parameters (such as air volume and temperature) to achieve efficient combustion and emission control.

[0100] The sensing layer includes a combustion field sensing network, an emission monitoring module, and an online coal quality detection array.

[0101] Among them, the combustion field perception network can monitor the combustion status (such as temperature distribution and flame shape) in real time.

[0102] The emission monitoring module can detect pollutant (such as NOx and SO2) emission data; the online coal quality detection array can analyze key parameters such as coal composition and calorific value in real time.

[0103] The perception layer collects data in real time through the combustion field perception network, emission monitoring module, and online coal quality detection array, and uploads it to the decision-making layer through real-time data stream.

[0104] The decision-making layer includes a dynamic coal blending model, a multi-objective optimization engine, and an adaptive compensation system.

[0105] Among them, the dynamic coal blending model generates the optimal coal blending ratio based on real-time coal quality data and operational feedback; the multi-objective optimization engine can balance objectives such as combustion efficiency, emission control, and cost, and output strategies for adjusting coal blending and combustion optimization; the adaptive compensation system can dynamically adjust compensation parameters according to real-time operating conditions (such as coal quality fluctuations and equipment status), and make real-time corrections to the strategies to ensure system robustness.

[0106] The results of the execution layer (such as combustion efficiency and emission levels) are returned to the perception layer through operational feedback, triggering a new round of optimization iterations to ensure continuous dynamic optimization of the system.

[0107] The process of the dynamic coal blending model is as follows: Figure 2 As shown, its specific implementation is as follows:

[0108] (1) Feature extraction: Real-time raw coal quality data (including 12 indicators such as calorific value, volatile matter, sulfur content, and ash fusion point) are obtained from the perception layer and input into the dynamic coal blending model. Then, physical features are extracted, and a deep feature encoder (3-layer CNN+GRU network) is used to extract temporal features, outputting a 32-dimensional normalized feature vector:

[0109] ;

[0110] (2) DRL agent decision-making: Input the 32-dimensional normalized feature vector into the DRL agent to make preliminary decisions.

[0111] The DRL agent employs the Proximal Policy Optimization (PPO) algorithm, and its state space, action space, and reward function are calculated using the following formulas:

[0112] State space:

[0113] ;

[0114] Where V represents a 32-dimensional normalized eigenvector; Q real Indicates the real-time heat load of the boiler (MW); η curr ENOx represents the current thermal efficiency (%); ENOx represents E NOx Emission concentration (mg / m³).

[0115] Action space:

[0116] ;

[0117] The proportion adjustment amount for the i-th type of coal represents the amount by which the proportion of the i-th type of coal needs to be increased or decreased relative to the current coal blending ratio; n is the number of coal types.

[0118] Reward function:

[0119] ;

[0120] Where η represents thermal efficiency, C coal Q represents the cost of coal. tar Q represents the target heat load. real This represents the real-time heat load. The reward function R encourages the DRL agent to improve thermal efficiency, reduce coal costs, reduce NOx emissions, and bring the real-time heat load closer to the target heat load.

[0121] (3) Physical constraint layer verification: The extracted physical features and the preliminary decision input of the DRL agent are verified by the physical constraint layer.

[0122] The constraint types are shown in Table 1 below:

[0123] Table 1

[0124]

[0125] Constraint handling:

[0126] Feasibility adjustments were made using the modified Lagrange multiplier method:

[0127] ;

[0128] in, Here, xi is the adjustment factor, and xi represents the equipment constraint capability.

[0129] (4) Hybrid decision generation: calculate the conflict degree between DRL decision and physical constraints, then dynamically adjust the weight allocation, and output the final coal blending scheme.

[0130] The formula for calculating the conflict degree δ is as follows:

[0131] ;

[0132] Among them, a DRL Let a be the decision vector given to the DRL agent. phy The decision vector to satisfy physical constraints.

[0133] The formula for calculating the dynamic weight allocation w is as follows:

[0134] ;

[0135] Where k is a constant used to adjust the sensitivity of weight allocation, that is, the degree of influence of conflict degree δ on weight w.

[0136] The final coal blending scheme calculation formula is as follows: .

[0137] In this embodiment, the DRL policy network parameters are updated every 24 hours.

[0138] Automatically update constraint boundaries when equipment is under maintenance or coal yard inventory changes:

[0139] ;

[0140] Where, N outage This indicates the number of coal feeders currently out of service.

[0141] See Figure 3 The multi-objective optimization engine consists of an input layer, a core algorithm layer, a decision layer, and an output layer.

[0142] The input layer includes an objective function, real-time operating data, and constraints. The objective function is a dual objective function of economic efficiency and environmental protection. The constraints are categorized into three types: boiler safety, equipment capacity, and environmental regulations. The real-time data includes coal quality parameters, load demand, and emission monitoring values.

[0143] The objective function is:

[0144] ;

[0145] Among them, c i x is the unit price of the i-th type of coal (yuan / ton); i The blending ratio of the i-th type of coal (%) ej η is the conversion factor (mg / m³ / %) for the j-th emission type; act / tar The actual / target thermal efficiency (%) is given; α, β, γ are weighting coefficients (α+β+γ=1).

[0146] The constraints are:

[0147] ;

[0148] Where, x i Let q be the blending ratio of the i-th type of coal. i The lower heating value (MJ / kg) of the i-th type of coal; s i The sulfur content (%) of the i-th type of coal; P min / max S represents the boiler's permissible heat load range (MW); max The upper limit of sulfur content (%) is the environmental standard.

[0149] The core algorithm layer includes improved NSGA-III, adaptive reference point generation, and dynamic population management. Improved NSGA-III is the main framework for multi-objective optimization. Adaptive reference points are used to dynamically adjust the distribution density of reference points. The purpose of dynamic population management is to retain historical elite solutions and accelerate convergence.

[0150] The formula for adaptive generation of reference points is:

[0151] ;

[0152] Where Δfk is the rate of change of the kth objective function.

[0153] The decision-making layer includes Pareto front extraction and fuzzy decision-making modules. The Pareto front is used to visualize the non-dominated solution set, and the fuzzy decision-making is used to comprehensively allocate weights and screen the solution set. The output layer includes the optimal solution set and control instructions. The optimal solution set is the set of feasible solutions that satisfy all constraints, and the control instructions refer to the specific coal feed ratio and combustion parameters.

[0154] See Figure 4 The implementation steps of the multi-objective optimization engine are as follows:

[0155] (1) Initialize the population; generate N random solutions, each solution is an m-dimensional vector (m = number of coal types); the encoding method adopts real number encoding to represent the blending ratio of each coal type.

[0156] (2) Non-dominated sorting; calculate the dominance relationship for each solution, and the hierarchical sorting formula is as follows: ;

[0157] Where I=1, y dominates x, otherwise 0;

[0158] (3) Generate adaptive reference points;

[0159] The formula for adjusting the density of the reference point is:

[0160] ;

[0161] Where k=5, nobj=2, and t is the number of iterations.

[0162] (4) Crossover variation;

[0163] Using SBX crossover and polynomial mutation, the calculation formula is as follows:

[0164] ;

[0165] Where, η c =20, η c This is the distribution index.

[0166] (5) Dynamic population update; an elite retention strategy is adopted, retaining the top 20% of historical optimal solutions; the calculation formula for the elite retention strategy is: P t+1 =Select(P t ∪Q t N).

[0167] (6) Fuzzy decision-making: The weights are determined by the entropy weight method, and then the optimal solution is selected from the Pareto solution set by TOPSIS sorting.

[0168] The formula for calculating the weights using the entropy weight method is as follows: ;

[0169] ;

[0170] The TPOSIS sort is calculated as follows:

[0171] ;

[0172] The ideal solution distance is:

[0173] ;

[0174] ;

[0175] Among them, w j E represents the target weighting coefficient. j p is the information entropy of the j-th index; ij is the proportion of the standardized value of the i-th sample under the j-th indicator; m is the total number of samples, and n is the total number of indicators; Let be the Euclidean distance from the i-th solution to the positive ideal solution (optimal solution); z is the Euclidean distance from the i-th solution to the negative ideal solution (worst solution); ij These are the standardized indicator values; Let j be the index value of the positive ideal solution; The j-th index value is the negative ideal solution.

[0176] The adaptive compensation system operates as follows:

[0177] (1) Data acquisition: Real-time reception of 200+ parameters such as coal quality, combustion, and equipment status (sampling rate 1Hz);

[0178] (2) Anomaly identification: Multi-dimensional anomaly detection is performed every 5 seconds (computation time <50ms);

[0179] (3) Compensation decision: Match the best compensation type from the policy library and generate specific parameters through the transfer learning model; The transfer learning model uses MMD (maximum mean difference) to minimize the distribution difference between domains to realize feature pairs; The model parameters are updated once every 100 new feedback data are received.

[0180] Its incremental learning formula is:

[0181] ;

[0182] in, To compensate for the loss of effect (such as thermal efficiency deviation); λ is the domain adaptation loss; λ=0.5 is the tradeoff coefficient.

[0183] (4) Security verification: if the verification passes, an instruction is issued; if it fails, a downgrade strategy is triggered.

[0184] (5) Effect feedback: Record the boiler efficiency and emission data after compensation for online updates of the transfer learning model.

[0185] In this embodiment, tasks are divided into three levels according to time scale: minute-level tasks, hour-level tasks, and day-level tasks.

[0186] Minute-level tasks include:

[0187] The coal feeding ratio is adjusted dynamically based on real-time coal quality testing data;

[0188] The combustion air-coal ratio is optimized, and the opening of the secondary air damper is adjusted through oxygen feedback (adjustment accuracy ±1%).

[0189] Abnormal operating condition compensation triggers emergency control for events such as sudden changes in coal quality and load fluctuations.

[0190] Hourly tasks include:

[0191] Coal type combination optimization: generating new coal blending schemes based on inventory, coal prices, and emission constraints;

[0192] Environmental parameters are recalculated, and the parameters of the desulfurization and denitrification system are dynamically adjusted based on current emission data.

[0193] Equipment condition assessment, monitoring the health status of key equipment such as coal mills and coal feeders.

[0194] Heavenly-level missions include:

[0195] The coal procurement plan has been updated, and procurement recommendations have been generated by combining market conditions and inventory forecasts.

[0196] Long-term inventory optimization balances coal yard storage costs with blending demand;

[0197] Environmental protection strategies are adjusted, and emission control thresholds are updated in accordance with changes in environmental policies.

[0198] The daily plan provides boundary conditions (such as upper and lower limits of inventory) for hourly optimization; the results of hourly optimization set target ranges for minute-level control; minute-level abnormal events trigger hourly strategy re-optimization (marked with red crit); and hourly equipment evaluation results affect the formulation of the daily procurement plan.

[0199] When coal quality fluctuations exceed 2%, minute-level adjustments are immediately initiated; when NOx exceeds the standard for 10 minutes, hour-level strategy reconstruction is triggered; when daily coal price changes exceed 5%, daily plan updates are executed ahead of schedule.

[0200] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-coal blending dynamic optimization system, characterized in that, The execution layer, the perception layer, and the decision layer are included; The execution layer executes specific operations of hybrid blending combustion, receives control instructions generated by the decision layer, and feeds back blending operation data after execution according to the control instructions to the perception layer; The perception layer collects data in real time through a combustion field perception network, an emission monitoring module, and a coal quality online detection array, and uploads the real-time data stream to the decision layer; The decision layer includes a dynamic coal blending model, a multi-objective optimization engine, and a self-adaptive compensation system; The dynamic coal blending model generates optimal coal blending ratios based on real-time coal quality data and operation feedback; the multi-objective optimization engine balances combustion efficiency, emission control, and cost targets, and outputs adjustment coal blending and combustion optimization strategies; The self-adaptive compensation system dynamically adjusts compensation parameters according to real-time working conditions and makes real-time corrections to the strategies; The operation results of the execution layer are returned to the perception layer through operation feedback, triggering a new round of optimization iteration; The specific implementation of the dynamic coal blending model is as follows: (1) Feature extraction: Obtain real-time raw coal quality data from the perception layer, input the real-time raw coal quality data into the dynamic coal blending model, then extract physical features, and use a deep feature encoder to extract time sequence features, output a 32-dimensional normalized feature vector: ； Where v1, v2,..., v32 refer to each feature value in a 32-dimensional feature vector; (2) DRL agent decision: input the 32-dimensional normalized feature vector into the DRL agent for preliminary decision; (3) Physical constraint layer verification: input the extracted physical features and the preliminary decision produced by the DRL agent into the physical constraint layer for verification; Use the modified Lagrange multiplier method for feasibility adjustment: ； wherein, is an adjustment factor, and xi is the equipment constraint capacity; (4) Hybrid decision generation, calculate the conflict degree δ of DRL decision and physical constraint, then dynamically adjust the weight w distribution, and output the final coal blending scheme; The calculation formula of the conflict degree δ is as follows: ； wherein a DRL is the decision vector given by the DRL agent, a phy is the decision vector satisfying the physical constraints; The dynamic weight w distribution calculation formula is as follows: ； Where k is a constant, used to adjust the sensitivity of weight distribution, that is, the influence degree of conflict degree δ on weight w; The final coal blending scheme calculation formula is: .

2. The multi-coal blending dynamic optimization system according to claim 1, wherein, The execution layer includes an intelligent coal feeder unit, a three-dimensional coal mixing device, and a burner collaborative controller; the intelligent coal feeder unit adjusts the coal supply according to the control instructions to ensure that the fuel supply matches the demand; the three-dimensional coal mixing device accurately mixes different coal types according to the instructions of the dynamic coal blending model; The burner collaborative controller adjusts the burner parameters to achieve efficient combustion and emission control.

3. The multi-coal blending dynamic optimization system according to claim 1, wherein The perception layer includes a combustion field perception network, an emission monitoring module, and a coal quality online detection array; the combustion field perception network can monitor the combustion state in real time; the emission monitoring module can detect pollutant emission data; the coal quality online detection array can analyze key parameters of coal types including composition and calorific value in real time.

4. The multi-coal blending dynamic optimization system of claim 1, wherein, The DRL agent uses the proximal policy optimization algorithm, and the calculation formulas of its state space, action space, and reward function are as follows: State space: ； wherein V represents a 32-dimensional normalized feature vector; Q real represents the real-time heat load of the boiler (MW); η curr represents the current thermal efficiency (%); ENOx represents the E NOx emission concentration (mg / m³); Action space: ； The proportional adjustment amount of the i-th coal indicates the amount by which the proportion of the i-th coal needs to be increased or decreased with respect to the current coal blending proportion; n is the number of coal types; Reward function: ； wherein η represents the thermal efficiency, C coal represents the coal cost, Q tar represents the target thermal load, Q real represents the real-time thermal load; the reward function R encourages the DRL agent to improve the thermal efficiency, reduce the coal cost, reduce the NOx emission, and make the real-time thermal load close to the target thermal load.

5. The multi-coal blending dynamic optimization system of claim 1, wherein, The DRL strategy network parameters are updated every 24 hours; when the equipment is overhauled or the coal yard inventory changes, the constraint boundary is automatically updated.

6. The multi-coal blending dynamic optimization system of claim 1, wherein, The multi-objective optimization engine comprises an input layer, a core algorithm layer, a decision layer and an output layer; The input layer comprises a target function, real-time working condition data and constraint conditions; the target function is a dual target function of economy and environmental protection, the constraint conditions are three types of constraints of boiler safety, equipment capacity and environmental protection regulations, and the real-time data comprises coal quality parameters, load demand and emission monitoring values; The target function is: ； Wherein, c i is the unit price of the i-th coal (Yuan / t); x i is the blending ratio of the i-th coal (%); e j is the conversion factor of the j-th emission (mg / m³ / %); η act / tar is the actual / target thermal efficiency (%); α, β, γ are weight coefficients; The constraint conditions are: ； Wherein, x i is the i-th coal blending ratio, q i is the i-th coal low heat value (MJ / kg); s i is the i-th coal sulfur content (%); P min / max is the boiler allowable heat load range (MW); S max is the environmental standard sulfur upper limit (%) The core algorithm layer comprises improved NSGA-III, adaptive reference point generation and dynamic population management; the improved NSGA-III is the main framework of multi-objective optimization, the adaptive reference point is used for dynamically adjusting the reference point distribution density, and the dynamic population management can retain historical elite solutions and accelerate convergence; The adaptive reference point generation formula is: ； Wherein, Δfk is the change rate of the kth target function; The decision layer comprises a Pareto frontier extraction and a fuzzy decision module; the Pareto frontier is a visualization of the non-dominated solution set, and the fuzzy decision is used for comprehensive weight distribution and solution set screening; the output layer comprises an optimal solution set and a control instruction, wherein the optimal solution set is a feasible scheme set satisfying all constraints, and the control instruction is a specific coal blending ratio and combustion parameter.

7. The multi-coal blending dynamic optimization system of claim 1, wherein, The implementation steps of the multi-objective optimization engine are as follows: (1) Initialize the population; generate N random solutions, each solution being an m-dimensional vector; the coding mode adopts real number coding, representing the blending ratio of each coal; (2) Non-dominated sorting; the dominance relation of each solution is calculated, and the hierarchical sorting formula is as follows: ; Wherein, I = 1, y dominates x, otherwise 0; (3) Generate adaptive reference points; The reference point density adjustment formula is: ； Wherein, k = 5, nobj = 2, t is the iteration number; (4) Cross mutation; SBX crossover and polynomial mutation are adopted, and the calculation formula is: ； where η c = 20, η c is the distribution exponent; (5) Dynamic population update; an elite preservation strategy is adopted to preserve the top 20% of historical optimal solutions; wherein, the calculation formula of the elite preservation strategy is: P t+1 = Select(P t ∪ Q t , N); (6) Fuzzy decision; the weight is determined by using the entropy weight method, and then the optimal solution is selected from the Pareto solution set by using TOPSIS sorting.

8. The multi-coal blending dynamic optimization system of claim 7, wherein, The calculation formula for determining the weight of the entropy weight method is: ​ ； The calculation of TPOSIS sorting is: ； The ideal solution distance is: ； ； where w j is the target weight coefficient, E j is the information entropy of the jth indicator; p ij is the proportion of the normalized value of the ith sample under the jth indicator; m is the total number of samples, and n is the total number of indicators; is the Euclidean distance from the ith scheme to the positive ideal solution; is the Euclidean distance from the ith scheme to the negative ideal solution; z ij is the normalized indicator value; is the jth indicator value of the positive ideal solution; is the jth indicator value of the negative ideal solution.

9. The multi-coal blending dynamic optimization system of claim 1, wherein, The working steps of the adaptive compensation system are as follows: (1) Data acquisition, real-time reception of coal quality, combustion and equipment state parameters; (2) Abnormality identification, multi-dimensional abnormality detection is performed every n seconds; (3) Compensation decision, the best compensation type is matched from the strategy library, and specific parameters are generated through a transfer learning model; the transfer learning model adopts MMD to minimize the distribution difference between fields to realize feature matching; the model parameters are updated once every n new feedback data is received; The incremental learning formula is: ； wherein, to compensate for loss of effect; to compensate for loss of domain adaptation. λ = 0.5, which is a weighting coefficient; (4) Safety verification, if the verification is passed, the instruction is issued, and if the verification fails, the degradation strategy is triggered; (5) Effect feedback, record the boiler efficiency and emission data after compensation, which is used for online updating of the transfer learning model.

Images