Industrial coal blending optimization method, system and program product based on multi-agent cooperation

Abstract

The application provides an industrial coal blending optimization method and system based on multi-agent cooperation and a program product, and relates to the technical field of industrial intelligent control and data processing. The method comprises the following steps: obtaining production state data of a current time step, wherein the data covers coal inventory, physicochemical indexes and raw material resource weight coefficients; inputting the data into a mechanism calculation module, using a linear programming algorithm to calculate a reference blending ratio vector with the minimum comprehensive resource consumption index as the target; simultaneously inputting the data into a strategy correction agent, outputting a correction action vector for the reference blending ratio and a confidence factor of the current working condition; and dynamically weighting and fusing the reference blending ratio and the correction action. The double-layer coupling architecture of mechanism guarantee and AI efficiency improvement is adopted, and the confidence factor is used as a safety valve to ensure the production safety, solve the problems of poor nonlinear adaptability and feedback lag of the traditional coal blending model, and realize the global optimization of industrial raw material configuration.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control and data processing technology, and in particular to an industrial coal blending optimization method, system and program product based on multi-agent collaboration. Background Technology

[0002] Industrial coal blending is a crucial step in industrial production such as thermal power and coking. Its core lies in combining various coal types with different sources and physical and chemical properties in a specific ratio, while meeting the quality constraints of the target product, in order to achieve a balance between production stability and resource utilization efficiency.

[0003] However, in actual production, coal quality is significantly affected by differences in mining locations, storage conditions, and changes in operating conditions. The coal blending process is characterized by multiple constraints, nonlinearity, and time-varying features. Existing coal blending schemes mostly rely on manual experience or optimization methods based on mechanistic models. Manual experience methods are highly subjective and have poor stability; mechanistic model methods are usually based on linear or weakly nonlinear assumptions, making it difficult to adapt to fluctuations in coal quality, parameter drift, and complex operating conditions, causing the optimization results to easily deviate from the target in actual operation.

[0004] Therefore, we propose an industrial coal blending optimization method, system, and program product based on multi-agent collaboration; the information disclosed in the background section is only for enhancing the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an industrial coal blending optimization method, system, and program product based on multi-agent collaboration, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The industrial coal blending optimization method based on multi-agent cooperation includes the following steps:

[0008] Step S1: Obtain the production status data for the current time step. The production status data includes the inventory of each type of coal, the physical and chemical indicators of raw coal, the weight coefficient of raw material resources, the quality constraints of the target product, and environmental characteristic data.

[0009] Step S2: Input the production status data into the preset mechanism calculation module, and use a linear programming algorithm with the goal of minimizing the comprehensive resource consumption index to calculate the benchmark ratio vector that satisfies the target product quality constraints. ;

[0010] Step S3: Input the production status data into the pre-trained policy correction agent, and output the correction action vector for the benchmark ratio vector. and the confidence factor of the current operating condition ;

[0011] Step S4: Based on the benchmark ratio vector Correcting motion vectors and confidence factor Perform weighted fusion to generate the final allocation vector. The final proportion vector is then sent to the coal blending actuator.

[0012] Step S5: Obtain the actual output quality data after the lag period, calculate the reward value based on the actual output quality data, and update the policy correction agent.

[0013] The policy correction agent includes an input layer, a mask layer, and an attention mechanism layer;

[0014] In steps S1 and S3, when there are missing or outlier values ​​in the raw coal physicochemical indicators of the production status data, the mask layer sets the corresponding feature value to zero, and the attention mechanism layer reallocates the weights of the remaining valid data to generate the correction action vector. .

[0015] In step S4, the final ratio vector is generated. The calculation formula is:

[0016] ;

[0017] Wherein, Normalize is a normalization function used to ensure that the total proportions are 100%; the confidence factor The value range is [0,1]; when the fluctuation of the production status data exceeds the preset threshold or a sensor malfunctions, the strategy correction agent outputs a value close to 0. This causes the system to degenerate into a control mode dominated by the aforementioned mechanism calculation module.

[0018] In step S5, the formula for calculating the reward value includes a quality penalty term, a consumption optimization term, and a stable operation term: ;

[0019] The consumption optimization item The difference between the baseline consumption index and the final consumption index is normalized using the hyperbolic tangent function.

[0020] The quality penalty item A second-order squared penalty is applied when the actual output quality data exceeds the target product quality constraint.

[0021] The stable operation item The negative feedback value is calculated based on the Euclidean distance between the final allocation vector at the current time step and the final allocation vector at the previous time step.

[0022] In step S5, obtaining the actual output quality data after the lag period specifically includes:

[0023] Establish a time-aligned data buffer; at time t, store the production status data and correction action vector into the data buffer; after obtaining the actual output quality data at time t+T, retrieve the data corresponding to time t from the data buffer, and calculate the reward value and update the parameters.

[0024] The mechanism calculation module contains a matrix of technical indicators used to characterize the contribution rate of each raw coal to the blended coal index.

[0025] The method further includes a mechanism model adaptive update step: calculating the residual between the actual output quality data and the quality estimated by the mechanism calculation module based on the current parameters;

[0026] Based on the residual, the gradient descent method is used to reverse-correct the technical indicator matrix inside the mechanism calculation module, and the nominal values ​​in the technical indicator matrix are updated to the effective values ​​under operating conditions.

[0027] In step S3, the corrected action vector Each component is constrained by a maximum allowable adjustment range to limit the degree of deviation of the policy-correcting agent from the baseline proportion vector.

[0028] The industrial coal blending optimization system based on multi-agent cooperation adopts the above-mentioned industrial coal blending optimization method based on multi-agent cooperation, including:

[0029] The data acquisition module is used to acquire production status data at the current time step. The production status data includes inventory of each type of coal, physical and chemical indicators of raw coal, weight coefficient of raw material resources, quality constraints of target products, and environmental characteristic data.

[0030] The mechanism calculation module is used to run a linear programming algorithm, with the objective function of minimizing the comprehensive resource consumption index, and calculates the baseline allocation vector based on the production status data. ;

[0031] The strategy correction module is used to run a strategy correction agent based on a deep neural network, and output correction action vectors based on the production status data. and confidence factor ;

[0032] The decision fusion module is used to calculate based on the formula. Calculate the final mix ratio vector and control the coal feeder to execute it through the control interface;

[0033] The feedback update module is used to collect actual output quality data after the lag period, calculate the comprehensive reward value to update the strategy correction module, and calculate the quality prediction residual to update the internal parameters of the mechanism calculation module.

[0034] A computer program product, comprising computer program code, which, when run on a computer, enables the computer to implement the aforementioned industrial coal blending optimization method based on multi-agent collaboration.

[0035] The beneficial effects of this invention are as follows:

[0036] This invention employs linear programming to solve the problem through a mechanism calculation module. Under the premise of satisfying process boundaries such as target product quality constraints and chemical composition conservation, it outputs a benchmark ratio, giving the coal blending scheme an inherent "compliant and controllable" safety foundation. It introduces a strategy correction agent based on deep neural networks to dynamically fine-tune the ratio. When the data is complete and the operating conditions are stable, it enters a normal mode to improve the optimization effect under complex nonlinear operating conditions, thereby improving resource allocation efficiency and overall optimization level.

[0037] This invention introduces a confidence factor and a safety fallback mode in the decision fusion stage. When drastic fluctuations in production status data or sensor malfunctions are detected, the agent outputs a correction close to zero, causing the system to degenerate into a control mode dominated by mechanism calculation. This ensures that production safety and quality baselines can still be maintained under abnormal operating conditions. A mask layer and an attention mechanism layer are set at the input end of the strategy correction agent. When online analyzer data is missing or abnormal, the missing channel is set to zero and the effective feature weights are dynamically reallocated. This still generates an effective correction action vector, significantly improving the robustness of the system under conditions of incomplete data in the industrial field.

[0038] On the one hand, this invention sets a maximum allowable adjustment range constraint on the correction action vector to avoid system oscillation caused by agent decision-making; on the other hand, it introduces a stable operation term into the reward value to apply negative feedback to frequent changes in the ratio, thereby reducing wear on actuators such as coal feeders and improving operational stability.

[0039] This invention addresses the issue of a lag between coal blending and output quality detection. It utilizes a time alignment buffer mechanism to backtrack and match historical state data with output quality data, calculates a comprehensive reward value, and updates the agent's strategy. Simultaneously, based on the residual between actual output quality and mechanistic prediction quality, it uses gradient descent to reverse-calibrate the technical indicator matrix within the mechanistic model, causing the mechanistic model parameters to gradually converge from nominal values ​​to effective values ​​under operating conditions, thereby achieving closed-loop adaptive and long-term stable optimization. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the industrial coal blending optimization method based on multi-agent collaboration of the present invention;

[0041] Figure 2 This is a schematic diagram of the industrial coal blending optimization system framework based on multi-agent collaboration of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0043] Example 1: As Figure 1 As shown, this embodiment provides an industrial coal blending optimization method based on multi-agent collaboration. Its core lies in using a "mechanism model" to ensure physical boundary safety and using "agents" to optimize resource allocation efficiency. The specific steps are as follows:

[0044] Step S1: Construction and preprocessing of production status data:

[0045] The system constructs a production status data vector for the current time step t at a set period (e.g., every 15 minutes). :

[0046] ;

[0047] in:

[0048] : Current physical inventory levels of each coal bunker;

[0049] : Physicochemical index matrix of each raw coal (including moisture, ash, sulfur, etc.);

[0050] (Raw Material Resource Weighting Coefficient): This is a key control parameter introduced in this embodiment. In this system, it is defined as the resource consumption cost per unit mass of raw material. According to preset rules, the system assigns a higher weighting coefficient to high-grade or scarce coal types and a lower weighting coefficient to low-grade or surplus coal types.

[0051] Raw material resource weighting coefficient The calculation formula is: ;in is the energy value coefficient of the i-th type of coal (usually taken as the ratio of the received lower heating value of the coal to the calorific value of standard coal, which characterizes its grade). This is the inventory scarcity index for this type of coal (the lower the inventory, the lower the scarcity index). The higher the value, the range is [0,1]. This is a scarcity penalty factor (a constant, such as 0.5). Using this formula, the system automatically suppresses the overuse of high-grade coal with low inventory levels.

[0052] Physical constraints on the quality of the target product (such as blended coal);

[0053] Environmental characteristic data (such as humidity and temperature).

[0054] Furthermore, regarding the handling of missing sensor data:

[0055] This embodiment incorporates a masking layer at the input of the neural network. When a signal loss is detected in the online analyzer (reading is NaN or 0), the masking layer generates a corresponding mask vector to set that feature channel to zero. The subsequent attention layer dynamically calculates the correlation weights between the remaining valid features, refocusing on reliable data to generate an effective correction action vector. This mechanism ensures the system's robustness in the event of partial sensing failure.

[0056] Step S2: Baseline calculation based on minimum resource cost:

[0057] The mechanism calculation module runs a linear programming algorithm. Its objective function is set as "minimizing the comprehensive resource consumption index":

[0058] ;

[0059] in For the proportions, Let be the resource weighting coefficient of the i-th raw material, with the constraint that chemical composition is conserved:

[0060] ;

[0061] in, This represents the value of the j-th physicochemical index (such as sulfur or ash content) of the i-th type of raw coal. This represents the maximum allowable constraint value for the j-th physicochemical index of the target product. Solving for this yields the baseline mix ratio vector. This step ensures the chemical compliance of the formulation.

[0062] Step S3, Dynamic Correction Decision of the Agent: The policy correction module makes decisions based on the input... Output two key quantities:

[0063] Correcting motion vectors : Each component This represents the direction of fine-tuning the baseline ratio. To prevent system oscillations caused by AI decision-making, this embodiment sets a maximum allowable adjustment range. (e.g., ±2%), i.e., limit ;

[0064] Confidence factor Value range [0,1]. This factor characterizes the agent's understanding of the current operating conditions.

[0065] Specifically, the policy correction agent adopts a structure combining a multilayer perceptron (MLP) and a self-attention mechanism. The number of nodes in the input layer is consistent with the dimension of the production state data (e.g., 32-dimensional); after the masking layer, a multi-head attention layer is connected with 4 heads to capture the coupling relationship between different coal quality indicators; then, three fully connected layers are connected with 128, 64, and 32 nodes respectively, using ReLU as the activation function. The output layer uses the Tanh activation function to map the action to the [-1,1] interval, and finally multiplies by the maximum allowable adjustment range. .

[0066] Step S4, Dual-modal fusion and execution: The decision fusion module generates the final matching vector according to the formula. :

[0067] ;

[0068] Normal mode ( When the data is complete and the operating conditions are stable, the system adopts the agent's correction suggestions to explore the potential of nonlinear optimization.

[0069] Safe rollback mode ( When detected When there are severe fluctuations or widespread sensor failures, the agent automatically outputs a value close to 0. , making Degenerate into This is equivalent to the system automatically switching back to pure mechanism control to ensure production safety.

[0070] Where Normalize is the normalization function used to ensure that the total proportions are 100%; confidence factor The value range is [0,1]; when the fluctuation of the production status data exceeds the preset threshold or a sensor malfunction is detected, the strategy correction agent outputs a value close to 0. This causes the system to degenerate into a control mode dominated by the aforementioned mechanism calculation module;

[0071] Specifically, the processing logic of the Normalize function is as follows: First, ... The negative values ​​in the calculated intermediate vector are set to 0 (ReLU processing). Then, all components in the vector are summed to obtain the total sum Sum. Finally, each component in the vector is divided by Sum to ensure that each component in the final output proportion vector is non-negative and sums to 1.

[0072] Step S5: Parameter update based on lag feedback: Due to the time lag T (e.g., 8 hours) between raw material input and output quality detection, the system uses a buffer to align data and calculate the comprehensive reward value. :

[0073] ;

[0074] Consumption optimization items : The Index is the resource consumption index (i.e., the weighted sum of the allocation ratio and resource weight). This term is used to reward the system for successfully reducing the total resource consumption cost of the system without compromising quality.

[0075] Quality penalty items When the actual output quality exceeds the constraint range, a second-order squared penalty is applied. This forces the model to strictly adhere to quality standards; among them The actual output quality data obtained in step S5 after the lag period. To constrain the quality of the target product.

[0076] In a preferred embodiment, the penalty coefficient The value range is [10, 50], depending on the stringency of the target product quality constraints. For example, for the ash content index, when the actual value exceeds the constraint by 0.1%, set... It can generate sufficient gradient signals to correct the action.

[0077] Stable operation item :calculate It provides negative feedback to the frequent changes in the ratio, so as to reduce the wear of the coal feeder.

[0078] Furthermore, the mechanistic model features adaptive calibration: the system possesses self-evolution capabilities. The mechanistic calculation module internally maintains a "technical index matrix" (which defines the theoretical contribution rate of each raw coal to the final quality).

[0079] In actual production, theoretical values ​​often deviate from actual values ​​due to coal oxidation or changes in mine locations. The system periodically performs the following calibration: calculating the residual between the actual output quality and the quality predicted by the mechanism model. The coefficients in the "technical indicator matrix" are updated in reverse using the gradient descent algorithm.

[0080] For example, if the theoretical sulfur contribution of a certain type of coal is 1.0%, but actual feedback shows that it leads to higher sulfur residue, the system will automatically correct this parameter to the "operating condition effective value" (e.g., 1.1%). This process allows the mechanistic model to dynamically adapt to real changes in the physical world, ensuring a continuous supply of accurate values. .

[0081] Example 2: Figure 2 As shown, this embodiment provides an industrial coal blending optimization system based on multi-agent collaboration. This system is mainly applied to the raw material pretreatment stage of thermal power plants or coking plants, and mainly includes the following functional modules:

[0082] Data Acquisition Module: Responsible for acquiring real-time physical status data of the production site. This module connects to the DCS and LIMS systems via an industrial bus (such as OPC UA) to collect data including coal bunker level (inventory), online elemental analyzer data (physicochemical indicators of raw coal), and raw material resource weight coefficients (i.e., numerical weights characterizing the ease or scarcity of obtaining each type of coal).

[0083] Mechanism Calculation Module: Includes a built-in linear programming solver. This module stores chemical mechanism models for coal blending, used to calculate basic proportioning schemes while meeting process limits.

[0084] Strategy correction module: Equipped with a pre-trained deep neural network. This module acts as an agent, responsible for dynamically fine-tuning the basic scheme based on nonlinear operating conditions.

[0085] Decision fusion module: responsible for executing the "baseline + correction" fusion algorithm, generating the final coal feeder control instructions, and issuing them for execution through the PLC interface.

[0086] Feedback update module: Used to collect output quality data in the lag period and build a closed-loop control loop.

[0087] The lag period T is determined based on the conveyor belt length and sampling analysis time, typically ranging from 4 to 8 hours. The data buffer employs a first-in, first-out (FIFO) queue structure, with a capacity set to... ,in To control the step size (e.g., 15 minutes).

[0088] Example 3: This example provides a computer program product, which includes computer program code. When the computer program code is run on a computer, the computer implements the industrial coal blending optimization method based on multi-agent cooperation as described in Example 1.

[0089] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0090] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0091] In conclusion, 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An industrial coal blending optimization method based on multi-agent collaboration, characterized in that, Includes the following steps: Step S1: Obtain the production status data for the current time step. The production status data includes the inventory of each type of coal, the physical and chemical indicators of raw coal, the weight coefficient of raw material resources, the quality constraints of the target product, and environmental characteristic data. Step S2: Input the production status data into the preset mechanism calculation module, and use a linear programming algorithm with the goal of minimizing the comprehensive resource consumption index to calculate the benchmark ratio vector that satisfies the target product quality constraints. ; Step S3: Input the production status data into the pre-trained policy correction agent and output the correction action vector for the benchmark ratio vector. and the confidence factor of the current operating condition ; Step S4: Based on the benchmark ratio vector Correcting motion vectors and confidence factor Perform weighted fusion to generate the final allocation vector. The final proportion vector is then sent to the coal blending actuator. Step S5: Obtain the actual output quality data after the lag period, calculate the reward value based on the actual output quality data, and update the policy correction agent.

2. The industrial coal blending optimization method based on multi-agent collaboration according to claim 1, characterized in that, The policy correction agent includes an input layer, a mask layer, and an attention mechanism layer; In steps S1 and S3, when there are missing or outlier values ​​in the raw coal physicochemical indicators of the production status data, the mask layer sets the corresponding feature value to zero, and the attention mechanism layer reallocates the weights of the remaining valid data to generate the correction action vector. .

3. The industrial coal blending optimization method based on multi-agent collaboration according to claim 1, characterized in that, In step S4, the final ratio vector is generated. The calculation formula is: ； Wherein, Normalize is a normalization function used to ensure that the total proportions are 100%; the confidence factor The value range is [0,1]; when the fluctuation of the production status data exceeds the preset threshold or a sensor malfunctions, the strategy correction agent outputs a value close to 0. This causes the system to degenerate into a control mode dominated by the aforementioned mechanism calculation module.

4. The industrial coal blending optimization method based on multi-agent collaboration according to claim 1, characterized in that, In step S5, the formula for calculating the reward value includes a quality penalty term, a consumption optimization term, and a stable operation term: ; The consumption optimization item The difference between the baseline consumption index and the final consumption index is normalized using the hyperbolic tangent function. The quality penalty item A second-order squared penalty is applied when the actual output quality data exceeds the target product quality constraint. The stable operation item The negative feedback value is calculated based on the Euclidean distance between the final allocation vector at the current time step and the final allocation vector at the previous time step.

5. The industrial coal blending optimization method based on multi-agent collaboration according to claim 1, characterized in that, In step S5, obtaining the actual output quality data after the lag period specifically includes: Establish a time-aligned data buffer; at time t, store the production status data and correction action vector into the data buffer; after obtaining the actual output quality data at time t+T, retrieve the data corresponding to time t from the data buffer, and calculate the reward value and update the parameters.

6. The industrial coal blending optimization method based on multi-agent collaboration according to claim 1, characterized in that, The mechanism calculation module contains a matrix of technical indicators used to characterize the contribution rate of each raw coal to the blended coal index. The method further includes a mechanism model adaptive update step: calculating the residual between the actual output quality data and the quality estimated by the mechanism calculation module based on the current parameters; Based on the residual, the gradient descent method is used to reverse-correct the technical indicator matrix inside the mechanism calculation module, and the nominal values ​​in the technical indicator matrix are updated to the effective values ​​under operating conditions.

7. The industrial coal blending optimization method based on multi-agent collaboration according to claim 1, characterized in that, In step S3, the corrected action vector Each component is constrained by a maximum allowable adjustment range to limit the degree of deviation of the policy-correcting agent from the baseline proportion vector.

8. An industrial coal blending optimization system based on multi-agent collaboration, employing the industrial coal blending optimization method based on multi-agent collaboration as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire production status data at the current time step. The production status data includes inventory of each type of coal, physical and chemical indicators of raw coal, weight coefficient of raw material resources, quality constraints of target products, and environmental characteristic data. The mechanism calculation module is used to run a linear programming algorithm, with the objective function of minimizing the comprehensive resource consumption index, and calculates the baseline allocation vector based on the production status data. ; The strategy correction module is used to run a strategy correction agent based on a deep neural network, and output correction action vectors based on the production status data. and confidence factor ; The decision fusion module is used to calculate based on the formula. Calculate the final mix ratio vector and control the coal feeder to execute it through the control interface; The feedback update module is used to collect actual output quality data after the lag period, calculate the comprehensive reward value to update the strategy correction module, and calculate the quality prediction residual to update the internal parameters of the mechanism calculation module.

9. A computer program product, characterized in that, The computer program product includes computer program code, which, when run on a computer, enables the computer to implement the industrial coal blending optimization method based on multi-agent collaboration as described in any one of claims 1 to 7.

Images