A coking co-production multi-objective optimization method and system based on a global optimization strategy

Abstract

The application discloses a coking co-production multi-objective optimization method and system based on a global optimization strategy, and the multi-objective optimization method first constructs a historical data set and trains a coking co-production prediction model; then a global multi-objective optimization engine is constructed, the global multi-objective optimization engine predicts the coke yield and the methanol yield based on the trained coking co-production prediction model, a multi-objective optimization function with a constraint condition is used in the multi-objective optimization model to perform multi-objective optimization, and a Pareto optimal solution set is obtained, the Pareto optimal solution set corresponds to control parameters of all sections of the coking co-production, the control parameters of all sections of the coking co-production constitute an optimal production decision scheme, and the control system of all sections of the coking co-production performs the coordinated automatic production of the multiple sections in the coking co-production according to the optimal production decision scheme. The application solves the problem that the overall benefit is limited due to local optimization in traditional coking co-production, and significantly improves the overall operation efficiency, economic benefit and environmental protection level of the coking co-production system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coking cogeneration, in particular to a coking cogeneration multi-objective optimization method and system based on a global optimization strategy. BACKGROUND

[0002] As an important basic industry of heavy industry, in recent years, coking enterprises have started to build optimization systems to optimize process operation, production management and business management in response to cost control, efficiency improvement, resource utilization and environmental compliance.

[0003] At the same time, large-scale coking enterprises are developing towards group operation and lean management, gradually forming a new production mode of coking cogeneration (as shown in Figure 1 The production process shoulders the mission of converting coal into coke, coal gas and methanol and other high value-added chemical products in order.

[0004] Under this background, the managers of coking enterprises are trapped in a new dilemma of difficult organization and coordination of production, and also face the challenge of the need to simultaneously globally optimize key production variables distributed in different sections.

[0005] The optimization strategy of the existing coking automatic production system is based on a local optimization strategy, focusing on single-section variable optimization, ignoring the dynamic coupling relationship between multiple sections, and unable to be applied to the global optimization scene of multi-section coordinated production. And the traditional optimization strategy is usually a single optimization target, but in actual production, coking cogeneration is not only a global optimization problem involving multi-section coordination, but also a multi-objective optimization problem that needs to consider multiple different optimization targets. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a coking cogeneration multi-objective optimization method and system based on a global optimization strategy, to solve the problem of limited overall efficiency in traditional coking cogeneration caused by local optimization, and to significantly improve the overall operation efficiency, economic benefit and environmental protection level of the coking cogeneration system.

[0007] The technical solution of the present application is:

[0008] A coking cogeneration multi-objective optimization method based on a global optimization strategy, specifically comprising the following steps:

[0009] (1) Collecting historical production variable parameters and production indexes of all sections in coking cogeneration, and then performing data preprocessing to obtain a historical data set; the production indexes include coke yield, coal gas yield and methanol yield;

[0010] (2) Building a coking cogeneration prediction model, and training the coking cogeneration prediction model using the historical data set;

[0011] (3) constructing a global multi-objective optimization engine, the global multi-objective optimization engine comprising a trained coking co-production prediction model and a multi-objective optimization model, the global multi-objective optimization engine predicting the coke production and the methanol production based on the trained coking co-production prediction model, the multi-objective optimization model performing multi-objective optimization by using a multi-objective optimization function with constraints to obtain a Pareto optimal solution set, the Pareto optimal solution set corresponding to control parameters of all sections of the coking co-production, the control parameters of all sections of the coking co-production forming an optimal production decision scheme, and a control system of all sections of the coking co-production performing collaborative automatic production of multiple sections in the coking co-production according to the optimal production decision scheme;

[0012] The multi-objective optimization function with constraints is specifically shown in the following formula (1):

[0013] (1);

[0014] In formula (1), represents the multi-objective optimization function, that is, simultaneously satisfying the maximum coke production, the maximum methanol production and the minimum coal blending raw material cost; represents the coke production per unit time, represents the methanol production per unit time, represents the coal blending raw material cost per unit production; represents the constraint condition of the multi-objective optimization function, represents the minimum quality requirement of coke, represents the process quality requirement of coke; represents the coal gas production per unit time, represents a function that the coal gas production determines the methanol production, represents the flue gas side temperature of the coke oven section, represents the flue coke side temperature of the coke oven section, represents the gas collecting pipe pressure of the coke oven section, including the input end pressure and the output end pressure of the gas collecting pipe, represents the coking time of the coke oven section, represents a function determined by four variable parameters determining the coal gas production; represents the minimum value of the variable parameter, represents the maximum value of the variable parameter, represents that the values of the four variable parameters are constrained.

[0015] The global multi-objective optimization engine is based on real-time production data of coking co-production to perform multi-objective dynamic optimization. The real-time production data of coking co-production includes multiple variable parameters. After the feature values ​​of multiple variable parameters are normalized, data verification is performed. After data verification, all verified and retained variable parameters are uploaded to the global multi-objective optimization engine for multi-objective dynamic optimization.

[0016] The data verification specifically includes the following steps:

[0017] S11. Set the weights of all variable parameters in the real-time production data. And set activation thresholds for all variable parameters. ;

[0018] S12. Calculate the activation score for each variable parameter. , see the following formula (2) for details:

[0019] (2);

[0020] In equation (2), The eigenvalues ​​representing the i-th variable parameter, This represents the eigenvalue of the i-th variable parameter after normalization.

[0021] S13, when If the i-th variable parameter is selected, it is retained; otherwise, it is removed.

[0022] S14. Summarize all retained variable parameters and upload them to the global multi-objective optimization engine for multi-objective dynamic optimization.

[0023] The real-time production data for the coking and co-production process includes variable parameters for four stages, specifically:

[0024] For the coal blending section: moisture content, volatile matter, sulfur content, ash content, caking index, and number of coal charges per hole;

[0025] For the coke oven section: coke oven gas temperature before preheating, coke oven gas temperature after preheating, pressure at the input end of the gas collecting pipe, pressure at the output end of the gas collecting pipe, suction force on the fan side of the flue, suction force on the coke side of the flue, temperature on the fan side of the flue, temperature on the coke side of the flue, and coking time.

[0026] For the gas section: primary cooler resistance, average gas temperature after primary cooler, electrostatic precipitator resistance, gas pressure after mist eliminator, average blower oil temperature, average blower oil pressure, average gas pressure after desulfurization tower, gas temperature after benzene washing tower, gas pressure after benzene washing tower, and lean oil temperature of benzene washing tower.

[0027] For the methanol section: gas consumption, R104 catalyst bed temperature, compressor low-pressure cylinder inlet gas flow rate, compressor circulation section inlet flow rate, converter inlet coke gas temperature, converter coke oven gas and outlet converter gas pressure difference, converter outlet converter gas temperature, RT synthesis tower inlet temperature, and RP synthesis tower pressure difference.

[0028] The coking cogeneration prediction model includes a data preprocessing module, a PDM module, and an FMM module;

[0029] The data preprocessing module aligns the variable parameters of all work segments for each data sample according to the sampling time axis, forming a unified multivariate input matrix. In other words, each data sample, after processing by the data preprocessing module, outputs a time-series data tensor with multiple work segments, multiple variable parameters, and different time steps. , , Represents the time dimension, i.e., the sequence length. Total number of representative work sections , The total number of variable parameters, time series data tensor The elements in are denoted as , Represents time step Time, Section under the variable parameters The value;

[0030] The PDM module includes an L-layer structure, with the input of the first layer being a time-series data tensor. The input for each layer is , Represents the current layer. This represents the output of the previous layer structure. Total number of representative work sections Representing the The temporal characteristics of the work section, the output of each layer of the structure is The output of the last layer structure This is the output of the PDM module;

[0031] The input of the FMM module is The FMM module utilizes an independent predictor for each work section. Predicting the historical characteristics of this work section yields future predictions. Predictor As a linear layer, directly from a length of Historical features mapped to a length of The future timeline, subsequently mapped to The predictions for all work sections are weighted and summed to form the final prediction, which is based on several variable dimensions. , the output of is the future production index of coking cogeneration predicted by the coking cogeneration prediction model.

[0032] The processing procedure of each layer structure of the PDM module is as follows:

[0033] a. Extracting seasonal components and mixing:

[0034] Firstly, starting from the residual, see the following formula (3):

[0035] (3);

[0036] In formula (3), represents the initial residual, represents the initial trend component;

[0037] Then, the residual is periodically averaged according to the minimum period to obtain the seasonal component of the first layer coarse period , see the following formula (4):

[0038] (4);

[0039] In formula (4), , represents the number of periods of the first layer coarse period;

[0040] Then, multi-scale seasonal layering is performed, and the period set is defined: For each coarse period , the residual is calculated, see the following formula (5):

[0041] (5);

[0042] In formula (5), represents the residual of the m-1 layer coarse period, represents the residual of the m-2 layer coarse period, represents the seasonal component of the m-1 layer coarse period;

[0043] Subsequently, the seasonal component of the m layer coarse period is calculated, see the following formula (6):

[0044] (6);

[0045] In formula (6), , represents the number of periods of the m layer coarse period;

[0046] Finally, the mixed seasonal component output by the first layer structure of the PDM module is calculated , specifically as follows:

[0047] (7) ;

[0048] In formula (7), represents the m-th layer, The layer structure satisfies the normalization constraint for the learnable weight coefficient of the seasonal component of the m-th layer coarse period of section s and variable parameter v: ;

[0049] b. Extract the trend component and mix:

[0050] First, initialize the trend component, specifically as follows:

[0051] (8) ;

[0052] In formula (8), represents the moving average operation, and the initial window size is ;

[0053] Then, perform layer-by-layer trend refinement of H layers, specifically as follows:

[0054] (9) ;

[0055] In formula (9), , , represents the trend component of the h-th layer, represents the trend component of the h-1-th layer, represents the window size of the h-th layer; represents the learnable weight coefficient;

[0056] Finally, calculate the mixed trend component of the PDM module of the m-th layer output by the H-th layer structure , specifically as follows:

[0057] (10) ;

[0058] In formula (10), represents the trend component of the H-th layer, which is the mixed trend component of the m-th layer output by the PDM module of the H-th layer structure ;

[0059] c. Cross-section coupling:

[0060] Regarding the section as a graph node, the set of section nodes is , an adjacency matrix is constructed, and the elements in the adjacency matrix ​ This indicates the intensity of the influence of the i-th section on the j-th section;

[0061] Then, graph convolution is used to fuse features across work sections, as shown in equation (11) below:

[0062] (11);

[0063] In equation (11), It is an adjacency matrix with self-loops. Represents the identity matrix; represent The degree matrix, the diagonal elements of the degree matrix ; Represents the PDM module number The output of the layer structure; Represents the PDM module number The weight matrix of the layer structure, Represents the PDM module number The dimensionality of the input temporal features in the layered structure. Represents the weight matrix The dimensionality of the output features after transformation; Represents a non-linear activation function; Represents the PDM module number The results of cross-section integration of the layered structure;

[0064] Finally, the PDM module... The results of cross-section integration of layered structures With mixed seasonal portions Mixed trend components The PDM module is then merged to obtain the first... Layered output , see the following formula (12) for details:

[0065] (12);

[0066] In equation (12), This represents a splicing operation.

[0067] The multi-objective optimization model is solved using the non-dominated sorting genetic algorithm III. That is, the non-dominated sorting genetic algorithm III is used to embed the real-time production data of coking co-production into a multi-objective optimization function with constraints and solve it to obtain a Pareto optimal solution set, in which any one solution is the optimal solution, thus forming a set of optimal production decision schemes.

[0068] A coking co-production multi-objective optimization system based on a global optimization strategy comprises a data acquisition and verification module, a global multi-objective optimization engine, a collaborative production decision module and a multi-section collaborative control module.

[0069] The data acquisition and verification module is used for acquiring real-time production data of the coking co-production, performing data verification on characteristic values of multiple variable parameters in the real-time production data after normalization processing, and uploading all verified and retained variable parameters to the global multi-objective optimization engine.

[0070] The global multi-objective optimization engine comprises a trained coking co-production prediction model and a multi-objective optimization model, is used for predicting the coke yield and the methanol yield based on the trained coking co-production prediction model, performing multi-objective optimization by using a multi-objective optimization function with a constraint condition, and obtaining a Pareto optimal solution set.

[0071] The collaborative production decision module comprises control parameters of all sections of the coking co-production corresponding to the Pareto optimal solution set, and the control parameters of all sections of the coking co-production form an optimal production decision scheme.

[0072] The multi-section collaborative control module is used for collaborative automatic production of multiple sections in the coking co-production according to the optimal production decision scheme, and re-feedbacks real-time production data to the data acquisition and verification module.

[0073] Advantages of the present application:

[0074] (1) The real-time production data of the coking co-production is uploaded to the global multi-objective optimization engine for multi-objective dynamic optimization after data verification, thereby reducing data transmission amount, reducing the load of the global multi-objective optimization engine, and improving the efficiency of prediction and the overall efficiency of evolutionary calculation.

[0075] (2) The coking co-production prediction model constructed in the present application is a PDM module and a FMM module based on a TimeMixer model, utilizes multi-scale observation to decouple seasonal and trend changes, integrates the advantages of multiple predictors, and jointly predicts the yield and quality indicators of each section, thereby significantly improving the overall prediction accuracy and response speed.

[0076] (3) The PDM module of the present application gradually accumulates and mixes the seasonal components of all variable parameters of all sections from fine to coarse in different scales, and strengthens the capture of short-period local dynamics of each section; the PDM module gradually transmits the global long-term trend from coarse to fine, and integrates into the fine-grained changes of each section and each variable, so that the historical information of multiple sections is deeply layered and mixed, and finally in the cross-section coupling, each section is no longer isolated, and the dynamic coupling and transition influence between sections are automatically mined through parameter sharing and high-dimensional embedding.

[0077] (4), the global multi-objective optimization engine constructed by the application predicts the coke yield and methanol yield based on the trained coking co-production prediction model, the multi-objective optimization model adopts a multi-objective optimization function with a constraint condition to perform multi-objective optimization, obtains a Pareto optimal solution set, solves the optimization problem of multiple conflicting objectives in coking co-production, and makes the future production indexes (coke yield and methanol yield) of coking co-production as high as possible and the coal blending raw material cost as low as possible under the premise of not reducing the coke quality, so that the comprehensive benefit is maximized. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 is a process flow diagram of coking co-production.

[0079] Figure 2 is a flow chart of the multi-objective optimization method of coking co-production.

[0080] Figure 3 is a flow chart of the prediction of the coking co-production prediction model.

[0081] Figure 4 is a processing flow chart of each layer structure of the PDM module.

[0082] Figure 5 is a structure block diagram of the multi-objective optimization system of coking co-production. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0084] See Figure 2 A multi-objective optimization method of coking co-production based on a global optimization strategy, specifically comprising the following steps:

[0085] (1) Collect all historical production variable parameters and production indexes of coking co-production, then perform data preprocessing to obtain a historical data set; the production indexes include coke yield, coal gas yield and methanol yield;

[0086] (2) Construct a coking co-production prediction model, and train the coking co-production prediction model using the historical data set;

[0087] See Figure 3 The coking co-production prediction model comprises a data preprocessing module, a PDM module and a FMM module;

[0088] The data preprocessing module aligns the variable parameters of all work segments for each data sample according to the sampling time axis, forming a unified multivariate input matrix. In other words, each data sample, after processing by the data preprocessing module, outputs a time-series data tensor with multiple work segments, multivariate parameters, and different time steps. , , Represents the time dimension, i.e., the sequence length. Total number of representative work sections , The total number of variable parameters, time series data tensor The elements in are denoted as , Represents time step Time, Section under the variable parameters The value;

[0089] The PDM module consists of an L-layer structure, with the first layer taking a time-series data tensor as its input. The input for each layer is , Represents the current layer. This represents the output of the previous layer structure. Total number of representative work sections Representing the The temporal characteristics of the work section, the output of each layer of the structure is The output of the last layer structure This is the output of the PDM module;

[0090] See Figure 4 The processing procedure for each layer of the PDM module is as follows:

[0091] a. Extract seasonal portions and mix:

[0092] First, start with the residuals, as shown in equation (3) below:

[0093] (3);

[0094] In equation (3), Represents the initial residual. Represents the initial trend component;

[0095] Then, the residuals are processed according to the minimum period. By performing periodic averaging, the seasonal components of the first coarse period are obtained. , see the following formula (4) for details:

[0096] (4);

[0097] In equation (4), , representing the number of cycles in the first coarse cycle;

[0098] Next, multi-scale seasonal layering is performed, and a set of periods is defined: For each coarse cycle Calculate the residual , see the following formula (5) for details:

[0099] (5);

[0100] In equation (5), Represents the residual of the (m-1)th coarse period. Represents the residual of the (m-2)th coarse period. Represents the seasonal component of the coarse period of the (m-1)th layer;

[0101] Then the seasonal components of the coarse period of the m-th layer are calculated. , see the following formula (6) for details:

[0102] (6);

[0103] In equation (6), , representing the period number of the m-th coarse period;

[0104] Finally, the PDM module's first... Hybrid seasonal components output by layer structure , see the following formula (7) for details:

[0105] (7);

[0106] In equation (7), Representing the The layer structure, for the learnable weighting coefficients of the seasonal components of the coarse period in the m-th layer for the work section s and variable parameter v, satisfies the normalization constraint: ;

[0107] b. Extract the trend components and blend them:

[0108] First, initialize the trend components, as shown in equation (8) below:

[0109] (8);

[0110] Equation (8), This represents a moving average operation, with an initial window size of [size missing]. ;

[0111] Next, the trend of layer H is refined layer by layer, as shown in the following formula (9):

[0112] (9);

[0113] In equation (9), , , The trend component representing the h-th layer, Represents the trend component of the (h-1)th layer. This represents the window size of the h-th layer; Represents the learnable weight coefficients;

[0114] Finally, the PDM module's first... Mixing trend component of layer structure output , see the following formula (10) for details:

[0115] (10);

[0116] In equation (10), The trend component representing layer H is used as the first layer of the PDM module. Mixing trend component of layer structure output ;

[0117] c. Cross-section coupling:

[0118] Treating work sections as graph nodes, the set of work section nodes is as follows: Construct an adjacency matrix Adjacency matrix elements in This indicates the intensity of the influence of the i-th section on the j-th section;

[0119] Then, graph convolution is used to fuse features across work sections, as shown in equation (11) below:

[0120] (11);

[0121] In equation (11), It is an adjacency matrix with self-loops. Represents the identity matrix; represent The degree matrix, the diagonal elements of the degree matrix ; Represents the PDM module number The output of the layer structure; Represents the PDM module number The weight matrix of the layer structure, Represents the PDM module number The dimensionality of the input temporal features in the layered structure. Represents the weight matrix The dimensionality of the output features after transformation; Represents a non-linear activation function; Represents the PDM module number cross-section fusion results of layer structure;

[0122] Finally, the PDM module outputs cross-section fusion results of layer structure with mixed seasonal components , mixed trend components are fused to obtain the PDM module outputs layer structure , which can be specifically seen in the following formula (12):

[0123] (12);

[0124] In formula (12), represents a splicing operation;

[0125] The input of the FMM module is , and the FMM module uses an independent predictor to predict the historical characteristics of each section to obtain a future prediction , and the predictor is a linear layer that directly maps the historical characteristics with a length of to a future time sequence with a length of , then maps to variable dimensions, and finally, the prediction results of all sections are weighted and summed to form a final prediction: , and the output is the future production index of the coking polygeneration predicted by the coking polygeneration prediction model;

[0126] (3) Collect real-time production data of the coking polygeneration, which specifically includes variable parameters of four sections (see Figure 1 ), and specifically includes:

[0127] For the coal blending section: blending coal moisture, blending coal volatile matter, blending coal sulfur content, blending coal ash content, blending coal caking index, and coal loading number per hole;

[0128] For the coke oven section: pre-heating temperature of coke oven gas, post-heating temperature of coke oven gas, input end pressure of gas collecting pipe, output end pressure of gas collecting pipe, suction force of flue machine side, suction force of coke side, temperature of flue machine side, temperature of coke side, and coking time;

[0129] For the gas section: resistance of primary cooler, average temperature of gas after primary cooler, resistance of electric trap, pressure of gas after mist collector, average oil temperature of air blower, average oil pressure of air blower, average pressure of gas after desulfurization tower, temperature of gas after benzene washing tower, pressure of gas after benzene washing tower, and temperature of lean oil after benzene washing tower;

[0130] For the methanol section: coal gas usage, R104 catalyst bed temperature, compressor low-pressure cylinder inlet gas flow, compressor circulation section inlet flow, conversion furnace inlet coke coal gas temperature, conversion furnace coke oven gas and outlet conversion gas pressure difference, conversion furnace outlet conversion gas temperature, RT synthesis tower inlet temperature, RP synthesis tower pressure difference;

[0131] Then the characteristic values of the plurality of variable parameters are normalized and subjected to data verification, which specifically includes the following steps:

[0132] S11, setting the weight of all variable parameters in real-time production data , and setting an activation threshold for all variable parameters ;

[0133] S12, calculating the activation score of each variable parameter , specifically as follows:

[0134] (2);

[0135] In formula (2), represents the characteristic value of the i-th variable parameter, represents the normalized characteristic value of the i-th variable parameter;

[0136] S13, when , the i-th variable parameter is screened and retained; otherwise, it is screened out;

[0137] S14, all retained variable parameters are aggregated and uploaded to the global multi-objective optimization engine for multi-objective dynamic optimization;

[0138] (3) constructing a global multi-objective optimization engine, which includes a trained coking co-production prediction model and a multi-objective optimization model. The global multi-objective optimization engine predicts coke production and methanol production based on the trained coking co-production prediction model. The multi-objective optimization model uses a non-dominated sorting genetic algorithm III to solve, i.e. the non-dominated sorting genetic algorithm III embeds the real-time production data of the coking co-production into a multi-objective optimization function with constraints and solves it to obtain a set of Pareto optimal solutions, any solution (corresponding to the control parameters of each section of the coking co-production) is an optimal solution, thereby forming a set of optimal production decision schemes. The control system of all sections of the coking co-production performs coordinated automatic production of multiple sections in the coking co-production according to the optimal production decision scheme;

[0139] The multi-objective optimization function with constraints is specifically as follows:

[0140] (1);

[0141] In formula (1), representing a multi-objective optimization function, i.e. simultaneously satisfying the maximum coke production, the maximum methanol production and the minimum cost of coal blending raw materials; representing the coke production per unit of time, representing the methanol production per unit of time, representing the cost of coal blending raw materials per unit of production; representing the constraint condition of the multi-objective optimization function, representing the minimum quality requirement of coke, representing the process quality requirement of coke; representing the coal gas production per unit of time, representing the function of the coal gas production determining the methanol production, representing the flue gas side temperature of the coke oven section, representing the flue coke side temperature of the coke oven section, representing the gas collecting main pressure of the coke oven section, including the gas collecting main input end pressure and the gas collecting main output end pressure, representing the coking time of the coke oven section, representing the function of the four variable parameters determining the coal gas production; representing the minimum value of the variable parameter (control parameter affecting the section production), representing the maximum value of the variable parameter (control parameter affecting the section production), representing the value of the four variable parameters being constrained.

[0142] see Figure 5 a coking co-production multi-objective optimization system based on a global optimization strategy, comprising a data acquisition and verification module 1, a global multi-objective optimization engine 2, a collaborative production decision module 3 and a multi-section collaborative control module 4;

[0143] The data acquisition and verification module 1 is used to acquire real-time production data of the coking co-production. After the characteristic values of multiple variable parameters in the real-time production data are normalized, data verification is performed. All the verified and retained variable parameters are uploaded to the global multi-objective optimization engine;

[0144] The global multi-objective optimization engine 2 comprises a trained coking co-production prediction model and a multi-objective optimization model. The coke production and the methanol production are predicted based on the trained coking co-production prediction model. The multi-objective optimization model adopts a multi-objective optimization function with constraint conditions to perform multi-objective optimization, and obtains a Pareto optimal solution set;

[0145] The collaborative production decision module 3 corresponds the Pareto optimal solution set to the control parameters of all sections of the coking co-production, and the control parameters of all sections of the coking co-production form an optimal production decision scheme;

[0146] The multi-stage synergic control module 4 is used for the synergic automatic production of the multi-stage in the coking cogeneration according to the optimal production decision scheme, and the real-time production data is re-fed to the data acquisition and verification module.

[0147] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for coking co-production multi-objective optimization based on global optimization strategy, characterized in that: Specifically comprising the following steps: (1) Collecting historical production variable parameters and production indexes of all sections of the coking co-production, and then performing data preprocessing to obtain a historical data set; the production indexes include coke yield, coal gas yield and methanol yield; (2) Constructing a coking co-production prediction model, and training the coking co-production prediction model by using the historical data set; The coking co-production prediction model comprises a data preprocessing module, a PDM module and an FMM module; The data preprocessing module is to align all variable parameters of each data sample in all sections according to a sampling time axis to form a unified multivariate input matrix, that is, each data sample is output after being processed by the data preprocessing module as a time series data tensor of multiple sections, multiple variable parameters and different time steps , , represents a time dimension, that is, a sequence length, represents a total number of sections, , represents a total number of variable parameters, and an element in the time series data tensor is represented as , represents a variable parameter at a time step moment, section ; and represents a value of the variable parameter. The PDM module comprises an L-layer structure, an input of a first layer structure is a time series data tensor , an input of each layer structure is , , represents a current layer, , represents an output of a previous layer structure, , represents a total number of process sections, , represents a time series feature of a first process section, , an output of each layer structure is , and an output of a last layer structure is , that is, an output of the PDM module. The input of the FMM module is For each section, the FMM module utilizes an independent predictor to predict the historical features of the section to obtain a future prediction The predictor is a linear layer that directly maps the historical features of length to a future time series of length , then maps to variable dimensions, and finally weights and sums the prediction results of all sections to form the final prediction: The output is the future production index of the coking cogeneration predicted by the coking cogeneration prediction model. (3) Constructing a global multi-objective optimization engine, wherein the global multi-objective optimization engine comprises the trained coking co-production prediction model and a multi-objective optimization model; the global multi-objective optimization engine predicts the coke yield and the methanol yield based on the trained coking co-production prediction model; the multi-objective optimization model performs multi-objective optimization by using a multi-objective optimization function with a constraint condition to obtain a Pareto optimal solution set; the Pareto optimal solution set corresponds to control parameters of all sections of the coking co-production; the control parameters of all sections of the coking co-production constitute an optimal production decision scheme; and a control system of all sections of the coking co-production performs coordinated automatic production of multiple sections in the coking co-production according to the optimal production decision scheme.

2. The method of claim 1, wherein the method is based on a global optimization strategy. The multi-objective optimization function with the constraint condition is specifically shown in the following formula (1): （1）； In formula (1), represents a multi-objective optimization function, i.e., simultaneously satisfying the maximum coke production, the maximum methanol production, and the minimum cost of coal blending raw materials; represents the coke production per unit time, represents the methanol production per unit time, represents the cost of coal blending raw materials per unit production; represents a constraint condition of the multi-objective optimization function, represents the minimum quality requirement of coke, represents the process quality requirement of coke; represents the coal gas production per unit time, represents a function in which the coal gas production determines the methanol production, represents the flue gas side temperature of the coke oven section, represents the flue coke side temperature of the coke oven section, represents the gas collecting main pressure of the coke oven section, including the gas collecting main input end pressure and the gas collecting main output end pressure, represents the coking time of the coke oven section, represents a function in which the four variable parameters determine the coal gas production; represents the minimum value of the variable parameter, represents the maximum value of the variable parameter, represents that the values of the four variable parameters are constrained.

3. The method of claim 2, wherein the method is based on a global optimization strategy. The global multi-objective optimization engine performs multi-objective dynamic optimization based on real-time production data of the coking co-production; the real-time production data of the coking co-production comprises a plurality of variable parameters; characteristic values of the plurality of variable parameters are normalized and then subjected to data verification; after the data verification, all the verified and retained variable parameters are uploaded to the global multi-objective optimization engine for multi-objective dynamic optimization.

4. The method of claim 3, wherein the method is characterized by: The data verification specifically comprises the following steps: S11, set the weight of all variable parameters in real-time production data and set the activation threshold for all variable parameters ; S12, calculating an activation score for each variable parameter See, in particular, equation (2) below: （2）； In formula (2), an eigenvalue representing an i-th variable parameter, an eigenvalue representing the i-th variable parameter after normalization processing S13, when then the i-th variable parameter is retained; Otherwise, the variable parameters are filtered out; S14, all the retained variable parameters are uploaded to the global multi-objective optimization engine for multi-objective dynamic optimization.

5. The method of claim 3, wherein the method is characterized by: The real-time production data of the coking co-production comprises variable parameters of four sections, specifically: For the coal blending section: moisture content of blended coal, volatile content of blended coal, sulfur content of blended coal, ash content of blended coal, caking index of blended coal, and coal loading number per hole; For the coke oven section: temperature before preheating of coke oven gas, temperature after preheating of coke oven gas, input end pressure of gas collecting pipe, output end pressure of gas collecting pipe, suction force of flue machine side, suction force of coke side, temperature of flue machine side, temperature of coke side, and coking time; For the coal gas section: resistance of primary cooler, average temperature of coal gas after primary cooler, resistance of electric trap, pressure of coal gas after mist collector, average oil temperature of air blower, average oil pressure of air blower, average pressure of coal gas after desulfurization tower, temperature of coal gas after benzene washing tower, pressure of coal gas after benzene washing tower, temperature of lean oil after benzene washing tower; For the methanol section: coal gas usage, temperature of R104 catalyst bed, gas flow at the inlet of low-pressure cylinder of compressor, inlet flow of circulation section of compressor, temperature of coking coal gas at the inlet of conversion furnace, pressure difference between coking gas and outlet conversion gas of conversion furnace, temperature of outlet conversion gas of conversion furnace, inlet temperature of RT synthesis tower, pressure difference of RP synthesis tower.

6. The method of claim 1, wherein the method is based on a global optimization strategy. The processing process of each layer structure of the PDM module is as follows: a, extracting seasonal components and mixing: Firstly, starting from the residual, specifically shown in the following formula (3): （3）； In formula (3), represents the initial residual, represents the initial trend component; The residual is then period-averaged with the minimum period to obtain the seasonal component of the first layer of coarse periods , as shown in the following equation (4): （4）； In formula (4), , represents the number of periods of the first coarse period. Then, the multi-scale seasonal layering is performed to define a set of periods: For each coarse period The residual is calculated as shown in the following equation (5): （5）； In formula (5), a residual representative of a coarse period of the m-1th layer, a residual representative of a coarse period of the m-2th layer, a seasonal component representative of a coarse period of the m-1th layer; The seasonal component of the mth rough cycle is then calculated and is given by the following equation (6): （6）； In formula (6), , represents the number of periods of the mth coarse period. Finally, the hybrid seasonal component of the output of the PDM module at level / is calculated as shown in equation (7) below: （7）； In formula (7), represents the learnable weight coefficient of the seasonal component of the mth layer coarse period for section s and variable parameter v for the first layer structure, satisfying the normalization constraint: ; b, extracting trend components and mixing: First, the trend component is initialized, see the following formula (8): （8）； Formula (8), representing a moving average operation, the initial window size is ; Then, the trend refinement of each layer of H layer is carried out, see the following formula (9): （9）； In formula (9), , , represents a trend component of the hth layer, represents a trend component of the (h-1)th layer, represents a window size of the hth layer; represents a learnable weight coefficient; Finally, the mixing tendency component of the output of the PDM module at the lth level structure is calculated and is given by the following equation (10): （10）； In formula (10), representing the trend component of the Hth layer, as the mixed trend component output by the PDM module of the lth layer structure ; c. Cross-section coupling: Treating work sections as graph nodes, the set of work section nodes is as follows: Construct an adjacency matrix Adjacency matrix elements in This indicates the intensity of the influence of the i-th section on the j-th section; Then, the cross-section feature fusion is carried out by using the graph convolution method, see the following formula (11): （11）； In formula (11), is the adjacency matrix of the graph with self-loops, represents the identity matrix; represents the degree matrix, the diagonal elements of the degree matrix ; represents the output of the PDM module of the (l-1)-th layer structure; represents the weight matrix of the PDM module of the l-th layer structure, represents the dimension size of the input time sequence feature of the PDM module of the l-th layer structure, represents the dimension size of the output feature after the transformation of the weight matrix ; represents the nonlinear activation function; represents the cross-section fusion result of the PDM module of the l-th layer structure; Finally, the cross-section fusion result of the PDM module's 1st layer structure with the mixed seasonal component , the mixed trend component is fused to obtain the output of the PDM module's 1st layer structure , which is specifically shown in the following formula (12): （12）； In formula (12), represents a concatenation operation.

7. The method of claim 3, wherein the method is based on a global optimization strategy. The multi-objective optimization model adopts the non-dominated sorting genetic algorithm III to solve, that is, the non-dominated sorting genetic algorithm III is used to embed the real-time production data of the coking co-production into the multi-objective optimization function with constraints and solve, to obtain a set of Pareto optimal solutions, any solution of which is an optimal solution, thereby forming a set of optimal production decision schemes.

8. A coking co-production multi-objective optimization system for implementing the coking co-production multi-objective optimization method of claim 1, characterized by: It comprises a data acquisition and verification module, a global multi-objective optimization engine, a collaborative production decision module and a multi-section collaborative control module. The data acquisition and verification module is used to acquire the real-time production data of the coking co-production, normalize the characteristic values of multiple variable parameters in the real-time production data, and then perform data verification, and upload all the verified and retained variable parameters to the global multi-objective optimization engine; The global multi-objective optimization engine comprises a trained coking co-production prediction model and a multi-objective optimization model, and is based on the trained coking co-production prediction model to predict the coke production and methanol production, and the multi-objective optimization model adopts a multi-objective optimization function with constraints to perform multi-objective optimization, to obtain a set of Pareto optimal solutions; The collaborative production decision module uses the control parameters corresponding to all sections of the coking co-production in the set of Pareto optimal solutions to form an optimal production decision scheme; The multi-section collaborative control module is used to perform collaborative automatic production of multiple sections in the coking co-production according to the optimal production decision scheme, and re-feed the real-time production data to the data acquisition and verification module.

Images