CN117313554B

CN117313554B - Multi-section combined multi-objective optimization method, system, equipment and medium for coking production

Info

Publication number: CN117313554B
Application number: CN202311597782.7A
Authority: CN
Inventors: 李晓洁; 黄庚; 沈青华; 陈恩红
Original assignee: University of Science and Technology of China USTC
Current assignee: University of Science and Technology of China USTC
Priority date: 2023-11-28
Filing date: 2023-11-28
Publication date: 2024-03-29
Anticipated expiration: 2043-11-28
Also published as: CN117313554A

Abstract

The invention discloses a multi-section combined multi-target optimization method, a system, equipment and a medium for coking production, which are one-to-one corresponding schemes, wherein: the joint optimization of a plurality of continuous working sections in coking production is performed, so that the method is closer to the actual situation of an industrial site; and the optimal solution set is searched under the constraint and multi-target conditions, so that the maximum coke yield and the minimum coking energy consumption can be considered on the premise of ensuring the coke quality, and the quality index, economic index and environmental protection index of actual production are met; in addition, the invention selects the only optimal solution from the optimal solution set, thereby not only meeting the real-time requirement of industrial field operation on decision making, but also avoiding subjectivity of manual selection; in practical engineering application, the invention can provide more optimal schemes of coal blending schemes for coal blending operators, and can also provide unique optimal schemes for Jiao Luwen control operators, pressure control operators and coke pushing operators, so that the invention can be conveniently combined with a service system of a production line, and is easy to practice and popularize.

Description

Multi-section combined multi-objective optimization method, system, equipment and medium for coking production

Technical Field

The invention relates to the technical field of coking production optimization, in particular to a multi-section combined multi-objective optimization method, a system, equipment and a medium for coking production.

Background

The core production scenario of the coking industry is coke oven coking. The coking production has a plurality of working sections and is continuous from front to back, and mainly comprises a coal blending working section, a Jiao Luwen pressure control working section and a coke pushing operation working section.

The traditional coking production has the problems of industrial index test data hysteresis, complex process mechanism, difficult and inaccurate coal blending, strong subjectivity depending on past experience in the setting of the coke oven production environment, and the like, so the establishment of the data-driven coking production AI (artificial intelligence) optimization capability is always a research hot spot and difficulty. In recent years, many research efforts have emerged to train production prediction models using coking historical production data and to optimize them using intelligent algorithms, focusing mainly on two aspects:

firstly, coal blending optimization. Document 1 (ZiJK, jin F, zhao J, et al, A multi-objective simulated annealing algorithm-based coal blending optimization approach in coking process [ C ]// 2020 IEEE Intl Conf on Dependable, autonomic and Secure Computing, intl Conf on Pervasive Intelligence and Computing, intl Conf on Cloud and Big Data Computing, intl Conf on Cyber Science and Technology Congress (DASC/Picom/CBDCom/CyberSciTech). Calgary, AB, canada: IEEE, 2020:103-109.) uses coke ash, sulfur and coal blending cost as optimization targets, establishes a coal blending optimization model of a multi-target simulated annealing algorithm, calculates the optimal coal blending proportion of the existing raw coal, and gives a plurality of pareto raw coal matching schemes; document 2 (Zhao Xiaoyan. Study on coking coal blending ratio optimization model and realization of [ J ]. Metallurgical automation, 2013,37 (02): 6-8+19.). The model is established based on single coal quality analysis data and inventory by taking the quality of blended coal and coke as constraint conditions and the cost of blended coal as objective function, and the model is solved by adopting a differential evolution algorithm and expert coal blending experience; document 3 (Liu Jiguang. Coking and blending optimization model [ J ]. Coal technology, 2012,31 (09): 224-226.) establishes a mathematical model between the coke performance index and the blending ratio, and performs optimization solution on the established model to obtain the optimal blending ratio for producing high quality coke.

Secondly, the coke oven energy consumption is optimized. Document 4 (Tian Z, li S, wang y. The multi-objective optimization model of flue aimed temperature of coke oven [ J ]. Journal of Chemical Engineering of Japan,2018,51 (8): 683-694.) analyzes a correlation model between coke quality, coke yield, coking energy consumption and a target temperature of a flue, and builds a multi-target genetic algorithm optimization model of coke oven flue temperature with maximum coke yield and minimum coking energy consumption as optimization objects and with coke quality and production boundary as constraints to obtain a target temperature optimal value of The coke oven flue, thereby effectively reducing energy consumption; document 5 (Li Ailian, meng Guanjie. Coke oven flame path temperature stability optimization control simulation [ J ]. Computer simulation, 2018,35 (07): 265-268+309.) proposes a flame path temperature closed-loop control scheme based on soft measurement and set value optimization, establishes an optimization model by combining mass field data and an expert system to obtain a target flame path standard temperature value with low energy consumption and high yield, and adopts a least square support vector machine model to realize flame path temperature soft measurement; document 6 (Li Ailian, bi Zewei. Coke oven gas-collecting tube pressure set-point multi-objective optimization study [ J ]. Computer simulation, 2020,37 (04): 260-264+351.) uses coking production history data to respectively establish correlation models of coking energy consumption, coke quality and yield and gas-collecting tube pressure, and calculates an optimal production objective through the optimization model to obtain a corresponding gas-collecting tube pressure set-point.

However, the above documents have mainly the following three-point problems:

first, research efforts have focused on predictive modeling and optimization of a single process. For example, coal blending optimization focuses on the selection of optimal coal blending ratios under complex process constraint conditions. However, the coal blending operation occurs before the blended coal enters the coke oven, so that the resonance influence of the production environment setting of the coke oven and the blending ratio of the blended coal on the product quality, the production cost and the environmental factors is not considered.

And secondly, the existing multi-objective evolutionary method cannot well approach to the real pareto front when solving the multi-objective optimization problem, and the optimization process also easily loses population diversity.

Thirdly, the existing research work aims at searching an optimal solution set, and in actual engineering application, one solution is still selected from the optimal solution set manually to serve as a production scheme which is actually put into execution, and the manual selection has a certain subjectivity on a working section with high requirements on real-time property and uniqueness of instructions in the way of coke oven operation control.

In the heating process of coking production, the production environment settings such as flame path temperature, gas collecting pipe pressure, coking time and the like are in resonance with the quality indexes of the raw material coal, the flame path temperature, the gas collecting pipe pressure, the coking time and the like are mutually coupled to determine the quality indexes such as sulfur content, ash content, crushing strength, wear resistance and the like of the final product coke and the economic indexes such as coking energy consumption and environmental indexes, but the problems of the documents exist, so that the quality indexes, the economic indexes and the environmental indexes of actual production have huge optimization space.

In view of this, the present invention has been made.

Disclosure of Invention

The invention aims to provide a multi-section combined multi-objective optimization method, a system, equipment and a medium for coking production, which can improve the coke yield and reduce the energy consumption of a coke oven on the premise of ensuring the coke quality, and finally the obtained optimal decision variable can be conveniently combined with a service system of a production line, and is easy to practice and popularize.

The invention aims at realizing the following technical scheme:

a multi-section combined multi-target optimization method for coking production comprises the following steps:

collecting historical production data and carrying out data preprocessing to obtain a historical production data training set;

training a coking production prediction model by using the historical production data training set, wherein the training set comprises the following steps: a coke quality prediction model, a coking energy consumption prediction model and a coke yield prediction model;

taking a coke quality prediction model as a constraint condition, taking a coking energy consumption prediction model and a coke yield prediction model as objective functions to be optimized, and establishing a constrained multi-objective optimization problem model, wherein decision variables of the constrained multi-objective optimization problem model are a union of decision variables of the coke quality prediction model, the coking energy consumption prediction model and the coke yield prediction model, and the decision variables of the coke quality prediction model, the coking energy consumption prediction model and the coke yield prediction model are characteristic data selected in data preprocessing and used as input of corresponding models;

Solving the constrained multi-objective optimization problem model by using a multi-objective particle swarm optimization algorithm based on multi-swarm dynamic co-evolution to obtain a pareto optimal solution set meeting constraint conditions;

and selecting an optimal solution from the pareto optimal solution set as an optimal decision variable of the constrained multi-objective optimization problem model, generating an optimal solution based on the optimal decision variable, and using the optimal solution for production deployment.

A multi-section combined multi-objective optimization system for coking production, comprising:

the data acquisition and preprocessing unit is used for acquiring historical production data and preprocessing the data to obtain a historical production data training set;

the model training unit is used for training a coking production prediction model by using the historical production data training set, and comprises the following steps: a coke quality prediction model, a coking energy consumption prediction model and a coke yield prediction model;

the system comprises a constrained multi-objective optimization problem model construction unit, a coking energy consumption prediction model and a coking yield prediction model, wherein the constrained multi-objective optimization problem model construction unit is used for constructing a constrained multi-objective optimization problem model by taking the coke quality prediction model as constraint conditions and taking the coking energy consumption prediction model and the coking yield prediction model as objective functions to be optimized, decision variables of the constrained multi-objective optimization problem model are a union set of decision variables of the coke quality prediction model, the coking energy consumption prediction model and the coke yield prediction model, and the decision variables of the coke quality prediction model, the coking energy consumption prediction model and the coke yield prediction model are characteristic data selected in data preprocessing and used as input of corresponding models;

The problem solving unit is used for solving the constrained multi-objective optimization problem model by using a multi-objective particle swarm optimization algorithm based on multiple swarm dynamic co-evolution to obtain a pareto optimal solution set meeting constraint conditions;

the optimal solution selecting and deploying unit is used for selecting an optimal solution from the pareto optimal solution set as an optimal decision variable of the constrained multi-objective optimization problem model, generating an optimal solution based on the optimal decision variable, and using the optimal solution for production deployment.

A processing apparatus, comprising: one or more processors; a memory for storing one or more programs;

wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the aforementioned methods.

A readable storage medium storing a computer program which, when executed by a processor, implements the method described above.

The technical scheme provided by the invention can be seen that (1) the invention is not only the optimization of the local working section, but also the joint optimization of a plurality of continuous working sections in the coking production, so that the invention is more close to the actual situation of an industrial site; (2) The invention searches the optimal solution set under the constraint and multi-objective conditions, so that the multi-objective problem model constructed by the invention can achieve the maximum coke yield and minimum coking energy consumption on the premise of not reducing the coke quality, and meets the quality index, economic index and environmental protection index of actual production; (3) According to the invention, a unique optimal solution is selected from the optimal solution set and used as an optimal decision variable for actual deployment, so that the real-time requirement of industrial field operation on decision is met, and the subjectivity of manual selection is avoided; in practical engineering application, when actual deployment is carried out based on the optimal decision variables, more optimal schemes of the coal blending schemes can be provided for coal blending operators, and only optimal schemes can be provided for Jiao Luwen control operators, pressure control operators and coke pushing operators.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the description of the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a multi-section combined multi-objective optimization method for coking production provided by an embodiment of the invention;

FIG. 2 is a flowchart of a multi-objective particle swarm optimization algorithm based on multiple swarm dynamic co-evolution according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of selecting an optimal solution according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a multi-section combined multi-objective optimization system for coking production provided by an embodiment of the present invention;

fig. 5 is a schematic diagram of a processing apparatus according to an embodiment of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to fall within the scope of the invention.

The terms that may be used herein will first be described as follows:

the terms "comprises," "comprising," "includes," "including," "has," "having" or other similar referents are to be construed to cover a non-exclusive inclusion. For example: including a particular feature (e.g., a starting material, component, ingredient, carrier, formulation, material, dimension, part, means, mechanism, apparatus, step, procedure, method, reaction condition, processing condition, parameter, algorithm, signal, data, product or article of manufacture, etc.), should be construed as including not only a particular feature but also other features known in the art that are not explicitly recited.

The multi-section combined multi-objective optimization method, system, equipment and medium for coking production provided by the invention are described in detail below. What is not described in detail in the embodiments of the present invention belongs to the prior art known to those skilled in the art. The specific conditions are not noted in the examples of the present invention and are carried out according to the conditions conventional in the art or suggested by the manufacturer. The reagents or apparatus used in the examples of the present invention were conventional products commercially available without the manufacturer's knowledge.

Example 1

The embodiment of the invention provides a multi-section combined multi-target optimization method for coking production, which can be used in the optimization application of the coking production process, so as to perform combined AI optimization on a plurality of continuous sections such as coal blending, jiao Luwen pressure control, coke pushing operation and the like, and mainly comprises the following steps as shown in figure 1:

step 1, acquiring historical production data and carrying out data preprocessing to obtain a historical production data training set.

In the embodiment of the invention, the data preprocessing mainly comprises the following steps: data cleaning, data standardization, feature selection and other operations, and finally obtaining a historical production data training set for model training.

And step 2, training a coking production prediction model by using the historical production data training set.

In the embodiment of the invention, the coking production prediction model mainly comprises: coke quality prediction model) Coking energy consumption prediction model (++>) With coke yield prediction model (+)>）。

In the embodiment of the invention, a coke quality prediction model, a coking energy consumption prediction model and a coke yield prediction model are respectively trained by using the historical production data training set based on an improved Transformer network structure (document 7: xiaojie L, runlong Y, guiquan L, lei C, enhong C, shengjun L. Research on Multi-objective Optimization Algorithm for Coal blending, in: china National Conference on Big Data and Social Computing, xinjiang, july 15-17, 2023. Berlin: springer, 2023, pp., 37-60.).

Wherein: the characteristic data category input by the coke quality prediction model comprises: setting quality index parameters of the matched coal and the production environment of the coke oven; outputting the coke quality prediction model as quality index parameters of the coke; the characteristic data category input by the coking energy consumption prediction model comprises: setting quality index parameters of the matched coal and the production environment of the coke oven; outputting the coking energy consumption prediction model as the usage amount of coke oven gas; the characteristic data categories input by the coke yield prediction model comprise: setting class index parameters in the production environment of the coke oven; and outputting the coke yield prediction model as the coke tapping number.

And 3, establishing a constrained multi-objective optimization problem model.

In the embodiment of the invention, a coke quality prediction model is used as a constraint condition, a coking energy consumption prediction model and a coke yield prediction model are used as objective functions to be optimized, and a constrained multi-objective optimization problem model is established.

The decision variables of the constrained multi-objective optimization problem model are the union of the decision variables of a coke quality prediction model, a coking energy consumption prediction model and a coke yield prediction model, and the decision variables of the coke quality prediction model, the coking energy consumption prediction model and the coke yield prediction model are characteristic data which are selected in the data preprocessing and are input as corresponding models.

And 4, solving the constrained multi-objective optimization problem model.

In the embodiment of the invention, the constrained multi-objective optimization problem model is solved by using a multi-objective particle swarm optimization algorithm based on multiple swarm dynamic co-evolution, and the pareto optimal solution set meeting constraint conditions is obtained. The preferred embodiment of the solution process is as follows:

step S1, carrying out algorithm initialization, setting the population number N, wherein the dimension of each individual is the number K of decision variables of a constrained multi-objective optimization problem model, initializing a matrix of N rows and K columns to represent an initial population, each row to represent an individual, each individual is a K-dimensional decision variable, and initializing the value of each dimension of each individual; setting the maximum iteration number T _max The current number of iterations is t=0.

S2, calculating fitness of an initial population, and carrying out rapid non-dominant sorting on the initial population to determine a pareto level; and establishing an external archive set storage initial population with the size of N.

S3, generating a new population based on a particle swarm optimization algorithm of the dynamic co-evolution of a plurality of swarms; the current population comprises a plurality of sub-populations which are obtained by dividing in an initial population and are fixed, the pareto level of the current population and all sub-populations in the current population is calculated, the current global optimal solution, the optimal solution found by each sub-population and the historical optimal solution of each sub-population are determined based on the pareto level of the current population and all sub-populations in the current population, and the positions of each individual are updated according to the current global optimal solution, so that a new solution is generated and a new population is generated; if the current iteration number is t=0, the current population is the initial population, otherwise, the current population is a new population generated in the last iteration.

In the embodiment of the present invention, each particle involved in the particle swarm optimization algorithm corresponds to an individual, and is therefore collectively referred to as an individual for the purpose of unified expression of the whole text.

The preferred embodiment of step S3 is as follows:

(1) The current population comprises three sub-populations; the first sub-population is an exploration sub-population, and the task of the first sub-population is to access the solution space of the objective function of the multi-objective optimization problem and concentrate on exploring the solution space of the problem to be optimized; the second sub-population is an developing sub-population, and the task of the second sub-population is to perform local search in a solution space; the third sub-population is a dynamic balance sub-population, which dynamically switches the search state between exploration and development through adaptive parameters, i.e. it can adaptively perform the tasks of exploration and development of the sub-population.

In the embodiment of the invention, the three sub-populations are already divided during the initial population, and the sub-populations are not required to be divided again in the subsequent process.

(2) Respectively carrying out rapid non-dominant sorting on the current population and three sub-populations inside the current population, and determining a corresponding pareto level; designating an individual at the first pareto level and at the most middle position in the current population as a current global optimal solution; secondly, designating individuals at the first pareto level in each sub-population and at the most middle position as the optimal solutions found by each sub-population; the history optimal solution of each individual is determined according to the non-dominant relationship, and if the current solution and the history optimal solution are not dominant to each other, one solution is randomly selected as the history optimal solution; if the current solution does not govern the current historical optimal solution, the current historical optimal solution is still selected; if the current solution dominates the current historical optimal solution, selecting the current solution as the historical optimal solution, wherein the solution of the individual is the corresponding position.

In an embodiment of the present invention, each individual stores two locations (i.e., two solutions): the new position obtained after the current update is the current position; the other is the location where its own history is optimal. Briefly, a new historical optimal solution criterion: it is determined whether the new position (i.e., new solution) obtained after the current individual is updated is better than its historical optimal position (historical optimal solution). If the position is better than the current position, the current position is updated to be the current position, otherwise, the current position is not updated.

In the embodiment of the present invention, if the current population is the initial population, the corresponding pareto levels (the step 2 is already completed) need not to be repeatedly calculated, and only the pareto levels corresponding to the inner sub-population are calculated in the part.

In the embodiment of the invention, each individual records the historical optimal position (namely, the historical optimal solution) discovered by iterative optimization of the algorithm, namely, each individual records the current position and the optimal position of the history.

(3) And according to the determined current global optimal solution, the optimal solution found by each sub-population and the historical optimal solution of each individual, carrying out position update by combining the sub-populations to which the individuals belong, and carrying out co-evolution among the three sub-populations in each iteration process, exchanging information of each other, further updating the positions of the appointed individuals, and finally generating a new solution and generating a new population.

Those skilled in the art will appreciate that each individual's location is updated to produce a new solution, and that all of the individuals after the updated location constitute a new population. More specifically, each sub-population updates its own internal individual location according to a different strategy, and all updated sub-populations constitute (generate) a new population.

And S4, calculating the adaptability of the new population.

In the embodiment of the invention, after the new population is generated, the fitness of each individual in the generated new population needs to be calculated and used for calculating the distance between the subsequent pareto level and the crowding degree.

And S5, merging the new population with an external archive set, performing rapid non-dominant sorting to determine the pareto level of the merged population, and then calculating the crowding degree distance of each individual in the merged population.

And S6, selecting N individuals with the best pareto level and crowding degree distances from the combined population to update the external archive set.

Step S7, judging whether the cycle end condition is reached, if t=t _max Then the maximum iteration times are reachedThe number of the N individuals contained in the external archive set at the moment is the pareto optimal solution set meeting the constraint condition; if T is not equal to T _max The process returns to step S3.

And 5, selecting an optimal solution, generating an optimal solution and using the optimal solution for production deployment.

In the embodiment of the invention, a multi-ideal point-based minimum distance method is provided, and is used for selecting a most suitable solution from the pareto optimal solution set as the optimal solution of the constrained multi-objective optimization problem model. The optimal solution will be used for production deployment as an optimal production solution.

The scheme provided by the embodiment of the invention comprises the following steps: (1) The coal blending working section, the Jiao Luwen pressure control working section and the coke pushing operation working section are regarded as a complete industrial production process to carry out joint optimization. On the premise of not reducing the coke quality, the coke yield is improved, the energy consumption of the coke oven is reduced, and the optimal solution set of a coal blending scheme, a temperature control set value, a pressure control set value and a coking time combined production scheme is sought; (2) Based on a non-dominant ordering framework commonly used in multi-objective optimization, a novel evolutionary algorithm is provided, namely a multi-objective particle swarm optimization algorithm based on multiple kinds of swarm dynamic co-evolution. The algorithm is used for generating a new solution to improve the solving performance of multi-section combined multi-objective optimization in coking production. The improved multi-objective evolutionary algorithm further improves the optimizing performance in the multi-objective optimizing process through methods such as multiple group strategies, co-evolution, optimal solution disturbance based on random Gaussian vectors and the like; (3) A minimum distance algorithm based on multiple ideal points is used for solving a unique solution from the optimal solution set, and the method is an optimal and unique joint production scheme, so that the subjectivity problem of manual selection is avoided.

In order to more clearly demonstrate the technical scheme and the technical effects provided by the invention, the method provided by the embodiment of the invention is described in detail below by using specific embodiments.

1. Data collection and preprocessing.

In the embodiment of the invention, historical production data is adopted, and then pretreatment is carried out.

(1) Data cleaning: and selecting historical production data meeting the requirements, namely removing data which do not meet the requirements, such as repeated values, missing values, abnormal values and the like.

(2) Data normalization: and respectively carrying out standardization processing on each type of historical production data, wherein the data obtained through processing is called characteristic data.

In the embodiment of the invention, the data is standardized, so that the difference between different data can not influence the training of the model. For a particular data categoryAnd any data thereof, the standardized processing mode is as follows:

；

wherein,for normalizing the processed data (called characteristic data), the data is processed by the method of the present invention>And->Respectively is a data category->Maximum data and minimum data of the data set.

(3) Feature selection: the most representative characteristic data is selected from the original data and used as input data for model training (namely three types of models mentioned later) so as to reduce the complexity of the model and improve the generalization capability of the model.

2. And (5) building a deep learning network structure.

In an embodiment of the present invention, a classical transducer network structure was simplified and modified to obtain an improved transducer network structure (document 7: xiaojie L, runlong Y, guiquan L, lei C, enhong C, shengjun L. Research on Multi-objective Optimization Algorithm for Coal blending, in: china National Conference on Big Data and Social Computing, xinjiang, july 15-17, 2023. Berlin: springer, 2023, pp. 37-60.) to accommodate the context of the present invention. The improvement of the above document 7 is as follows: (1) The invention reduces the number of self-attention layers and feedforward layers in the model because of lower feature dimension and smaller data volume; (2) Because the invention does not relate to sequence order information, the part related to position coding in the model is removed; (3) Because the invention does not need to embed discrete symbols into a continuous space, the part of the model relevant to the embedded layer is simplified; (4) The invention eliminates the related part of the normalization operation in the model because the normalization processing of the input and the output in the module is not needed. In summary, the improved transducer network structure mainly comprises two parts, namely an Encoder and a Decoder.

3. And training a prediction model.

Through intensive coking industry site investigation and analysis, the invention provides a method for converting coking production prediction problems into building three prediction models, namely a coke quality prediction model, a coking energy consumption prediction model and a coke yield prediction model.

1. Coke quality prediction model）。

(1) Constructing coke quality prediction model based on improved Transformer network）。

When the improved transducer network is used for establishing a coke quality prediction model, the input characteristic data categories comprise: setting quality index parameters of the matched coal and the production environment of the coke oven; the model output is the quality index parameter of the coke.

Exemplary: the input quality index parameters of the blended coal can comprise: sulfur content%) Ash ()>) Moisture (+)>) Adhesive index (+)>) Volatile component (++>) And fineness (Q'); the input production environment setting class index parameters may include: the temperature of the flame path during heating (+)>) Header pressure (P'), and coking time (L); the outputted coke quality index parameters may include: the crushing strength (M40), the abrasion resistance (M10), the reactivity index (CRI), the post-reaction strength (CSR), the Coke sulfur content (Coke_Std) and the Coke ash (Coke_ad) of the Coke.

Based on the index parameters provided by the above examples, the input vector of the coke quality prediction modelThe method comprises the following steps:the method comprises the steps of carrying out a first treatment on the surface of the Output vector of coke quality prediction model +.>The method comprises the following steps:。

the internal process flow of the coke quality prediction model is expressed as:

；

wherein the first two equations are encoder processes in the coke quality prediction model,and->For the treatment of the result->The output vector of the prediction result of the decoder, namely the coke quality prediction model.

(2) Model training and parameter updating methods.

In order to determine the parameter values of the coke quality prediction model, the coke quality prediction model is provided with the ability to fit the prediction, and a random gradient descent algorithm is used here to repeatedly and iteratively update the model.

Calculating a prediction error valueExpressed as:

；

wherein,refers to all pending parameters of the coke quality prediction model.

Calculating the gradient: for a specific parameterIn terms of this parameter->Gradient values in iteration->The method comprises the following steps:

；

wherein,refers to the partial derivative.

Parameter updating: in order to correct the coke quality prediction model, the prediction error is reduced, and the parameters are updated:

wherein,is the learning rate of the coke quality prediction model, +.>、Parameters->A parameter value before update and a parameter value after update.

2. Coking energy consumption prediction model）。

(1) Constructing a coking energy consumption prediction model based on an improved transducer network）。

In the embodiment of the invention, when the improved transducer network is used for establishing the coking energy consumption prediction model, the input characteristic data categories comprise: setting quality index parameters of the matched coal and the production environment of the coke oven; the model output is the coke oven gas usage amount (Mj).

Exemplary: the input quality index parameters of the blended coal can comprise: moisture (Mt) and volatile matters [ (Mt ]) The method comprises the steps of carrying out a first treatment on the surface of the The input production environment setting class index parameters may include: the temperature of the flame path during heating (+)>) Header pressure (P'), coking time (L), flue side suction (+.>) Suction of flue coke side (+)>) Pre-heating temperature of coke oven gas (+.>) Temperature after coke oven gas preheating (++)>）。

Based on the index parameters provided by the above examples, the coking energy consumption prediction model input vectorThe method comprises the following steps:the method comprises the steps of carrying out a first treatment on the surface of the Coking energy consumption prediction model output vector +.>Is that。

The coking energy consumption prediction model also comprises an encoder and a decoder, and the internal processing flow is the same as that of the coke quality prediction model, so that the description is omitted.

(2) Model training and parameter updating methods.

In order to determine the parameter values of the coking energy consumption prediction model, the model is provided with the ability to fit the prediction, and a random gradient descent algorithm is used here, and the model is repeatedly and iteratively updated. The specific scheme is the same as the coke quality prediction model, so that the description is omitted.

3. Model for predicting coke yield）。

(1) Constructing a coke yield prediction model based on an improved Transformer network）。

In the embodiment of the invention, when the improved transducer network is used for establishing a coke yield prediction model, the input characteristic data categories comprise: setting class index parameters in the production environment of the coke oven; the coke yield prediction model is output as the coke tapping number (S).

Exemplary: the input production environment setting class index parameters may include: header pressure (P'), coking time (L), and coal charging belt weighing data (M).

Based on the index parameters provided by the above examples, a coke yield prediction model input vectorThe method comprises the following steps:the method comprises the steps of carrying out a first treatment on the surface of the Output vector->The method comprises the following steps:。

Similarly, the coke yield prediction model also includes an encoder and a decoder, and the internal process flow is the same as that of the coke quality prediction model, so that description thereof will be omitted.

(2) Model training and parameter updating methods.

To determine the parameter values of the coke yield prediction model, the model is provided with the ability to fit the predictions, where a random gradient descent algorithm is used, iteratively updating the model repeatedly. The specific scheme is the same as the coke quality prediction model, so that the description is omitted.

4. Multi-objective optimization scheme.

1. And (5) constructing a multi-section combined multi-target optimization model for coking production.

In the embodiment of the invention, a trained coke quality prediction model is adopted) As constraint conditions, a coking energy consumption prediction model (++>) And a coke yield prediction model (++>) And establishing a constrained multi-objective optimization problem model for the two objective functions to be optimized.

Wherein the number of decision variables of the overall multi-objective optimization problem model is three models、And->Union of the respective decision variables. In summary, the multi-objective optimization problem model established in the invention is shown in the following formula:

；

wherein st is a constraint condition symbol,optimizing questions for multiple objectivesQuestion model (I/O)>Representing a coke quality prediction model,/->Representing a coking energy consumption prediction model, < >>Represents a coke yield predictive model,/->Representing a multi-objective optimization problem objective function, with min () being a minimization function;And->The model is expressed as two objective functions to be optimized, < ->The model is a constraint.

In the embodiment of the invention, the number of decision variables defining the multi-objective optimization problem model is K.

According to the examples provided above: the number of decision variables of the coke yield prediction model is 3, and the decision variables respectively set class indexes (gas collecting tube pressure, coal charging belt weighing data and coking time) for the coke oven production environment. The number of decision variables of the coking energy consumption prediction model is 9, and the decision variables are respectively matched with coal quality indexes (volatile matters and moisture) and coke oven production environment setting indexes (flue temperature, gas collecting pipe pressure, flue side suction, flue coke side suction, coking time, coke oven gas preheating temperature and coke oven gas preheating temperature in the heating process). The number of decision variables of the coke quality prediction model is 9, and the decision variables are respectively set up as the quality indexes (moisture, ash, sulfur, volatile, bonding index and fineness) of the matched coal and the production environment of the coke oven (gas collecting tube pressure, coking time and flame path temperature). Therefore, the number K of decision variables in the multi-objective optimization problem is the union of the three model decision variables, and is equal to 14, and the meaning is respectively matched with the quality index (moisture, ash, sulfur, volatile matters, bonding index and fineness) of coal and the setting index (the temperature of a flame path, the pressure of a gas collecting pipe, the data of a coal charging belt, the side suction of a flue machine, the side suction of a flue, the coking time, the temperature before preheating of coke oven gas and the temperature after preheating of the coke oven gas) of the coke oven production environment. Of course, the specific values of K and the specific meanings contained therein may be set by the user according to actual situations or experience in practical applications by way of example only.

2. Multi-target particle swarm optimization algorithm based on multi-swarm dynamic co-evolution.

Particle swarm optimization (Particle Swarm Optimization, PSO), one of the classical evolutionary algorithms widely used, document 8 (J. Kennedy, R. Eberhart, particle swarm optimization, in: 1995 IEEE International Conference on Neural Networks (ICNN 95), univ W Austraia, perth, australia, 1995, pp. 1942-1948.) describes a related algorithm scheme. In order to solve the problem of multi-section combined multi-objective optimization in coking production, the invention provides a multi-objective particle swarm optimization algorithm based on multiple-swarm dynamic co-evolution based on a rapid non-dominant ordering framework (document 9: deb K, pratap A, agarwal S, et al.A. fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE Trans. On Evol [ J ]. IEEE Transactions on Evolutionary Computation, 2002, 6.), and the flow of the multi-objective particle swarm optimization algorithm is shown in figure 2.

Step S1, initializing an algorithm.

In the embodiment of the invention, algorithm initialization is performed based on classical circle map chaotic mapping.

The formula for initializing the algorithm population based on classical circle map is as follows:

；

wherein,a value representing the kth dimension of the decision variable, +. >Representing the value of the k+1 dimension produced using the circle map formula from the value of the k dimension, mod () is a remainder function. Setting the lower bound of decision variables as LB, the upper bound as UB, LB and UB as two K-dimensional arrays respectively, LB _k And UB _k Respectively the value of their kth dimension.

The method comprises the steps of initializing an algorithm, setting the population quantity as N, namely, N individuals, wherein the dimension of each individual is the number K of decision variables of a constrained multi-objective optimization problem model, initializing a matrix of N rows and K columns to represent the initial population, each row to represent an individual, each individual is a K-dimensional decision variable to represent a solution of a K-dimensional problem, each decision variable generates an initial position according to a formula of a circle map, and the value of the initial position cannot cross the upper boundary and the lower boundary of the decision variable. Setting the maximum iteration number T _max The current number of iterations is t=0.

The initial population is expressed as:

；

the subscript in the matrix represents the position in the matrix, the 1 st numerical value represents the row number, the 2 nd numerical value represents the column number, and each row of corresponding individuals calculates the initialized value by adopting the formula.

Step S2, calculating the fitness of the initial population, carrying out rapid non-dominant sorting to determine the pareto hierarchy, and establishing an external archive set with the scale of N.

In the embodiment of the invention, the fitness value of each individual in the initial population is calculated, namely the function value and constraint violation condition of each sub-objective in the multi-objective model are calculated respectively, namely each individual is calculated、And->Function values. Then, using the rapid non-dominant ranking framework of document 10, all individuals are rapidly non-dominant ranked to determine different pareto levels for all individuals in the population by the calculated fitness, and an external archive set of size N is created to archive the population. Wherein non-dominant ranking is used to divide individuals in a population into different levels (ranks) according to their dominant relationship, wherein higher levels of individuals are more superior (not being dominant by other individuals). A scheme for calculating the pareto hierarchy is provided below.

Substep S21, initializing an empty Pareto layer level set, for example called pareto_ranks.

Substep S22, traversing each individual in the population. The following is performed for each individual:

(1) The individual's dominant count, doming_count, is initialized to 0, and an empty dominant individual set, doming_set.

(2) Traversing each individual in the population again, and for each traversed individual j, performing the following operations: if individual i (the currently traversed individual) dominates individual j, then individual j is added to the minimized set, while the minimizing_count of individual i is incremented. If the individual j dominates the individual i, the determination_count is cleared and the internal loop is skipped, and the next individual goes to step S22.

(3) If the minimization_count is 0, indicating that the individual i is not at the discretion of any other individual, then the individual i is added to the first hierarchy of pareto_ranks.

And a substep S23, constructing an individual set of the next hierarchy. From the first hierarchy of pareto_ranks, the dominant individual set of all individuals is merged and the dominant count, doming_count, is decremented. If the individual's domination_count becomes 0, it is added to the second level of pareto_ranks. This process continues until all individuals are assigned to a certain pareto tier.

Substep S24 returns pareto_ranks, where each hierarchy contains a set of individuals that are not dominated by other individuals, and the higher the hierarchy, the more advantageous the individuals. The external archive set is used to store pareto_ranks.

And S3, generating a new population based on a particle swarm optimization algorithm of the dynamic co-evolution of the multiple swarms.

Substep S31, introducing an original PSO location update formula.

Each individual in the population represents a solution to the problem to be optimized. The objective function of the problem to be optimized is a K-dimensional function and has N individuals, the firstThe location codes of the individual individuals are:each element in the position code represents a position value for one dimension. Each individual has the concept of speed, let +. >The individual speed codes are:each element in the velocity code represents a velocity value for one dimension. In addition, the historical optimal solution of each individual from optimization and the global optimal solution found so far by the whole individual group are used for guiding the position update of each individual, the individual with the best adaptability found by the algorithm from iteration is the global optimal solution, and the position of each individual at the moment with the best adaptability in the position update process is the historical optimal solution of the individual, mainly because the position of the individual is possibly degraded after the update. Is provided with->Is->Historical optimal solutions for individual individuals:each element in the historical optimal solution represents a value of the historical optimal solution for one dimension;The optimal solution found for the entire individual group (the current global optimal solution) is encoded as:Each element in the current globally optimal solution represents a value of the current globally optimal solution for one dimension. The original location update formula is as follows:

；

。

each individual continuously and iteratively updates own speed and position according to the formula, so as to realize searching in a problem space, and find a high-quality approximate optimal solution for the problem to be optimized. Indicating +.>Speed code of individual->Representing the updated (i.e. t+1st iteration) velocity code,/for>Position coding representing the current iteration of the individual, +.>Representing the updated position code of the individual.Is a routineThe sexual weight parameter represents inheritance of the current individual speed, which is typically between 0 and 2.And->Learning factors, respectively, represent +.>Learning of individual historical and global optimal positions is typically set at a constant 2.And->Is two random numbers with values between 0 and 1.

And a substep S32, PSO based on dynamic co-evolution of multiple groups.

The multiple populations are a common framework for improving the performance of the algorithm in the evolutionary algorithm (literature 10: ma H P, shen S G, yu M, et al, multi-population techniques in nature inspired optimization algorithms: A comprehensive survey [ J ]. Swarm Evol Comput, 2019, 44: 365-87.), and the invention provides a novel three-population co-evolution strategy. The improved PSO position updating formula and operation flow based on multiple group dynamic co-evolution are as follows:

(1) Multiple group partitions and improved location update formulas.

After the initialization of the population is completed in the aforementioned step S1, the population is divided into three sub-populations, and then, the division is not required, and for simplicity, the scale of each sub-population is set to be the same. Each sub-population is designated to play a different role in the optimization process.

Exploring a sub-population: the sub-population is used for exploring the solution space of the problem, and the main task of the sub-population is to quickly access the solution space of the objective function of the problem to be optimized so as to find a high-quality solution and avoid the premature sinking into local optimum in the optimization process. The location update formula for this sub-population improvement is as follows:

；

in the above formula, PB _Rand A historical optimal solution for a randomly selected individual in the population. The individuals in the exploring sub-population can avoid the loss of algorithm exploring performance caused by the fact that the current global optimal solution GB falls into the local optimal by learning from the randomly selected historical optimal solution instead of the current global optimal solution GB of the whole population. In addition, in the case of the optical fiber,the dimension of the multi-element Gaussian variable generated randomly is K, and the value of each dimension is limited to be [ -1,1]Is a kind of medium.As a random disturbance item, the ability of exploring random variation of each individual in the sub-population is given, the probability of the algorithm jumping out of the local optimum is increased, and the population diversity is maintained.

Developing sub-populations: the sub-population is used for improving the optimization effect and precision of the algorithm through rapid local search. The method has the main tasks of carrying out depth local search in a given solution space region, rapidly improving algorithm precision, accelerating algorithm convergence and finding out an approximate optimal solution. The location update formula for this sub-population improvement is as follows:

；

In the above-mentioned formula(s),for an adaptive scaling factor, +.>For an algorithm iteration, the variable from 2 linearly decreases to 0, ++>Is in the range of 0,1]Is a random variable of (a). Through the formula, the individuals in the development sub-population are concentrated on searching the area near the current global optimal solution GB, so that algorithm convergence is quickened, and solving precision is improved.

Dynamic balance sub-population: the remaining population is called a dynamic balance sub-population, and the sub-population dynamically switches the search state between exploration and development through self-adaptive parameters so as to realize dynamic balance of PSO exploration and development and maintain population diversity. The location update formula for this sub-population improvement is as follows:

；

wherein,is [0,1]Random number in between, A is an adaptive scaling factor,>for an algorithm iteration, the variable from 1 linearly decreases to 0, ++>T is an adaptive parameter which is nonlinearly decreased from 1 to 0 _max The maximum iteration number of the algorithm is calculated, and t is the current iteration number. For the evolutionary algorithm, the algorithm needs to have good exploratory performance in the early iteration stage, and the solution space of the problem is spread as much as possible, so as to expect to find the area where the globally optimal solution is located; and the algorithm at the later stage of iteration should be focused on the development operation based on local search so as to accelerate algorithm convergence and improve the solving precision.

Because of the monotonicity of the cosine function,the early stage slowly decreases from 1, so that the dynamic balance sub population performs exploration operation with higher probability, and the whole PSO algorithm is more prone to exploring the solution space of the problem to be optimized. Monotonicity of the later-term factor cosine function>Rapidly decreasing to a value close to 0 and because of the pre-coefficient +.>Also gradually approaching 0, thus->The value of (2) becomes smaller. At this time, random number +.>Easily be greater than->Resulting in a greater probability of dynamically balancing sub-populations to perform local search-based development operations. From the above formula, it can be seen that the position update formula of the dynamic balance sub-population combines the exploration and development sub-population, and the self-adaptive parameter can dynamically switch the search state between exploration and development to play a role in balancing exploration and development, and maintain population diversity. />

(2) Co-evolution.

As described previously, the entire population is divided into three sub-populations of equal size, and a new location update formula is designed for each sub-population, each having a different search function. In addition, the sub populations co-evolve and exchange information with each other, so that PSO can achieve better optimization effect. The novel co-evolution strategy proposed in the present invention is:

(a) During each iteration, each sub-population records the optimal solution LGB found by its optimization iteration, and each sub-population shares the optimal solutions with each other. The optimal solutions found by each of the searching, developing and dynamic balancing sub-populations are respectively as follows: LGB1, LGB2, and LGB3.

(b) After each iteration, the worst fitness individual within each sub-population changes its own position by learning to LGB1, LGB2 and LGB3. The location update formula is as follows:

；

wherein,、、is an intermediate variable calculated by the worst fitness individual by independently using LGB1, LGB2 and LGB 3;、And (5) the updated speed codes and the updated position codes.

Through the formula, individuals with worst internal fitness of each sub-population can learn from each other to the optimal solutions of different sub-populations, so that co-evolution is realized.

(3) And (5) an optimal position disturbance strategy.

After each iteration is completed, the current global optimal solution GB found by the PSO algorithm is randomly disturbed to increase the capability of the GB to jump out of the local optimal. The disturbance method is to apply a random Gaussian disturbance to GB, and the formula is as follows:

；

wherein,is a K-dimensional random Gaussian vector with each dimension value of [ -0.5, 0.5 ]Between them;For assigning symbols, a globally optimal solution is obtained when applying random Gaussian perturbation>When the global optimal solution GB is superior to the current global optimal solution GB, the global optimal solution is +.>And the current global optimal solution is used for the next iteration, otherwise, the current global optimal solution GB is still kept unchanged.

And S33, generating an operation flow of a new population based on PSO of dynamic co-evolution of multiple populations.

(1) In single-objective PSO, its GB and PB of each individual are unique, but in multi-objective optimization, GB, LGB and PB of each individual in multi-objective PSO based on multiple group dynamic co-evolution are determined by a random assignment method, since multiple functions are to be optimized simultaneously, and multiple solutions may be at the first pareto level in the population.

First, the pareto levels of the whole population and the inner sub-population are calculated according to the pareto level calculation step in step S2. Individuals in the whole population at the first pareto level and at the most intermediate positions are designated GB. Second, individuals within each sub-population at the pareto first level, and the most intermediate location, are designated as their respective LGBs. Finally, the PB of each individual is determined according to the non-dominant relationship, and if the new PB and the original old PB are not dominant to each other, one PB is randomly selected; if the new PB does not dominate the original old PB, the old PB is still selected; if the new PB dominates the old PB, then the new PB is selected.

(2) Based on the determined GB, LGB and PB for each individual, the location of each individual is updated using the approach in sub-step S32, generating a new solution and generating a new population.

And S4, calculating the adaptability of the new population.

In the embodiment of the invention, after the new population is generated, the fitness of each individual in the generated new population needs to be calculated and used for calculating the pareto level and the crowding degree distance subsequently.

And S5, non-dominant sorting and calculating the crowding degree distance.

In the embodiment of the invention, the new population is combined with the external archive set, the pareto level of the combined population is calculated according to the pareto level calculation step in the step S2, and then the crowding degree distance of each individual in the combined population is calculated. The following is a specific step of calculating the crowding degree distance:

in the substep S51, the individuals in the population are ranked according to a certain target value of each individual in the target space (generally, the result after the non-dominant ranking).

Substep S52, initializing the crowding degree distance number for each targetGroup distances is 0, and the array length is the population size N; the objective here is to mean that the multi-objective optimization problem model removes two terms other than the constraint terms, i.eIs- >And->。

In a substep S53, for each target, the difference (i.e., the maximum value minus the minimum value of the target value) for each individual on the target is calculated and stored in an array (delta).

Substep S54, for each target, sets the crowding distances of the 1 st and nth individuals in the sorted population to infinity (or a larger value), indicating that they are at boundary positions and cannot be eliminated.

Step S55, for each target, traversing the 2 nd to N-1 st individuals in the sorted population, and calculating the crowding degree distance of each target:

(1) Adding a value equal to the fitness value of the individual i+1 on the target r to the crowding degree distance (distance i) of the ith individual on the target r, subtracting the fitness value of the individual i-1 on the target r, and dividing by delta [ r ], wherein delta [ r ] is the difference between the maximum fitness and the minimum fitness on the target r, so as to perform distance standardized scaling and overcome the influence caused by the inconsistent sizes of different sub-objective functions.

(2) And repeating the steps, and sequentially calculating the crowding degree distances on all the targets.

And a substep S56, for each individual, summing the crowding distances on all the targets to obtain the total crowding distance value of the individual.

And S57, sorting individuals in the sorted population according to the total crowding degree distance value from large to small.

And S6, updating the external archive set.

N individuals with the best pareto grade and crowding degree distance are selected from the combined population to form a new external archive set, and the old external archive set is updated. Specific: n are selected from the combined population, the pareto grades are selected from the highest to the highest in sequence, and when the pareto grades belong to the same pareto grade, the priority of individuals with the largest crowding degree is selected.

Step S7, judging whether the cycle end condition is reached, if t=t _max The maximum iteration times are reached, and the population of N individuals contained in the external archive set at the moment is the pareto optimal solution set obtained through optimization; if T is not equal to T _max The process returns to step S3.

5. A unique production scheme is selected from the optimal solution set.

In the embodiment of the invention, the final production scheme is selected from the pareto solution set based on a multi-ideal point minimum distance method. Specific: a multi-objective solution set meeting the conditions is obtained by using a multi-objective optimization algorithm, and most solutions in the solution set can be used for production deployment.

But in order to maximize coke yield and minimize coke energy consumption during coking production and meet coke quality constraint requirements, it is necessary to select a most appropriate solution (a joint production scheme) from the pareto solution set for production deployment, specifically: the pareto optimal solution set obtained by storage optimization in the archiving set is N individuals, each individual is K-dimensional, and corresponds to decision variables of the multi-objective problem model, namely N groups of decision variables, but the invention aims at the multi-objective optimization problem with constraint, so that the solution set obtained by optimization does not necessarily meet the constraint, and partial solution set may violate the constraint. Our goal is to choose the most appropriate solution (i.e., the most appropriate individual K-dimensional decision variable) from the solutions that satisfy the constraint, or that have the least illegal constraint, to generate the optimal solution for the actual production deployment. 1. And calculating ideal points in the multi-objective optimization problem model of coking production.

In the embodiment of the invention, the constrained multi-objective optimization problem model is decomposed into two constrained single-objective optimization problem models, which are expressed as:

；

wherein,and->Optimizing the problem model for two single objectives, +.>And->Is two single objective functions. Because the optimization result of the single-objective optimization problem is usually only one and the pareto solution set is not required, the method can respectively obtain +.>And->In meeting->Approximately optimal solution under constraint conditions, +.>The result of the solution is marked as X ₁ ，The result of the solution is marked as X ₂ . These two solutions represent, respectively, the ++in case of not taking into account the fact that the multi-objective optimizations dominate each other>And->The best results each can achieve if the constraints are met. X is to be ₁ And X ₂ Set to->And->As shown in fig. 3.

Furthermore, X in FIG. 3 ₃ Is a hypothetical ideal point, called a hypothetical point. Its coordinates are (P) _max ，E _min ) The specific meaning is as follows: it is assumed that there is such a solution that the yield is infinitely large and the energy consumption is infinitely small. The imaginary point is either theoretically or practically absent, so that a large product value P can be predefined _max And a smaller energy consumption value E _min To approximately replace the imaginary point.

2. And calculating the distance between each solution in the pareto optimal solution set and the ideal point and the imaginary point.

Firstly, eliminating solutions which violate constraint conditions in the pareto optimal solution set, obtaining a set Z of solutions which meet the constraint, and calculating the distance between each solution in the set Z of solutions which meet the constraint and an ideal point and an imaginary point to be expressed as:

；

wherein,to calculate the sum of the distances of a single solution from two ideal points in the set Z of solutions satisfying the constraint,to calculate the distance of a single solution from a hypothetical point in the set Z of solutions satisfying the constraint, dis () is a distance calculation function,the ith solution in the set of solutions to satisfy the constraint Z, Z being PaThe set of solutions satisfying the constraint in the rector optimal solution set.

3. And selecting an optimal solution.

Selecting the solution with the smallest Dis1 value, which represents the distance F ₁ And F ₂ The distance and minimum of the ideal points can reflect that the ideal points fully considerAnd->And a best solution is selected in the trade-off. If only one solution exists, selecting the solution as an optimal decision variable of the constrained multi-objective optimization problem model, and further generating an optimal solution according to the optimal decision variable, namely the specific numerical value of each decision variable. If a plurality of solutions exist, selecting the solution with the minimum Dis2 value in the plurality of solutions as the optimal solution to be used as the optimal decision variable of the constrained multi-objective optimization problem model, wherein the optimal decision variable has the following meaning: at multiple solution distances- >And->When the sum of distances from the ideal points is the same, a solution closest to the virtual point is selected.

The scheme provided by the embodiment of the invention comprises the following key technologies:

(1) Based on continuous working sections of coking production, a set of multi-section combined multi-objective optimization method is established, namely a prediction model is established for coking production by adopting a deep learning technology, and combined multi-objective optimization is carried out on a plurality of continuous working sections of coking production, wherein the method is a bottom key technology for promoting global production optimization by coking enterprises;

(2) The multi-objective PSO algorithm based on the dynamic co-evolution of multiple groups is used for solving the multi-objective optimization problem of multi-section combination of coking production. The improved multi-objective algorithm is based on various group ideas, and an adaptive switching strategy and random Gaussian variation operation which are explored and developed are added, wherein the strategies are used for improving the performance of PSO for solving the multi-objective optimization problem;

(3) A minimum distance method based on multiple ideal points is provided for selecting a most suitable joint production scheme from the pareto solution set. In the working section with high requirement on implementation operability, the method has higher application value and popularization demonstration value;

Based on the key technology, the invention can improve the coke yield and reduce the energy consumption of the coke oven on the premise of ensuring the coke quality, and finally the obtained optimal decision variable can be conveniently combined with a service system of a production line, thereby being easy to practice and popularize.

From the description of the above embodiments, it will be apparent to those skilled in the art that the above embodiments may be implemented in software, or may be implemented by means of software plus a necessary general hardware platform. With such understanding, the technical solutions of the foregoing embodiments may be embodied in a software product, where the software product may be stored in a nonvolatile storage medium (may be a CD-ROM, a U-disk, a mobile hard disk, etc.), and include several instructions for causing a computer device (may be a personal computer, a server, or a network device, etc.) to perform the methods of the embodiments of the present invention.

Example two

The invention also provides a multi-working section combined multi-target optimizing system for coking production, which is mainly used for realizing the method provided by the previous embodiment, and as shown in fig. 4, the system mainly comprises:

It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-described division of the functional modules is illustrated, and in practical application, the above-described functional allocation may be performed by different functional modules according to needs, i.e. the internal structure of the system is divided into different functional modules to perform all or part of the functions described above.

Example III

The present invention also provides a processing apparatus, as shown in fig. 5, which mainly includes: one or more processors; a memory for storing one or more programs; wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the methods provided by the foregoing embodiments.

Further, the processing device further comprises at least one input device and at least one output device; in the processing device, the processor, the memory, the input device and the output device are connected through buses.

In the embodiment of the invention, the specific types of the memory, the input device and the output device are not limited; for example:

the input device can be a touch screen, an image acquisition device, a physical key or a mouse and the like;

The output device may be a display terminal;

the memory may be random access memory (Random Access Memory, RAM) or non-volatile memory (non-volatile memory), such as disk memory.

Example IV

The invention also provides a readable storage medium storing a computer program which, when executed by a processor, implements the method provided by the foregoing embodiments.

The readable storage medium according to the embodiment of the present invention may be provided as a computer readable storage medium in the aforementioned processing apparatus, for example, as a memory in the processing apparatus. The readable storage medium may be any of various media capable of storing a program code, such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a magnetic disk, and an optical disk.

The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions easily contemplated by those skilled in the art within the scope of the present invention should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A multi-section combined multi-target optimization method for coking production is characterized by comprising the following steps:

Selecting an optimal solution from the pareto optimal solution set as an optimal decision variable of the constrained multi-objective optimization problem model, generating an optimal solution based on the optimal decision variable, and using the optimal solution for production deployment;

the method for solving the constrained multi-objective optimization problem model by using a multi-objective particle swarm optimization algorithm based on multiple swarm dynamic co-evolution comprises the following steps:

step S1, carrying out algorithm initialization, setting the population quantity as N, namely N individuals, wherein the dimension of each individual is the number K of decision variables of the constrained multi-objective optimization problem model, initializing a matrix of N rows and K columns to represent an initial population, each row to represent an individual, each individual is a decision variable of K dimension, and initializing the initial position of each individual; setting the maximum iteration number T _max The current iteration number is t=0;

s2, calculating fitness of an initial population, performing rapid non-dominant sorting on the initial population to determine a pareto level, and establishing an external archive set with a scale of N;

s3, generating a new population based on a particle swarm optimization algorithm of the dynamic co-evolution of a plurality of swarms; the current population comprises a plurality of sub-populations which are obtained by dividing in an initial population and are fixed, the pareto level of the current population and all sub-populations in the current population is calculated, the current global optimal solution, the optimal solution found by each sub-population and the historical optimal solution of each sub-population are determined based on the pareto level of the current population and all sub-populations in the current population, and the positions of each individual are updated according to the current global optimal solution, so that a new solution is generated and a new population is generated; if the current iteration number is t=0, the current population is the initial population, only the pareto levels of all the sub-populations in the interior are calculated at the moment, otherwise, the current population is a new population generated by the last iteration;

S4, calculating the adaptability of the new population;

s5, merging the new population with an external archive set, performing rapid non-dominant sorting to determine the pareto level of the merged population, and then calculating the crowding degree distance of each individual in the merged population;

s6, selecting N individuals with the best pareto level and crowding distance from the combined population to update the external archive set;

step S7, judging whether the cycle end condition is reached, if t=t _max The maximum iteration times are reached, and the population of N individuals contained in the external archive set at the moment is the pareto optimal solution set obtained through optimization; if T is not equal to T _max Returning to the step S3;

the step of generating a new population based on the particle swarm optimization algorithm of the dynamic co-evolution of the multiple swarms comprises the following steps:

the current population comprises three sub-populations; the first sub-population is an exploring sub-population, and the task is to access a solution space of a multi-objective optimization problem objective function; the second sub-population is an developing sub-population, and the task of the second sub-population is to perform local search in a given solution space; the third sub-population is a dynamic balance sub-population, which dynamically switches the search state between exploration and development through adaptive parameters;

respectively carrying out rapid non-dominant sorting on the current population and three sub-populations thereof, determining a corresponding pareto level, and if the current population is an initial population, only calculating the pareto level corresponding to the inner sub-population at the moment; designating an individual at the first pareto level and at the most middle position in the current population as a current global optimal solution; secondly, designating individuals at the first pareto level in each sub-population and at the most middle position as the optimal solutions found by each sub-population; the history optimal solution of each individual is determined according to the non-dominant relationship, and if the current solution and the current history optimal solution are not dominant to each other, one solution is randomly selected as the history optimal solution; if the current solution does not govern the current historical optimal solution, the current historical optimal solution is still selected; if the current solution dominates the current historical optimal solution, selecting the current solution as the historical optimal solution, wherein the solution of the individual is the corresponding position;

And according to the determined current global optimal solution, the optimal solution found by each sub-population and the historical optimal solution of each individual, carrying out position update by combining the sub-populations to which the individuals belong, and carrying out co-evolution among the three sub-populations in each iteration process, exchanging information of each other, further updating the positions of the appointed individuals, and finally generating a new solution and generating a new population.

2. The multi-stage joint multi-objective optimization method for coking production according to claim 1, wherein the data preprocessing comprises:

data cleaning: selecting historical production data meeting requirements;

data normalization: respectively carrying out standardized processing on each type of historical production data, wherein the processed data are called characteristic data;

feature data selection: characteristic data input as a coke quality prediction model, a coking energy consumption prediction model and a coke yield prediction model are respectively selected.

3. The multi-segment joint multi-objective optimization method for coking production according to claim 1, wherein training out the coking production prediction model by using the historical production data training set comprises:

based on the improved transducer network structure, respectively training a coke quality prediction model, a coking energy consumption prediction model and a coke yield prediction model by utilizing the historical production data training set; wherein, the transducer is a transforming neural network;

The characteristic data category input by the coke quality prediction model comprises: setting quality index parameters of the matched coal and the production environment of the coke oven; outputting the coke quality prediction model as quality index parameters of the coke;

the characteristic data category input by the coking energy consumption prediction model comprises: setting quality index parameters of the matched coal and the production environment of the coke oven; outputting the coking energy consumption prediction model as the usage amount of coke oven gas;

the characteristic data categories input by the coke yield prediction model comprise: setting class index parameters in the production environment of the coke oven; and outputting the coke yield prediction model as the coke tapping number.

4. The multi-segment joint multi-objective optimization method for coking production according to claim 1, wherein the constrained multi-objective optimization problem model is expressed as:

；

wherein st is a constraint condition symbol,optimizing problem model for multiple objectives->Representing a coke quality prediction model,/->Representing a coking energy consumption prediction model, < >>Represents a coke yield predictive model,/->Representing a multi-objective optimization problem objective functionMin () is a minimization function.

5. The multi-segment joint multi-objective optimization method for coking production according to claim 4, wherein selecting the optimal solution from the pareto optimal solution set as the optimal decision variable of the constrained multi-objective optimization problem model comprises:

Decomposing the constrained multi-objective optimization problem model into two constrained single-objective optimization problem models, expressed as:

；

wherein,and->Optimizing the problem model for two single objectives, +.>And->Is two single objective functions;

solving two constrained single-objective optimization problem models,the result of the solution is marked as X ₁ ，The result of the solution is marked as X ₂ X is taken as ₁ And X is ₂ All are referred to as ideal points;

defined as a virtual point X ₃ Eliminating solutions which violate constraint conditions in the pareto optimal solution set, obtaining a set Z of solutions meeting the constraint, and calculating ideal points and imaginary points of each solution in the set Z of solutions meeting the constraintIs expressed as:

；

wherein,to calculate the sum of the distances of a single solution from two ideal points in the set Z of solutions satisfying the constraint,to calculate the distance of a single solution from a hypothetical point in the set Z of solutions satisfying the constraint, dis () is a distance calculation function,selecting the solution with the minimum Dis1 value from the ith solution in the set Z of solutions meeting the constraint, and if only one solution exists, selecting the solution as the optimal solution to be used as the optimal decision variable of the constrained multi-objective optimization problem model; if a plurality of solutions exist, selecting the solution with the minimum Dis2 value in the plurality of solutions as the optimal solution to be used as the optimal decision variable of the constrained multi-objective optimization problem model.

6. A multi-section combined multi-objective optimization system for coking production, comprising:

The optimal solution selecting and deploying unit is used for selecting an optimal solution from the pareto optimal solution set as an optimal decision variable of the constrained multi-objective optimization problem model, generating an optimal solution based on the optimal decision variable, and using the optimal solution for production deployment;

S4, calculating the adaptability of the new population;

7. A processing apparatus, comprising: one or more processors; a memory for storing one or more programs;

wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any of claims 1-5.

8. A readable storage medium storing a computer program, which when executed by a processor implements the method of any one of claims 1-5.