A method for quantitative prediction of combustion state of flare and optimization of optimal combustion steam quantity

Abstract

A method for quantitatively predicting the combustion state of a flare and optimizing the amount of combustion-supporting steam is disclosed, belonging to the field of intelligent environmental protection technology. The method includes the following steps: based on computational fluid dynamics, simulation modeling is performed on a dual-steam combustion-supporting flare to obtain data on exhaust gas composition and velocity, combustion-supporting steam quantity, and burnout rate; based on the simulation data, an LSTM network is used to establish a flare combustion state prediction model; the NSGA-III algorithm is used to optimize the amount of combustion-supporting steam required for the flare; and the optimization algorithm is improved into a dynamic optimization algorithm. This invention solves the problem of scarce actual flare data by obtaining a large amount of reliable data through simulation of actual working conditions in software, solves the problem of slow simulation model calculation and difficulty in practical application through neural network modeling, and solves the coupling relationship between energy consumption and combustion state through optimization algorithms, enabling rapid and accurate quantitative determination of the flare combustion state and the required amount of combustion-supporting steam.

Description

Technical fields:

[0001] This invention pertains to a flare gas emission prediction and monitoring system for petrochemical enterprises. It utilizes data modeling and optimization algorithms to accurately determine the flame combustion state of the flare and optimize set values, thereby saving combustion steam while ensuring complete combustion of exhaust gas, improving the combustion efficiency of the flare, and saving a significant amount of energy. Background technology:

[0002] Vent systems are specialized combustion facilities in petrochemical plants, refineries, and other factories for combustible and toxic gases that cannot be recovered or reprocessed from the equipment. The exhaust gases from vent systems, released into the air after complete combustion, can cause significant harm to the environment and the lives of people in surrounding areas. Traditional smoke and chemical sensors are difficult to operate normally for extended periods under high temperatures and are challenging to install and maintain. Long-range intelligent image sensing can only detect black smoke produced by the vent; if the exhaust gas is not completely combusted and no black smoke is produced, this method is of limited effectiveness. With the rapid development of artificial intelligence neural networks, the integration of artificial intelligence technology with the petrochemical industry is being extensively studied.

[0003] David et al. applied computational fluid dynamics to quantitatively analyze the flare combustion process based on fundamental physics and turbulence and chemical models. They performed simulation modeling on Fluent software, conducting parameter analysis on the combustion steam / fuel and air / fuel ratios, and providing combustion efficiency calculated from temperature, species distribution, and material balance. They observed how hydrocarbons undergo incomplete combustion under high steam / fuel or air / fuel ratios until the flame extinguishes. The study shows that applying computational fluid dynamics to simulate flare burners can obtain realistic prediction data; however, the simulation model runs slowly, and the composition and velocity of the flare gas change in real time under actual operating conditions, making it difficult for simulations to meet the requirements of real-time performance and flexibility.

[0004] Qiao et al. designed a deep dual-channel neural network (DCNN) for smoke detection, sequentially connecting multiple convolutional layers and max-pooling layers, selectively attaching batch normalization layers to each convolutional layer to reduce overfitting and accelerate training. Two important components were introduced: skip connections to avoid gradient vanishing and improve feature propagation, and global average pooling to reduce the number of parameters and alleviate overfitting. A cascaded operation was deployed to combine the two deep sub-networks, complementing each other. However, this method can only identify the presence or absence of black smoke during flare burning, enabling only a qualitative study of the flare's combustion state, and cannot qualitatively determine the flare's combustion state and various parameters.

[0005] The inventors discovered that in existing research, actual operating data of flares is extremely scarce and difficult to obtain. There are two existing studies on the combustion state of flares: one is flare combustion modeling and simulation based on computational fluid dynamics (CFD). Although this method is accurate, the calculation is extremely slow and cannot be used for actual control, making it inflexible; the other is flare combustion state judgment based on image recognition. Although this method is fast, it can only judge the black smoke produced by the flare, and can make a qualitative judgment but cannot make a quantitative and accurate judgment. Summary of the Invention:

[0006] To address the aforementioned problems, this invention proposes a method for quantitatively predicting the combustion state of a flare and optimizing the amount of combustion-supporting steam. The main objective of this invention is to combine computational fluid dynamics modeling and simulation with artificial intelligence neural networks to construct a neural network data model to solve the problem of quantitatively determining the flame burnout rate. The NSGA-Ⅲ optimization algorithm is used to optimize the amount of combustion-supporting steam required for complete and clean combustion of the flare, and the optimization algorithm is further improved to make it more suitable for actual dynamic working conditions.

[0007] A method for quantitatively predicting the combustion state of a flare and optimizing the amount of combustion-supporting steam is characterized by: acquiring flare combustion data through CFD modeling and simulation; establishing a flare combustion state prediction model using an LSTM neural network; and optimizing the setpoint of the combustion-supporting steam during the flare combustion process using the NSGA-Ⅲ multi-objective optimization algorithm. The method includes the following steps:

[0008] Step 1: Use CFDFluent to build a model and perform simulation calculations for the dual-steam combustion type flare to obtain combustion state data: exhaust gas composition and flow rate, combustion steam quantity and burnout rate;

[0009] Step 2: Establish a combustion transition prediction model for the flare using an LSTM neural network. The network topology consists of three layers: an input layer, a hidden layer, and an output layer. The inputs are the exhaust gas component concentration, exhaust gas velocity, combustion steam I velocity at time T1, combustion steam II velocity at time T1, and combustion zone temperature. The outputs are the exhaust gas burnout rate, combustion steam I velocity at time T2, and combustion steam II velocity at time T2.

[0010] Step 3: Design optimization objectives based on the flare combustion steam energy consumption and flare combustion state. Flare combustion needs to ensure maximum complete combustion of exhaust gas while minimizing combustion steam consumption. These two objectives are conflicting, therefore, flare combustion can be viewed as a multi-objective optimization problem.

[0011] The optimization objectives for the combustion of the flare are f1(t) and f2(t), where f1(t) is the energy consumption index, i.e., the unit steam and electricity consumption for combustion of the flare. E d For electricity consumption, D d The amount of steam produced. f2(t) is the combustion efficiency of the flare (1-η). c ),

[0012] The CO2 considered here refers only to the CO2 produced by the flame, not the CO2 that is part of the fuel stream.

[0013] Based on the multi-objective optimization model min{f1(t),f2(t)}, the optimization setpoints for the flow rates of combustion steam I and combustion steam II at time T2 of the flare are sought.

[0014] Step 4: Optimize the key parameter setpoints of the flare stack using the NSGA-Ⅲ multi-objective optimization algorithm. The output values ​​of the LSTM combustion state prediction model—burnout rate, combustion steam I velocity at time T2, and combustion steam II velocity at time T2—are used as input values ​​for the optimization algorithm. The optimal combustion steam setpoints for the next time step are then optimized based on multi-objective optimization.

[0015] Step 5: Incorporate the environmental monitoring and response mechanisms into NSGA-III and improve them into a dynamic multi-objective optimization algorithm to match the actual operating conditions of the flare combustion. The steady-state detection method is selected as the environmental change detection mechanism, and linear regression model prediction is used as the environmental change response mechanism.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] The method described in this invention uses CFD modeling and simulation technology to solve the problems of difficult acquisition and insufficient data volume of combustion state data for flare torches, obtaining a large amount of combustion state data. It uses neural network modeling technology to establish an LSTM combustion state prediction model for the flare torch, solving the problems of overly complex quantitative analysis of combustion state and slow computation. It uses multi-objective genetic algorithm optimization technology to obtain the amount of combustion-supporting steam required for "clean combustion" of the flare torch, so as to control the steam valve of the flare torch. This invention can achieve rapid quantitative judgment of the combustion state of flare torch exhaust gas, while determining the optimal amount of combustion-supporting steam to achieve combustion standards and save steam as much as possible. It can effectively help petrochemical plant flare torch systems reduce air pollution and achieve energy conservation and emission reduction. Attached image description:

[0018] Figure 1 Schematic diagram of steam injection in a dual-stream steam combustion-supporting flare.

[0019] Figure 2 : NSGA-Ⅲ algorithm flowchart.

[0020] Figure 3: A schematic diagram of the reference plane for the four-part, three-objective problem.

[0021] Figure 4 : Schematic diagram of the relationship between population individuals and reference points. Detailed implementation method:

[0022] The following is a detailed description of the implementation examples of the present invention. These implementation examples are carried out under the premise of the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the protection scope of the present invention is not limited to the following implementation examples.

[0023] Step 1: A model was built using CFDFluent for the dual-stream steam-assisted flare, and simulation calculations were performed to obtain combustion state data: exhaust gas composition and flow rate, combustion steam quantity, and burnout rate. A schematic diagram of the steam injection for the dual-stream steam-assisted flare is attached. Figure 1 .

[0024] Choose, but not limited to, ANSYS's ICEMCFD software to build a physical model of a dual-steam combustion-assisted flare. After building the model, perform boundary delineation, generate tables for Fluent calculations, save and import the model into Fluent software for simulation calculations.

[0025] After running the computational model, the mass fraction of key substances in the combustion zone of the flare torch can be obtained based on the mass balance of the flame region. In Fluent, this is calculated by summing density, mass fraction of substances, velocity vector, and the projected surface area of ​​the region.

[0026] m i =∑ cells ρx i vA (2)

[0027] Where m i Let ρ be the boundary mass flow rate, ρ be the density, and x be the mass flow rate. i Let v be the mass fraction of the substance, v be the velocity vector, and A be the surface area of ​​the region.

[0028] After obtaining the mass flow rate of the material in the combustion zone of the flare, the burnout rate (CE) and destruction removal rate (DRE) of the flare exhaust gas can be accurately calculated.

[0029]

[0030] In the formula, M is the total amount of hydrocarbons in the flare gas, and m is the amount of unburned hydrocarbons in the flue gas.

[0031] Step 2: Use an LSTM neural network to establish a combustion transition prediction model for the flare.

[0032] Based on the combustion reactions of common main components in flares, petrochemical industry standards, and the chemical reaction CO + H2O → CO2 + H2 after adding steam, the amount of combustion-supporting steam required at the current moment can be preliminarily calculated.

[0033] Among them G steam For the amount of combustion-supporting steam, q cm M is the flare gas flow rate. c This represents the average molecular weight of hydrocarbons in the exhaust gas. The calculated combustion vapor volume and the internationally required burnout rate of 98% are used as standards to exclude unrealistic or outlier data.

[0034] Based on expert experience and actual operational data from the petrochemical industry, the data was screened to exclude outliers. The data was then filtered and normalized.

[0035] Where X represents the original data, X max X min These represent the maximum and minimum values ​​of the data, X. * This is a scaled-down version of the original data obtained after linear normalization.

[0036] In the formula, μ and σ are the mean and standard deviation of the input data, respectively, and z is the output data mapped to a mean of 0 and a standard deviation of 1.

[0037] An LSTM neural network model was built in the Python environment. The preprocessed data was divided into three parts: 70% training data, 20% test data, and 10% validation data. Supervised learning was used, and the model accuracy was corrected after training to obtain a prediction model for the combustion state of the flare.

[0038] Step 3: Design optimization objectives based on the energy consumption of combustion steam in the flare and the combustion state of the flare. Flare combustion needs to ensure the maximum complete combustion of exhaust gas while minimizing the consumption of combustion steam. These two objectives are conflicting, so flare combustion can be viewed as a multi-objective optimization problem.

[0039] The optimization objectives for the combustion of the flare are f1(t) and f2(t), where f1(t) is the energy consumption index, i.e., the unit steam and electricity consumption for combustion of the flare. E d For electricity consumption, D d The amount of steam produced. f2(t) is the combustion efficiency of the flare (1-η). c ).

[0040] The CO2 considered here refers only to the CO2 produced by the flame, not the CO2 that is part of the fuel stream.

[0041] Based on the multi-objective optimization model min{f1(t),f2(t)}, the optimization setpoints for the flow rates of combustion steam I and combustion steam II at time T2 of the flare are sought.

[0042] Step 4: Optimize the setpoints of key parameters for the flare stack using the NSGA-Ⅲ multi-objective optimization algorithm. Use the output values ​​of the LSTM combustion state prediction model as the input values ​​for the optimization algorithm, and optimize the setpoints according to the multi-objective optimization model.

[0043] The overall flowchart of the NSGA-Ⅲ algorithm can be found in the appendix. Figure 2 As shown, the population size is set to N=100, and the binary crossover probability p is simulated. c =1, crossover distribution index η1=10. The specific steps are as follows: polynomial mutation probability p m =1 / n, the variation distribution index η2 = 20, and the proportion of random diversity introduced is 20%.

[0044] (1) Assume a population A of size 100. Perform genetic operations on population A to obtain a population B of size N. Mix population A and population B to obtain a population C of size 200.

[0045] (2) Perform non-dominated sorting on population C to obtain individuals at non-dominated levels 1, 2, 3, 4... Add the individuals at non-dominated levels 1, 2, 3... to the next generation set D sequentially until the size of set D is greater than 100. Record the current non-dominated level L, and select K individuals from level L such that the sum of K and all individuals from previous levels equals N.

[0046] (3) Function scaling operation: A scaling operation is performed on the multi-objective functions to facilitate the next step of associating reference points. First, the minimum value of each objective dimension i in the M objective functions needs to be calculated, obtaining the minimum value Z corresponding to the i-th objective. i This Z i The set of points is called the ideal point set. Having obtained this ideal point set, we have the scalarization formula for the first step, as shown below:

[0047] In the formula f i v(x) is the target value in the target dimension i, f i (x) represents the original target value in target dimension i. Let i be the ideal point in the target dimension. The extreme point is found using the ASF (achievement scalarizing function) function, as shown in the formula below. This formula also applies to the target function in each dimension.

[0048] In the formula, f i ′(x) is the target value in the target dimension i, w i Let be the function weights along the target dimension i. Iterate through each function, find the individual with the smallest ASF value (the extreme point), calculate the intercept on the corresponding coordinate axis, and record the coordinate value of each point on the corresponding axis as 'a'. i . Get a i and z i After obtaining the specific values, normalization can be performed using the formula shown below, as in Formula 11. The intercepts on each normalized target axis are now located at... The hyperplane constructed using these intercept points will make

[0049]

[0050] (4) Individual reference points are associated with a four-partition, three-objective-function approach. The specific distribution of reference points can be found in the appendix. Figure 3 For each individual in the population, all vectors are traversed to find the nearest reference point, and the information of the reference point and the corresponding shortest distance are recorded. The distance from the individual to the reference point vector is described using the perpendicular distance; a 3D example is shown in the appendix. Figure 4 .

[0051] (5) Selecting offspring and deleting reference points. After non-dominated sorting, assuming that the total number of population members from the first non-dominated level to the FL level exceeds the population size N for the first time, then define S. t+1 Let S be the set containing all individuals in FL. t+1 The population size exceeds the pre-defined number of members, therefore a screening process is necessary. The first step in the screening is to iterate through each reference point and check for S points that do not contain FL. t+1 The number of references is counted, and the reference point with the fewest references is found, that is, the reference point associated with the fewest individuals in the population. Its number of references is recorded as p. j If p j =0, but if an individual in FL is associated with this reference point vector, then find the point with the smallest distance from it, extract it from FL, and add it to the selected next generation population. Set p j =p j +1; If no individual in FL is referenced to this reference point, then delete the reference point vector; if pj If the value is greater than 0, then the nearest reference point is selected until the population size is N.

[0052] Step 5: Incorporate environmental monitoring and response mechanisms into NSGA-III. Steady-state monitoring is selected as the environmental change monitoring mechanism, and linear regression model prediction is used as the environmental change response mechanism.

[0053] The main idea of ​​re-evaluation is to select a subset of individuals from the population and re-evaluate them. If there is a difference between the objective function values ​​of two adjacent iterations, then the environment is considered to have changed. The steady-state detection method involves randomly arranging all individuals and then re-evaluating them one by one. If a difference is found in the weight evaluation of a particular individual, it is determined that the environment has changed, and the remaining individuals do not need to be evaluated.

[0054] The prediction mechanism establishes time series data for individuals associated with the same reference point, and uses a linear regression model to predict the initial population under new environments, thereby responding to environmental changes. Each reference point is configured according to the Pareto optimal solution at different times. (i = 1, 2, ..., N) r The time series is composed of these components, which reflects the changing patterns of the optimal solution. This represents the optimal solution at times t, t-1, etc., where i represents the reference point, and N represents the optimal solution at times t, t-1, etc. r This represents the total number of reference points.

[0055] Within the same environment, individuals associated with different reference points reflect the distribution of the Pareto optimal solution set in the current environment. Therefore, the time series associated with each reference point describes the movement of the Pareto solution set, and the prediction of the Pareto solution set position at time t+1 can be expressed as...

[0056] In the formula, i = 1, 2, ..., N r Let f represent the reference point, and f represent the linear regression prediction model: the prediction model is built using information from the previous two time points as historical information. Starting from the third time point, the following model is used to predict new individuals for individuals associated with the same reference point.

[0057] in, It is the prediction error at time t. This is the theoretical Pareto optimal solution. The prediction error at time t is fed back into the prediction at time t+1 to improve the accuracy of the prediction.

Claims

1. A method for quantitatively predicting the combustion state of a flare and optimizing the amount of combustion-supporting steam, characterized in that: Combustion data of the flare is obtained through CFD modeling and simulation; a combustion state prediction model for the flare is established using an LSTM neural network; and the NSGA-Ⅲ multi-objective optimization algorithm is used to optimize the setpoint of the combustion-supporting steam during the flare combustion process. The method includes the following steps: Step 1: Use CFDFluent to build a model and perform simulation calculations for the dual-steam combustion type flare to obtain combustion state data: exhaust gas composition and flow rate, combustion steam quantity and burnout rate; Step 2: Use an LSTM neural network to establish a prediction model for the combustion state of the flare. Step 3: Design optimization targets based on the combustion steam energy consumption of the flare and the combustion status of the flare; Step 4: Optimize the setpoint for combustion steam in the flare stack using the NSGA-Ⅲ multi-objective optimization algorithm; Step 5: Incorporate the environmental monitoring mechanism and environmental response mechanism into NSGA-Ⅲ and improve them into a dynamic multi-objective optimization algorithm to fit the actual working conditions of the flare combustion. In step 3, optimization objectives are designed based on the energy consumption of combustion steam in the flare and the combustion state of the flare; the combustion of the flare is regarded as a multi-objective optimization problem. The optimization objectives for the combustion of the flare are f1(t) and f2(t); f1(t) is the energy consumption index, i.e., the unit steam electricity consumption for combustion of the flare. E d For electricity consumption, D d f2(t) represents the amount of steam generated; f2(t) represents the combustion efficiency of the flare (1-η). c ), The CO2 considered here refers only to the CO2 produced by the flame, not the CO2 that is part of the fuel stream; Based on the multi-objective optimization model Optimize the setpoints for the flow rates of combustion steam I and combustion steam II at time T2 of the flare venting. Step 4: Use the NSGA-Ⅲ multi-objective optimization algorithm to optimize the setpoint of the combustion steam for the flare. Use the output values ​​of the LSTM combustion state prediction model—burnout rate, combustion steam I flow rate at time T2, and combustion steam II flow rate at time T2—as the input values ​​of the optimization algorithm, and optimize the optimal setpoint of the combustion steam for the next time step according to the multi-objective optimization.

2. The method for quantitative prediction of combustion state and optimization of optimal combustion steam quantity of a flare according to claim 1, characterized in that the steps are as follows: In the simulation modeling of the flare, a dual-stream steam combustion type flare is selected in CFDFluent for modeling and simulation calculation; After running the model, the mass fraction of important substances in the combustion zone of the flare torch is obtained based on the material balance of the flame zone; in Fluent, it is calculated by summing density, mass fraction of substances, velocity vector and surface area projection of the region. m i =∑ cells ρx i vA (1) Where m i Let ρ be the boundary mass flow rate, ρ be the density, and x be the mass flow rate. i Let v be the mass fraction of the substance, v be the velocity vector, and A be the surface area of ​​the region. After obtaining the mass flow rate of the material in the combustion zone of the flare, the burnout rate (CE) and destruction removal rate (DRE) of the flare exhaust gas are accurately calculated. In the formula, M is the total amount of hydrocarbons in the flare gas, and m is the amount of unburned hydrocarbons in the flue gas.

3. The method for quantitative prediction of combustion state and optimization of optimal combustion-supporting steam quantity of a flare according to claim 1, characterized in that, Step 2: Establish an LSTM combustion state prediction model based on the flare simulation data. The network topology consists of three layers: input layer, hidden layer, and output layer. The inputs are the exhaust gas component concentration, exhaust gas velocity, combustion steam I velocity at time T1, combustion steam II velocity at time T1, and combustion zone temperature. The outputs are the exhaust gas burnout rate, combustion steam I velocity at time T2, and combustion steam II velocity at time T2.

4. The method for quantitative prediction of combustion state and optimization of optimal combustion-supporting steam quantity of a flare according to claim 1, characterized in that, Step 5: Incorporate the environmental monitoring and response mechanisms into NSGA-Ⅲ; select the steady-state detection method as the environmental change monitoring mechanism, and use the linear regression model prediction as the environmental change response mechanism.

5. The method for quantitative prediction of combustion state and optimization of optimal combustion-supporting steam quantity of a flare according to claim 1, characterized in that, the first step: For a dual-steam combustion-assisted flare, a model was built using CFDFluent and simulation calculations were performed to obtain combustion state data: exhaust gas composition and flow rate, combustion-assisted steam quantity and burnout rate. Select the ICEMCD software to build a physical model of a dual-steam combustion-assisted flare. After the model is built, perform boundary division and generate a table for Fluent calculation. Save and import the model into Fluent software for simulation calculation. After running the calculation model, the mass fraction of important substances in the combustion zone of the flare torch flame is obtained based on the material balance of the flame zone; in Fluent, the mass fraction of substances is calculated by summing density, mass fraction of substances, velocity vector and projection of the region surface area. m i =∑ cells ρx i vA (2) Where m i Let ρ be the boundary mass flow rate, ρ be the density, and x be the mass flow rate. i Let v be the mass fraction of the substance, v be the velocity vector, and A be the surface area of ​​the region. After obtaining the mass flow rate of the material in the combustion zone of the flare, the burnout rate (CE) and the destruction and removal rate (DRE) of the flare exhaust gas are accurately calculated. In the formula, M is the total amount of hydrocarbons in the flare gas, and m is the amount of unburned hydrocarbons in the flue gas; Step 2: Use an LSTM neural network to establish a prediction model for the combustion transition of the flare; The required amount of combustion steam at the current moment is calculated. Among them G steam For the amount of combustion-supporting steam, q cm M is the flare gas flow rate. c The average molecular weight of hydrocarbons in the exhaust gas; using the calculated amount of combustion vapor and the internationally required burnout rate of 98% as standards, and excluding abnormal data that do not conform to reality; Data filtering and normalization; Where X represents the original data, X max X min These represent the maximum and minimum values ​​of the data, X. * This is a proportionally scaled version of the original data obtained after linear normalization. In the formula, μ and σ are the mean and standard deviation of the input data, respectively, and z is the output data mapped to a mean of 0 and a standard deviation of 1. An LSTM neural network model was built in the Python environment, and the preprocessed data was divided into data segments. Supervised learning was used to train the model and obtain a prediction model of the combustion state of the flare. Step 3: Design optimization objectives based on the combustion steam energy consumption and combustion status of the flare; treat the flare combustion as a multi-objective optimization problem. The optimization objectives for the combustion of the flare are f1(t) and f2(t), where f1(t) is the energy consumption index, i.e., the unit steam and electricity consumption for combustion of the flare. E d For electricity consumption, D d f2(t) represents the amount of steam generated; f2(t) represents the combustion efficiency of the flare (1-η). c ); The CO2 considered here refers only to the CO2 produced by the flame, not the CO2 that is part of the fuel stream; Based on the multi-objective optimization model Optimize the setpoints for the flow rates of combustion steam I and combustion steam II at time T2 of the flare venting. Step 4: Use the NSGA-Ⅲ multi-objective optimization algorithm to optimize the setpoints of key parameters of the flare; use the output value of the LSTM combustion state prediction model as the input value of the optimization algorithm, and optimize the setpoints according to the multi-objective optimization model; Set the population size N = 100 and simulate the binary crossover probability p. c =1, crossover distribution index η1=10. The specific steps are as follows: polynomial mutation probability p m =1 / n, the variation distribution index η2 = 20, and the proportion of random diversity introduced is 20%; (1) Assume a population A of size 100. Perform genetic operations on population A to obtain a population B of size N. Mix population A and population B to obtain a population C of size 200. (2) Perform non-dominated sorting on population C to obtain individuals at non-dominated levels 1, 2, 3, 4...; add the individuals at non-dominated levels 1, 2, 3... to the set D of the next generation of offspring in sequence until the size of set D is greater than 100; record the non-dominated level L at this time, and select K individuals from level L such that the sum of K and all individuals at previous levels equals N; (3) Function scaling operation: The multi-objective function is scalarized to facilitate the next step of associating reference points; First, the minimum value of each objective dimension i in the M objective functions needs to be calculated to obtain the minimum value Z corresponding to the i-th objective. i This Z i The set of points is the ideal point set; after obtaining this ideal point set, the scalarization formula for the first step is obtained, as shown below: In the formula f i ′(x) is the target value in the target dimension i, f i (x) represents the original target value in target dimension i. Let i be the ideal point in the target dimension; use the ASF function to find the extreme point, as shown in the formula below. This formula also applies to the target function in each dimension. In the formula, f i ′(x) is the target value in the target dimension i, w i Let a be the function weight on the target dimension i; iterate through each function, find the individual with the smallest ASF value (i.e., the extreme point), calculate the intercept on the corresponding coordinate axis, and record the coordinate value of each point on the corresponding axis as a. i ; get a i and z i After obtaining the specific values, normalization is performed according to the formula shown below, as in Formula 11; the intercept on each normalized target axis is now located at f. i n =1, the hyperplane constructed using these intercept points will make (4) Individual association reference point, select four partitions and three objective functions; for each individual in the population, traverse all vectors to find the reference point closest to each individual in the population, and record the information of the reference point and the corresponding shortest distance; (5) Selecting offspring and deleting reference points; After non-dominated sorting, assuming that the total number of population members from the first non-dominated level to the FL level exceeds the population size N for the first time, then define S t+1 Let S be the set containing all individuals in FL. t+1 The population size exceeds the pre-defined number of members, therefore a screening process is necessary; the first step in the screening is to iterate through each reference point and check the S points that do not contain FL. t+1 The number of references is counted, and the reference point with the fewest references is found, that is, the reference point associated with the fewest individuals in the population. Its number of references is recorded as p. j If p j =0, but if an individual in FL is associated with this reference point vector, then find the point with the smallest distance from it, extract it from FL, and add it to the selected next generation population. Set p j =p j +1; If no individual in FL is referenced to this reference point, then delete the reference point vector; if p j If the value is greater than 0, then select the nearest reference point from among them until the population size is N; Step 5: Incorporate environmental monitoring and response mechanisms into NSGA-III; select steady-state detection as the environmental change monitoring mechanism and linear regression model prediction as the environmental change response mechanism; The prediction mechanism establishes time series for individuals associated with the same reference point, and uses a linear regression model to predict the initial population under new environments, thereby responding to environmental changes; each reference point at different times follows the Pareto optimal solution. The time series is composed of these components, which reflects the changing patterns of the optimal solution. Let N represent the optimal solution at times t and t-1, where i represents the reference point and N is the reference point. r This represents the total number of reference points. In the same environment, individuals associated with different reference points reflect the distribution of the Pareto optimal solution set in the current environment; therefore, the time series associated with each reference point describes the movement of the Pareto solution set, and the prediction of the Pareto solution set position at time t+1 is expressed as follows: In the formula, i = 1, 2, ..., N r Let f represent the reference point, and f represent the linear regression prediction model: the prediction model is built using information from the previous two time points as historical information; starting from the third time point, the following model is used to predict new individuals for individuals associated with the same reference point. in, It is the prediction error at time t. This is the theoretical Pareto optimal solution.

Images