A multi-objective decision analysis method for hydrocracking process based on transfer preference learning

Abstract

The application discloses a kind of based on the multi-objective decision analysis method of hydrogenation cracking process of transferable preference learning.This method belongs to the field of chemical process modeling and multi-objective optimization, aims at solving the problem of high computational cost, slow model convergence and decision maker preference difficult to efficiently integrate in multi-objective optimization in hydrogenation cracking process.This method constructs the multi-objective evolutionary algorithm (TDP-MOEA) driven by transferable preference learning, introduces the preference information of decision maker in the optimization process, divides the target space into region of interest (ROI) and non-focus area, and uses Kriging (Kriging) surrogate model to approximate the sub-problems in ROI.To improve the robustness and generalization ability of surrogate model, the algorithm realizes the mutual information transmission between sub-models through sample migration and parameter migration.Experiments show that this method can realize efficient optimization and accurate decision of hydrogenation cracking process under fewer expensive evaluation times, and has wide industrial application potential.

Description

Technical Field

[0001] This invention relates to the fields of multi-objective optimization technology and complex industrial process optimization technology, specifically but not limited to a preference-heuristic multi-objective evolutionary algorithm (TDP-MOEA) based on transfer learning for oil refining hydrocracking processes, which is suitable for solving expensive multi-objective optimization problems. Background Technology

[0002] Hydrocracking, a core process in the oil refining industry, converts heavy oil fractions (such as vacuum gas oil) into low-boiling-point, high-value products (such as low-sulfur diesel and high-smoke-point jet fuel) under high temperature and pressure conditions, which is crucial for improving the quality of refined products. This process typically employs a dual-bed catalytic reactor. The first reactor decomposes sulfur- and nitrogen-containing compounds, while the second reactor performs hydroisomerization and hydrocracking on the processed liquid fraction.

[0003] From an optimization perspective, hydrocracking is a typical multi-objective optimization problem, requiring optimization of operating conditions such as feed flow rate, reaction pressure, and temperature to balance multiple objectives including product yield and cost. However, this process suffers from two major technical challenges: First, the evaluation cost is high. Hydrocracking simulation models built using process simulation software are highly nonlinear, making performance evaluation under new operating conditions time-consuming and costly, hindering the efficient application of traditional optimization methods. Second, the search efficiency is low. Traditional multi-objective evolutionary algorithms require a complete Pareto solution set and Pareto front, but in real-world scenarios, Pareto dominance relationships are difficult to distinguish in high-dimensional objective spaces, and numerous evaluation operations exceed industrially acceptable costs. Furthermore, decision-makers struggle to sift through the massive solution set to find the optimal solution for the region of interest.

[0004] While existing preference-driven multi-objective evolutionary algorithms can focus on regions of interest, they suffer from drawbacks such as insufficient flexibility of preference models, weak generalization ability of surrogate models, and ineffective utilization of information between subproblems, making it difficult to meet the optimization needs of complex industrial processes such as hydrocracking. Therefore, there is an urgent need for a multi-objective optimization method that can efficiently integrate decision-maker preferences, reduce evaluation costs, and improve optimization accuracy. Summary of the Invention

[0005] To address the pain points of high evaluation costs and low search efficiency in existing technologies, this invention proposes a multi-objective decision analysis method for hydrocracking processes based on transferable preference learning. By constructing a novel preference model with m+1 reference vectors, the method accurately defines the region of interest. It integrates the Kriging surrogate model with the expectation-improved filling criterion to reduce expensive evaluation costs. Furthermore, it innovatively introduces a sample + parameter dual-layer transfer learning strategy to enhance model robustness and population diversity, achieving efficient solutions for multi-objective optimization of the hydrocracking process. This provides oil refining companies with an operational optimization solution that is both economical and practical.

[0006] The technical solution to achieve the purpose of this invention is as follows:

[0007] A multi-objective decision analysis method for hydrocracking processes based on transfer preference learning includes:

[0008] S1. Preference Expression and Region of Interest (ROI) Definition: Based on the preference vector given by the decision-maker. and desired region size parameter Generate the Region of Interest (ROI) and construct within the ROI. There are reference vectors, where The number of objective functions;

[0009] S2. Proxy Model Establishment: An initial sample set is generated based on Latin hypercube sampling. The samples are evaluated using the true objective function. Multiple Kriging proxy models are constructed based on the reference vector to approximate the fitness of different sub-problems.

[0010] S3. Model-driven evolutionary search: Candidate solutions are generated using the Expected Improvement (EI) criterion, and sample selection and evolutionary updates are performed by combining the predicted values ​​and uncertainties of the Kriging model.

[0011] S4. Transfer learning mechanism: Sample transfer and parameter transfer are introduced during the evolution process. During sample transfer, high-performance individuals are selected from different sub-problems to be used for training other sub-models. Parameter transfer is based on feature importance analysis and adaptive weight strategy to realize mutual information transfer between sub-model parameters.

[0012] S5. Determining the optimal solution set: Obtain the optimal non-dominated solution set within the ROI through non-dominated sorting, forming the optimal decision scheme that satisfies the preference constraints.

[0013] Preferably, the boundary determination step of the region of interest (ROI) includes:

[0014] The decision-maker provides a unitized preference vector. Set area control parameters , The boundary vector set is calculated by representing the decision-maker's preference weight for the m-th objective. :

[0015]

[0016] Through linear interpolation formula Generate a set of reference vectors uniformly distributed within the ROI. This is used to guide the search direction of the algorithm. This step can be based on parameters. The value of can be flexibly adjusted to adjust the range of the region of interest, and the equiangular distribution characteristics of the reference vector are maintained in the high-dimensional target space, thereby ensuring that the search process has good balance and representativeness while meeting the decision-maker's preferences.

[0017] Preferably, the establishment of the proxy model includes the following steps:

[0018] The Latin hypercube sampling method (LHS) is used to generate N initial sample points in the decision space; for each sample point, the true objective function is called to calculate its multi-objective response value F(x). i Based on the angular relationship between the sample points and each reference vector, the samples are divided into corresponding sub-problem subgroups; a kriging model is independently constructed for each subgroup, and its correlation function uses a weighted Gaussian kernel:

[0019]

[0020] in This represents the correlation parameter of the k-th input variable. For smoothing coefficients, and These are any two input variables from the k-th group of input variables.

[0021] The model hyperparameters are obtained by maximizing the log-likelihood function:

[0022]

[0023] Where S is the sample covariance matrix and N is the number of input variables. This represents the sample standard deviation.

[0024] Preferably, the sampling and evolution process employs a desired improved EI sampling function. As a search criterion, it is calculated as follows:

[0025]

[0026] in It is the normal cumulative distribution function. Let be the probability density function. This represents the minimum fitness value of the current population. and These represent the predicted value and uncertainty of the Kriging model, respectively. This function, while maintaining a balance between exploration and development, prioritizes sampling in high-potential improvement regions, thereby improving the convergence speed of the ROI region.

[0027] Preferably, the transfer learning mechanism includes two levels of knowledge transfer:

[0028] (1) Sample transfer: In each generation of evolution, the top 11d−1+25 sample individuals with the best fitness values ​​are selected from multiple sub-problems as a shared sample pool; each sub-model introduces data from the shared sample pool during training to achieve data transfer across sub-problems, thereby improving the fitting accuracy of local models and global exploration capabilities.

[0029] (2) Parameter transfer: Identify the key parameter index set of each sub-model through feature selection algorithm. Based on group proportion With adaptive weights Update parameters:

[0030]

[0031] in The number of evaluations performed. To maximize the number of evaluations, , Indicates the first The first Kriging model One parameter.

[0032] 6. The method according to claim 1:

[0033]

[0034] in The angle between the individual and the reference vector. For a fixed angle between reference vectors, Indicates the first The distance from the target vector of each individual to the origin. This is the penalty function.

[0035] This design promotes the solution set to quickly approach the Pareto front in the early stage and enhances the diversity of the solution set in the later stage, ensuring the uniform distribution of the final ROI region.

[0036] Beneficial effects of the present invention

[0037] 1. Significant advantages in efficiency and cost: By accurately defining the ROI using m+1 reference vectors and combining the Kriging surrogate model and EI filling criterion, the actual number of evaluations of the hydrocracking process is controlled within 500, which is more than 40% less than traditional algorithms. It can still output effective solutions stably in small ROI scenarios, while most existing algorithms are prone to having no solution, which greatly reduces the time and economic cost of industrial optimization.

[0038] 2. Excellent performance and outstanding industrial value: The two-layer transfer learning strategy fully explores the common information between sub-problems, making the distribution of non-dominated solutions within the ROI more uniform. It is suitable for multi-objective conflict scenarios in hydrocracking, and the output results can directly guide industrial operations. It is universally applicable to various expensive multi-objective optimization scenarios. Attached Figure Description

[0039] Figure 1 A process diagram of the hydrocracking process of the present invention is shown.

[0040] Figure 2 A schematic diagram of the preference model is shown.

[0041] Figure 3 The parallel coordinate plot of the Pareto front obtained by the TDP-MOEA algorithm in the hydrocracking optimization problem is shown. Detailed Implementation

[0042] To further understand the present invention, preferred embodiments of the present invention are described below in conjunction with examples. However, it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the scope of the claims of the present invention.

[0043] The description in this section pertains only to typical embodiments, and the present invention is not limited to the scope of the embodiments described. Combinations of different embodiments, substitution of some technical features in different embodiments, and substitution of similar or identical prior art with some technical features in the embodiments are also within the scope of the description and protection of the present invention.

[0044] Taking the hydrocracking process as an example, this paper introduces a multi-objective decision analysis method for the hydrocracking process based on transferable preference learning.

[0045] Step 1: Define the optimization objective and the boundary of decision variables

[0046] according to Figure 1 The process diagram and reaction principle of the hydrocracking process are shown in Table 1-2. The selected optimization objectives and optimization decision variables are also shown in Table 1-2.

[0047] Table 1 Optimization Targets for Hydrocracking

[0048]

[0049] Table 2 Optimization Decision Variables for Hydrocracking

[0050]

[0051] Step 2: Preference Expression and ROI Definition

[0052] 1) Input decision-maker preference parameters: unit preference vector Wpre = {0.2, 0.2, 0.2, 0.2, 0.2}, ROI size parameter ε=0.5. See the preference model. Figure 2 .

[0053] 2) Generate the extreme vector W b Since the target number m=5, the extreme vectors are (1,0,0,0,0), (0,1,0,0,0), (0,0,1,0,0), (0,0,0,1,0), (0,0,0,0,1), (0,0,0,0,1).

[0054] 3) Calculate the ROI boundary vector: using the formula This yields 5 boundary vectors;

[0055] 4) Generate the ROI reference vector set W ROI : Through linear interpolation formula Calculate 5 reference vectors and W pre The merging results in six uniformly distributed reference vectors, which define the ROI range.

[0056] Step 3: Initialize the population and build the Kriging proxy model

[0057] 1) Generate initial population: Using the Latin hypercube sampling method, 126 initial individuals are generated within the decision variable boundary (population size adapted to 5-objective optimization scenario).

[0058] 2) Realistic evaluation of the initial population: Through the interface between MATLAB and Aspen HYSYS, the decision variables of the initial individuals are input into the simulation model to obtain the 5 objective function values ​​of each individual and calculate the individual fitness;

[0059] 3) Divide the population into subpopulations and train the model: Based on the angle between the individual and the 6 reference vectors, the initial population is divided into 6 subpopulations. 11×8-1+25=112 individuals with the best fitness are selected from each subpopulation (d=8 is the dimension of the decision variable). 6 Kriging surrogate models are trained to approximate the fitness value of each subproblem. At the same time, the initial population is stored in the archive set and the current number of evaluations FEs=126 is recorded.

[0060] Step 4: Evolutionary Iteration and Transfer Learning Optimization

[0061] Set the iteration termination condition: Maximum number of evaluations Max FEs = 500. When FEs ≤ 500, perform the following operations in a loop:

[0062] 1) Offspring generation: Based on the Kriging model of each subpopulation, the expected improvement (EI) filling criterion is used to screen high-potential individuals and generate offspring individuals for each subpopulation. The cumulative number of offspring is consistent with the initial population (126).

[0063] 2) Realistic evaluation of offspring: Input the decision variables of offspring individuals into the Aspen HYSYS simulation model, obtain the objective function value and fitness, and store them in a temporary set R;

[0064] 3) Archive update: Merge descendant individuals in R into the archive to ensure that the archive covers all valid individuals generated in the iteration;

[0065] 4) Two-layer transfer learning update model:

[0066] Sample transfer: The 112 individuals with the best fitness in each subpopulation were incorporated into the Kriging model training samples of other subpopulations to enrich data diversity;

[0067] Parameter migration: Every 5 generations, the key parameter set for each subproblem is selected using the binary particle swarm optimization algorithm, based on the formula. Update the model hyperparameters, where Adaptive weight adjustment;

[0068] 5) Iterative update: FEs accumulates the number of offspring evaluations (126 times per iteration), and the number of generations Gen = Gen + 1.

[0069] Step 5: ROI Optimal Solution Selection and Output

[0070] 1) Non-dominated sorting: Perform non-dominated sorting on all individuals (500 in total) in the archive set to obtain non-dominated levels F1 (optimal level), F2, etc., and the Pareto front is as follows. Figure 3 As shown.

[0071] 2) Filtering the optimal solution within the ROI: Extract individuals located within the ROI boundary from the F1 level to form the final optimal solution set Pop. ROI ;

[0072] The description and application of the present invention herein are illustrative and not intended to limit the scope of the invention to the embodiments described above. The effects or advantages described in the specification may not be apparent in actual experimental cases due to uncertainties in specific conditions or other factors, and such descriptions are not intended to limit the scope of the invention. Variations and modifications to the embodiments disclosed herein are possible, and various substitutions and equivalents of the components in the embodiments are well known to those skilled in the art. It should be understood by those skilled in the art that the invention can be implemented in other forms, structures, arrangements, proportions, and with other components, materials, and parts without departing from the spirit or essential characteristics of the invention. Other variations and modifications can be made to the embodiments disclosed herein without departing from the scope and spirit of the invention.

Claims

1. A multi-objective decision analysis method for hydrocracking processes based on transfer preference learning, characterized in that, include: S1. Preference Expression and Region of Interest (ROI) Definition: Based on the preference vector given by the decision-maker. and desired region size parameter Generate the Region of Interest (ROI) and construct within the ROI. There are reference vectors, where The number of objective functions; S2. Proxy Model Establishment: An initial sample set is generated based on Latin hypercube sampling, the samples are evaluated using the true objective function, and multiple Kriging proxy models are constructed based on the reference vector to approximate the fitness of different sub-problems. S3. Model-driven evolutionary search: Candidate solutions are generated using the expected improvement criterion, and sample selection and evolutionary updates are performed by combining the predicted values ​​and uncertainties of the Kriging model. S4. Transfer learning mechanism: Sample transfer and parameter transfer are introduced during the evolution process. During sample transfer, high-performance individuals are selected from different sub-problems to be used for training other sub-models. Parameter transfer is based on feature importance analysis and adaptive weight strategy to realize mutual information transfer between sub-model parameters. S5. Determining the optimal solution set: Obtain the optimal non-dominated solution set within the ROI through non-dominated sorting, forming the optimal decision scheme that satisfies the preference constraints.

2. The multi-objective decision analysis method according to claim 1, characterized in that, The steps for determining the boundary of the Region of Interest (ROI) include: The decision-maker provides a unitized preference vector. Set area control parameters , The boundary vector set is calculated by representing the decision-maker's preference weight for the m-th objective. : Through linear interpolation formula Generate a set of reference vectors uniformly distributed within the ROI. This is used to guide the search direction of the algorithm.

3. The method according to claim 1, characterized in that, The establishment of the proxy model includes the following steps: The Latin hypercube sampling method (LHS) is used to generate N initial sample points in the decision space; for each sample point, the true objective function is called to calculate its multi-objective response value F(x). i Based on the angular relationship between the sample points and each reference vector, the samples are divided into corresponding sub-problem subgroups; a kriging model is independently constructed for each subgroup, and its correlation function uses a weighted Gaussian kernel: in This represents the correlation parameter of the k-th input variable. For smoothing coefficients, and These are any two input variables from the k-th group of input variables; The model hyperparameters are obtained by maximizing the log-likelihood function: Where S is the sample covariance matrix and N is the number of input variables. This represents the sample standard deviation.

4. The method according to claim 1, characterized in that, The sampling and evolution process employs an expected improvement EI acquisition function. As a search criterion, it is calculated as follows: in It is the normal cumulative distribution function. Let be the probability density function. This represents the minimum fitness value of the current population. and These represent the predicted values ​​and uncertainties of the Kriging model, respectively. This function prioritizes sampling in high-potential improvement regions while ensuring a balance between exploration and development, thereby improving the convergence speed of ROI regions.

5. The method according to claim 1, characterized in that, The transfer learning mechanism includes two levels of knowledge transfer: (1) Sample transfer: In each generation of evolution, the top 11d−1+25 sample individuals with the best fitness values ​​are selected from multiple sub-problems as a shared sample pool; each sub-model introduces data from the shared sample pool during training to achieve data transfer across sub-problems, thereby improving the fitting accuracy of local models and global exploration capabilities. (2) Parameter transfer: Identify the key parameter index set of each sub-model through feature selection algorithm. Based on group proportion With adaptive weights Update parameters: in The number of evaluations performed. To maximize the number of evaluations, , Indicates the first The first Kriging model One parameter.

6. The method according to claim 1: in The angle between the individual and the reference vector. For a fixed angle between reference vectors, Indicates the first The distance from the target vector of each individual to the origin. This is the penalty function.

Images