Robust optimization method and robust optimization device for emission uncertainty of cracking furnace

Abstract

The invention discloses a robust optimization method and a robust optimization device for emission uncertainty of a cracking furnace. The method comprises the following steps: respectively acquiring sampling data under corresponding distribution of uncertainty variables and decision variables of the cracking furnace; performing numerical simulation on the sampling data based on a turbulent combustion coupling model to obtain training set data; taking the sampling data as an input variable, taking the training set data as an output variable, and constructing a neural network agent model based on particle swarm optimization as an uncertain quantification model; quantizing decision variables based on an uncertain quantization model, and constructing a robust optimization model; and performing robust optimization on the decision variable through a dynamic adaptive genetic algorithm by means of a robust optimization model so as to obtain an ideal output working condition of the uncertainty variable. According to the method, the output uncertainty can be quantified on the basis of the input uncertainty of the fixed working condition, and the output quantity is optimized, so that the robust efficiency and precision of the emission of the cracking furnace under the influence of the uncertainty are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of combustion numerical simulation, and particularly relates to a robust optimization method for the emission uncertainty of a cracking furnace, a robust optimization device for the emission uncertainty of a cracking furnace, and a computer-readable storage medium. Background Art

[0002] China is a major ethylene-producing country in the world, and ethylene cracking furnaces are important equipment for producing ethylene. The radiant section in the cracking furnace is the main area where combustion reactions occur, generating a large amount of heat for heat transfer in the furnace tubes. Raw materials such as naphtha in the furnace tubes are then cracked to produce products such as ethylene and propylene. During the intense combustion process in the radiant section, a large amount of nitrogen oxides (NO X ) emissions are produced.

[0003] During the operation of industrial ethylene cracking furnaces, due to various industrial practical reasons, there is always a certain gap between the actual input values and the calibrated values. Therefore, the distribution of the actual input values and the calibrated values shows a slight fluctuation, which is the uncertainty existing in industrial inputs. This input with uncertain fluctuations enters the system and will lead to the uncertainty of the output. For industrial ethylene cracking furnaces, the uncertainty of the input will result in the non-fixed emissions of nitrogen oxides and ethylene yields. In addition, these uncertainties will also cause the lack of robustness in multiple outputs of industrial ethylene cracking furnaces, making it impossible to achieve better values for multiple objectives simultaneously, and at the same time, it will also affect the anti-interference ability of the system to fluctuations. Therefore, considering this phenomenon, it is necessary to perform robust optimization on the entire system in the presence of input uncertainty.

[0004] In this regard, traditional computational fluid dynamics (CFD) optimization research uses discrete methods to analyze the results by changing variables. However, this method also has limitations. For example, the optimal result may exist at a certain point, and it is impossible to evaluate the overall model, nor does it consider the interaction between variables. Therefore, the limitation of this deterministic modeling optimization method is that it does not consider the uncertainty in the actual production process, which is very likely to lead to an optimization result with poor system robustness.

[0005] At present, when conducting multi-objective optimization on the robust optimization objective of a system, the NSGA-II genetic algorithm is mostly used. In terms of the convergence speed, there is a situation where the population does not converge to the true Pareto front fast enough. Especially in problem scenarios such as complex multi-peaks, it may require a large number of iterations to approach the ideal front. Moreover, in terms of the approximation ability, for a Pareto front with a complex shape and discontinuity, it is sometimes difficult to accurately approximate, and there may be deficiencies such as uneven individual distribution and poor approximation effect in local areas, resulting in the inability to well and completely present the true optimal solution front shape.

[0006] In order to solve the above problems existing in the prior art, there is an urgent need in the art for a robust optimization technology for the emission uncertainty of a cracking furnace, which can quantify the output uncertainty under the condition of input uncertainty of a fixed working condition and optimize the output quantity, so as to improve the robust efficiency and accuracy of the cracking furnace under the influence of uncertainty. Summary of the Invention

[0007] The following gives a brief overview of one or more aspects to provide a basic understanding of these aspects. This overview is not an exhaustive survey of all conceived aspects, and neither aims to identify the key or decisive elements of all aspects nor attempts to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description given later.

[0008] In order to overcome the above-mentioned defects existing in the prior art, the present invention provides a robust optimization method for the emission uncertainty of a cracking furnace, a robust optimization device for the emission uncertainty of a cracking furnace, and a computer-readable storage medium, which can quantify the output uncertainty under the condition of input uncertainty of a fixed working condition and optimize the output quantity, so as to improve the robust efficiency and accuracy of the cracking furnace emission under the influence of uncertainty.

[0009] In addition, the above-mentioned robust optimization device for the emission uncertainty of a cracking furnace provided according to the second aspect of the present invention. This kind of robust optimization device for the emission uncertainty of a cracking furnace includes a memory and a processor. The processor is connected to the memory and is configured to implement the above-mentioned robust optimization method for the emission uncertainty of a cracking furnace provided according to the first aspect of the present invention.

[0010] In addition, according to the third aspect of the present invention, there is also provided a computer-readable storage medium, on which computer instructions are stored. When the computer instructions are executed by a processor, the above-mentioned robust optimization method for the emission uncertainty of a cracking furnace provided according to the third aspect of the present invention is implemented. Brief Description of the Drawings

[0011] After reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings, the above features and advantages of the present invention can be better understood. In the drawings, the components are not necessarily drawn to scale, and components having similar relevant characteristics or features may have the same or similar reference numerals.

[0012] Figure 1 A schematic flowchart of a robust optimization method for cracking furnace emissions uncertainty provided according to some embodiments of the present invention is shown;

[0013] Figure 2 A schematic diagram of the structural model of an ethylene cracking furnace provided according to some embodiments of the present invention is shown;

[0014] Figure 3A A schematic flowchart of obtaining sampling data provided according to some embodiments of the present invention is shown;

[0015] Figure 3B A schematic diagram of the result of a sample set after Latin hypercube sampling provided according to some embodiments of the present invention is shown;

[0016] Figure 4 A schematic flowchart of obtaining training set data provided according to some embodiments of the present invention is shown;

[0017] Figure 5A A schematic flowchart of constructing an uncertainty quantification model provided according to some embodiments of the present invention is shown;

[0018] Figure 5B A schematic flowchart of iteratively updating the velocity and position of particles to obtain the weights and thresholds of an optimized BP neural network provided according to some embodiments of the present invention is shown;

[0019] Figure 6 A schematic flowchart of the completion of constructing an uncertainty quantification model provided according to some embodiments of the present invention is shown;

[0020] Figure 7 A schematic flowchart of obtaining a robust optimization model provided according to some embodiments of the present invention is shown;

[0021] Figure 8 A schematic flowchart of dynamic adaptive robust optimization provided according to some embodiments of the present invention is shown;

[0022] Figure 9 A schematic diagram of the result of dynamic adaptive robust optimization provided according to some embodiments of the present invention; and

[0023] Figure 10 A schematic block diagram of a robust optimization device for cracking furnace emissions uncertainty provided according to some embodiments of the present invention is shown.

[0024] Reference numerals:

[0025] Steps S110 to S150;

[0026] Ethylene cracking furnace 200;

[0027] Sidewall burner 210;

[0028] Bottom burner 220;

[0029] Steps S121 to S124;

[0030] Steps S131 to S133;

[0031] Steps S310 to S340;

[0032] Steps S610 to S654;

[0033] Steps S141 to S144;

[0034] Steps S151 to S156;

[0035] Robust optimization device for cracking furnace emissions uncertainty 10;

[0036] Memory 101; and

[0037] Processor 102. Detailed implementation manners

[0038] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Although the description of the present invention will be introduced in conjunction with the preferred embodiments, this does not mean that the features of this invention are limited to this implementation manner. On the contrary, the purpose of introducing the invention in conjunction with the implementation manner is to cover other alternatives or modifications that may be extended based on the claims of the present invention. In order to provide a deep understanding of the present invention, many specific details will be included in the following description. The present invention can also be implemented without using these details. In addition, in order to avoid confusing or obscuring the key points of the present invention, some specific details will be omitted in the description.

[0039] In the description of the present invention, it should be noted that, unless otherwise clearly defined and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0040] In addition, the terms "upper", "lower", "left", "right", "top", "bottom", "horizontal", and "vertical" used in the following description should be understood as the orientations shown in this section and the related drawings. Such relative terms are only for convenience of description and do not represent that the devices described need to be manufactured or operated in a specific orientation, and thus should not be construed as a limitation on the present invention.

[0041] It is understood that although terms such as "first", "second", "third", etc. may be used herein to describe various components, regions, layers, and / or parts, these components, regions, layers, and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers, and / or parts. Therefore, the first component, region, layer, and / or part discussed below may be referred to as the second component, region, layer, and / or part without departing from some embodiments of the present invention.

[0042] As described above, traditional Computational Fluid Dynamics (CFD) optimization studies are carried out using discrete methods, and the results are analyzed by changing variables. However, this method also has limitations. For example, the optimal result may exist at a certain point, the overall model cannot be evaluated, and the interaction between variables is not considered. Therefore, the limitation of this deterministic modeling optimization method is that it does not consider the uncertainty in the actual production process, which is very likely to lead to an optimization result with poor system robustness. Currently, for multi-objective optimization of the robust optimization target of the system, the NSGA-II genetic algorithm is mostly used. In terms of the convergence speed, there is a situation where the population does not converge fast enough to the true Pareto front. Especially in problem scenarios such as complex multi-peaks, it may take more iteration times to approach the ideal front. Moreover, in terms of the approximation ability, for a Pareto front with a complex shape and discontinuity, it is sometimes difficult to accurately approximate, and there may be deficiencies such as uneven individual distribution and poor approximation effect in local areas, resulting in the inability to well present the true optimal solution front morphology.

[0043] To solve the above problems existing in the prior art, the present invention provides a robust optimization method for the uncertainty of the cracking furnace emissions, a robust optimization device for the uncertainty of the cracking furnace emissions, and a computer-readable storage medium, which can quantify the output uncertainty based on the input uncertainty under fixed operating conditions and optimize the output quantity, thereby improving the robust efficiency and accuracy of the cracking furnace emissions under the influence of uncertainty.

[0044] Please refer to Figure 1 , Figure 1The flowchart shows a robust optimization method for the emissions uncertainty of a cracking furnace according to some embodiments of the present invention.

[0045] As Figure 1 shown, in some embodiments of the present invention, the robust optimization method for the emissions uncertainty of a cracking furnace mainly may include the following steps. First, step S110 may be executed: sampling data corresponding to the uncertainty variables and decision variables of the cracking furnace are obtained respectively.

[0046] In some non - restrictive embodiments, the implementation object of the present invention may be an ethylene cracking furnace in an industrial cracking furnace. Specifically, an ethylene cracking furnace equipped with a multi - stage fuel burner may be selected as the research object of this embodiment. Since the present invention mainly studies the optimization of the generation of pollutants during combustion, therefore, the pollutant may be selected as the emission of nitrogen oxides with output characteristics, which is the main component of pollutant emissions.

[0047] Specifically, please refer to Figure 2 , Figure 2 The schematic diagram shows the structural model of an ethylene cracking furnace according to some embodiments of the present invention.

[0048] As Figure 2 shown, in some embodiments, the ethylene cracking furnace 200 mainly may include a side - wall burner 210 and a bottom burner 220. As key components in the heating system of the cracking furnace, they play an important role in the heat distribution in the furnace, combustion efficiency, and environmental protection performance. In view of the fact that both the side - wall burner 210 and the bottom burner 220 are located at the lower end in the ethylene cracking furnace 200, therefore, preferably, the position of the sampling point may be determined at the highest point A at the center of the furnace bottom, so as to obtain stable flue gas data.

[0049] In some alternative embodiments, multiple uncertainty input variables in the multi - stage fuel bottom burner 220 may be selected, for example, it may include the pre - heated air temperature, the primary gas inlet flow rate, the secondary gas inlet flow rate, and the side - wall burner flow rate, to perform uncertainty quantification (UQ). These uncertainty variables are independent of each other.

[0050] Further, please refer to Figure 3A , Figure 3A The flowchart shows the process of obtaining sampling data according to some embodiments of the present invention. As Figure 3A shown, the above - mentioned step S110 may be embodied as steps S111 - S112.

[0051] As Figure 3AAs shown, in some embodiments, step S111 may be first executed: by distribution assumption, define the uncertainty variables and decision variables corresponding to the input variables of the ethylene cracking furnace, and obtain the distributions of all input variables.

[0052] Specifically, in some alternative embodiments, the probability distribution of the uncertain input variables may be selected as the commonly used Gaussian distribution. The measurement uncertainty of the bottom burner 220 of the multi-stage fuel is usually tested to be 3%, so in this embodiment, the fluctuation range of the uncertainty variable can be selected as 3%. If the i-th uncertainty variable can be characterized as a Gaussian distribution with an average value and a standard deviation σ i = 1%μ i then under the Gaussian distribution, a 3% fluctuation can be approximately assumed to be 3σ. The distribution selection of the deterministic parameter, i.e., the decision variable, can be a uniform distribution. Therefore, the distributions of the respective uncertainty variables in this embodiment can be as shown in Table 1 below:

[0053] Table 1

[0054]

[0055]

[0056] Continuing as Figure 3A shown, then step S112 may be executed: by the Latin hypercube method, perform data sampling on the uncertainty variables to obtain sampling data that can completely represent the distributions of the uncertainty variables and the decision variables based on a small number of sample points.

[0057] In some embodiments, after defining the above uncertainty variables, the Latin hypercube method can be used to perform data sampling on the uncertainty variables. In this embodiment, by the Latin hypercube method, it is possible to stably and uniformly define, according to the experimental input variables, the requirement to completely represent the input variable distribution through a small number of sample points, greatly reducing the computational pressure of subsequent numerical simulations.

[0058] For example, refer to Figure 3B , Figure 3B which shows a schematic diagram of the sample set result after Latin hypercube sampling provided according to some embodiments of the present invention.

[0059] In Figure 3B the embodiment shown, the abscissa represents the input variable ξ i and the ordinate represents the input variable ξ j , where the input variables include 4 uncertainty variables and 1 decision variable, and ξ1, ξ2, ξ3, ξ4 respectively represent the above 4 uncertainty variables of the preheating temperature, the fuel flow rate at the primary port, the fuel flow rate at the secondary port, and the sidewall fuel flow rate, while ξ5 represents the deterministic decision variable.

[0060] Under specific working conditions, input variables can be divided into uncertain input variables and deterministic input variables. Among them, decision variables are deterministic input variables. In some alternative embodiments, the decision variable can be selected as the excess air coefficient of an ethylene cracking furnace, which is a variable determined by the operator and does not have fluctuations similar to uncertain variables, so it can be used as a decision variable.

[0061] As Figure 3B shown, when i = j, it represents a sampling histogram of an uncertain variable, and when i ≠ j, it represents a two-dimensional sampling scatter plot of two variables. In Figure 3B , whether from the sampling results of single variables or the scatter plots presented by multiple variables, through the Latin hypercube sampling method, the probability distribution space can be efficiently explored with relatively few sample points. It samples based on interval division, enabling the sample points to evenly cover the main areas within the entire distribution range. Compared with some irregular sampling methods, it can capture the key features of the distribution faster.

[0062] Continuing as Figure 1 shown, the robust optimization method for the uncertainty of the emissions of the cracking furnace provided by the present invention may further include step S120: Based on the turbulent combustion coupling model, perform numerical simulation on the sampling data to obtain training set data.

[0063] The data set obtained by the above sampling can be calculated through high-precision numerical simulation, so as to obtain the pollution emission data of all samples. Specifically, please refer to Figure 4 , Figure 4 shows a schematic flow chart of obtaining training set data according to some embodiments of the present invention. As Figure 4 shown, the above step S120 can be embodied as steps S121 to S124.

[0064] As Figure 4 shown, in some embodiments, step S121 can be executed first: Obtain the structural model of the cracking furnace and perform mesh division on it. Since CFD simulation is preferably used in this embodiment to perform numerical simulation on the sampling data. The basic principle of CFD simulation is to solve the control equations of fluid flow based on numerical methods. Before calculation, the continuous computational domain can be preferably discretized into many small units, and the density and quality of the mesh will directly affect the calculation accuracy.

[0065] Next, step S122 can be executed: According to the structural model after mesh division, determine the turbulent model, combustion model, and radiation heat transfer model during the cracking process of the cracking furnace.

[0066] Specifically, in some alternative embodiments, the turbulence model can be the standard k-ε model, the combustion model can be the finite rate / eddy dissipation-post processing NOx combustion model, and the radiative heat transfer model can be the DO radiation model. Those skilled in the art can understand that the specific selection schemes of the above-mentioned turbulence model, combustion model, and radiative heat transfer model are only a non-restrictive implementation manner provided by the present invention, aiming to clearly demonstrate the main concept of the present invention and provide a specific scheme convenient for the public to implement, rather than limiting the protection scope of the present invention.

[0067] After that, step S123 can be executed: Based on the turbulence model, combustion model, and radiative heat transfer model, a coupled turbulence combustion model is established.

[0068] Specifically, in some embodiments, the CFD simulation software Fluent 19.2 can be selected. Fluid calculations are performed on the Fluent 19.2 platform, the completed meshes are imported, the corresponding calculation models mentioned above are selected, the model calculation parameters are set, and a coupled turbulence combustion model is established.

[0069] After that, step S124 can be executed: Based on the coupled turbulence combustion model, the sampled data is numerically simulated by computational fluid dynamics to obtain training set data.

[0070] Specifically, in some embodiments, preferably, the sampled data is preprocessed and used as boundary conditions and input into the coupled turbulence combustion model for numerical simulation. Here, the preprocessing mainly involves unit conversion of these sampled data used as boundary conditions before inputting into the calculation. For example, converting kg / h to kg / s, or dividing the input quantity by 6 to distribute it to 6 input ports.

[0071] Furthermore, in some preferred embodiments, the output data of the coupled turbulence combustion model can be compared with the actual industrial operation data to ensure the simulation accuracy. Here, the actual data can refer to the real output data records of the cracking furnace operating under the same conditions provided by the factory.

[0072] Through the above step S120, the present invention can complete effective CFD calculations of the coupled turbulence combustion model, thereby achieving the optimization effect of accurately predicting pollutant emissions.

[0073] Continue as Figure 1 As shown, the robust optimization method for cracking furnace emission uncertainty provided by the present invention may further include step S130: Using the sampled data as input variables and the training set data as output variables, a neural network surrogate model optimized by the particle swarm algorithm is constructed as an uncertainty quantification model.

[0074] Specifically, please combineFigure 5A and Figure 6 are understood together that Figure 5A shows a schematic flow chart of constructing an uncertainty quantification model provided according to some embodiments of the present invention, Figure 6 shows a schematic completion flow chart of constructing an uncertainty quantification model provided according to some embodiments of the present invention.

[0075] As Figure 5A shown, the above step S130 can be embodied as steps S131 to S133.

[0076] As Figure 6 shown, before performing step S131, step S610 can be first executed: reading problem data. Through step S610, the training set data can be transmitted to the model. Then, step S620 can be executed: setting parameters of the Particle Swarm Optimization (PSO) algorithm.

[0077] Specifically, as Figure 5A shown, in some embodiments, step S131 can be first executed: creating a BP neural network and determining its topological structure.

[0078] Specifically, in some embodiments, the topological structure of the BP neural network is determined according to the number of input / output parameters of the samples, and can include the number of neurons in the input layer, hidden layer and output layer, as well as the selection of activation functions.

[0079] For example, in the above embodiment, the number of input variables is 5 (4 uncertainty variables and 1 decision variable), and the number of output variables is 1, so an input layer with 5 and an output layer with 1 can be constructed. There is an approximate relationship n2 = 2×n1 + 1 between the number n2 of neurons in the hidden layer neural network and the number n1 of neurons in the input layer. Therefore, the number of hidden layers can be selected as 11 here. Further, training parameters can be set. Optionally, the number of training times can be set to 1000 times, the required training accuracy can be set to 1×10 -6 , and the learning rate can be set to 0.1. In the embodiments of the present invention, the neural network surrogate model plays a role in predicting values, and a linear function can be selected as the activation function to directly output the predicted values.

[0080] After that, step S132 can be executed: initializing the parameters of the particle swarm algorithm to obtain an initial population.

[0081] Specifically, in some embodiments, the parameters of the particle swarm algorithm can at least include the number of particles, inertia weight and learning factor, and the position and velocity of each particle are randomly initialized to obtain an initial population (corresponding to step S630).

[0082] The inertia weight controls the influence of the particle's last velocity on the current velocity. A larger value is conducive to the particle's global search in the solution space, while a smaller value is conducive to the particle's local fine search. Usually, in the early stage of the algorithm, the inertia weight value can be set to a larger value, such as 0.9, and gradually reduced as the iteration proceeds. For example, it can be set to 0.4 in the late iteration to achieve the transition from global search to local search.

[0083] Determine the learning factors c1 and c2, which represent the degree of learning of the particle to its own historical optimal position and the group's historical optimal position, respectively. Specifically, c1 is usually called the individual learning factor, which mainly affects the speed and degree of the particle's learning to its own historical optimal position. c2 is usually called the social learning factor, which mainly affects the speed and degree of the particle's learning to the global optimal position of the population. Different values ​​have different effects on the particle search behavior. A larger c1 will make the particle more inclined to conduct local searches near itself and pay more attention to its own experience, while a larger c2 will make the particle more inclined to move closer to the global optimal position and pay more attention to the experience of the group. In this embodiment, through these two learning factors c1 and c2, the particle swarm algorithm can find a balance between local search and global search. c1 controls the particle to learn from its own optimal position, which helps the particle to conduct a fine search in the local area and dig out the local optimal solution. c2 controls the particle to learn from the global optimal position, prompting the particle to explore the entire search space, avoid falling into the local optimum, and find the global optimal solution faster. In addition, by adjusting the size of c1 and c2, the algorithm can flexibly adjust the focus between local search and global search, thereby improving the search efficiency and accuracy of the algorithm.

[0084] Optionally, in some embodiments, the learning factors c1 and c2 may be initialized to 4.5 so that the particle can achieve a better balance between its own experience and the group experience, making full use of the better solutions discovered by itself and moving closer to better solutions in the group.

[0085] Afterwards, step S133 may be performed: iteratively updating the speed and position of the particle to obtain the weight and threshold of the optimized BP neural network.

[0086] Specifically, in some embodiments, when the current iteration number is less than the preset maximum iteration number, the following loop calculation may be performed to perform particle swarm iteration (corresponding to steps S640 to S654). Figure 5B , Figure 5B FIG. 1 is a flow chart showing an iterative update of the particle speed and position according to some embodiments of the present invention to obtain the weights and thresholds of the optimized BP neural network. Figure 5B As shown, the above step S133 can be further embodied as steps S310 to S340.

[0087] As shown Figure 5B In some embodiments, as shown, step S310 may be performed first: substituting the positions of the particles in the particle swarm into the BP neural network as the weights and thresholds of the BP neural network (corresponding to step S651). Then, step S320 may be performed: performing forward propagation calculation via the training set data to obtain the output of the BP neural network (corresponding to step S652).

[0088] After that, step S330 may be performed: obtaining the particle position corresponding to the global optimal fitness according to the fitness value corresponding to each particle. Specifically, the fitness value corresponding to each particle may be calculated. Then, the current fitness of each particle is compared with the individual historical optimal fitness, and the individual extreme value is updated. At the same time, the fitness of all particles may also be compared to obtain the particle position corresponding to the global optimal fitness. Then, the velocity (corresponding to step S641) and position (corresponding to step S642) of the particle are iteratively updated. The update formula for the particle velocity may be as follows:

[0089]

[0090] The update formula for the particle position may be as follows:

[0091]

[0092] where pbest is the individual extreme value, gbest is the global extreme value, w is the inertia weight, c is the learning factor, where k is the iteration number, i is the particle number, d is the dimension of the particle, specifically including the two dimensions of the weights and thresholds of the neural network represented by the particle, and r1 and r2 are random numbers in the interval [0, 1]. In this embodiment, the particle velocity V id and the particle position X id are the formulas for the implementation process of improving and optimizing on the basis of the basic BP neural network using the particle swarm algorithm.

[0093] It should be emphasized that the above "fitness" is the position of the particle, and the position of the particle represents the weights and thresholds of the neural network. Therefore, the "comparing fitness" in the embodiment can be understood as that during the iteration process, the weights and thresholds represented by the fitness are brought into the neural network to calculate the fitting degree and error of the neural network model, and what is actually compared is the superiority of the neural network, and the weights and thresholds that make the neural network more superior are retained. (Specifically, it may correspond to steps S610 - S654).

[0094] Furthermore, since the data of the position of the particle swarm needs to be decoded before it can be used as the weights and thresholds of the neural network, as Figure 6 shown, step S650: particle decoding may be performed.

[0095] As the velocity and position of the particles are continuously updated, step S340 can be executed: in response to the current iteration number reaching the preset maximum iteration number, end the iteration (corresponding to step S655), and based on the particle positions corresponding to each global optimal fitness, obtain the particle position corresponding to the global extreme value as the weights and thresholds of the optimized BP neural network.

[0096] Through the above step S130, the present invention can establish a surrogate model based on the data regression prediction method of optimizing the BP neural network by the particle swarm algorithm, thereby achieving the optimization effect of reducing the time and workload of frequently performing CFD calculations.

[0097] Continue as Figure 1 shown, the robust optimization method for the cracking furnace emission uncertainty provided by the present invention may further include step S140: quantize the decision variables based on the uncertainty quantification model and construct a robust optimization model.

[0098] Specifically, please refer to Figure 7 , Figure 7 which shows a schematic flow chart of obtaining a robust optimization model according to some embodiments of the present invention. As Figure 7 shown, the above step S140 can be embodied as steps S141 to S144.

[0099] As Figure 7 shown, in some embodiments, step S141 can be executed first: based on the sampling data, obtain the uncertain output corresponding to the decision variables via the uncertainty quantification model. Since all input variables have been sampled by the Latin hypercube method previously, the uncertain output corresponding to the sampled deterministic decision variables can be solved through the uncertainty quantification model.

[0100] Then, step S142 can be executed: quantify the fluctuation amount of the uncertain output by the Monte Carlo method.

[0101] Specifically, in the embodiments of the present invention, the uncertainty quantification model constructed by optimizing the BP neural network by the particle swarm algorithm is used to predict the corresponding output data via the sampled input data. For each deterministic decision variable input, a sufficient amount of uncertainty input variable matrices under the Latin hypercube sampling are matched, and a series of output variable values are predicted and solved under the fixed decision variable considering the fluctuations of the uncertainty input variables. Optionally, the uncertainty of the output variables can be calculated by the statistical method of the Monte Carlo method.

[0102] Then, step S143 can be executed: obtain the yields of the cracking products under each decision variable as supplementary optimization data.

[0103] Specifically, in some embodiments, when the cracking furnace selected is an ethylene cracking furnace, the cracking product is ethylene. The ethylene production corresponding to the deterministic decision variables can be calculated through coilsim (a software tool for simulating and optimizing ethylene cracking furnaces and related processes) as supplementary optimization data. Specifically, via the training set data obtained in step S120 above, the heat flux distribution of the furnace tube part of the combustion process instance output by the CFD software for each working condition is used as a parameter input, and iterative calculations are performed until the furnace tube temperature distribution of the instance is stable, and the corresponding ethylene production is obtained.

[0104] From the obtained ethylene production data, it can be observed that under various input working conditions, the ethylene production only fluctuates when the deterministic decision variables change, while the impact of the uncertain input variables on the ethylene production is almost negligible. Therefore, only the impact of the deterministic decision variables on the ethylene production needs to be considered, and the uncertainty quantification of the ethylene production is no longer carried out. Instead, it is used as one of the objectives of multi-objective robust optimization. The objectives to be achieved by the "multi-objective robust optimization" here respectively include, first, reducing the mean value of nitrogen oxide emissions, second, reducing the standard deviation of nitrogen oxide emissions. Reducing the standard deviation also represents controlling the robustness of emissions when there are uncertain fluctuations in the input, and third, increasing the ethylene production. Therefore, in this embodiment, the ethylene production can be used as one of the objectives of multi-objective robust optimization.

[0105] Then, step S144 is executed: Using the fluctuation amount of the uncertain output and the production of the cracked product as supplementary optimization data as the output of the surrogate model, and again based on the BP neural network optimized by the particle swarm algorithm, a robust optimization model is constructed.

[0106] Specifically, in some embodiments, the input of the constructed surrogate model is the decision variable, such as the value of the excess air coefficient, and its output quantities are the mean value and variance (square of the standard deviation) of nitrogen oxides, which can represent the uncertainty of pollutant emissions from the ethylene cracking furnace and reflect its robustness, and the ethylene production reflecting the working efficiency of the ethylene cracking furnace. It can be understood that the robust optimization model here refers to a calculation model that puts the genetic algorithm to optimize the decision variables, and it is established based on the BP neural network and belongs to a black box model.

[0107] Continuing back to Figure 1 As shown, the robust optimization method for the emission uncertainty of the cracking furnace provided by the invention may further include step S150: Via the robust optimization model, through the dynamic adaptive genetic algorithm, the decision variables are robustly optimized to obtain the ideal output working condition of the uncertainty variables.

[0108] Specifically, in combination with Figure 8 Understood together, Figure 8The flowchart of the dynamic adaptive robust optimization provided according to some embodiments of the present invention is shown. Specifically, as Figure 8 shown, step S150 can be embodied as steps S151 to S156.

[0109] As Figure 8 shown, in some embodiments, step S151 can be executed first: determining the value range distribution of the decision variables, the optimization objectives of the robust optimization model, the population size N, the crossover probability Pc, the mutation probability Pm, the number of optimization objectives, and the boundary range of the decision variables.

[0110] Specifically, in some embodiments, the decision variable of the ethylene cracking furnace can be the excess air coefficient, and its value range can be [1, 1.3]. The optimization objectives of the robust optimization model are as described above and can include three, namely, the mean value of nitrogen oxides, the standard deviation, and the ethylene production.

[0111] Furthermore, in some embodiments, the population size N can be set to 100, the crossover probability Pc and the mutation probability Pm can be set to 1, the number of optimization objectives is 3, which represents multi-objective optimization belonging to three tables, and the variable boundary range can be set according to the distribution assumption.

[0112] Then, step S152 can be executed: initializing the population P0 through global average search.

[0113] In the genetic algorithm, population initialization is the first step of the algorithm, and its quality directly affects the subsequent search process. Although random initialization is simple, it is likely to cause the individuals in the population to be unevenly distributed in the solution space, making it possible for the algorithm to not fully explore the entire solution space in the initial stage of the search, increasing the risk of falling into a local optimal solution. Global average search can distribute points relatively evenly in the space, so that when the population is calculated or optimized, it can better cover the entire search space and reduce the influence of local deviation on the results. Specifically, the uniform distribution property of global average search ensures that when generating the initial population, individuals can be relatively evenly scattered in the solution space, avoiding the problem that individuals may be concentrated in certain specific regions caused by random initialization. Its characteristic of small deviation enables the good point set to have higher accuracy when approaching the true optimal solution, and can more effectively guide the algorithm to search in the direction of the global optimum.

[0114] Furthermore, in some embodiments, the construction method of the initial population P0 initialized by global average search can adopt the following steps. First, it is known that the spatial dimension where the population P0 is located is n, that is, the dimension of the multi-objective optimized by the algorithm, and the population size is m. Calculate a = (a1, a2,..a n .), where Let \(j\) represent the \(j\)-th spatial dimension, where \(1\leq j\leq n\), \(a\) represent the point selected as an individual in population \(P_0\), \(prime\_number\_min\) represent the smallest prime number satisfying \((prime\_number\_min - 3) / 2\geq n\), and \(i\) represent the \(i\)-th individual, where \(1\leq i\leq m\). It can be understood that since \(a_1\), \(a_2\), etc. represent the dimensions within point \(a\), thus, in combination with the above specific embodiments, \(a_1\), \(a_2\), \(a_3\) can specifically be the mean value of nitrogen oxides, the standard deviation, and the ethylene production respectively.

[0115] Then, a point set \(P\) with quantity \(m\) can be constructed n \((i)=\{(a_{1i} 1 ,a_{2i} 2 ,\cdots,a n _i n )\}\), where \(i = 1,2,3,\cdots,n\).

[0116] After that, \(P\) n can be mapped to the feasible region where population \(P_0\) is located, \(X\) ij =l j +P n (i)(u j -l j ), where \(X\) ij is the mapping value representing \(P\) n , \(l\) j represents the lower limit of the current dimension, and \(u\) j represents the upper limit of the current dimension. The key purpose of adopting this step is to smoothly convert the value from the theoretical unit cube space to the feasible region of the actual problem, so as to effectively ensure that the individuals in the generated initial population satisfy the constraint conditions of the problem.

[0117] Through the above step S152, the present invention innovatively improves the commonly used random initialization in the non-dominated genetic algorithm to global average search initialization. The population generated by traditional random initialization is unevenly distributed over the entire space, which may cause the algorithm to prematurely fall into a local optimal solution during the search process and has a low utilization rate of the search space. Specifically, there may be a situation where individuals are overly dense in some regions while sparse in other regions. This makes it easy for the algorithm to be limited to local regions during the selection, crossover, and mutation operations and unable to comprehensively explore the potential optimal solutions in the solution space. In contrast, the global average search initialization in the present invention can make the individuals in the initial population more evenly distributed in the solution space, increasing the diversity of the population and providing the algorithm with more opportunities to explore the global optimal solution. The uniform distribution characteristic of global average search enables the initial population to cover all regions of the solution space, so that a more comprehensive perception of the entire solution space can be obtained at an early stage of the algorithm, which helps to avoid the algorithm from prematurely converging to a local optimal solution and improves the possibility of the algorithm finding the global optimal solution.

[0118] Continue as Figure 8 shown, and then step S153 can be executed: calculate the crowding degree of each pair of individuals in population P0, and divide population P0 into multiple non-dominated levels.

[0119] In some embodiments, through crowding degree calculation, it can be used as a basis for subsequent non-dominated sorting of individuals. Specifically, for each pair of individuals i and q in population P0, compare their performances on the objective functions corresponding to all optimization objectives. Among them, in response to individual i being not inferior to individual q on all objective functions and being superior to individual q on at least one objective function, it can be determined that individual i dominates individual q. In the specific embodiment of the above ethylene cracking furnace, there are three objective functions for each working condition, namely the mean value, standard deviation of nitrogen oxides, and ethylene production. Therefore, the "performance on the objective function" here is the output of the model.

[0120] After that, the number of times n i that each individual is dominated can be marked, that is, how many other individuals can dominate it. At the same time, the set S i of individuals dominated by each individual can be recorded. Then, based on the individuals in population P0 with a domination count of 0, these individuals can be grouped into the first non-dominated level F1.

[0121] Furthermore, based on the set S i of individuals dominated by each individual i in the first non-dominated level F1, the number of times n i that the individuals in the set S q are dominated can be set to n i minus 1, where i ∈ F1 and q ∈ S i . Then, the individuals with a domination count of n q =0 can be grouped into the second non-dominated level F2. Repeat this process until the entire population P0 is divided into multiple different non-dominated levels.

[0122] In this embodiment, through the above steps, non-dominated sorting is performed on all individuals. The more individuals an individual dominates, the higher its rank, so a higher domination rank should be assigned accordingly. For the individuals with an initial domination count of 0, they are the individuals with the highest rank, and no other individuals can dominate them. Then, mainly observe the individuals they dominate. If among these individuals, there is an individual with a domination count of 1, that is, they are only dominated by the individuals with the highest rank, then they are the individuals of the second rank. In the algorithm, according to this thinking to judge the individual levels, the following calculation can be performed, that is, the individuals with a domination count n i of 0 are classified into the first rank and no longer participate in the comparison. Subtract one from the domination counts of all the remaining individuals, that is, nq = n i -1, and at this time, n q with a value of 0 is classified into the second level. Then, the remaining values are decreased by 1. At this time, those with n being 0 are classified into the third level... and so on, until the entire population P0 is divided into multiple different non-dominated levels.

[0123] Continue as Figure 8 shown. After that, step S154 can be executed: Based on the sorting of non-dominated levels or fitness, an adaptive ranking probability hit parent selection strategy for parent selection is performed.

[0124] Specifically, in some alternative embodiments, the individuals in the population P0 can be sorted according to their fitness values. Alternatively, in some other embodiments, this step can directly use the non-dominated sorting of the previous step. The "adaptive" here means that instead of randomly selecting individuals as parents, according to the order of fitness or non-dominated ranking, better individuals have a greater probability of being selected as parents. Then, the probability of each individual being selected as a parent is based on the superiority of this individual itself, and there is no artificial parameter to interfere with this probability, so it is an adaptation.

[0125] Then, according to the ranking of each individual's non-dominated sorting, the probability of each individual being hit and selected as a parent individual can be calculated. Among them, the expression of the probability calculation formula can be as follows:

[0126]

[0127] Among them, η - and η + are specified constants, and it is required that η - + η + = 2, and 0 ≤ η - ≤ 1. That is to say, in the embodiments of the present invention, the probability of each individual being hit and selected as a parent individual is based on the ranking, which is the meaning of "ranking probability hit".

[0128] Furthermore, in the probability calculation formula, different value combinations of the two parameters η - and η + will result in different results of the probability calculation formula. When η - = 0 and η + = 2, the selection pressure reaches the maximum value. At this time, the algorithm will highly tend to select high-quality individuals with forward fitness values, thereby accelerating the convergence speed and enabling the algorithm to more quickly focus on the search in the region close to the optimal solution. The "selection pressure" here refers to the strength of the correlation between the selection probability and the individual's own ranking. And when η - = η +When = 1, the selection mode degenerates into a completely random selection, and the probability of each individual being selected is 1 / N, which is a random selection without any pressure. This can effectively expand the search scope in the early stage of the algorithm, avoid falling into the local optimal solution prematurely, and provide a broader space for global search.

[0129] In this embodiment, by flexibly adjusting η - and η + values, the adaptive ranking probability hit selection strategy can accurately control the selection pressure according to different stages of the algorithm and the characteristics of the problem, and achieve the dynamic balance between the search ability and the convergence speed.

[0130] Through the above step S154, the present invention innovatively improves the commonly used tournament parent selection in the non-dominated genetic algorithm to the adaptive ranking probability hit parent selection. In the traditional tournament selection, when selecting parents each time, only a small number of individuals are randomly selected for comparison, and the individual with the highest fitness is selected as the parent. Although this method is simple and easy to implement, due to its strong randomness, some potential high-quality individuals may be frequently missed in the initial stage of evolution, resulting in an inefficient search process. Moreover, in the later stage of the algorithm, when the fitness of individuals in the population gradually approaches, the selection pressure of the tournament selection is relatively insufficient, and it is difficult to quickly guide the algorithm to converge to the global optimal solution. In contrast, the adaptive ranking probability hit parent selection can more reasonably guide the search direction throughout the evolution process through the global sorting of population individuals and fine probability allocation. In the initial stage, it gives individuals with relatively low fitness but potential value a certain chance of survival and participation in genetic operations, thus effectively maintaining the diversity of the population and avoiding falling into the local optimal solution prematurely; in the later stage, as the algorithm progresses and the population fitness improves, by appropriately adjusting the selection pressure, the adaptive ranking probability hit parent selection can more efficiently screen out high-quality individuals, accelerate the convergence of the algorithm to the global optimal solution, and significantly improve the stability and accuracy of convergence.

[0131] Continue as Figure 8 shown, and then execute step S155: Based on the parent individuals, perform particle swarm optimization crossover and dynamic mutation to optimize the offspring population.

[0132] Specifically, in some embodiments, the particle swarm optimization crossover can be performed through the following steps. First, compare the non-dominated rankings of the parent individuals. Then, the offspring individuals can be generated according to the evolution direction formula, where the evolution direction formula can be as follows:

[0133]

[0134] where V p1 ,V p2respectively represent the evolutionary directions of two offspring individuals, corresponding to the speed of updating the individual fitness value in the particle swarm algorithm. f is the inertia weight, which controls the influence degree of the particle's previous speed on the current speed. h1 and h2 are learning factors, which are respectively used to adjust the degree of the particle approaching the individual extreme value and the global extreme value. f, h1, and h2 are all constants. r1, r2, r3, and r4 are random numbers from 0 to 1, p1 and p2 are the selected parent individuals, c p is the gene difference point, and c p = p1 - p2, z p is the gene common point, and The formula for generating the offspring individuals can be as follows:

[0135] In this embodiment, the present invention innovatively improves the polynomial crossover commonly used in the non-dominated genetic algorithm to particle swarm optimization crossover. This improvement method can make the evolutionary direction of the offspring determined according to the optimal direction and common genes in the two parent individuals, thereby effectively improving the convergence efficiency of the algorithm.

[0136] Furthermore, in some preferred embodiments, the dynamic mutation can be performed through the following steps. First, for the individual to be mutated x = (x1, x2,..., x n ), a random integer l ∈ [1, n] is generated, and the l-th gene of x is dynamically mutated to obtain a new gene y l . The formula for dynamic mutation can be as follows. When r ≤ 0.5, y l = x l + Δ(t, u l - x l ); and when r ≥ 0.5, y l = x l - Δ(t, u l - x l ), where r is a random number between (0, 1), t is the current evolutionary generation, u l is the upper limit of the gene, and in the function , T is the maximum number of evolutions, and b is the parameter affecting the uniformity.

[0137] In this embodiment, the innovation of the present invention improves the uniform mutation commonly used in the non-dominated genetic algorithm to dynamic mutation. Different from uniform mutation, the characteristic of dynamic mutation is that the mutation amplitude is not constant, but gradually decreases according to the progress of the evolutionary generation or other parameters. Initially, the mutation amplitude is large, which helps to explore the search space from the perspective of diversity; as the number of evolutionary generations increases, the mutation range becomes smaller and smaller, increasing the fine-tuning ability of the algorithm. This mutation method enables the algorithm to have a large-range search ability in the early stage and small-range adjustment in the later stage. Its effect is greatly affected by the maximum number of iterations, and the mutation range is small in the later stage. Dynamic mutation can introduce different degrees of randomness at different stages to better balance diversity and convergence, thereby improving the performance of the algorithm and finding better solutions.

[0138] After that, continue as Figure 8 shown, step S156 can be executed: merge the offspring population Qt and the parent population Pt to obtain the merged population Rt, and replace the population P0 with the merged population Rt, and repeat the above steps to obtain the iteratively updated offspring population Qt+1 until the iteration termination condition is reached.

[0139] Specifically, in some embodiments, the offspring population Qt and the parent population Pt are merged to form a population Rt of size 2N. According to the above parent selection strategy, non-dominated sorting is performed on the population Rt to generate a non-dominated set, and the crowding degree is calculated, thereby generating a new offspring population Qt+1 until the iteration termination condition is reached.

[0140] Furthermore, in some embodiments, in response to reaching the termination condition, the dynamic adaptive genetic algorithm can be run to perform robust optimization on the fixed operating conditions of the ethylene cracking furnace. The algorithm can generate a Pareto solution set to provide decision variables with the highest domination level that can be selected and the corresponding ideal output operating conditions at this time. That is to say, based on the Pareto solution set, the decision variables with the highest domination level can be determined, and the ideal output operating conditions of the uncertainty variables corresponding to the decision variables with the highest domination level can be obtained.

[0141] It can be combined with Figure 9 to understand together, Figure 9 which is a schematic diagram of the result of dynamic adaptive robust optimization provided according to some embodiments of the present invention.

[0142] As in Figure 9 the embodiment shown, the curves in the figure represent the Pareto solution set with the highest non-dominated level. Each point above corresponds to the selection of the decision variable of "excess air coefficient". At the same time, the values corresponding to the point on the three coordinate axes are the mean, standard deviation of nitrogen oxides (NO), and ethylene production corresponding to the current decision variable considering input uncertainty. Appropriate operating conditions can be selected from these points and applied to actual industry.

[0143] So far, the above-mentioned robust optimization method for the cracking furnace emission uncertainty provided by the first aspect of the present invention has been introduced. The decision maker can select a set of non-dominated solutions according to requirements and preferences to guide the design of decision variables for the actual required industrial models.

[0144] Next, please refer to Figure 10 , Figure 10 which shows a structural block diagram of a robust optimization device for cracking furnace emission uncertainty provided according to some embodiments of the present invention.

[0145] In some non-limiting embodiments, the above-mentioned robust optimization device 10 for cracking furnace emission uncertainty provided by the second aspect of the present invention can be used to implement the above-mentioned robust optimization method for cracking furnace emission uncertainty provided by the first aspect of the present invention. The robust optimization device 10 for cracking furnace emission uncertainty may include a memory 101 and a processor 102. When the processor 102 executes the computer program stored in the memory 101, the above-mentioned robust optimization method for cracking furnace emission uncertainty can be implemented.

[0146] Specifically, in some non-limiting embodiments, the above-mentioned computer-readable storage medium provided by the third aspect of the present invention stores computer instructions. When the computer instructions are executed by the processor 102, they can be used to implement the above-mentioned robust optimization method for cracking furnace emission uncertainty provided by the first aspect of the present invention.

[0147] Those skilled in the art can understand that the embodiments of the above-mentioned robust optimization device 10 for cracking furnace emission uncertainty are only some non-limiting implementation manners provided by the present invention, aiming to clearly show the main concept of the present invention and provide some specific solutions convenient for the public to implement, rather than limiting all working modes or all functions of the robust optimization device 10 for cracking furnace emission uncertainty. Similarly, the robust optimization device 10 for cracking furnace emission uncertainty is also only a non-limiting implementation manner provided by the present invention, and does not limit the implementation subject of each step in these robust optimization methods for cracking furnace emission uncertainty.

[0148] Although the above methods are illustrated and described as a series of actions for simplicity of explanation, it should be understood and appreciated that these methods are not limited by the order of the actions, because according to one or more embodiments, some actions may occur in a different order and / or concurrently with other actions shown and described herein or not shown and described herein but understood by those skilled in the art.

[0149] Those skilled in the art will further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability of hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

[0150] The various illustrative logical modules and circuits described in connection with the embodiments disclosed herein can be implemented using a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0151] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read from, and write to, the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.

[0152] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functions may be stored on or transmitted via a computer-readable medium as one or more instructions or code. The computer-readable medium includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. The storage media may be any available media that can be accessed by a computer. By way of example and not limitation, such computer-readable media can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a web site, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. As used herein, disk and disc include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk, and Blu-ray disc where disks typically reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0153] In summary, the present invention provides a robust optimization method for cracking furnace emissions uncertainty, a robust optimization device for cracking furnace emissions uncertainty, and a computer-readable storage medium, which can quantify output uncertainty based on the input uncertainty under fixed operating conditions and optimize the output quantity, thereby improving the robust efficiency and accuracy of cracking furnace emissions under the influence of uncertainty.

[0154] The foregoing description of the disclosure has been provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A robust optimization method for the emissions uncertainty of a cracking furnace, characterized in that, It includes the following steps: Respectively obtain the sampling data under the distributions corresponding to the uncertainty variables and decision variables of the cracking furnace; Based on the turbulent combustion coupling model, perform numerical simulation on the sampling data to obtain the training set data; Using the sampling data as the input variables and the training set data as the output variables, construct a neural network surrogate model optimized by the particle swarm algorithm as the uncertainty quantification model; Based on the uncertainty quantification model, quantify the decision variables and construct a robust optimization model; And Via the robust optimization model, through the dynamic adaptive genetic algorithm, robustly optimize the decision variables to obtain the ideal output condition of the uncertainty variables.

2. The robust optimization method according to claim 1, wherein The cracking furnace includes an ethylene cracking furnace. The step of respectively obtaining the sampling data under the distributions corresponding to the uncertainty variables and decision variables of the cracking furnace includes: Through distribution assumptions, define the uncertainty variables and decision variables corresponding to the input variables of the ethylene cracking furnace, and obtain the distributions of the uncertainty variables and decision variables; Through the Latin hypercube method, perform data sampling on the uncertainty variables to obtain the sampling data that can completely represent the distributions of the uncertainty variables and decision variables based on a small number of sample points.

3. The robust optimization method according to claim 1, wherein, The step of performing numerical simulation on the sampling data based on the turbulent combustion coupling model to obtain the training set data includes: Obtain the structural model of the cracking furnace and perform mesh division on it; According to the structural model after mesh division, determine the turbulent model, combustion model, and radiation heat transfer model during the cracking process of the cracking furnace; Based on the turbulent model, the combustion model, and the radiation heat transfer model, establish the turbulent combustion coupling model; and Based on the turbulent combustion coupling model, perform simulation numerical simulation on the sampling data through computational fluid dynamics to obtain the training set data.

4. The robust optimization method according to claim 1, wherein The step of using the sampling data as the input variables and the training set data as the output variables to construct a neural network surrogate model optimized by the particle swarm algorithm as the uncertainty quantification model includes: Create a BP neural network and determine its topological structure, where the topological structure includes the number of neurons in the input layer, hidden layer, and output layer, as well as the activation function; Initialize the parameters of the particle swarm algorithm to obtain the initial population, where the parameters of the particle swarm algorithm at least include the number of particles, inertia weight, and learning factor; and Iteratively update the velocity and position of the particles to obtain the weights and thresholds of the optimized BP neural network.

5. The robust optimization method according to claim 4, wherein The step of iteratively updating the velocity and position of the particles to obtain the weights and thresholds of the optimized BP neural network includes: Substitute the positions of the particles in the particle swarm into the BP neural network as the weights and thresholds of the BP neural network; Perform forward propagation calculation through the training set data to obtain the output of the BP neural network; According to the fitness values corresponding to the particles, obtain the particle position corresponding to the global optimal fitness; and In response to the current iteration number reaching the preset maximum iteration number, the iteration is terminated, and based on the particle positions corresponding to each of the global optimal fitness values, the particle position corresponding to the global extreme value is obtained as the weights and thresholds of the optimized BP neural network.

6. The robust optimization method according to claim 5, wherein, The step of obtaining the particle position corresponding to the global optimal fitness according to the fitness values corresponding to each particle includes: Calculating the fitness values corresponding to each of the particles; Comparing the current fitness of each particle with the individual historical optimal fitness and updating the individual extreme value; Comparing the fitness of all particles to obtain the particle position corresponding to the global optimal fitness; and Iteratively updating the velocity and position of the particles, where the update formula for the particle velocity is as follows: The update formula for the particle position is as follows: Where pbest is the individual extreme value, gbest is the global extreme value, w is the inertia weight, c is the learning factor, where k is the iteration number, i is the particle number, d is the dimension of the particle, including the two dimensions of the weights and thresholds of the neural network represented by the particle, and r1 and r2 are random numbers in the interval [0, 1].

7. The robust optimization method according to claim 5, characterized in that The step of quantifying the decision variables based on the uncertainty quantification model and constructing a robust optimization model includes: Based on the sampling data, obtaining the uncertain output corresponding to the decision variables via the uncertainty quantification model; Quantifying the fluctuation amount of the uncertain output by the Monte Carlo method; Obtaining the yield of the lysate under the decision variables as supplementary optimization data; and Using the fluctuation amount of the uncertain output and the supplementary optimization data as the output of the surrogate model, and constructing the robust optimization model again based on the BP neural network optimized by the particle swarm algorithm.

8. The robust optimization method according to claim 5, wherein The step of robustly optimizing the decision variables via the robust optimization model by the dynamic adaptive genetic algorithm to obtain the ideal output condition of the uncertainty variables includes: Determining the value range distribution of the decision variables, the optimization objective of the robust optimization model, the population size N, the crossover probability Pc, the mutation probability Pm, the number of optimization objectives, and the boundary range of the decision variables; Initializing the population P0 by global average search; Calculating the crowding degree of each pair of individuals in the population P0 and dividing the population P0 into multiple non-dominated layers; Based on the ranking of the non-dominated layers or the fitness, performing a parent selection strategy of adaptive ranking probability hit parent selection; Based on the parent individuals, performing particle swarm optimization crossover and dynamic mutation to optimize the offspring population; and Merging the offspring population Qt and the parent population Pt to obtain the merged population Rt, and replacing the population P0 with the merged population Rt, repeating the above steps to obtain the iteratively updated offspring population Qt+1 until the iteration termination condition is reached.

9. The robust optimization method according to claim 8, wherein, The step of initializing the population P0 by global average search includes: Based on the fact that the spatial dimension where population P0 is located is n and the population size is m, calculate a = (a1, a2,..., a n ), where j represents the j-th spatial dimension, 1 ≤ j ≤ n, a represents the point selected as an individual in population P0, prime_number_min represents the smallest prime number that satisfies (prime_number_min - 3) / 2 ≥ n, and i represents the i-th individual, 1 ≤ i ≤ m; Construct a point set \(P\) with the number of elements \(m\). n \((i)=\{(a_{1i}\ 1 ,a_{2i}\ 2 ,\cdots,a\ n i\ n )\}\), where \(i = 1, 2, 3,\cdots,n\); and Map P n to the feasible region where the population P0 is located. X ij = l j + P n (i)(u j - l j ), where X ij is the mapping value representing P n , l j represents the lower bound of the current dimension, and u j represents the upper bound of the current dimension.

10. The robust optimization method according to claim 8, characterized in that, The step of calculating the crowding degree of each pair of individuals in the population P0 and dividing the population P0 into multiple non-dominated layers includes: For each pair of individuals i and q in the population P0, compare their performance on the objective functions corresponding to all the optimization objectives. Wherein, in response to individual i not being inferior to individual q on all the objective functions and being superior to individual q on at least one objective function, determine that individual i dominates individual q; Mark the number of times n that each individual is dominated i and obtain the set S of individuals dominated by each of said individuals i Based on the individuals in the population P0 with a domination count of 0, form the first non-dominated layer F1; Based on the set S of individuals dominated by each individual i in the first non-dominated layer F1 i , the number of times the individuals in the set S i are dominated, n q is set to n i minus 1, where i ∈ F1 and q ∈ S i ; The individuals with the number of domination times being n q = 0 are grouped into the second non-dominated layer F2; and Repeat the above steps to divide the entire population P0 into multiple different non-dominated layers.

11. The robust optimization method according to claim 8, wherein, The step of performing the parent selection strategy of adaptive ranking probability hit parent selection based on the ranking of the non-dominated layer or the fitness includes: Sort the individuals in the population P0 according to the non-dominated layer or the fitness; Based on the ranking of the sorting, determine the probability of each individual being hit and selected as a parent individual, wherein the expression of the probability calculation formula is as follows: where η - and η + are specified constants, and it is required that η - + η + = 2, and 0 ≤ η - ≤ 1.

12. The robust optimization method according to claim 11, wherein, The step of performing particle swarm optimization crossover and dynamic mutation based on the parent individuals to optimize the offspring population includes: Compare the rankings of the non-dominated layers of the respective parent individuals; Generate offspring individuals according to the evolution direction formula, wherein the evolution direction formula is as follows: Among them, V p1 , V p2 respectively represent the evolutionary directions of two offspring individuals, corresponding to the velocities for updating the individual fitness values in the particle swarm algorithm. f is the inertia weight, h1 and h2 are learning factors, and f, h1, h2 are all constants. r1, r2, r3, r4 are random numbers from 0 to 1, p1, p2 are the selected parent individuals, c p is the gene difference point, and c p = p1 - p2, z p is the gene common point, and and the formula for generating the said offspring individuals is as follows:

13. The robust optimization method according to claim 12, characterized in that, The step of performing particle swarm optimization crossover and dynamic mutation based on the parent individuals to optimize the offspring population further includes: For the individual x = (x1, x2,..., x n ), randomly generate an integer l ∈ [1, n], and perform dynamic mutation on the l-th gene of x to obtain a new gene y l , where the formula for the dynamic mutation is as follows: When r ≤ 0.5, y l = x l + Δ(t, u l - x l ), When r ≥ 0.5, y l = x l - Δ(t, u l - x l ), where r is a random number between (0, 1), t is the current generation number of evolution, u l is the upper limit of the gene, and in the function , T is the maximum number of generations of evolution, and b is the parameter affecting the uniformity.

14. The robust optimization method according to claim 8, wherein The step of reaching the iteration termination condition further includes: In response to reaching the termination condition, obtain the Pareto solution set generated based on the dynamic adaptive genetic algorithm; and Based on the Pareto solution set, determine the decision variable with the highest domination level and obtain the ideal output working condition of the uncertainty variable corresponding to the decision variable with the highest domination level.

15. A robust optimization device for the emission uncertainty of a cracking furnace, characterized in that, The robust optimization device includes a memory and a processor, and the processor implements the robust optimization method according to any one of claims 1 to 14 when executing the computer program stored on the memory.

16. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When the computer instruction is executed by the processor, the robust optimization method according to any one of claims 1 to 14 is implemented.

Images