Combustion mechanism simplification and optimization method based on two-stage deep neural network algorithm

Abstract

A combustion mechanism simplification and optimization method based on a two-stage deep neural network algorithm belongs to the field of combustion dynamics simulation and comprises three stages of component simplification based on a first-stage deep neural network, deep reaction simplification and coefficient optimization based on a second-stage deep neural network. Firstly, a deep neural network model with component retention states and initial conditions as input and combustion characteristics as output is constructed, and efficient screening and gradual simplification of mechanism components are achieved. Secondly, deeply simplifying a reaction path, setting a time scale threshold value, eliminating non-key rapid reactions, and constructing a reaction number controlled simplification mechanism; and finally, constructing a deep neural network model taking the key reaction kinetic parameters as input, and performing global optimization of mechanism parameters in combination with a genetic algorithm to improve the prediction precision of the mechanism on key combustion performance indexes. According to the method, a combustion mechanism with high simplification and high fidelity can be realized, and the method is suitable for rapid calculation in complex three-dimensional combustion simulation.

Description

Technical Field

[0001] The present invention belongs to the field of combustion dynamics simulation and relates to a combustion mechanism simplification and optimization method based on a two-stage deep neural network algorithm. Background Art

[0002] Accurate and efficient chemical reaction mechanisms can reveal in detail the key chemical reactions, dominant reaction pathways, and intermediate products in the fuel combustion process, thus providing a solid scientific basis for the optimization of the combustion process. However, complete chemical reaction mechanisms usually involve hundreds to thousands of chemical species and corresponding reactions, which not only increases the difficulty of understanding the kinetic mechanism, but also significantly increases the computational burden in three-dimensional applications. Therefore, it is particularly important to simplify the mechanism, while minimizing the demand for computing resources, maintaining the accuracy and practicality of the mechanism, and ensuring that it can accurately predict key combustion characteristics and emission generation.

[0003] Traditional mechanism simplification methods mainly include computational singular perturbation method, principal component analysis, direct relationship diagram method and its derivative methods, path flux analysis method, etc. These methods achieve reaction network simplification by selectively eliminating minor reaction paths or components, aiming to improve the efficiency of numerical simulation while maintaining the accuracy of predicting combustion characteristics. However, these methods still have certain defects in practical applications, and the degree of simplification is limited. For example, in the prior art, Xu Huaping (CN116432418A) discloses a method for constructing and simplifying the detailed mechanism of methanol / biodiesel mixed fuel. This method uses an error-based direct relationship diagram method to simplify the mechanism, but this is highly sensitive to the selection of target components. If the initial conditions are improperly set, the simplified mechanism will lose the key reaction path, making it difficult to meet the accuracy requirements under specific working conditions.

[0004] In the field of combustion mechanism modeling and optimization, deep neural networks, with their powerful nonlinear modeling capabilities and high-dimensional feature extraction advantages, offer a new approach for the efficient simplification of complex reaction networks. Compared to traditional mechanism simplification methods that rely on manual experience to screen key species, deep neural networks, through end-to-end learning, can directly establish a nonlinear mapping between input parameters and combustion characteristics, significantly improving the degree of simplification while ensuring accurate prediction of combustion characteristics. However, currently, there are relatively few methods for combustion mechanism simplification and optimization based on deep neural networks, and their specific application methods and scope require further exploration. Summary of the Invention

[0005] The purpose of the present invention is to address the above-mentioned problems and provide a method for simplification and optimization of combustion mechanisms based on a two-stage deep neural network algorithm. First, a deep neural network model with component retention states and initial conditions as input and combustion characteristics as output is constructed to replace traditional simulation tools, thereby achieving efficient screening and gradual simplification of mechanism components to obtain the simplest mechanism of components. Secondly, the computational singular perturbation method is used to deeply simplify the reaction path of the simplest mechanism of components, set a time scale threshold, eliminate non-critical rapid reactions, and construct a simplest mechanism for the number of reactions. Finally, a deep neural network model with key reaction kinetic parameters as input is constructed for the simplest mechanism for the number of reactions, and a genetic algorithm is used to perform global optimization of the mechanism parameters to obtain a final mechanism with accurate predictions. The method can achieve a combustion mechanism that is both highly simplified and high-fidelity, and is suitable for rapid calculations in complex three-dimensional combustion simulations.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions to achieve it.

[0007] A combustion mechanism simplification and optimization method based on a two-stage deep neural network algorithm includes the following steps:

[0008] Step 1: Build and train a deep neural network model. The input of the deep neural network model is the component retention state and initial conditions such as temperature and pressure, and the combustion characteristics are used as output. It is used to replace traditional reaction kinetics simulation software to achieve rapid screening and gradual simplification of components in the reaction mechanism, thereby obtaining a combustion mechanism with the most simplified components. Specifically:

[0009] Step 1.1: Randomly delete k components from the detailed chemical reaction mechanism to generate multiple sets of simplified mechanisms. These mechanisms are then simulated using traditional reaction kinetics simulation software under different initial conditions to obtain corresponding combustion characteristics data and construct an initial dataset, which is divided into an 80% training set, a 10% validation set, and a 10% test set.

[0010] Step 1.2: Based on the initial dataset constructed in step 1.1, the first-stage deep neural network model I is trained. The model takes temperature, pressure, equivalence ratio, blending ratio, and a 01 vector (where "1" represents component retention and "0" represents component deletion) as input, and outputs the combustion characteristics to be predicted.

[0011] In step 1.3, using the deep neural network model I trained in step 1.2, the combustion characteristics of all simplified mechanisms generated in step 1.1 are quickly predicted and compared with the corresponding experimental values to calculate the average error, and the set of mechanisms whose errors are within the set threshold is extracted; the set threshold is 10%.

[0012] Step 1.4: Use the traditional dynamics simulation software Cantera to perform computational verification under different operating conditions on the mechanism set screened in step 1.3 to obtain more accurate combustion characteristic data and construct a secondary data set;

[0013] Step 1.5: Use the secondary dataset constructed in step 1.4 to train the deep neural network model I again for the next stage of mechanism screening and evaluation, further screening out redundant components and retaining key components;

[0014] Steps 1.6, 1.2, and 1.5 constitute an iterative process. Each iteration gradually reduces the number of components, k, to be deleted. As the number of iterations increases, the intensity of component deletion decreases. When the average error between the predicted and experimental values for the combustion characteristics exceeds 10%, the construction of the simplified component mechanism for that threshold condition is complete.

[0015] Step 2: Based on the simplest component mechanism obtained in step 1, the computational singular perturbation method is used to deeply simplify the reaction paths in the mechanism. By setting a time scale threshold, non-critical rapid reactions are identified and eliminated, and finally a simplified mechanism that meets the requirements of the limited number of reactions is generated. Specifically:

[0016] Step 2.1, setting the initial threshold, step increment, and termination threshold for reaction simplification;

[0017] Step 2.2: For different combustion characteristics, the computational singular perturbation method is used to perform perturbation analysis to obtain the corresponding simplified combustion mechanism;

[0018] Step 2.3, combine the simplified mechanisms obtained in step 2.2 based on different combustion characteristics to obtain the simplest mechanism for the reaction number.

[0019] Step 3: To address the prediction bias that may be introduced by the deep simplification of the reaction path in Step 2, a deep neural network model II is constructed with key reaction kinetic parameters as input, and a genetic algorithm is combined to perform global optimization of the mechanism parameters to restore or improve the prediction accuracy of the mechanism in the corresponding combustion characteristics. Specifically:

[0020] Step 3.1: Perform sensitivity analysis on the simplest mechanism of the reaction number obtained in step 2.3 to determine the key reactions that have a greater impact on the combustion characteristics;

[0021] Step 3.2: Randomly modify the pre-exponential factors of key reactions to generate the combustion characteristics under different initial conditions as a data set through mechanism simulation. The pre-exponential factors vary in the range of [0.001, 1000].

[0022] Step 3.3, train the deep neural network model II, taking temperature, pressure, equivalence ratio, blending ratio, modified pre-exponential factor and activation energy as inputs, and outputting the desired predicted combustion characteristics;

[0023] Step 3.4: Generate an initial population sample using a hybrid initialization method. The initial population sample generated by the hybrid initialization method includes multiple random samples, including: 33% of the samples are uniformly distributed, 33% of the samples are Gaussian distributed, and 34% of the samples are constant individuals.

[0024] Step 3.5: Use the deep neural network model II trained in step 3.3 to evaluate the fitness of each sample obtained in step 3.4, calculate the fitness function value, and retain the sample with the smallest fitness function value;

[0025] Step 3.6, perform selection and crossover operations to reorganize elite individuals with smaller fitness function values with the retained samples;

[0026] Step 3.7, by setting a certain probability of mutation operation to increase the diversity of the population, to prevent the genetic algorithm from falling into the local optimum;

[0027] In step 3.8, when the set maximum number of iterations is reached or the early termination condition is met, the optimal parameter coefficients of the key reactions under the current working conditions are output, and the final simplified mechanism is constructed accordingly.

[0028] Furthermore, steps 1, 2, and 3 further include the following steps:

[0029] The combustion characteristics: ignition delay period, laminar flame speed, and nitrogen oxide concentration were selected as verification parameters. The ignition delay period, laminar flame speed, and nitrogen oxide concentration of different simplified mechanisms and optimized mechanisms were verified using experimental and simulation data.

[0030] Furthermore, steps 1 and 3 further include the following steps respectively:

[0031] The determination coefficient greater than 0.99 and the minimum mean square loss error less than 0.02 are selected as the basis for judging whether the deep neural network model I in step 1.5 has achieved accuracy.

[0032] The determination coefficient greater than 0.99 and the minimum mean square loss error less than 0.02 are selected as the basis for judging whether the deep neural network model II in step 3.3 has achieved accuracy.

[0033] Furthermore, the step 2 further comprises the following steps:

[0034] The initial threshold, step increment and termination threshold of the singular perturbation method are calculated as 0.02, 0.01 and 0.2 respectively.

[0035] Furthermore, the step 3 further comprises the following steps:

[0036] The fitness function value of each random sample is calculated using the trained deep neural network model. The calculation formula is as follows:

[0037]

[0038] Where f represents the fitness function value; τ, μ, and ω represent the laminar combustion velocity, ignition delay time, and nitrogen oxide concentration, respectively. Then τ0 represents the ignition delay time predicted by the neural network model, τ i represents the corresponding experimental value of ignition delay time found in the literature, μ0 represents the laminar combustion velocity predicted by the neural network model, and μ i represents the corresponding laminar combustion velocity experimental value found in the literature, ω0 represents the nitrogen oxide concentration predicted by the neural network model, ω i represents the nitrogen oxide concentration; n, m and k represent the operating condition number corresponding to the flow combustion velocity, ignition delay time and nitrogen oxide concentration, respectively.

[0039] Furthermore, the step 3 further comprises the following steps:

[0040] In the relevant settings of the genetic algorithm, the population with the largest fitness function value is directly eliminated according to the principle of survival of the fittest, and the elite population retains 2% of the total population. The crossover method is single-point crossover, and the probability of crossover between the two samples is 0.5. Bidirectional mutation is used, and the mutation probability is 0.15.

[0041] The beneficial effects of the present invention are:

[0042] This invention provides a progressive mechanism simplification and optimization method based on a two-stage deep neural network model: component screening, deep reaction simplification, and parameter optimization. This method uses a deep neural network model to replace traditional kinetic simulation software to predict combustion processes, significantly improving computational efficiency and search capabilities in the process of exploring simplified chemical reaction kinetic mechanisms. Testing has verified that this method can effectively maintain high prediction fidelity for key combustion characteristics such as laminar combustion velocity and ignition delay time while constructing a highly simplified combustion mechanism. It also exhibits good versatility and adaptability, making it suitable for simulating the combustion of various fuel types. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Overview of the mechanistic simplification and optimization approach constructed for the present invention.

[0044] Figure 2 This is a simplified flowchart of the components of step (1) in the method constructed by the present invention based on the deep neural network model.

[0045] Figure 3 A structural diagram of the first-stage deep neural network model I for simplifying the ammonia / methanol / hydrogen mechanism using the mechanism simplification and optimization method constructed by the present invention.

[0046] Figure 4 The present invention is used to simplify the ammonia / methanol / hydrogen mechanism by using the mechanism simplification and optimization method constructed in the present invention. ...

[0047] Figure 5 This is a flow chart of step (3) in the method constructed by the present invention for optimizing the combustion mechanism based on a deep neural network model combined with a genetic algorithm.

[0048] Figure 6 A schematic diagram of sensitivity analysis under multiple operating conditions when optimizing the ammonia / methanol / hydrogen mechanism using the mechanism simplification and optimization method constructed by the present invention; Figure 6 (a) is a sensitivity analysis diagram based on laminar burning velocity; Figure 6 (b) is a sensitivity analysis diagram based on ignition delay time.

[0049] Figure 7 Structural diagram of the second-stage deep neural network model II for optimizing the ammonia / methanol / hydrogen mechanism using the mechanism simplification and optimization method constructed by the present invention.

[0050] Figure 8 The present invention is used to optimize the ammonia / methanol / hydrogen mechanism by using the mechanism simplification and optimization method constructed in the present invention, and the fitness function changes with the genetic generation number.

[0051] Figure 9 A comparison chart of the laminar combustion velocity and ignition delay time simulation results of the ammonia / methanol / hydrogen combustion mechanism constructed by the present invention and the mechanism generated by the traditional simplified method, as well as the corresponding experimental data; Figure 9 (a) is a diagram for verifying the laminar combustion velocity of ammonia hydrogen and methanol hydrogen; Figure 9 (b) is the laminar combustion velocity verification diagram of ammonia methanol and hydrogen methanol; Figure 9 (c) Ignition delay time verification diagram for pure ammonia and ammonia-hydrogen; Figure 9 (d) is the ignition delay time verification diagram of methanol and ammonia methanol; Figure 9 (e) is the verification diagram of nitrogen oxide concentration of ammonia hydrogen. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention, but the protection scope of the present invention is not limited to the embodiments.

[0053] The present invention simplifies and optimizes the ammonia / methanol / hydrogen (NH3 / CH3OH / H2) mechanism, which involves 59 components and 344 reactions. The mechanism is compared with the mechanism obtained using a traditional simplified approach in the study "Study on ammonia / methanol blends with ammonia cracking for low-carbon combustion and NO2 reduction" by Meng et al. The present invention is described in detail through analysis of this example.

[0054] Attachment Figure 1 As shown, a combustion mechanism simplification and optimization method based on a two-stage deep neural network algorithm includes the following steps:

[0055] Step (1): Construct and train a deep neural network model I (DNN I) with component retention state and initial conditions including temperature, pressure, etc. as input and combustion characteristics as output, to replace the traditional reaction kinetics simulation software Cantera to achieve rapid screening and gradual simplification of components in the reaction mechanism, thereby obtaining the most simplified combustion mechanism of components. Figure 2 As shown, the specific content and method steps include:

[0056] Step 1.1: Randomly delete 6 components from the detailed mechanism to obtain multiple simplified mechanisms. Use Cantera to simulate the combustion characteristics under different initial conditions, such as laminar burning velocity (LBV), ignition delay time (IDT) and nitrogen oxides (NO X ) concentration to obtain the initial data set, where the LBV operating range is: pressure (P) 1-7atm, temperature (T) 298-343K, equivalence ratio (Ф) 0.6-1; ignition delay time test conditions: P 1.2-40atm, T 920-2490K, NO X Test conditions: NO, P is 1-4atm, T is 368K, Φ is 0.8-1.2. Finally, the results for LBV, IDT and NO were obtained. X The datasets contain 170,000, 340,000, and 56,000 sets of data respectively, where the datasets are divided into 80% training set, 10% validation set, and 10% test set;

[0057] Step 1.2: Train DNNⅠ, where the input is the initial temperature (T0), pressure (P0) and the corresponding "0" and "1", and the output is the corresponding combustion characteristics. Figure 3As shown in the figure, after multiple comparisons and debugging, it was finally determined that the model contains 4 hidden layers, each hidden layer contains 1500 neurons, the activation function uses the ReLU function, the optimizer uses Adam, the learning rate is set to 0.0002, and the training rounds are set to 1600 rounds. Finally, the minimum mean square error of the two models on the test set is less than 0.02, and the determination coefficient is greater than 0.99, and the training is completed;

[0058] Step 1.3: Use the trained DNNⅠ to select simplified mechanisms whose predictions from the deep neural network are within 10% of the experimental values.

[0059] Step 1.4: Use Cantera to calculate the combustion characteristics under different working conditions for the simplified NH3 / CH3OH / H2 mechanism obtained in the third step, and perform a secondary screening;

[0060] Step 1.5: Use the combustion characteristic dataset obtained in step 4 to train DNNⅠ again and continue to screen the simplified mechanism for the next stage;

[0061] Step 1.6: Repeat steps 1.2 to 1.5 to gradually delete the components. The relationship between the number of components, the number of reactions, and the corresponding errors based on LBV and IDT as a function of the number of iterations in the simplified NH3 / CH3OH / H2 mechanism using this invention is shown in the attached figure. Figure 4 As shown, the final simplified mechanism containing 30 components and 168 reactions was obtained.

[0062] Step (2): Based on the simplified mechanism of components obtained in step (1), the computational singular perturbation method (CSP) is used to deeply simplify the reaction paths in the mechanism. By setting the time scale threshold, non-critical fast reactions are identified and eliminated, and finally a simplified mechanism with a large number of reactions is generated:

[0063] The threshold for applying CSP was initially set to 0.01 and gradually adjusted from 0.02 to 0.2. Finally, the LBV-based mechanism containing 30 components and 73 reactions and the IDT-based mechanism containing 30 substances and 78 reactions were simplified using CSP, and finally a mechanism containing 30 components and 92 reactions was obtained. Compared with the simulated values of the detailed mechanism, the LBV, IDT and NO X The average errors of concentration were 23.8%, 10.24% and 15.36%, respectively.

[0064] Step (3): In order to address the prediction deviation that may be introduced by the simplification of the reaction path in step (2), a deep neural network model II (DNN II) with key reaction kinetic parameters as input is constructed, and a genetic algorithm (GA) is combined to perform global optimization of the mechanism parameters to restore or improve the prediction accuracy of the mechanism on key combustion performance indicators. Figure 5As shown, the specific content and method steps include:

[0065] Step 3.1: Perform sensitivity analysis on the mechanism obtained in step (2) under various working conditions based on LBV and IDT, as shown in the attached figure. Figure 6 As shown, 17 key reactions were finally identified for optimization;

[0066] Step 3.2: Randomly modify the pre-exponential factors (A) of 17 key reactions, where the change range of A is [0.001, 1000]. Use Cantera simulation to obtain LBV, IDT and NO under different initial conditions. X The initial data sets obtained for concentration contained 150,000, 270,000, and 40,000 sets of data, respectively;

[0067] Step 3.3: Train DNNⅡ with inputs T0, P0 and A1-A 17 The output is the corresponding combustion characteristics. Figure 7 As shown in the figure, after multiple comparisons and debugging, it was finally determined that the model contains 3 hidden layers, each hidden layer contains 600 neurons, the activation function uses the ReLU function, the optimizer uses Adam, the learning rate is set to 0.0008, and the number of training rounds is set to 1600. Finally, the minimum mean square error of the two models on the test set is less than 0.02, and the determination coefficient is greater than 0.99, and the training is completed;

[0068] Step 3.4: Initialize the population and use the hybrid initialization method to generate samples. The initial population is set to 200.

[0069] Step 3.5: Use the trained DNNⅡ to calculate the fitness function value and evaluate the fitness function of 200 samples to determine the inheritance and elimination status;

[0070] Step 3.6: selection and crossover, where the crossover probability is set to 0.5;

[0071] Step 3.7: Bidirectional mutation, where the mutation probability is set to 0.15;

[0072] Step 3.8: As attached Figure 8 As shown in the figure, after 600 iterations, the fitness function value of the best individual was 0.089, which is within the acceptable range. Therefore, the optimal A for 17 key reactions was found, and the mechanism optimization was completed. The specific key reaction change multiples are shown in the table below.

[0073] Table 1

[0074]

[0075] The NH3 / CH3OH / H2 combustion mechanism finally constructed by the present invention, which includes 30 components and 92 reactions, was simulated and compared with the mechanism obtained by the traditional simplified method, which includes 42 components and 200 reactions. The results were verified by combining experimental data. Figure 9 The comparison results show that the mechanism constructed by the present invention has a great effect on LBV, IDT and NO under the premise of maintaining a small number of components and reactions. X The simulation error of key combustion characteristics such as concentration is small, which can more accurately reflect the actual combustion behavior of the mixed fuel.

[0076] The above-described embodiments merely express the implementation methods of the present invention, but should not be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A combustion mechanism simplification and optimization method based on a two-stage deep neural network algorithm, characterized in that: The combustion mechanism simplification and optimization method comprises the following steps: Step 1: Build and train a deep neural network model. The input of the deep neural network model is the component retention state and initial conditions, and the combustion characteristics are output. The model can achieve rapid screening and gradual simplification of the components in the reaction mechanism, thereby obtaining a combustion mechanism with the most simplified components. Step 2: Based on the simplest component mechanism obtained in step 1, the computational singular perturbation method is used to deeply simplify the reaction paths in the mechanism. By setting a time scale threshold, non-critical fast reactions are identified and eliminated, and finally a simplified mechanism that meets the requirements of the limited number of reactions is generated; In step 3, in order to address the prediction deviation that may be introduced due to the deep simplification of the reaction path in step 2, a deep neural network model II is constructed with key reaction kinetic parameters as input, and a genetic algorithm is combined to perform global optimization of the mechanism parameters to restore or improve the prediction accuracy of the mechanism in the corresponding combustion characteristics.

2. The method for simplifying and optimizing the combustion mechanism based on a two-stage deep neural network algorithm according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1: Randomly delete k components from the chemical reaction mechanism to generate multiple sets of simplified mechanisms. These mechanisms are then simulated under different initial conditions using traditional reaction kinetics simulation software to obtain corresponding combustion characteristic data and construct an initial dataset, which is divided into a training set, a validation set, and a test set. Step 1.2: Based on the initial dataset constructed in step 1.1, the first-stage deep neural network model I is trained. The model takes temperature, pressure, equivalence ratio, blending ratio, and a 01 vector (where "1" represents a component to be retained and "0" represents a component to be deleted) as input, and outputs the combustion characteristics to be predicted. In step 1.3, using the deep neural network model I trained in step 1.2, quickly predict the combustion characteristics of all simplified mechanisms generated in step 1.1, compare them with the corresponding experimental values, calculate the average error, and extract the set of mechanisms whose errors are within the set threshold; Step 1.4: Perform computational verification on the mechanism set screened in step 1.3 under different operating conditions to obtain more accurate combustion characteristic data and construct a secondary data set. Step 1.5: Use the secondary dataset constructed in step 1.4 to train the deep neural network model I again for the next stage of mechanism screening and evaluation, further screening out redundant components and retaining key components; Step 1.6, step 1.2 to step 1.5 constitute an iterative process. Each round of iteration gradually reduces the number of deleted components k. As the number of iterations increases, the intensity of component deletion gradually decreases, and finally the process of constructing the simplest component mechanism is completed.

3. The method for simplifying and optimizing the combustion mechanism based on a two-stage deep neural network algorithm according to claim 2 is characterized in that: In step 1: In step 1.3, the threshold is set to 10%; In step 1.4, the mechanism set obtained by screening in step 1.3 is calculated and verified under different working conditions using the traditional dynamics simulation software Cantera; In step 1.5, the coefficient of determination greater than 0.99 and the minimum mean square loss error less than 0.02 are selected as the basis for judging whether the deep neural network model I in step 1.5 has achieved accuracy; In step 1.6, when the average error calculated by comparing the predicted value of the mechanism for combustion characteristics with the experimental value exceeds 10%, the process of constructing the simplest component mechanism under the threshold condition is completed.

4. The method for simplifying and optimizing the combustion mechanism based on a two-stage deep neural network algorithm according to claim 2 is characterized in that: The step 2 is specifically as follows: Step 2.1, setting the initial threshold, step increment, and termination threshold for reaction simplification; Step 2.2: For different combustion characteristics, the computational singular perturbation method is used to perform perturbation analysis to obtain the corresponding simplified combustion mechanism; Step 2.3, combine the simplified mechanisms obtained in step 2.2 based on different combustion characteristics to obtain the simplest mechanism for the reaction number.

5. The method for simplifying and optimizing the combustion mechanism based on a two-stage deep neural network algorithm according to claim 4 is characterized in that: In step 2.2, the initial threshold, step increment and termination threshold of the singular perturbation method are calculated to be 0.02, 0.01 and 0.2 respectively.

6. The method for simplifying and optimizing the combustion mechanism based on a two-stage deep neural network algorithm according to claim 4 is characterized in that: The step 3 is specifically as follows: Step 3.1: Perform sensitivity analysis on the simplest mechanism of the reaction number obtained in step 2.3 to determine the key reactions that have a greater impact on the combustion characteristics; Step 3.2: Randomly modify the pre-exponential factors of key reactions to generate a mechanism simulation to obtain the combustion characteristics under different initial conditions as a data set; Step 3.3: Train the deep neural network model II, taking temperature, pressure, equivalence ratio, blending ratio, modified pre-exponential factor, and activation energy as inputs, and outputting the desired predicted combustion characteristics. A coefficient of determination greater than 0.99 and a minimum mean square error less than 0.02 are used as the basis for determining whether the deep neural network model II in step 3.3 has achieved accuracy. Step 3.4, using a hybrid initialization method to generate an initial population sample, the initial population sample generated by the hybrid initialization method includes multiple random samples; Step 3.5: Use the deep neural network model II trained in step 3.3 to evaluate the fitness of each sample obtained in step 3.4, calculate the fitness function value, and retain the sample with the smallest fitness function value; Step 3.6, perform selection and crossover operations to reorganize elite individuals with smaller fitness function values with the retained samples; Step 3.7: Increase the diversity of the population by setting mutation operations to prevent the genetic algorithm from falling into local optimality; In step 3.8, when the set maximum number of iterations is reached or the early termination condition is met, the optimal parameter coefficients of the key reactions under the current working conditions are output, and the final simplified mechanism is constructed accordingly.

7. The method for simplifying and optimizing the combustion mechanism based on a two-stage deep neural network algorithm according to claim 6 is characterized in that: In step 3, the fitness function value of each random sample obtained in step 3.4 is calculated using the trained deep neural network model II. The calculation formula is as follows: Where f represents the fitness function value; τ, μ, and ω represent the laminar combustion velocity, ignition delay time, and nitrogen oxide concentration, respectively. Then τ0 represents the predicted ignition delay time, τ i represents the experimental value of ignition delay time, μ0 represents the predicted laminar combustion velocity, μ i represents the experimental value of laminar combustion velocity, ω0 represents the predicted nitrogen oxide concentration, ω i represents the nitrogen oxide concentration; n, m and k represent the operating condition number corresponding to the flow combustion velocity, ignition delay time and nitrogen oxide concentration, respectively.

8. The method for simplifying and optimizing combustion mechanism based on a two-stage deep neural network algorithm according to claim 6, characterized in that: In step 3: In step 3.2, the range of the pre-exponential factor is [0.001, 1000]; In step 3.4, the multiple random samples are: 33% of the samples are uniformly distributed, 33% of the samples are Gaussian distributed, and 34% of the samples are constant individuals.

9. The method for simplifying and optimizing combustion mechanism based on a two-stage deep neural network algorithm according to claim 6, characterized in that: In the step 3.7, in the relevant settings of the genetic algorithm, the population with the largest fitness function value is directly eliminated according to the principle of survival of the fittest, and the elite population retains 2% of the total population. The crossover method is single-point crossover, and the probability of crossover between the two samples is 0.

5. The two-way mutation has a mutation probability of 0.

15.

10. The method for simplifying and optimizing combustion mechanism based on a two-stage deep neural network algorithm according to claim 1, characterized in that: Combustion characteristics are selected as verification parameters, including ignition delay period, laminar flame speed, and nitrogen oxide concentration. The ignition delay period, laminar flame speed, and nitrogen oxide concentration of different simplified mechanisms and optimized mechanisms are verified using experimental and simulation data.

Images