A method to improve the accuracy of dynamic simulation of kinetic reactors

Abstract

A method for improving the accuracy of dynamic simulation of a kinetic reactor includes establishing a basic simulation model of the kinetic reactor, designing a training framework, and using the training framework as an agent and the basic simulation model. The agent is used as the training target, the basic simulation model as the learning environment, and the values ​​of key parameters as the action space of the agent. Based on the training framework, the agent is trained using actual operating data of the kinetic reactor to obtain the optimized key parameter value selection strategy under different operating conditions. The basic simulation model is then optimized to improve its accuracy. This invention constructs a dynamic model of a kinetic reactor that covers all operating conditions and possesses high accuracy, high stability, and high computational efficiency.

Description

Technical Field

[0001] This invention relates to the field of kinetic reactor simulation technology, and in particular to a method for improving the accuracy of dynamic simulation of kinetic reactors. Background Technology

[0002] Kinetic reactors have wide applications in various fields such as chemical engineering, energy, and environment, playing a crucial role, especially in fine chemicals, syngas production, and catalytic cracking. Due to the highly nonlinear and time-varying characteristics of heat transfer, mass transfer, and reaction processes within the reactor, their dynamic behavior directly affects the stability and safety of the entire process. Therefore, establishing accurate and reliable dynamic simulation models of kinetic reactors is of great significance for achieving efficient process control, process optimization, and safety analysis.

[0003] Current mainstream kinetic reactor modeling methods still face many challenges in practical applications. Existing models mainly fall into two categories: mechanistic modeling and data-driven modeling. Mechanistic modeling is based on first principles and constructs mathematical models through fundamental chemical engineering theories such as material balance, energy balance, phase equilibrium, and reaction kinetics. Data-driven modeling relies on historical operational and running data and uses machine learning or deep learning methods (such as neural networks, support vector machines, and random forests) to establish nonlinear mapping relationships between input and output variables.

[0004] Both of these modeling methods currently have certain limitations: Mechanistic modeling involves a large number of parameters that need to be measured or estimated (such as rate constants, activation energies, mass transfer coefficients, etc.) due to the complex chemical reaction processes, making model parameter tuning quite complex. Simulation results may be affected by uncertainties, and its ability to cover all operating conditions is limited. The main shortcomings of data-driven modeling lie in its extrapolation capabilities, with prediction accuracy decreasing under operating conditions outside the training data range. In addition, the model presents a "black box" structure, with weak interpretability, making it difficult to fully meet the requirements of process safety, controllability, and transparency.

[0005] Therefore, the existing technology has problems and needs further improvement and development. Summary of the Invention

[0006] (I) Purpose of the invention: In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a method to improve the accuracy of dynamic simulation of kinetic reactors. By combining the advantages of mechanism modeling and intelligent algorithms, the problem of difficulty in accurately determining key parameters and insufficient model accuracy in dynamic simulation models of kinetic reactors can be solved.

[0007] (II) Technical Solution: To solve the above-mentioned technical problems, this technical solution provides a method for improving the accuracy of dynamic simulation of kinetic reactors, including the following steps:

[0008] Step 1: Establish a basic simulation model of the kinetic reactor and determine the key parameters in the basic simulation model;

[0009] Step 2: Design a training framework, which includes an agent and a basic simulation model. The agent is used as the training target, the basic simulation model is used as the learning environment, and the values ​​of the key parameters are used as the action space of the agent.

[0010] Step 3: Based on the training framework, train the agent using the actual operating data of the kinetic reactor to obtain the optimization strategy for the key parameter values ​​of the agent under different working conditions.

[0011] Step 4: Apply the trained agent to the basic simulation model and conduct simulation tests. The agent, through simulation tests, adopts different strategies according to the real-time operating conditions of the kinetic reactor, outputs optimized key parameter values, and optimizes the parameters of the basic simulation model.

[0012] Furthermore, step 1 includes the following key parameters in the basic simulation model: the pressure-affected reaction rate coefficient D, the activation energy Ea, and the pre-exponential factor A.

[0013] Furthermore, step 2 includes the agent adopting a dual-network structure of an Actor network and a Critic network based on deep deterministic policy gradients.

[0014] Furthermore, step 2 includes introducing long short-term memory layers into both the Actor network and the Critic network to capture temporal characteristics during the dynamic process of the reactor.

[0015] Furthermore, step 2 also includes constructing a reward function using the negative sum of squared total system errors, the reward function being M = -(y sim - y true ) 2 In the formula, M is the reward value, and y sim For simulation output variables, y true For input variables.

[0016] Furthermore, step 3 includes introducing a physical constraint layer during training to perform boundary checks on the key parameters output by the agent, forcibly restricting the key parameters to within the thermodynamically feasible domain, Ea∈[40,400] kJ / mol, A∈[10... 6 10 14 ]s -1 .

[0017] Furthermore, step 3 also includes introducing a noise adaptation mechanism during the training of the agent, adding Ornstein-Uhlenbeck noise to the output layer of the Actor network, with the noise intensity decaying exponentially with the training rounds, the initial noise intensity being 0.2 and the decay rate being 0.01.

[0018] Furthermore, step 3 also includes introducing a smoothing filter into the output layer of the Actor network to suppress the action output oscillations generated by the agent during continuous training. The smoothing filter uses an exponentially weighted moving average (EWMA) to perform low-pass filtering on the adjustment of key parameters in continuous time steps, with the filter coefficient μ∈[0.7,0.95], which adaptively decays with each training round.

[0019] Furthermore, step 3 also includes collecting actual operating data of the kinetic reactor under different operating conditions, using the actual operating data under different operating conditions as input variables, and training the agent through the training framework of step 2 to train the agent's value selection strategy for key parameters under different operating conditions.

[0020] Furthermore, step 4 includes calculating the optimized independent error and the total system error, if ∈[0,v] and all e i All less than or equal to v i If the optimization is successful, the simulation output variables will be output as the simulation result; otherwise, return to step 2 to retrain the agent.

[0021] (III) Beneficial Effects: This invention utilizes a trained intelligent agent, coupled with a basic simulation model, and leverages historical data to achieve accurate modeling of the dynamic behavior of complex reactors. It constructs a dynamic model of a kinetic reactor that covers all operating conditions and possesses high precision, high stability, and high computational efficiency. This provides a new technical path for process simulation, optimization, and control in engineering practice, effectively solving the problems of difficulty in accurately determining key parameters and insufficient model accuracy in dynamic simulation models of kinetic reactors. It provides a more reliable simulation tool for in-depth research and practical application of kinetic reactors. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the method flow of the present invention;

[0023] Figure 2 is a schematic diagram of the training framework of the present invention;

[0024] Figure 3 This is a schematic diagram of the simulation test process of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to preferred embodiments. More details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention can obviously be implemented in many other ways different from those described herein. Those skilled in the art can make similar extensions and derivations based on actual application situations without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0026] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that these drawings are for illustrative purposes only and are not drawn to scale, and should not be construed as limiting the actual scope of protection of the present invention.

[0027] like Figure 1 As shown, a method for improving the accuracy of dynamic simulation of a kinetic reactor includes the following steps:

[0028] Step 1: Establish a basic simulation model of the kinetic reactor and determine the key parameters in the basic simulation model;

[0029] Step 2: Design a training framework, which includes an agent and a basic simulation model. The agent is used as the training target, the basic simulation model is used as the learning environment, and the values ​​of the key parameters are used as the action space of the agent.

[0030] Step 3: Based on the training framework, train the agent using the actual operating data of the kinetic reactor to obtain the optimization strategy for the key parameter values ​​of the agent under different working conditions.

[0031] Step 4: Apply the trained agent to the basic simulation model and conduct simulation tests. The agent, through simulation tests, adopts different strategies according to the real-time operating conditions of the kinetic reactor, outputs optimized key parameter values, and optimizes the parameters of the basic simulation model.

[0032] Step 1 specifically includes:

[0033] Step 101: Based on the fundamental mechanism of a kinetic reactor, construct a basic simulation model of the kinetic reactor by describing its physical and chemical behavior using mathematical equations. The basic simulation model of the kinetic reactor is as follows:

[0034] The basic simulation model of a kinetic reactor is as follows:

[0035] ;

[0036] in, .

[0037] In the formula, For reactor temperature, For feed temperature, The density of the reaction mixture, The specific heat capacity of the reaction mixture is... D is the residence time, and D is the coefficient of pressure affecting the reaction rate. The effect of pressure on the reaction rate constant, The reaction rate constant is affected by temperature. The heat of reaction, Where is the reactant concentration, and K is the reaction rate constant. For pressure, For temperature, Here, Ea is the pre-exponential factor, R is the activation energy, U is the gas constant, A is the heat transfer area, and Tc is the temperature of the cooling / heating medium.

[0038] The key parameters in the basic simulation model described in steps 102 and 101 include the pressure-affected reaction rate coefficient D, the activation energy Ea, and the pre-exponential factor A. These key parameters are divided into three independent parameter levels: level D, level Ea, and level A. Each level corresponds to a specific physical influence factor. Level D corresponds to the pressure influence factor, representing the pressure correction coefficient for the reaction rate; level Ea corresponds to the temperature influence factor, representing the reaction's sensitivity to temperature; and level A corresponds to the intrinsic reaction factor, representing the inherent rate characteristics of the reaction.

[0039] like Figure 2 As shown, the training framework for step 2 specifically includes:

[0040] Step 201: Construct a trained agent, which includes a neural network for learning the mapping relationship between the state of the basic simulation model and the key parameter values.

[0041] The agent employs a dual-network structure of an Actor network and a Critic network based on deep deterministic policy gradients, and introduces a Long Short-Term Memory (LSTM) layer to capture the temporal characteristics of the dynamic simulation process of the kinetic reactor. The specific structure of the Actor network includes the input [s] {t-n+1} , ..., s t The state sequence, after being processed by an LSTM layer, is then passed through a fully connected Tanh activation function to output the correction values ​​for each parameter level, thus realizing decision-making in the action space. The specific structure of the Critic network includes the input [s...]. {t-n+1} , ..., s t ]+[a {t-n+1} , ..., a tThe concatenated sequence of state and action sequences is processed by an LSTM layer and then output as an action-value function Q(s) through a fully connected Tanh layer. t ,a t ), used to evaluate the quality of the Actor network's output actions.

[0042] Meanwhile, both the Actor and Critic networks synchronize the parameters of the main and target networks through a soft update mechanism to ensure training stability. Preferably, the soft update coefficient τ of the soft update mechanism is 0.005.

[0043] Step 202: Design a hierarchical activation mechanism. First, collect operating parameters, including feed temperature fluctuation. Pressure change rate and reactant concentration deviation Then, preset operating condition thresholds are used to trigger the activation / dormancy of the corresponding parameter levels. When the fluctuations of all levels are less than or equal to the preset operating condition thresholds, it is set to a steady-state condition. At this time, level A is activated to maintain the stability of the intrinsic response; levels D and Ea are dormant, and the key parameter values ​​remain the optimization results of the previous moment. When the fluctuation of any level exceeds the preset operating condition threshold, it is set to a transient condition. At this time, levels D and Ea are activated simultaneously to quickly suppress the fluctuations, while level A remains active.

[0044] Step 203: Use the basic simulation model as the learning environment; use the values ​​of the key parameters as the action space of the agent; use the historical production data of the kinetic reactor as input variables, which are extracted from historical operation records and used as the initial conditions for simulation. The input variables include parameters such as the feed temperature, reactant concentration, and pressure of the kinetic reactor; input the state sequence of the input variables into the agent. Input the corrected key parameter values ​​as the continuous action space into the basic simulation model. The basic simulation model generates simulation output variables based on the corrected key parameter values ​​and input variables, and the parameters of the simulation output variables correspond one-to-one with the parameters of the input variables.

[0045] Step 204: Compare the simulation output variables with the input variables, calculate the independent error at each level and the total system error; perform corrections at each parameter level, and the calculation formulas for the correction amounts at each parameter level are as follows:

[0046] ,

[0047] In the formula, i represents the parameter level. The hierarchical learning rate is set to 0.3 for steady-state conditions and 0.8 for transient conditions. The total system error, Let i be the independent error of the i-th level. This represents the correction amount for each parameter level.

[0048] in, , This refers to the hierarchical weight.

[0049] Step 205: Dynamic correction of target parameter values. Based on real-time operating conditions and errors, target parameter values ​​are dynamically generated.

[0050] ，

[0051] In the formula, Let i be the target value of the i-th level at time t. These are the initial values ​​for the basic simulation model at level i. Let be the error correction coefficient for the i-th level at time t. Let be the independent error of the i-th level at time t. Let be the operating condition sensitivity coefficient of the i-th level at time t. Let be the operating condition disturbance value at time t.

[0052] Step 206: Construct a parameter-level collaborative coupling mechanism. Determine the hierarchical influence weight matrix: W=[w ij ], w ij This represents the influence coefficient of level j on level i (0≤w). ij ≤1), diagonal w ij =1 and has self-influence.

[0053] The initial matrix is:

[0054] ,

[0055] When P(t) > 1.2P, the system is updated to high-pressure condition to enhance the influence of Ea on D. The high-pressure condition matrix is ​​as follows:

[0056] .

[0057] Step 207: The parameters of each level in the kinetic reactor are not completely independent. For example, under high pressure, the influence of temperature on pressure is enhanced. The coupling term dynamically assigns influence coefficients between levels through the weight matrix W, so that the parameter correction value output by the Actor network is not only based on its own level state, but also integrates the optimization results of other related levels, avoiding system imbalance caused by single parameter optimization. This is especially suitable for simulation scenarios of transient fluctuations or extreme conditions. Therefore, a coupling term is introduced into the output layer of the Actor network to achieve hierarchical collaboration. The following is an example of D-level coupling, with the following formula:

[0058] ,

[0059] In the formula, Here, represents the corrected value for the D-level parameter, and j represents other levels besides D. This is a state sequence at level D. For based on Input variables, Let W be the influence coefficient of the j-th level on the D level in the weight matrix. This is the corrected value for the j-th level.

[0060] Step 208: Dynamically adjust the action space boundary based on the hierarchical error. If there are 5 consecutive rounds e i If the decline rate is less than 5%, then exploration is enhanced, and the action space boundary is expanded by 20%; if e i Less than or equal to 0.5V i If v > 0, then convergence accelerates, and the action space boundary shrinks by 10%. i This is the hierarchical error threshold.

[0061] Step 209: Construct the reward function using the negative sum of squared total system errors. Reward function: M = -(y sim - y true ) 2 In the formula, M is the reward value, and y sim For simulation output variables, y true The input variable is defined as follows: the smaller the total system error, the greater the reward value. This variable is used to measure the difference between the simulation output variable and the actual historical production data, guiding the agent to learn in the direction of reducing error.

[0062] Step 3 specifically includes:

[0063] Collect actual operating data of the kinetic reactor under different operating conditions. Use this data as input variables and train the agent using the training framework described in step 2. Gradually train the agent's optimal key parameter value selection strategy under different operating conditions. Simultaneously, introduce a physical constraint layer during training to perform boundary checks on the key parameters output by the agent, forcibly restricting the key parameters to within the thermodynamically feasible domain. Preferably, Ea∈[40,400] kJ / mol, A∈[10... 6 10 14 ]s -1 .

[0064] The agent training incorporates a noise adaptation mechanism, adding Ornstein-Uhlenbeck noise to the output layer of the Actor network. The noise intensity decays exponentially with each training round. Preferably, the initial noise intensity is 0.2, and the decay rate is 0.01.

[0065] Here is a specific example to illustrate the training process, which is as follows:

[0066] In each training round, the agent starts from the initial state state0 and constructs a state sequence state of length 5. sequence The input is then fed into the Actor network, which outputs action values, which are then perturbed by Gaussian noise. These actions are assigned to key parameters in the basic simulation model. The kinetic reactor then runs and outputs simulation output variables. The difference between the simulation output variables and the input variables is used to adjust the action space boundary, and the interaction data (s, a, r, s', done) is stored in the experience playback buffer.

[0067] When the amount of data in the buffer is greater than the batch size, BATCH SIZE When the value is 16, the system begins to train the network using the experience replay mechanism: the Critic network is updated according to the target Q-value function, and the target value is calculated using the target Actor network and the target Critic network; the Actor network updates its policy by maximizing the Q-value output of the current Critic network.

[0068] Meanwhile, the target network is updated using a soft update coefficient θ=0.005 to maintain learning stability.

[0069] In addition, to enhance the exploration capability during training, an Ornstein-Uhlenbeck noise mechanism is introduced. The noise intensity gradually decreases as training progresses, linearly decreasing from 0.2 to 0.01, thus transitioning the strategy from early exploration to later convergence, balancing exploration and exploitation.

[0070] To suppress the oscillations in the action output of the agent during continuous training, a smoothing filter is introduced into the output layer of the Actor network. The smoothing filter uses an exponentially weighted moving average (EWMA) to perform low-pass filtering on the adjustment of key parameters in continuous time steps. The filter coefficient μ∈[0.7,0.95] decays adaptively with training rounds to ensure smooth output actions and fast convergence.

[0071] When the total systematic error of 100 consecutive rounds If the value is less than or equal to v, where v is a preset threshold, then training is stopped and the key parameter values ​​of the agent are saved.

[0072] According to the reward value curve during the training process, the reward value gradually increases with the number of iterations, indicating that the basic simulation model continuously tries to learn better strategies in the learning environment exploration, and finally finds a better strategy in the later stage. That is, the reward value keeps increasing. The larger the reward value, the smaller the deviation between the simulation prediction value and the real value. The simulation output variable corrected by training gradually reduces the error with the real value until the optimal strategy is found.

[0073] like Figure 3 As shown, step 4 specifically includes:

[0074] Step 401: Apply the trained agent to the basic simulation model and conduct simulation tests.

[0075] Step 401 specifically includes:

[0076] Step 4011: Input the state sequence of the input variables into the trained agent. The trained agent outputs the optimal key parameter values ​​according to the optimal strategy obtained in step 3 and inputs them into the basic simulation model.

[0077] Step 4012: Input the real-time operating condition data as an input variable into the basic simulation model.

[0078] Step 4013: Based on the parameters of the input variables, activate the D and Ea levels and put the A level into dormancy; couple the correction values ​​of Ea and D through the weight matrix W, and output the optimized D and Ea values.

[0079] Step 4014: Calculate the optimized independent error and the total system error. If ∈[0,v] and all e i All less than or equal to v i If the optimization is successful, the simulation output variables will be output as the simulation results; otherwise, return to step 2 to retrain the agent.

[0080] Step 402: The successful optimization of the agent in step 4014 is defined as the agent passing the simulation test. The agent that passes the simulation test can dynamically switch the level activation state according to the real-time working conditions of the kinetic reactor at each simulation time step in subsequent actual production, and continuously output the optimized key parameter values ​​to optimize the parameters of the basic simulation model and improve the accuracy of the basic simulation model under all working conditions.

[0081] In this invention, hierarchical dynamic control is layered according to the physical meaning of parameters. Activation / dormancy is precisely triggered by operating condition fluctuation thresholds, avoiding redundant calculations in "one-size-fits-all" parameter optimization and improving efficiency. Simultaneously, the collaborative coupling mechanism dynamically adjusts the hierarchical influence coefficients using a weight matrix and corrects them through Actor network coupling terms, addressing the problem of neglecting correlations in single-parameter optimization and reducing cross-operating condition errors. The combination of these two approaches significantly reduces simulation errors across all operating conditions, substantially improves computational efficiency, and effectively covers transient and extreme operating conditions that traditional models struggle to handle.

[0082] This invention utilizes a trained intelligent agent, coupled with a basic simulation model, and leverages historical data to achieve accurate modeling of the dynamic behavior of complex reactors. It constructs a dynamic model of a kinetic reactor that covers all operating conditions and possesses high precision, high stability, and high computational efficiency. This provides a new technical path for process simulation, optimization, and control in engineering practice, effectively solving the problems of difficulty in accurately determining key parameters and insufficient model accuracy in dynamic simulation models of kinetic reactors. It provides a more reliable simulation tool for in-depth research and practical application of kinetic reactors.

[0083] The above description illustrates preferred embodiments of the present invention and helps those skilled in the art to more fully understand the technical solution of the present invention. However, these embodiments are merely illustrative and should not be construed as limiting the specific implementation of the present invention to these embodiments. For those skilled in the art, several simple deductions and modifications can be made without departing from the inventive concept, and all such modifications should be considered within the protection scope of the present invention.

Claims

1. A method for improving the accuracy of dynamic simulation of a kinetic reactor, characterized in that, Includes the following steps: Step 1: Establish a basic simulation model of the kinetic reactor and determine the key parameters in the basic simulation model; Step 2: Design a training framework, which includes an agent and a basic simulation model. The agent is used as the training target, the basic simulation model is used as the learning environment, and the values ​​of the key parameters are used as the action space of the agent. Step 3: Based on the training framework, train the agent using the actual operating data of the kinetic reactor to obtain the optimization strategy for the key parameter values ​​of the agent under different working conditions. Step 4: Apply the trained agent to the basic simulation model and conduct simulation tests. The agent, through simulation tests, adopts different strategies according to the real-time operating conditions of the kinetic reactor, outputs optimized key parameter values, and optimizes the parameters of the basic simulation model. The key parameters are divided into three independent parameter levels: level D, level Ea, and level A. Each level corresponds to a specific physical influence factor. Level D corresponds to the pressure influence factor, which represents the correction coefficient of pressure on the reaction rate; level Ea corresponds to the temperature influence factor, which represents the sensitivity of the reaction to temperature; and level A corresponds to the intrinsic reaction factor, which represents the inherent rate characteristics of the reaction. Collect operating parameters, including feed temperature fluctuation. Pressure change rate and reactant concentration deviation ; Preset operating condition thresholds trigger activation / dormancy at the corresponding parameter level; When all level fluctuations are less than or equal to the preset operating condition threshold, the condition is set to steady state. At this time, level A is activated to maintain the stability of the intrinsic response; levels D and Ea are dormant, and the key parameter values ​​remain the optimization results of the previous moment. When the fluctuation of any level exceeds the preset operating condition threshold, it is set as a transient operating condition. At this time, levels D and Ea are activated simultaneously to quickly suppress the fluctuation, while level A remains activated. In a kinetic reactor, the parameters at each level are not completely independent. Under high-pressure conditions, the influence of temperature on pressure is amplified. Coupling terms are dynamically assigned influence coefficients between levels through a weight matrix W, allowing the parameter correction values ​​output by the Actor network to be based on its own level state and integrate the optimization results of other related levels. This avoids system imbalance caused by single-parameter optimization and is used for simulation scenarios of transient fluctuations or extreme conditions. Coupling terms are introduced into the output layer of the Actor network to achieve hierarchical collaboration. The D-level coupling formula is as follows: ， In the formula, Here, represents the corrected value for the D-level parameter, and j represents other levels besides D. The state sequence is at level D. Based on Input variables, Let W be the influence coefficient of the j-th level on the D level in the weight matrix. This is the corrected value for the j-th level.

2. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 1 includes key parameters in the basic simulation model, including the pressure-affected reaction rate coefficient D, activation energy Ea, and pre-exponential factor A.

3. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 2 includes the agent adopting a dual-network structure of an Actor network and a Critic network based on deep deterministic policy gradients.

4. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 3, characterized in that, Both the Actor network and the Critic network incorporate long short-term memory layers to capture temporal characteristics during the reactor's dynamic process.

5. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 2 further includes constructing a reward function using the negative sum of squared total system errors, the reward function being M = -(y sim - y true ) 2 In the formula, M is the reward value, and y sim For simulation output variables, y true For input variables.

6. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 3 includes introducing a physical constraint layer during training to perform boundary checks on the key parameters output by the agent, forcibly restricting the key parameters to within the thermodynamically feasible domain, Ea∈[40,400] kJ / mol, A∈[10... 6 10 14 ]s -1 .

7. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 3 also includes introducing a noise adaptation mechanism during the training of the agent, adding Ornstein-Uhlenbeck noise to the output layer of the Actor network, with the noise intensity decaying exponentially with the training rounds, the initial noise intensity being 0.2 and the decay rate being 0.

01.

8. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 3 also includes introducing a smoothing filter into the output layer of the Actor network to suppress the action output oscillations generated by the agent during continuous training. The smoothing filter uses an exponentially weighted moving average (EWMA) to perform low-pass filtering on the adjustment of key parameters in continuous time steps, with the filter coefficient μ∈[0.7,0.95], which adaptively decays with each training round.

9. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 3 also includes collecting actual operating data of the kinetic reactor under different operating conditions, using the actual operating data under different operating conditions as input variables, and training the agent through the training framework of step 2 to train the agent's value selection strategy for key parameters under different operating conditions.

10. The method for improving the dynamic simulation accuracy of a kinetic reactor according to claim 1, characterized in that, Step 4 includes calculating the optimized independent error and the total system error. ∈[0,v] and all e i All less than or equal to v i If the optimization is successful, the simulation output variables will be output as the simulation result; otherwise, return to step 2 to retrain the agent.

Images