Method, medium and system for establishing liquid flow model in reactor

Abstract

The present invention provides a method, medium and system for establishing a liquid flow model in a reactor, which belongs to the field of computer model technology. The method for establishing a liquid flow model in a reactor of the present invention first constructs a basic fluid dynamics model including reactor geometric parameters, fluid parameters and reaction parameters, and then establishes a mathematical model including heat and mass transfer equations, fluid motion equations, energy conservation equations and component conservation equations. By introducing a neural network module and a convolution layer, a parameter contribution evaluation system is established to achieve dynamic optimization of model parameters. Based on a training data set, an initial parameter matrix of the convolution kernel is obtained, and a correlation relationship between weight coefficients and various parameters is established. The model is optimized through a verification data set, and finally a model that can accurately predict the flow characteristics of the liquid in the reactor is obtained. The present invention solves the technical problem in the prior art that the reactor fluid dynamics model is difficult to accurately characterize the flow characteristics of the fluid in a complex dynamic reaction process.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer models, and in particular relates to a method, medium and system for establishing a liquid flow model in a reactor. Background Art

[0002] In the fields of chemical industry, materials, pharmaceuticals, etc., reactors are important equipment for realizing chemical reactions. The flow characteristics of the fluid in the reactor have a decisive influence on the heat transfer, mass transfer and reaction efficiency of the reaction process. The traditional reactor fluid flow model is mainly based on the computational fluid dynamics (CFD) method, which describes the fluid flow characteristics by solving basic control equations such as the Navier-Stokes equations, the continuity equation and the energy equation. These models are usually solved by numerical methods such as the finite volume method and the finite element method, and are combined with turbulence models, multiphase flow models, etc. to deal with complex flow phenomena. In practical applications, researchers usually use commercial CFD software or self-developed programs to construct fluid flow models, and realize flow field calculations through steps such as meshing, boundary condition setting and solver parameter adjustment.

[0003] However, traditional fluid flow models have significant defects. First, these models are usually based on idealized assumptions and it is difficult to accurately describe the complex flow phenomena in the actual reaction process, such as local turbulence, secondary flow, dead zones, etc. Secondly, model parameters such as turbulent energy, dissipation rate, reaction rate, etc. usually use empirical values ​​or experimental fitting values, and lack an accurate description of the dynamic changes of parameters. Thirdly, traditional models are difficult to effectively deal with the time-varying characteristics of parameters such as fluid properties and reaction conditions during the reaction process, resulting in a decrease in model prediction accuracy over time. In addition, existing models often separate processes such as fluid flow, heat and mass transfer, and chemical reactions, ignoring the coupling between the processes, making it difficult for the model to reflect the overall behavior of the actual reaction system.

[0004] Faced with increasingly complex chemical reaction processes and ever-increasing process requirements, the limitations of traditional fluid flow models are becoming increasingly prominent. Especially when dealing with complex systems such as multiphase reactions and polymerization reactions, it is difficult for the model to accurately capture the dynamic changes in fluid flow characteristics, nor can it achieve real-time optimization and adjustment of model parameters. The lack of accuracy and poor adaptability of this model seriously restricts the process optimization and intelligent control of the reactor, and it is urgent to develop new modeling methods to break through this technical bottleneck. In other words, there is a technical problem in the existing technology that the reactor fluid dynamics model is difficult to accurately characterize the fluid flow characteristics in complex dynamic reaction processes. Summary of the Invention

[0005] In view of this, the present invention provides a method, medium and system for establishing a liquid flow model in a reactor, which can solve the technical problem in the prior art that the reactor fluid dynamics model is difficult to accurately characterize the fluid flow characteristics in complex dynamic reaction processes.

[0006] The present invention is implemented as follows: In a first aspect, the present invention provides a method for establishing a liquid flow model in a reactor, comprising the following steps: establishing a basic fluid dynamics model including reactor geometric parameters, fluid physical parameters and reaction process parameters; establishing a mathematical model of the reactor including a heat and mass transfer balance equation, a fluid motion balance equation, an energy conservation balance equation and a component conservation balance equation; constructing a deep neural network module to receive the calculation results of the basic fluid dynamics model and the reactor mathematical model; setting a flow feature convolution layer including multiple flow feature convolution kernels to calculate the contribution of the reaction process parameters to the process parameters of fluid flow; obtaining an initial parameter numerical matrix of the flow feature convolution kernel based on a model training data set, and establishing a mathematical correlation equation between the weight coefficient value of the flow feature convolution kernel and the reactor geometric parameters and reaction process parameters; optimizing the parameter values ​​of the flow feature convolution kernel according to the mathematical correlation equation and establishing a final liquid flow model in the reactor.

[0007] Among them, the contribution of the process parameters is obtained through the following steps: calculating the partial derivative value of each reaction process parameter to the fluid velocity field distribution; calculating the partial derivative value of each reaction process parameter to the temperature field distribution; calculating the partial derivative value of each reaction process parameter to the concentration field distribution; normalizing the partial derivative values ​​to obtain the influence degree value of each reaction process parameter on the fluid flow state; and calculating the contribution coefficient value of each reaction process parameter based on the influence degree value.

[0008] Among them, the mathematical correlation equations include: linear correlation equations that describe the mathematical relationship between the geometric parameters of the reactor and the numerical values ​​of the flow characteristic convolution kernel parameters; nonlinear correlation equations that describe the mathematical relationship between the reaction process parameters and the numerical values ​​of the flow characteristic convolution kernel parameters; and coupled correlation equations that describe the mathematical relationship between the contribution of the process parameters and the numerical values ​​of the flow characteristic convolution kernel parameters.

[0009] Among them, the geometric parameters of the reactor include the reactor diameter value, the reactor height value, the stirring paddle installation position value and the stirring paddle model parameter; the fluid physical parameters include the fluid density value, the fluid viscosity coefficient value, the fluid specific heat capacity value and the fluid thermal conductivity coefficient value; the reaction process parameters include the reaction temperature value, the reaction pressure value, the reaction conversion rate value and the reaction rate constant value.

[0010] Among them, the heat and mass transfer balance equation is used to calculate the temperature field distribution data inside the reactor; the fluid motion balance equation is used to calculate the fluid velocity field distribution data; the energy conservation balance equation is used to calculate the heat change data during the reaction process; and the component conservation balance equation is used to calculate the concentration field distribution data of each reactant.

[0011] Among them, before the step of establishing the final liquid flow model in the reactor, it also includes: collecting a model verification data set and calculating the model prediction result; comparing the error value between the model prediction result and the actual operation data, and when the error value is greater than the preset error threshold, returning to the step of recalculating the weight coefficient value of the flow feature convolution kernel.

[0012] Among them, the input parameters of the reactor mathematical model include: fluid temperature data, reactor wall temperature data and heat transfer coefficient value; fluid pressure data, fluid viscosity coefficient value and fluid density value; reaction heat value, stirring power value and heat dissipation value; reaction rate constant value, reactant initial concentration value and reaction conversion rate value.

[0013] The method further includes applying the final liquid flow model in the reactor to the reactor control system to achieve predictive control of the liquid flow in the reactor.

[0014] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, the program instructions are used to execute the above-mentioned method for establishing a liquid flow model in a reactor.

[0015] The third aspect of the present invention provides a system for establishing a liquid flow model in a reactor, comprising the above-mentioned computer-readable storage medium, wherein the system is any one of a computer, a server, and a single-chip microcomputer, and the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0016] Compared to the prior art, the present invention provides a method, medium, and system for modeling liquid flow within a reactor. By combining a neural network module with a traditional fluid dynamics model and introducing the concept of parameter contribution, this method accurately describes the flow characteristics of complex reaction systems and dynamically optimizes model parameters. The method first establishes a mathematical model encompassing fundamental equations such as heat and mass transfer, fluid motion, energy conservation, and component conservation. The model is then optimized using a neural network module, leveraging the mapping relationship between convolution kernels and reaction parameters to improve the model's adaptability and prediction accuracy.

[0017] Compared with traditional methods, the technical solution of the present invention has significant advantages. By introducing the concept of parameter contribution, a quantitative evaluation system for the influence of reaction parameters on the fluid flow state is established, so that the model can accurately identify and process the dynamic changes of key parameters. The parameter optimization mechanism based on the convolution kernel realizes the dynamic correlation between model parameters and actual process parameters, ensuring the adaptability of the model under different working conditions. In addition, the present invention unifies the processes of fluid flow, heat and mass transfer, and reaction kinetics, captures the coupling effect between the processes through the neural network module, and improves the overall prediction accuracy of the model.

[0018] Through these technological innovations, the present invention effectively addresses the limitations of traditional models when dealing with complex reaction systems. A dynamic optimization mechanism based on parameter contributions enables the model to accurately reflect changes in flow characteristics during the reaction process, providing a reliable theoretical foundation for process optimization and intelligent control of reactors. Therefore, the present invention addresses the technical problem that existing reactor fluid dynamics models struggle to accurately characterize fluid flow characteristics during complex dynamic reactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention.

[0020] Figure 2 This is the evolution curve of the loss function during the deep neural network training process in Example 2.

[0021] Figure 3 This is a radar chart showing the contribution of different process parameters in Example 2.

[0022] Figure 4 This is a diagram of the convergence process of the genetic algorithm optimization in Example 2.

[0023] Figure 5 This is a comparison chart of the prediction errors of the model in Example 2 under different working conditions. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] like Figure 1 FIG. 1 is a flow chart of a method for establishing a liquid flow model in a reactor according to the first aspect of the present invention. The method comprises the following steps:

[0026] S01. Establishing a basic fluid dynamics model, wherein the basic fluid dynamics model includes reactor geometric parameters, fluid physical parameters, and reaction process parameters;

[0027] S02. Establishing a mathematical model of the reactor, wherein the mathematical model of the reactor includes a heat and mass transfer balance equation, a fluid motion balance equation, an energy conservation balance equation, and a component conservation balance equation;

[0028] S03. Constructing a deep neural network module, wherein the deep neural network module comprises an input layer node, a hidden layer node, and an output layer node, wherein the input layer node receives calculation results of the basic fluid dynamics model and the reactor mathematical model;

[0029] S04. Setting a flow feature convolution layer in the deep neural network module, wherein the flow feature convolution layer includes a plurality of flow feature convolution kernels, and each of the flow feature convolution kernels corresponds to a set of reaction process parameters;

[0030] S05. Calculate the contribution of each reaction process parameter to the process parameter of fluid flow;

[0031] S06. Collecting reactor operation data, obtaining fluid temperature data, fluid pressure data, fluid flow rate data, and reactant concentration data, to form a model training data set;

[0032] S07. Inputting the model training data set into the deep neural network module to obtain the initial parameter numerical matrix of the flow feature convolution kernel;

[0033] S08. Calculating the weight coefficient value of the flow characteristic convolution kernel based on the initial parameter numerical matrix and the contribution of the process parameters, and establishing a mathematical correlation equation between the weight coefficient value and the geometric parameters of the reactor and the reaction process parameters;

[0034] S09. Optimizing the parameter values ​​of the flow characteristic convolution kernel according to the mathematical correlation equation to obtain optimized convolution kernel parameter values;

[0035] S10, modifying the basic fluid dynamics model based on the optimized convolution kernel parameter values ​​to obtain a modified fluid dynamics model;

[0036] S11, collecting a model verification data set, inputting the model verification data set into the modified fluid dynamics model, and calculating the model prediction result;

[0037] S12, comparing the error value between the model prediction result and the actual operation data, and when the error value is greater than a preset error threshold, returning to step S08;

[0038] S13. When the error value is less than the preset error threshold, a final liquid flow model in the reactor is obtained.

[0039] Optionally, the method further includes applying the final liquid flow model in the reactor to a reactor control system to implement predictive control of the liquid flow in the reactor.

[0040] The process parameter contribution is obtained by the following steps:

[0041] (1) Calculate the partial derivative values ​​of each reaction process parameter with respect to the fluid velocity field distribution;

[0042] (2) Calculate the partial derivative values ​​of each reaction process parameter with respect to the temperature field distribution;

[0043] (3) Calculate the partial derivative values ​​of each reaction process parameter with respect to the concentration field distribution;

[0044] (4) normalizing the partial derivative values ​​to obtain the influence of each reaction process parameter on the fluid flow state;

[0045] (5) calculating the contribution coefficient value of each reaction process parameter based on the influence degree value;

[0046] in:

[0047] The heat and mass transfer balance equation is used to calculate the temperature field distribution data inside the reactor. The input parameters include fluid temperature data, reactor wall temperature data and heat transfer coefficient value. The output result is temperature field distribution data.

[0048] The fluid motion balance equation is used to calculate the fluid velocity field distribution data. The input parameters include fluid pressure data, fluid viscosity coefficient value and fluid density value. The output result is the velocity field distribution data.

[0049] The energy conservation balance equation is used to calculate the heat change data during the reaction process. The input parameters include the reaction heat value, stirring power value and heat dissipation value, and the output result is the energy balance data;

[0050] The component conservation balance equation is used to calculate the concentration field distribution data of each reactant. The input parameters include the reaction rate constant value, the initial concentration value of the reactant and the reaction conversion rate value. The output result is the concentration field distribution data.

[0051] The geometric parameters of the reactor include the reactor diameter value, the reactor height value, the stirring paddle installation position value and the stirring paddle model parameter;

[0052] The fluid physical parameters include fluid density value, fluid viscosity coefficient value, fluid specific heat capacity value and fluid thermal conductivity value;

[0053] The reaction process parameters include reaction temperature, reaction pressure, reaction conversion rate and reaction rate constant;

[0054] The mathematical correlation equation includes:

[0055] Linear correlation equation: describes the mathematical relationship between the geometric parameters of the reactor and the numerical values ​​of the flow characteristic convolution kernel parameters;

[0056] Nonlinear correlation equation: describes the mathematical relationship between reaction process parameters and flow characteristic convolution kernel parameters;

[0057] Coupling correlation equation: describes the mathematical relationship between the contribution of process parameters and the numerical value of the flow characteristic convolution kernel parameters.

[0058] The flow feature convolution kernel is specifically a convolution operation unit for extracting fluid flow features;

[0059] The process parameter contribution is specifically a quantitative value of the influence of the reaction process parameter on the fluid flow state;

[0060] The fluid flow state includes the velocity field distribution, temperature field distribution and concentration field distribution of the fluid;

[0061] The preset error threshold is a preset allowable error range value between the model calculation result and the actual data.

[0062] The specific implementation of the above steps is described in detail below.

[0063] The specific implementation of step S01 is to establish a basic dynamic model that describes the movement of fluids within the reactor. First, the geometric parameters of the reactor are collected and organized, including the reactor diameter value within the range of 0.5 to 5 meters, the reactor height value within the range of 1 to 10 meters, the impeller installation position value generally located at 1 / 3 to 1 / 2 of the reactor height, and the impeller model parameters including the number of blades, blade angle, and blade diameter. Then, the fluid physical parameters are collected, including the fluid density value within the range of 800 to 2000 kilograms per cubic meter, the fluid viscosity coefficient value within the range of 0.001 to 0.1 Pascal seconds, the fluid specific heat capacity value within the range of 2000 to 5000 joules per kilogram Kelvin, and the fluid thermal conductivity value within the range of 0.1 to 0.5 watts per meter Kelvin. Finally, the reaction process parameters were collected, including reaction temperature values ​​ranging from 20 to 200 degrees Celsius, reaction pressure values ​​ranging from 0.1 to 1.0 MPa, reaction conversion values ​​ranging from 0.8 to 0.99, and reaction rate constant values ​​ranging from 0.001 to 0.1 per second. These parameters were input into computational fluid dynamics software to establish a basic fluid dynamics model.

[0064] The specific implementation of step S02 is to establish a mathematical model of the reactor, which contains four main equilibrium equations. First, a heat and mass transfer equilibrium equation is established. This equation is based on Fourier's law of heat conduction and calculates the temperature field distribution inside the reactor. The input fluid temperature data range is 20 to 200 degrees Celsius, the reactor wall temperature data range is 15 to 180 degrees Celsius, and the heat transfer coefficient value range is 100 to 1000 watts per square meter Kelvin. Then, a fluid motion equilibrium equation is established. This equation is based on the Navier-Stokes equations and calculates the fluid velocity field distribution. The input fluid pressure data range is 0.1 to 1.0 MPa, the fluid viscosity coefficient value range is 0.001 to 0.1 Pascal-second, and the fluid density value range is 800 to 2000 kilograms per cubic meter. Next, an energy conservation balance equation was established. This equation, based on the first law of thermodynamics, calculates the heat change during the reaction. Input values ​​for the heat of reaction ranged from -500 to 500 kilojoules per mole, stirring power values ​​ranged from 0.1 to 10 kilowatts, and heat dissipation values ​​ranged from 0.05 to 5 kilowatts. Finally, a component conservation balance equation was established. This equation, based on the law of conservation of mass, calculates the concentration field distribution of each reactant. Input values ​​for the reaction rate constant ranged from 0.001 to 0.1 per second, initial reactant concentrations ranged from 0.1 to 10 moles per liter, and reaction conversion rates ranged from 0.8 to 0.99.

[0065] The specific implementation of step S03 is to construct a deep neural network module using a multilayer perceptron network structure. The number of input layer nodes is the dimension of the calculation results of the basic fluid dynamics model and the reactor mathematical model. Three hidden layers are set, with the number of nodes in each layer being 2 times, 1.5 times, and 1 times the number of input layer nodes, respectively. The rectified linear unit activation function is used, and the number of output layer nodes is the predicted target dimension. Training uses the stochastic gradient descent optimization algorithm, with a learning rate of 0.001, a batch size of 32, and 1000 training rounds.

[0066] The specific implementation method of step S04 is to set a flow feature convolution layer in the deep neural network module, adopt a three-dimensional convolutional neural network structure, set multiple flow feature convolution kernels, each convolution kernel size is 3×3×3, the step size is 1, and the filling method is the same filling. The number of convolution kernels is the same as the number of reaction process parameter groups, and each convolution kernel corresponds to a group of reaction process parameters, including reaction temperature value, reaction pressure value, reaction conversion value and reaction rate constant value. A batch normalization layer is used for data standardization processing, and a maximum pooling layer is used for feature dimensionality reduction. The pooling kernel size is 2×2×2 and the step size is 2.

[0067] The specific implementation method of step S05 is to calculate the process parameter contribution of each reaction process parameter to the fluid flow, first calculate the partial derivative value of each reaction process parameter to the fluid velocity field distribution, and adopt the central difference method to calculate the numerical differential. Then calculate the partial derivative value of each reaction process parameter to the temperature field distribution, also adopt the central difference method. Then calculate the partial derivative value of each reaction process parameter to the concentration field distribution. These partial derivative values ​​are subjected to minimum and maximum normalization processing to obtain the influence degree value of each reaction process parameter on the fluid flow state, ranging between 0 and 1. Finally, based on the influence degree value, adopt the weighted average method to calculate the contribution coefficient value of each reaction process parameter.

[0068] The specific implementation of step S06 is to collect reactor operation data. Fluid temperature data is collected through a temperature sensor with a sampling frequency of 1 Hz and a sampling period of 8 hours. Fluid pressure data is collected through a pressure sensor with a sampling frequency of 1 Hz and a sampling period of 8 hours. Fluid flow rate data is collected through a flow meter with a sampling frequency of 1 Hz and a sampling period of 8 hours. Reactant concentration data is collected through a concentration detector with a sampling frequency of 0.1 Hz and a sampling period of 8 hours. The collected data is preprocessed, including outlier detection, missing value processing, and data smoothing, to form a model training data set.

[0069] The specific implementation of step S07 is to input the model training data set into the deep neural network module, use the backpropagation algorithm for network training, use the mean square error function as the loss function, use the adaptive moment estimation algorithm as the optimization algorithm, use the learning rate of 0.001, and train 1000 rounds. The initial parameter numerical matrix of the flow feature convolution kernel is obtained through training.

[0070] The specific implementation of step S08 is to calculate the weight coefficient value of the flow characteristic convolution kernel based on the initial parameter value matrix and the process parameter contribution, and use the linear regression method to establish the linear correlation equation between the reactor geometric parameters and the flow characteristic convolution kernel parameter values. A polynomial regression method is used to establish a nonlinear correlation equation between the reaction process parameters and the flow characteristic convolution kernel parameter values, and the polynomial order is 3. A neural network method is used to establish a coupled correlation equation between the process parameter contribution and the flow characteristic convolution kernel parameter values.

[0071] The specific implementation of step S09 is to optimize the parameter values ​​of the flow feature convolution kernel according to the mathematical correlation equation. A genetic algorithm is used for global optimization, with a population size of 100, an evolutionary number of 500, a crossover probability of 0.8, and a mutation probability of 0.1. A particle swarm optimization algorithm is used for local optimization, with a particle size of 50, a maximum number of iterations of 200, an inertia weight of 0.7, and an acceleration factor of 2.

[0072] The specific implementation of step S10 involves modifying the basic fluid dynamics model based on the optimized convolution kernel parameter values. This involves extracting fluid flow characteristics using a convolution operation and modifying the calculated results of the fluid velocity, temperature, and concentration field distributions. Using a residual network structure, the correction terms are added to the calculated results of the basic model to obtain a modified fluid dynamics model.

[0073] The specific implementation of step S11 is to collect a model validation data set, using the same method as step S06, with a sampling time of 4 hours. The model validation data set is input into the modified fluid dynamics model to calculate the model prediction results, including the fluid velocity field distribution, temperature field distribution, and concentration field distribution.

[0074] The specific implementation of step S12 is to compare the error values ​​of the model prediction results with the actual operating data, and evaluate them using a relative error index, with a preset error threshold set to 5%. If the relative error of the velocity field distribution is greater than 5%, the relative error of the temperature field distribution is greater than 5%, or the relative error of the concentration field distribution is greater than 5%, the process returns to step S08 and re-optimizes the convolution kernel parameters.

[0075] In step S13, when the relative errors of all prediction results are less than a preset error threshold, a final model of the liquid flow within the reactor is determined. This model includes optimized flow characteristic convolution kernel parameter values ​​and revised fluid dynamics equations. This model can be used to predict fluid flow states under different reaction process conditions, providing a decision-making basis for the reactor control system.

[0076] A specific implementation of optional step S14 involves applying the final reactor liquid flow model to the reactor control system to achieve predictive control of the reactor liquid flow. The model is first deployed to the reactor control system's computing unit, where online calculations are performed using a real-time data stream processing framework. Based on the model's prediction results, a model predictive control algorithm is then used to optimize and adjust the reactor's operating parameters. The prediction time domain is set to 60 minutes, the control time domain is set to 10 minutes, and the sampling period is 1 minute. Finally, the optimized operating parameters are distributed to the actuators via the control system's execution unit, achieving closed-loop control of the reactor liquid flow.

[0077] The equations or calculation processes involved in the present invention are described in detail below.

[0078] 1. Heat and mass transfer balance equation (the first equation in step S02):

[0079] The heat and mass transfer balance equation is specifically expressed as follows:

[0080]

[0081] Where T is the temperature field, unit is K; t is the time, unit is s; α is the temperature diffusion coefficient, Unit is m 2 / s; k is the thermal conductivity, unit is W / (m·K); ρ is the density, unit is kg / m 3 ;c p is the specific heat capacity, unit is J / (kg·K); q v is the volume heat source term, in W / m 3 ; ε T is the error term for temperature field calculation; x, y, z are spatial coordinates, and the unit is m.

[0082] Boundary conditions:

[0083]

[0084] Where n is the wall normal vector; h is the convective heat transfer coefficient, the unit is W / (m 2 ·K); T w is the wall temperature, in K; T f is the fluid temperature in K.

[0085] 2. Fluid motion equilibrium equation (the second equation in step S02):

[0086] The fluid motion balance equation is specifically expressed as follows:

[0087]

[0088] Where, is the velocity vector, unit is m / s; p is the pressure, unit is Pa; μ is the dynamic viscosity, unit is Pa·s; is the acceleration due to gravity, in m / s 2 ; ε v Compute the error term for the velocity field.

[0089] 3. Energy conservation balance equation (the third equation in step S02):

[0090] The energy conservation balance equation is specifically expressed as follows:

[0091]

[0092] Where Φ is the viscous dissipation function, and its unit is W / m 3 ;q r is the heat of reaction, in W / m 3 ;P s is the stirring power, in W; q l is the heat loss, in W; ε e is the energy calculation error term.

[0093] 4. Component conservation balance equation (the fourth equation in step S02):

[0094] The component conservation balance equation is specifically expressed as follows:

[0095]

[0096] Where c i is the concentration of component i, in mol / m 3 ;D i is the diffusion coefficient of component i, in m 2 / s;R i is the reaction rate of component i, in mol / (m 3 ·s); ε c Compute the error term for the concentration field.

[0097] 5. Partial derivative calculation (calculation process in step S05):

[0098] The partial derivative calculation is specifically expressed as follows:

[0099]

[0100] Where p j is the jth process parameter; Δp j is the parameter perturbation; ε p1 , ε p2 , ε p3 Compute the error term for the partial derivatives.

[0101] 6. Calculation of process parameter contribution (calculation process in step S05):

[0102] The process parameter contribution calculation is specifically expressed as follows:

[0103]

[0104] Where S j is the contribution of the jth process parameter; w1, w2, w3 are weight coefficients, and satisfy w1+w2+w3=1; ε s Calculate the error term for the contribution.

[0105] 7. Deep neural network structure (network structure in step S03):

[0106] The forward propagation calculation of the deep neural network is specifically expressed as follows:

[0107] h1=f1(W1x+b1);

[0108] h2=f2(W2h1+b2);

[0109] h3=f3(W3h2+b3);

[0110] y=W4h3+b4+ε n ;

[0111] Where x is the input vector; h1, h2, h3 are the hidden layer outputs; y is the output vector; W1, W2, W3, W4 are weight matrices; b1, b2, b3, b4 are bias vectors; f1, f2, f3 are activation functions; ε n is the neural network prediction error term.

[0112] 8. Flow feature convolution kernel calculation (convolution calculation in step S04):

[0113] The flow feature convolution kernel calculation is specifically expressed as follows:

[0114]

[0115] Where, F i，j，k is the convolution output feature; K l，m，n is the convolution kernel parameter; I i，j，k is the input feature; k Compute the error term for the convolution.

[0116] 9. Mathematical correlation equation (correlation equation in step S08):

[0117] The linear correlation equation is specifically expressed as follows:

[0118] K p =A1d+A2h+A3h s +A4n b +ε l ;

[0119] Where K p is the convolution kernel parameter; d is the reactor diameter; h is the reactor height; h s The installation position of the stirring paddle; n b is the number of stirring blades; A1, A2, A3, A4 are linear coefficients; ε l is the linearly correlated error term.

[0120] The nonlinear correlation equation is specifically expressed as follows:

[0121] K p =B1T 3 +B2T 2 +B3T+B4p 3 +B5p 2 +B6P+B7X 3 +B8X 2+B9X+B 10 k 3 +B 11 k 2 +B 12 k+ε n ;

[0122] Where, T is the reaction temperature; P is the reaction pressure; X is the reaction conversion rate; k is the reaction rate constant; B1, B2, . . . , B 12 is the nonlinear coefficient; ε n is the nonlinear correlation error term.

[0123] The coupling correlation equation is specifically expressed as follows:

[0124]

[0125] Where C j is the coupling coefficient; S j is the contribution of process parameters; φ j (p j ) is the parameter mapping function; ε c is the coupling correlation error term.

[0126] 10. Loss function calculation (training process in step S07):

[0127] The loss function calculation is specifically expressed as follows:

[0128]

[0129] Where N is the number of samples; y i is the true value; is the predicted value; λ is the regularization coefficient; W l is the network weight of the lth layer; ||·|| F is the Frobenius norm; ε l Compute the error term for the loss.

[0130] 11. Genetic algorithm optimization (optimization process in step S09):

[0131] The fitness function calculation is specifically expressed as follows:

[0132]

[0133] Where, MSE(x) is the mean square error; β is the balance coefficient; ψ j (x j ) is the parameter penalty function; ε g Calculate the error term for the fitness.

[0134] 12. Particle Swarm Optimization (Optimization process in step S09):

[0135] The speed update equation is specifically expressed as follows:

[0136]

[0137] The position update equation is specifically expressed as follows:

[0138]

[0139] Where, is the velocity of the i-th particle at the k-th iteration; for location; is the optimal position of the individual; is the global optimal position; w is the inertia weight; c1, c2 are acceleration coefficients; r1, r2 are random numbers; ε v , ε p is the update error term.

[0140] 13. Model predictive control (control process in step S14):

[0141] The predictive control objective function is specifically expressed as follows:

[0142]

[0143] Where N p is the prediction time domain; N c is the control time domain; y(t+k|t) is the k-step prediction output; r(t+k) is the reference trajectory; Δu(t+k|t) is the control increment; Q, R are weight matrices; ε c is the control error term.

[0144] Control constraints:

[0145] u min ≤u(t)≤u max ;

[0146] Δu min ≤Δu(t)≤Δu max ;

[0147] y min ≤y(t)≤y max ;

[0148] Where u min ,u max is the control variable constraint; Δu min , Δu max To control the incremental constraint; y min ,y max is the output variable constraint.

[0149] Now let's explain the principle and meaning of each equation:

[0150] 1. Heat and mass transfer balance equation:

[0151] This equation is constructed based on Fourier's law of heat conduction and the principle of conservation of energy. The term represents the rate of change of temperature over time, reflecting the transient characteristics of the system; The term represents heat conduction, and the second-order derivative form is used to describe the diffusion process of heat in space; the term is used to compensate for the error caused by model simplification. The boundary condition uses the third type of boundary condition to describe the wall of the reactor. The term represents the contribution of volume heat source, including reaction heat, stirring heat, etc.; ε T is the error convection heat transfer process.

[0152] 2. Fluid motion equilibrium equation:

[0153] This equation is based on the Navier-Stokes equations and describes the motion of the fluid. The term represents the inertial force of the fluid, including local acceleration and convective acceleration; The term represents the pressure gradient force; The term represents the viscous force; The term represents gravity; ε v is the error term. Continuity equation Reflects the conservation of mass.

[0154] 3. Energy conservation balance equation:

[0155] This equation comprehensively considers various energy transfer processes in the reactor. The term represents the change in energy carried by the fluid; The term represents heat conduction; the term Φ represents the contribution of mechanical energy into thermal energy; q r The term indicates whether the chemical reaction releases or absorbs heat; P s The term represents the stirring power input; q l The term represents the heat loss of the system; ε e is the error term.

[0156] 4. Component conservation balance equation:

[0157] This equation is based on the law of conservation of mass and describes the concentration distribution of reactants. The term represents the change of concentration over time; The term represents convective mass transfer; The term represents molecular diffusion; R i The term represents the chemical reaction rate; ε c is the error term.

[0158] 5. Calculation of process parameter contribution:

[0159] This calculation quantifies the impact of each process parameter on the fluid flow state through normalized partial derivatives. A three-term weighted sum is used to consider the impact on the velocity field, temperature field, and concentration field. The weight coefficients w1, w2, and w3 can be optimized based on experimental data.

[0160] 6. Deep Neural Network Structure:

[0161] The network uses a multi-layer perceptron structure, introducing nonlinear mapping capabilities through nonlinear activation functions. The output of each layer serves as the input to the next layer, extracting features layer by layer. The network parameters are optimized using a backpropagation algorithm to minimize prediction error.

[0162] 7. Flow feature convolution kernel calculation:

[0163] This calculation is based on the principles of three-dimensional convolution, leveraging local connectivity and weight sharing to extract flow features. Using a 3×3×3 convolution kernel captures flow patterns in local areas, and by stacking multiple convolution kernels, features at different scales are extracted. This method has the advantage of adaptively learning flow features without the need for manually designed feature extraction operators.

[0164] 8. Mathematical correlation equations:

[0165] The linear correlation equation describes the relationship between the reactor's geometric parameters and the convolution kernel parameters. The use of a linear combination is based on engineering experience, assuming that the effects of geometric parameters on flow characteristics are linearly additive. The nonlinear correlation equation uses a cubic polynomial to describe the influence of process parameters. This is because reaction processes often exhibit nonlinear characteristics, and cubic polynomials have sufficient expressive power. The coupled correlation equation incorporates parameter contributions as weights, reflecting the varying importance of different process parameters.

[0166] 9. Loss function calculation:

[0167] The loss function consists of two parts: the first is the mean squared error between the predicted and true values, reflecting the model's prediction accuracy; the second is the Frobenius norm of the weight matrix, which acts as a regularizer to prevent overfitting. The parameter λ balances the importance of these two components.

[0168] 10. Genetic Algorithm Optimization:

[0169] The fitness function design considers two aspects: the model's prediction accuracy (expressed as the inverse of the mean squared error) and a penalty term that weights the parameter contribution. This design ensures prediction accuracy while also considering the physical constraints of process parameters.

[0170] 11. Particle Swarm Optimization:

[0171] The velocity update equation contains three terms: inertia term Maintain the original movement trend of particles; cognitive items Guide particles to move to their optimal individual positions; social terms Guide the particles to move towards the global optimal position. The choice of parameters w, c1, c2 affects the convergence performance of the algorithm.

[0172] 12. Model Predictive Control:

[0173] The objective function consists of two items: output tracking error and control increments The former guarantees the control performance, while the latter suppresses the drastic change of the control quantity. The constraint conditions ensure that the control is within the practical feasible range. p and control time domain N c The choice of necessitates a balance between control performance and computational burden.

[0174] The innovations of these equations are mainly reflected in:

[0175] 1. Combining deep learning methods with traditional fluid dynamics models improves the accuracy and generalization ability of the model;

[0176] 2. The concept of process parameter contribution is introduced to quantify the importance of different parameters;

[0177] 3. Adopting multi-level optimization strategy to ensure the global optimality of model parameters;

[0178] 4. Various constraints in actual engineering are taken into consideration to improve the practicality of the model.

[0179] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, the program instructions are used to execute the above-mentioned method for establishing a liquid flow model in a reactor.

[0180] The third aspect of the present invention provides a system for establishing a liquid flow model in a reactor, comprising the above-mentioned computer-readable storage medium, wherein the system is any one of a computer, a server, and a single-chip microcomputer, and the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0181] Specifically, the principle of the present invention is as follows: The technical solution of the present invention is based on fluid dynamics theory and deep learning principles, and accurately describes the flow characteristics of the fluid in the reactor by constructing a multi-level model framework. Its core principle is to optimize the traditional fluid dynamics model by utilizing the nonlinear mapping capability of neural networks, and dynamically adjust the model parameters by calculating the parameter contribution. In the basic fluid dynamics model, the heat and mass transfer equations are used to describe the temperature field distribution, the fluid motion equations are used to characterize the velocity field characteristics, and the energy conservation and component conservation equations are used to characterize the material and energy changes in the reaction process.

[0182] The introduction of a neural network module imbues the model with adaptive learning capabilities. By setting up a multi-layer convolution kernel structure, each kernel corresponds to a specific set of reaction parameters, enabling a mapping between reaction parameters and model features. Parameter contributions are calculated based on the partial derivatives of each parameter with respect to the flow state, and the influence weights of different parameters are determined through normalization. This gradient-based contribution evaluation method enables the model to accurately identify key parameters and perform targeted optimization.

[0183] The model's optimization process follows the principle of feedback iteration. By comparing the error between the model's predictions and actual data, the convolution kernel parameters and weight coefficients are continuously adjusted until the preset accuracy requirements are achieved. This dynamic optimization mechanism ensures the model's adaptability and reliability in practical applications. Furthermore, by establishing a correlation between physical parameters and convolution kernel parameters, the model parameters are physically interpretable, providing a theoretical basis for process optimization and control.

[0184] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0185] The specific implementation of step S01 is to achieve high-precision prediction of liquid flow in the reactor by establishing a basic fluid dynamics model. First, the geometric parameters of the reactor need to be collected, including the reactor diameter value within the range of 0.5 to 5 meters, the reactor height value within the range of 1 to 10 meters, the stirring paddle installation position value located at 1 / 3 to 1 / 2 of the reactor height, the number of stirring paddle blades within the range of 2 to 8, the blade angle range of 30 to 60 degrees, and the blade diameter of 0.3 to 0.6 times the reactor diameter. Then, the fluid physical parameters are collected, and the fluid density value is obtained by experimental measurement or consulting relevant data manuals, ranging from 800 to 2000 kilograms per cubic meter. The fluid viscosity coefficient value is measured using a rotational viscometer, ranging from 0.001 to 0.1 Pascal seconds. The fluid specific heat capacity value is determined by differential scanning calorimetry, ranging from 2000 to 5000 joules per kilogram Kelvin. The fluid thermal conductivity value is measured using a transient hot wire method, ranging from 0.1 to 0.5 watts per meter Kelvin. Finally, the reaction process parameters were collected. The reaction temperature range was determined to be 20 to 200 degrees Celsius and the reaction pressure range was 0.1 to 1.0 MPa according to the process specifications. The reaction conversion rate was determined by gas chromatography to be in the range of 0.8 to 0.99. The reaction rate constant was determined by kinetic experimental methods to be in the range of 0.001 to 0.1 per second. These parameters were input into the computational fluid dynamics software, and the finite volume method was used for meshing. The number of meshes was not less than 1 million. A hexahedral structured mesh was used. The mesh quality required orthogonality greater than 0.8 and skewness less than 0.3. The Reynolds-averaged Navier-Stokes equations were used to describe the fluid motion. The standard k-ε model was selected to simulate turbulence, and the wall function was used to treat the near-wall region. The sliding mesh technique was used to consider the motion of the agitator, the time step was set to 0.001 seconds, and the total calculation time was not less than 60 seconds. The established basic fluid dynamics model needs to meet the principles of conservation of mass, conservation of momentum, and conservation of energy.

[0186] The specific implementation of step S02 is to establish a mathematical model of the reactor, which contains four main equilibrium equations. First, the heat and mass transfer equilibrium equation is established. Based on Fourier's heat conduction law, the finite difference method is used for discretization and solution. The specific form is:

[0187]

[0188] The parameters in the formula have been explained above. A second-order central difference scheme is used for spatial discretization, and a first-order forward difference scheme is used for temporal discretization, with a time step of 0.01 seconds and a spatial step of 0.01 meters. The fluid motion equilibrium equation is then established based on the Navier-Stokes equations:

[0189]

[0190] The pressure correction algorithm is used to solve the problem, and the iterative error is controlled within 0.0001. Then the energy conservation balance equation is established:

[0191]

[0192] The alternating direction implicit difference method is used to solve the problem. The convergence criterion is that the maximum relative error of the temperature field between two adjacent iterations is less than 0.001. Finally, the component conservation equilibrium equation is established:

[0193]

[0194] The flux balance method was used to ensure mass conservation, and the convergence criterion was that the component mass balance error was less than 0.1%.

[0195] The specific implementation of step S03 is to construct a deep neural network module using a multilayer perceptron network structure. The number of input layer nodes is determined by the dimensions of the calculation results of the basic fluid dynamics model and the reactor mathematical model. They are typically sampled values ​​of physical quantities such as flow field, temperature field, and concentration field at characteristic locations, and the number of nodes ranges from 100 to 1000. Three hidden layers are used, with the number of nodes in each layer being 2 times, 1.5 times, and 1 times the number of nodes in the input layer, respectively. The activation function uses the rectified linear unit function: f(x) = max(0, x). To prevent overfitting, batch normalization and dropout layers are added after each hidden layer, with the dropout ratio set to 0.2. The number of output layer nodes is the predicted target dimension, typically between 20 and 200. Network parameters are initialized using the He Kaiming initialization method, with the bias term initialized to 0. Training uses the stochastic gradient descent optimization algorithm, with the initial learning rate set to 0.001. The cosine annealing strategy is used to dynamically adjust the learning rate, with a minimum learning rate of 0.00001. The batch size is 32, the number of training rounds is 1000, and the model parameters are saved every 100 rounds. An early stopping strategy is used to prevent overfitting, and training is stopped if the validation set loss does not decrease for 10 consecutive rounds.

[0196] The specific implementation of step S04 is to set a flow feature convolution layer in the deep neural network module, using a three-dimensional convolutional neural network structure. Set multiple flow feature convolution kernels, and the calculation formula of each convolution kernel is:

[0197]

[0198] The convolution kernel size is 3×3×3, with a stride of 1 and identical padding. The number of convolution kernels is the same as the number of reaction process parameter groups, typically 4 to 8. A batch normalization layer is used for data normalization, and the normalization parameters are obtained through statistical analysis of the training data. A maximum pooling layer is used for feature dimensionality reduction, with a pooling kernel size of 2×2×2 and a stride of 2. Residual connections are added after the convolution layer, with every two layers forming a residual block. A total of three residual blocks are used, and finally a fully connected layer is used to map the features to the prediction target space. To improve feature extraction capabilities, an attention mechanism is used in each residual block to calculate the feature importance weights at different locations. The convolution kernel parameters are initialized using the Xavier method, with a standard deviation of 0.01. L2 regularization is used during training to prevent overfitting, with a regularization coefficient of 0.0001.

[0199] The specific implementation of step S05 is to calculate the contribution of each reaction process parameter to the process parameter of fluid flow, using a sensitivity analysis method. First, the partial derivative value of each reaction process parameter to the fluid velocity field distribution is calculated, and the specific formula is:

[0200]

[0201] The central difference method was used to calculate the numerical differential, and the parameter perturbation was set to 1% of the parameter baseline value. 100 perturbation calculations were performed for each parameter, and the average value was taken as the final result. The partial derivative values ​​of each reaction process parameter with respect to the temperature field distribution were then calculated:

[0202]

[0203] The central difference method is also used. Then the partial derivative values ​​of each reaction process parameter with respect to the concentration field distribution are calculated:

[0204]

[0205] These partial derivative values ​​are normalized to the minimum and maximum values:

[0206]

[0207] The influence of each reaction process parameter on the fluid flow state was determined, ranging from 0 to 1. The weight coefficients were determined using principal component analysis, generally satisfying the following equation: w1 > w2 > w3. Finally, the contribution coefficient of each reaction process parameter was calculated using a weighted average method. Parameters with a contribution coefficient greater than 0.5 were considered key process parameters.

[0208] The specific implementation method of step S06 is to collect reactor operation data and fluid temperature data through a temperature sensor. A platinum resistance temperature sensor is used with an accuracy of ±0.1 degrees Celsius, a sampling frequency of 1 Hz, and a sampling time of 8 hours. Five temperature measurement points are arranged at different heights of the reactor. Fluid pressure data is collected through a pressure sensor. A piezoresistive pressure sensor is used with an accuracy of ±0.1% of full scale, a sampling frequency of 1 Hz, and a sampling time of 8 hours. Three pressure measurement points are arranged at different positions of the reactor. Fluid flow rate data is collected through an electromagnetic flowmeter with an accuracy of ±0.5%, a sampling frequency of 1 Hz, and a sampling time of 8 hours. Reactant concentration data is collected through a concentration detector using an online gas chromatograph with an analysis cycle of 10 seconds and a sampling time of 8 hours. The collected data is preprocessed, first performing outlier detection, using the 3σ criterion to identify outliers, and correcting them using the local polynomial regression method. Missing values ​​were then processed using time series interpolation. For data with no more than 10 consecutive missing points, cubic spline interpolation was used. For data segments with more than 10 missing points, Kalman filtering was used for prediction. Finally, data smoothing was performed using the Savitzky-Golay filter with a window length of 11 and a polynomial order of 3. The preprocessed data was randomly divided into a training set and a validation set in a ratio of 8:2.

[0209] The specific implementation of step S07 is to input the model training data set into the deep neural network module and use the back propagation algorithm to perform network training. The loss function uses the mean square error function:

[0210]

[0211] The regularization coefficient λ is set to 0.0001. The optimization algorithm uses the adaptive moment estimation algorithm, and the main parameters β1 are set to 0.9, β2 is set to 0.999, and ε is set to 10 -8 The learning rate was 0.001, and the number of training rounds was 1000. After each round of training, the model performance was calculated on the validation set, and the model parameters with a validation set loss value less than 0.01 were recorded. The initial parameter matrix of the flow feature convolution kernel was obtained through training. The matrix dimension is determined by the convolution kernel size and is generally 3×3×3.

[0212] The specific implementation of step S08 is to calculate the weight coefficient value of the flow characteristic convolution kernel based on the initial parameter value matrix and the process parameter contribution, and establish a set of mathematical correlation equations. First, a linear regression method is used to establish a linear correlation equation between the geometric parameters of the reactor and the flow characteristic convolution kernel parameter values:

[0213] K p =A1d+A2h+A3h s +A4n b +ε l ,

[0214] The coefficient matrix is ​​estimated using the least squares method, and the coefficient of determination R 2 The requirement is greater than 0.9. Then the polynomial regression method is used to establish the nonlinear correlation equation between the reaction process parameters and the flow characteristic convolution kernel parameter values:

[0215] K p =B1T 3 +B2T 2 +B3T+B4P 3 +B5P 2 +B6P+B7X 3 +B8X 2 +B9X+B 10 k 3 +B 11 k 2 +B 12 k+ε n ,

[0216] Legendre polynomials are used as basis functions, and the coefficients are determined through fitting optimization. Then, a neural network method is used to establish the coupling correlation equation between the contribution of process parameters and the numerical value of the flow characteristic convolution kernel parameters:

[0217]

[0218] where φ j is the parameter mapping function, which is constructed using radial basis functions. Finally, the three equations are combined to form a complete set of mathematical equations to guide the optimization of convolution kernel parameters.

[0219] The specific implementation of step S09 is to optimize the parameter values ​​of the flow feature convolution kernel according to the mathematical correlation equation, using a two-stage optimization strategy. In the first stage, a genetic algorithm is used for global optimization, with the population size set to 100, the chromosome length equal to the number of parameters to be optimized, and a real number encoding method. The fitness function is:

[0220]

[0221] The mean square error threshold is set to 0.001, and the balance coefficient β is set to 0.1. The selection operation adopts the tournament selection method, and 3 individuals are selected for comparison each time. The crossover operation adopts arithmetic crossover with a crossover probability of 0.8. The mutation operation adopts Gaussian mutation with a mutation probability of 0.1, and the standard deviation decreases with the number of generations. The evolutionary generation is 500 generations, and the iteration is terminated when the optimal solution has not improved for 50 consecutive generations. In the second stage, the particle swarm optimization algorithm is used for local optimization, with a particle number of 50 and a maximum number of iterations of 200. The speed update equation is:

[0222]

[0223] The position update equation is:

[0224]

[0225] The inertia weight w decreases linearly with the number of iterations, with an initial value of 0.9 and a final value of 0.4. The acceleration coefficients c1 and c2 are both set to 2.0. The algorithm is considered to have converged when the improvement in the optimal solution is less than 0.0001 after 30 consecutive iterations.

[0226] The specific implementation method of step S10 is to correct the basic fluid dynamics model based on the optimized convolution kernel parameter values, and adopt the residual learning method. First, the optimized convolution kernel is used to extract the features of the flow field data to obtain a flow feature map. The feature map is then mapped back to the physical space through a deconvolution operation to obtain a flow field correction term. The correction term is then added to the calculation results of the basic model to correct the fluid velocity field distribution, temperature field distribution and concentration field distribution. The velocity field correction adopts the momentum correction method to introduce the eddy viscosity correction term; the temperature field correction adopts the thermal diffusion correction method to introduce the turbulent Prandtl number correction term; the concentration field correction adopts the mass diffusion correction method to introduce the turbulent Schmidt number correction term. The correction coefficient is determined by optimizing the model verification data, and the relative error between the corrected prediction result and the experimental data must be controlled within 5%.

[0227] The specific implementation of step S11 is to collect a model validation dataset. Data is collected using the same method as step S06, with a sampling time of 4 hours. The validation dataset includes operating data from five different operating conditions, with fluid temperature, pressure, flow rate, and concentration data recorded under each operating condition. To ensure data representativeness, the five operating conditions must cover the typical operating range of the reactor: Operating Condition 1 is a low-temperature, low-pressure condition with a reaction temperature of 50 degrees Celsius and a pressure of 0.2 MPa; Operating Condition 2 is a low-temperature, high-pressure condition with a reaction temperature of 50 degrees Celsius and a pressure of 0.8 MPa; Operating Condition 3 is a medium-temperature, medium-pressure condition with a reaction temperature of 120 degrees Celsius and a pressure of 0.5 MPa; Operating Condition 4 is a high-temperature, low-pressure condition with a reaction temperature of 180 degrees Celsius and a pressure of 0.2 MPa; and Operating Condition 5 is a high-temperature, high-pressure condition with a reaction temperature of 180 degrees Celsius and a pressure of 0.8 MPa. The validation dataset is input into the modified fluid dynamics model to calculate the flow field distribution prediction results under each operating condition.

[0228] The specific implementation of step S12 is to compare the error value between the model prediction result and the actual operation data, using a multi-index evaluation system. The relative error calculation formula of the velocity field distribution is:

[0229]

[0230] The relative error calculation formula of temperature field distribution is:

[0231]

[0232] The calculation formula for the relative error of concentration field distribution is:

[0233]

[0234] The preset error threshold is set to 5%. When the relative error of any indicator exceeds the threshold, it is necessary to return to step S08 to re-optimize the convolution kernel parameters. To improve optimization efficiency, the optimization target is dynamically adjusted based on the error analysis results, focusing on optimizing the prediction quantity with larger error.

[0235] The specific implementation method of step S13 is to determine the final liquid flow model in the reactor when the relative errors of all prediction results are less than the preset error threshold. The model includes the optimized flow feature convolution kernel parameter numerical matrix and the revised fluid dynamics calculation equations. When the model is stored, the network structure information, model parameters, normalization parameters, etc. are saved at the same time, and a hierarchical storage structure is adopted to facilitate the call and update of the model. When the model is deployed, model compression and quantization processing are required to reduce the computing resource usage. The final model can be used to predict the fluid flow state under different reaction process conditions, and the confidence interval of the prediction result is 95%. The model update cycle defaults to 3 months, and new operating data needs to be collected for model retraining during updating.

[0236] To better understand and implement the present invention, Example 2 of a specific application scenario is provided below: A research team, while studying process control for a large polycarbonate synthesis reactor, discovered that traditional flow models struggled to accurately predict fluid flow under complex reaction conditions, resulting in significant fluctuations in product quality. The research team decided to employ the method of the present invention to establish a high-precision model of the liquid flow within the reactor.

[0237] First, the research team collected the basic parameters of the reactor, as shown in Table 1:

[0238] Table 1 Basic parameters of reactor

[0239] Parameter Type Parameter name Parameter value Geometric parameters Reactor diameter 2.5m Geometric parameters Reactor height 4.8m Geometric parameters Agitator installation position 1.8m Geometric parameters Number of stirring blades 6 pieces Geometric parameters Angle of stirring blade 45 degrees Geometric parameters Stirring paddle diameter 1.2m

[0240] The fluid physical parameters were obtained through experimental measurement, as shown in Table 2:

[0241] Table 2 Fluid physical parameters

[0242] Physical parameters Numerical range Measurement method Fluid density <![CDATA[1150±50kg / m 3 ]]> Vibrating tube densitometer Fluid viscosity 0.045±0.005Pa·s Rotational viscometer Fluid specific heat capacity 2850±150J / (kg·K) Differential Scanning Calorimetry Fluid thermal conductivity 0.28±0.02W / (m·K) Transient hot wire method

[0243] The setting of reaction process parameters is shown in Table 3:

[0244] Table 3 Reaction process parameters

[0245] Process parameters Setting range Control accuracy Reaction temperature 180～200℃ ±0.5℃ Reaction pressure 0.6~0.8MPa ±0.01MPa Reaction conversion rate 0.92～0.96 ±0.01 Reaction rate constant 0.025～0.035s-1 ±0.001s-1

[0246] The research team established a basic fluid dynamics model based on the above parameters and solved it using commercial computational fluid dynamics software. The meshing used was a hexahedral structured grid with a total of 1.25 million grids and a minimum grid size of 0.01m. The orthogonality of the grid quality evaluation indicators was 0.85 and the skewness was 0.25. The turbulence model selected was the standard k-ε model, and the near-wall treatment used the standard wall function. + The value is controlled in the range of 30 to 300. The motion of the stirring paddle is realized by sliding grid technology, the time step is set to 0.001s, and the total calculation time is 180s.

[0247] The mathematical model of the reactor was established based on the heat and mass transfer balance equation, fluid motion balance equation, energy conservation balance equation and component conservation balance equation. The heat and mass transfer balance equation is:

[0248] The equilibrium equation of fluid motion is:

[0249] A deep neural network module with four hidden layers was constructed. The input layer had 256 nodes, corresponding to the sampled values ​​of the flow, temperature, pressure, and concentration fields at 64 feature locations. The number of hidden layer nodes was 512, 384, 256, and 128, respectively. The activation function used was the rectified linear unit function. The dropout ratio was set to 0.2, the batch size was 32, and the number of training epochs was 1000. The evolution of the loss function during network training is shown in Table 4:

[0250] Table 4 Network training loss function evolution table

[0251] Number of training rounds Training set loss Validation set loss Learning rate 100 0.0856 0.0921 0.001 300 0.0452 0.0487 0.0008 500 0.0287 0.0315 0.0005 700 0.0198 0.0225 0.0003 1000 0.0156 0.0182 0.0001

[0252] Figure 2 This figure shows the evolution of the loss function during deep neural network training. It includes three key metrics: training set loss, validation set loss, and learning rate. The dual Y-axis design visually demonstrates the changing trends of these three parameters over the number of training rounds. Eight flow feature convolution kernels, each 3×3×3 in size, were set in the deep neural network to correspond to different combinations of reaction process parameters. Sensitivity analysis calculated the contribution of each process parameter, as shown in Table 5:

[0253] Table 5 Process parameter contribution calculation results

[0254] Process parameters Velocity field contribution Temperature field contribution Concentration field contribution Comprehensive contribution Reaction temperature 0.385 0.452 0.325 0.412 Reaction pressure 0.278 0.215 0.245 0.246 Reaction conversion rate 0.156 0.185 0.285 0.198 rate constant 0.181 0.148 0.145 0.144

[0255] Figure 3Radar charts are used to display the contributions of different process parameters to the velocity, temperature, and concentration fields. Polar coordinates clearly illustrate the distribution of the contributions of the four process parameters (reaction temperature, reaction pressure, reaction conversion rate, and rate constant) to the three fields. A two-stage optimization strategy was used to optimize the convolution kernel parameters. The first stage employed a genetic algorithm with a population size of 100 and 500 evolution generations. The convergence process of the algorithm is shown in Table 6.

[0256] Table 6 Genetic algorithm optimization convergence process

[0257] Evolutionary Algebra Optimal fitness Average fitness Parameter standard deviation 100 0.856 0.725 0.158 200 0.912 0.845 0.125 300 0.945 0.898 0.086 400 0.968 0.935 0.052 500 0.982 0.965 0.028

[0258] Figure 4 The convergence process of the genetic algorithm optimization is demonstrated, including the changing trends of three indicators: optimal fitness, average fitness, and parameter standard deviation. A dual Y-axis design is used to intuitively display the changes in these indicators over the number of evolutionary generations. In the second phase, a particle swarm optimization algorithm was used for local optimization, with a particle size of 50 and a maximum number of iterations of 200. A comparison of the model prediction results obtained through optimization with the experimental data is shown in Table 7:

[0259] Table 7 Comparison of model prediction results and experimental data

[0260] Evaluation indicators Relative error of velocity field Relative error of temperature field Relative error of concentration field Working condition 1 3.25％ 2.85％ 3.56％ Working condition 2 3.68％ 3.12％ 3.82％ Working condition 3 3.45％ 2.98％ 3.65％ Working condition 4 3.86％ 3.25％ 3.92％ Working condition 5 3.92％ 3.35％ 4.15％

[0261] Figure 5 A grouped bar chart is used to show the comparison of the prediction errors of the model under different working conditions. The figure contains three relative errors of velocity field, temperature field and concentration field, which clearly shows the prediction accuracy of the model under different working conditions. The traditional reactor flow modeling method mainly uses computational fluid dynamics software to directly simulate or simplify the empirical model. The computational fluid dynamics method has a large amount of calculation, making it difficult to achieve real-time prediction, and the prediction error can reach more than 15% under complex working conditions. Although the empirical model has a fast calculation speed, it has low accuracy and the prediction error is usually more than 20%. The method of the present invention combines deep learning with the traditional fluid mechanics model, which significantly improves the prediction accuracy while ensuring computational efficiency, and the prediction error is controlled within 5%. In addition, the concept of process parameter contribution introduced in the present invention can quantitatively evaluate the influence of different parameters on the flow state, providing a theoretical basis for process parameter optimization. The two-stage optimization strategy is adopted to avoid local optimal solutions and ensure the global optimality of model parameters.

[0262] It should be noted that the variables involved in the present invention are explained in detail as shown in Table 8 below.

[0263] Table 8 Variable explanation table

[0264]

[0265]

[0266] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for establishing a liquid flow model in a reactor, characterized in that: include: Establish a basic fluid dynamics model including reactor geometric parameters, fluid physical parameters and reaction process parameters; establish a mathematical model of the reactor including heat and mass transfer balance equations, fluid motion balance equations, energy conservation balance equations and component conservation balance equations; construct a deep neural network module to receive the calculation results of the basic fluid dynamics model and the reactor mathematical model; set a flow feature convolution layer including multiple flow feature convolution kernels to calculate the contribution of the reaction process parameters to the process parameters of fluid flow, wherein the flow feature convolution kernel is specifically a convolution operation unit for extracting fluid flow characteristics; collect reactor operation data, obtain fluid temperature data, fluid pressure data, fluid flow rate data and reactant concentration data, form a model training data set, and obtain the flow feature convolution kernel based on the model training data set. An initial parameter numerical matrix is ​​obtained, and the weight coefficient value of the flow characteristic convolution kernel is calculated based on the initial parameter numerical matrix and the process parameter contribution, and a mathematical correlation equation is established, including: using a linear regression method to establish a linear correlation equation between the geometric parameters of the reactor and the flow characteristic convolution kernel parameter values, using a polynomial regression method to establish a nonlinear correlation equation between the reaction process parameters and the flow characteristic convolution kernel parameter values, using a neural network method to establish a coupling correlation equation between the process parameter contribution and the flow characteristic convolution kernel parameter values, and using the mathematical correlation equation to optimize the parameter values ​​of the flow characteristic convolution kernel; based on the optimized convolution kernel parameter values, the basic fluid dynamics model is modified, the fluid flow characteristics are extracted by using a convolution operation, and the calculation results of the fluid velocity field distribution, the temperature field distribution and the concentration field distribution are modified; The residual network structure is adopted to add the correction term to the calculation results of the basic model to obtain the corrected fluid dynamics model. After model verification, the final liquid flow model in the reactor is obtained.

2. The method for establishing a liquid flow model in a reactor according to claim 1, wherein: The process parameter contribution is obtained by the following steps: calculating the partial derivative value of each reaction process parameter to the fluid velocity field distribution; calculating the partial derivative value of each reaction process parameter to the temperature field distribution; calculating the partial derivative value of each reaction process parameter to the concentration field distribution; normalizing the partial derivative values ​​to obtain the influence degree value of each reaction process parameter on the fluid flow state; The contribution coefficient value of each reaction process parameter is calculated based on the influence degree value.

3. The method for establishing a liquid flow model in a reactor according to claim 1, wherein: The geometric parameters of the reactor include the reactor diameter value, the reactor height value, the stirring paddle installation position value and the stirring paddle model parameters; the fluid physical parameters include the fluid density value, the fluid viscosity coefficient value, the fluid specific heat capacity value and the fluid thermal conductivity coefficient value; the reaction process parameters include the reaction temperature value, the reaction pressure value, the reaction conversion rate value and the reaction rate constant value.

4. The method for establishing a liquid flow model in a reactor according to claim 1, wherein: The heat and mass transfer balance equation is used to calculate the temperature field distribution data inside the reactor; the fluid motion balance equation is used to calculate the fluid velocity field distribution data; the energy conservation balance equation is used to calculate the heat change data during the reaction process; and the component conservation balance equation is used to calculate the concentration field distribution data of each reactant.

5. The method for establishing a liquid flow model in a reactor according to claim 1, wherein: Before the step of establishing the final liquid flow model in the reactor, it also includes: collecting a model verification data set of the modified fluid dynamics model, calculating the prediction results of the modified fluid dynamics model; comparing the error value between the prediction result and the actual operation data, and when the error value is greater than a preset error threshold, returning to the step of recalculating the weight coefficient value of the flow feature convolution kernel.

6. The method for establishing a liquid flow model in a reactor according to claim 1, wherein: The input parameters of the reactor mathematical model include: fluid temperature data, reactor wall temperature data and heat transfer coefficient value; fluid pressure data, fluid viscosity coefficient value and fluid density value; reaction heat value, stirring power value and heat dissipation value; reaction rate constant value, reactant initial concentration value and reaction conversion rate value.

7. The method for establishing a liquid flow model in a reactor according to claim 1, wherein: The method further includes applying the final liquid flow model in the reactor to the reactor control system to achieve predictive control of the liquid flow in the reactor.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the method for establishing a liquid flow model in a reactor according to any one of claims 1 to 7.

9. A system for establishing a liquid flow model in a reactor, characterized in that: The computer-readable storage medium according to claim 8 is included, wherein the system is a computer, the computer-readable storage medium is provided in the system, and a microprocessor for executing program instructions stored in the computer-readable storage medium is provided in the system.

Images