Optimization Method and Related Device for Rapid Prediction Model of Flow Field Response of High Consistency Pulper

Abstract

This application provides a method and related apparatus for optimizing a rapid prediction model of the flow field response of a high-consistency pulper. The method includes: constructing a CFD flow field mechanism model that simulates the dynamic phase transition from Newtonian to non-Newtonian fluid based on cumulative strain using a user-defined function; constructing a response surface model and defining a parameter constraint space based on experimental design and CFD simulation of key parameters; constructing CFD data samples containing spatial features and time series; constructing a CNN-LSTM-Attention model and training it with the samples to obtain a rapid prediction model of the flow field response; and using a multi-objective whale optimization algorithm to perform parameter co-optimization on the rapid prediction model, outputting a Pareto optimal solution set as the optimal parameters of the model. This application integrates mechanism modeling, data-driven prediction, and intelligent optimization decision-making, which can achieve coordinated optimization of stirring time and energy consumption, effectively improve the simulation accuracy and optimization efficiency of the swelling process, and reduce equipment operating energy consumption and R&D trial and error costs.

Description

Technical Field

[0001] This application relates to the field of lyocell fiber technology, specifically to a method and apparatus for optimizing a rapid prediction model of the flow field response of a high-consistency pulper. Background Technology

[0002] In the production of lyocell fiber, the swelling stage is crucial. The high-consistency hydraulic pulper is the core equipment in the swelling process of lyocell fiber production. Its main function is to achieve full mixing and swelling of wood pulp and NMMO solution under high solids content and high viscosity conditions, so as to provide a uniform and stable pulp system for the subsequent spinning and dissolving process.

[0003] However, in actual industrial operation and simulation analysis, during the swelling process, the system gradually transforms from an initial Newtonian fluid system of "wood pulp particles + NMMO solution" into a non-Newtonian slurry system with high viscosity and shear-thinning characteristics. This phase transition process is accompanied by a change in viscosity by orders of magnitude and a sudden change in shear response. Traditional CFD simulations often use constant viscosity or a single non-Newtonian model, which makes it difficult to truly reflect the evolution of physical properties during the swelling process, resulting in a large deviation between the simulation results and actual operating conditions. Summary of the Invention

[0004] This application provides a method and related apparatus for optimizing a rapid prediction model of the flow field response of a high-consistency pulper, which can effectively improve the simulation accuracy and optimization efficiency of the swelling process, and reduce equipment operating energy consumption and R&D trial and error costs.

[0005] A first aspect of this application provides a method for optimizing a fast prediction model of the flow field response of a high-consistency pulper, the method comprising: Constructing a CFD flow field mechanism model for a high-consistency pulper; Based on the experimental design and CFD simulation of key parameters, a response surface model is constructed, and the parameter constraint space for subsequent optimization is defined according to the analysis results of the model. Construct CFD data samples; A CNN-LSTM-Attention model is constructed and trained using the CFD data samples to obtain a fast flow field response prediction model. The multi-objective whale optimization algorithm is used to perform parameter co-optimization on the fast prediction model of the flow field response, and the Pareto optimal solution set is output as the optimal parameters of the fast prediction model of the flow field response, thus obtaining the optimized fast prediction model of the flow field response.

[0006] In one possible implementation, the construction of the CFD flow field mechanism model for the high-consistency pulper includes: Based on the structure of the high-consistency hydraulic pulper used in the production of experimental lyocell fibers, a three-dimensional geometric model was constructed. The interaction between the wood pulp phase and the liquid phase was described using the Euler-Euler multiphase flow model, and a multiphase flow model was constructed. The SST k-ω model with rotational flow correction is used to describe the strong rotational and high shear characteristics of the flow field, and a turbulence model is constructed. The impeller rotation behavior is described using a motion mesh method, and the rotational speed is set for the rotation domain; The transformation process from Newtonian fluid to non-Newtonian fluid is simulated using user-defined functions based on Ansys Fluent.

[0007] In one possible implementation, the experimental design and CFD simulation based on key parameters, constructing a response surface model, and defining a parameter constraint space for subsequent optimization based on the analysis results of this model, includes: Rotor speed, number of spoilers, and number of rotor blades are selected as core independent variables, and the ranges of the rotor speed, number of spoilers, and number of rotor blades are defined. A three-factor, three-level, center-based composite test was used to construct the combination of rotor speed, number of spoilers, and number of rotor blades to obtain the test parameter set. For the aforementioned set of test parameters, transient simulations were conducted based on a CFD flow field mechanism model of a high-consistency pulper, and simulation results were obtained. Based on the simulation results, the stirring time and stirring energy consumption are determined as response variables; Based on the experimental parameter set and its corresponding response variables, variance analysis was performed to construct a quadratic response surface model for stirring time and stirring energy consumption. Based on the analysis results of the quadratic response surface model, the parameter constraint space is defined.

[0008] In one possible implementation, constructing the CFD data sample includes: Based on the parameter constraint space, the key parameters are sampled in the parameter space by combining the control variable method with Latin hypercube sampling. CFD simulations were performed on each set of experimental parameters to obtain flow field evolution data throughout the swelling process; By introducing the time dimension into the sample construction process and sampling key physical quantities at equal time steps, a high-dimensional sample structure of spatial features × time series is obtained, which serves as CFD data samples.

[0009] In one possible implementation, the expression formula for the CFD data sample is as follows: , In the formula, Represents the set of spatial flow field characteristics. (t) represents the temporal evolution characteristics, where t is the stirring duration and E is the stirring energy consumption.

[0010] In one possible implementation, a CNN-LSTM-Attention model is constructed, and the CNN-LSTM-Attention model is trained using the CFD data samples to obtain a fast flow field response prediction model, including: A multi-scale convolutional neural network subnetwork was constructed, and the spatial distribution characteristics of the flow field in the CFD data samples were extracted to obtain micro-scale shearing characteristics, meso-scale energy transport characteristics, and large-scale circulating flow characteristics. The microscale shearing features, mesoscale energy transport features, and large-scale circulating flow features are reorganized into a feature sequence according to time steps. Physical constraints based on energy dissipation or shear intensity are introduced to correct the state of the memory unit, thereby constructing the evolution law of the flow field over time. We weight the hidden states at each time step of the output of the Long Short-Term Memory Network, and introduce shear strength or energy dissipation level as a physical saliency index to guide the allocation of attention weights, thereby generating a global context vector that integrates features of key time periods. The global context vector is input into the fully connected layer to map the predicted stirring time and stirring energy consumption. The CNN-LSTM-Attention model is trained using the CFD data samples to obtain a trained fast prediction model for flow field response.

[0011] In one possible implementation, the step of employing a multi-objective whale optimization algorithm to collaboratively optimize the parameters of the fast flow field response prediction model and outputting a Pareto optimal solution set as the optimal parameters of the fast flow field response prediction model includes: The rotor speed, the number of spoilers, and the number of rotor blades are used as decision variables, and the parameter constraint space is used as the search boundary. The fast prediction model of the flow field response is used as the fitness evaluation function. For any candidate solution X, the corresponding stirring time and stirring energy consumption are quickly calculated through the prediction model. Initialize the whale population by randomly generating multiple candidate solutions within the search boundary as the initial population; Simulates the encirclement and predation behavior of whales and the bubble net attack mechanism. In each iteration, the candidate solution is updated by randomly selecting either a shrinking encirclement mechanism or a spiral update mechanism. An adaptive weighting strategy is introduced to enhance global exploration capabilities in the early stages of iteration and local development capabilities in the later stages of iteration, thereby avoiding getting trapped in local optima. Based on the Pareto dominance relation, a non-dominated solution set is constructed. The dominance relation of candidate solutions in each generation of the population is compared, and non-dominated solutions that are not dominated by any other solutions are selected and the Pareto front is updated. Once the maximum number of iterations or the convergence condition is reached, the Pareto optimal solution set is output as the optimal parameters for the fast flow field response prediction model.

[0012] This example provides a method for optimizing a rapid prediction model of the flow field response of a high-consistency pulper. First, a CFD flow field mechanism model of the high-consistency pulper is constructed by introducing a user-defined function based on cumulative strain, which can realistically simulate the dynamic transition from Newtonian to non-Newtonian fluid. Then, a response surface model is constructed based on experimental design and CFD simulation of key parameters, and a parameter constraint space is defined according to the analysis results. Next, a high-dimensional CFD data sample containing spatial features and time series is constructed, and a CNN-LSTM-Attention model is designed to train the sample to obtain a rapid prediction model of the flow field response. Finally, the flow field is extracted using a multi-scale convolutional neural network. The microscale shear, mesoscale energy transport, and large-scale circulation characteristics are analyzed. A physically constrained long short-term memory network is used to modify memory units by introducing energy dissipation or shear intensity to accurately model the temporal evolution of the flow field. A physical saliency-guided attention mechanism strengthens focus on key time stages, thereby improving the accuracy of single-time flow field response assessment while maintaining prediction accuracy. Finally, a multi-objective whale optimization algorithm is employed, using the aforementioned fast prediction model as the fitness function for parameter collaborative optimization. An adaptive weighting strategy balances global search and local exploitation capabilities, outputting a Pareto optimal solution set that minimizes both stirring time and energy consumption. This method organically integrates high-fidelity CFD simulation with deep learning prediction models, overcoming the shortcomings of traditional pure physical modeling, such as high computational cost and difficulty in iterative optimization, while avoiding the prediction instability problem caused by the lack of physical constraints in pure data-driven models. It can efficiently obtain multi-objective optimal parameter combinations that balance stirring efficiency and energy consumption while significantly reducing computational resource consumption. This provides a highly reliable and efficient solution for optimizing the process parameters of high-consistency pulpers in lyocell fiber production, effectively improving the simulation accuracy and optimization efficiency of the swelling process, and reducing equipment operating energy consumption and R&D trial and error costs.

[0013] A second aspect of this application provides a fast prediction model optimization system for the flow field response of a high-consistency pulper, the system comprising: The first processing unit is used to construct a CFD flow field mechanism model of a high-consistency pulper. The second processing unit is used for experimental design and CFD simulation based on key parameters, constructing a response surface model, and defining the parameter constraint space for subsequent optimization based on the analysis results of the model. The third processing unit is used to construct CFD data samples; The fourth processing unit is used to construct a CNN-LSTM-Attention model and train the CNN-LSTM-Attention model using the CFD data samples to obtain a fast flow field response prediction model. The fifth processing unit is used to perform parameter co-optimization on the fast prediction model of the flow field response using the multi-objective whale optimization algorithm, and output the Pareto optimal solution set as the optimal parameters of the fast prediction model of the flow field response, so as to obtain the optimized fast prediction model of the flow field response.

[0014] A third aspect of this application provides a terminal including a processor, an input device, an output device, and a memory, wherein the processor, input device, output device, and memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the steps of the fast prediction model optimization method for high-consistency pulper flow field response as described in the first aspect of this application.

[0015] A fourth aspect of this application provides a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in the fast prediction model optimization method for high-consistency pulper flow field response in the first aspect of this application.

[0016] A fifth aspect of this application provides a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in xx of the first aspect of this application. The computer program product may be a software installation package. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This application provides a schematic diagram of the overall process for optimizing a rapid prediction model of the flow field response of a high-consistency pulper. Figure 2 This application provides a schematic diagram of a three-dimensional geometric model of a high-consistency hydraulic pulper structure for an embodiment of a rapid prediction model optimization system for the flow field response of a high-consistency pulper. Figure 3 This application provides a schematic diagram of the overall structure of a fast prediction model optimization system for the flow field response of a high-consistency pulper. Figure 4 This application provides a schematic diagram of the structure of a terminal. Figure label: First processing unit-1, second processing unit-2, third processing unit-3, fourth processing unit-4, fifth processing unit-5. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0021] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0022] To better understand the optimization method for the rapid prediction model of the flow field response of a high-consistency pulper provided in the embodiments of this application, the following is a brief introduction to the scenarios in which the optimization method for the rapid prediction model of the flow field response of a high-consistency pulper is applied.

[0023] Currently, the main barriers to flow field simulation of high-consistency pulpers include: the strong nonlinearity of fluid properties and the difficulty in describing phase evolution; the difficulty in characterizing the dynamic evolution characteristics of the flow field using traditional models; the unclear mechanism of the synergistic effect between equipment structural parameters and process parameters; and the difficulty of using traditional optimization methods to deal with multi-objective and strongly nonlinear problems.

[0024] Among them, the fluid properties are highly nonlinear and the phase evolution is difficult to describe. During the swelling process, the system gradually transforms from the initial "wood pulp particles + NMMO solution" Newtonian fluid system into a non-Newtonian slurry system with high viscosity and shear thinning characteristics.

[0025] This phase transition process is accompanied by a change in viscosity by orders of magnitude and a sudden change in shear response. Traditional CFD simulations often use constant viscosity or a single non-Newtonian model, which makes it difficult to truly reflect the evolution of physical properties during the swelling process, resulting in a large deviation between the simulation results and the actual working conditions.

[0026] The dynamic evolution of the flow field is difficult to characterize using traditional models. The internal flow of a high-consistency pulper exhibits strong rotation, high shear, and strong turbulence, and the flow field continues to evolve over time. Existing studies are mostly based on steady-state or quasi-steady-state assumptions, which fail to reflect the differentiated flow characteristics in the initial, middle, and steady-state stages of mixing, and cannot reveal the essential impact of the flow field evolution path on mixing efficiency and energy consumption.

[0027] The mechanism of synergistic effect between equipment structural parameters and process parameters is unclear. There are significant interactions and secondary effects among parameters such as rotational speed, number of baffles, and number of rotor blades. Simple single-factor analysis or empirical design methods cannot reveal the laws of their synergistic influence, resulting in a lack of systematic theoretical basis for the optimization of equipment design and operating parameters.

[0028] Traditional optimization methods struggle to address multi-objective and strongly nonlinear problems. In engineering practice, stirring time and energy consumption are often interdependent, making it a typical multi-objective optimization problem. While response surface methodology can reveal the trends of parameter influence, it is limited by the assumption of low-order polynomials and is prone to getting trapped in local optima. Traditional genetic algorithms and particle swarm optimization algorithms suffer from slow convergence and poor stability in strongly nonlinear and multimodal spaces.

[0029] This invention aims to overcome the shortcomings of the prior art. This application proposes a rapid prediction model optimization technology for the flow field response of a high-consistency pulper that combines "mechanism modeling, data-driven prediction, and intelligent optimization decision-making". By modeling and learning the spatiotemporal evolution characteristics of the flow field, the optimal design of stirring time and stirring energy consumption is achieved.

[0030] A fast prediction model optimization method for the flow field response of a high-consistency pulper is applied to the fast prediction model optimization system for the flow field response of a high-consistency pulper. Figure 1 A schematic diagram of the overall process for optimizing a fast predictive model of the flow field response of a high-consistency pulper is shown. Figure 1 As shown, it includes: S1. Construct a CFD flow field mechanism model for a high-consistency pulper.

[0031] Step S1 includes the following steps: S101. Based on the structure of the high-consistency hydraulic pulper used in the production of experimental lyocell fibers, construct a three-dimensional geometric model.

[0032] Figure 2 A schematic diagram of a three-dimensional geometric model of a high-consistency hydraulic pulper structure is shown, illustrating a rapid prediction model optimization system for the flow field response of a high-consistency pulper. (See diagram for reference.) Figure 2 As shown, a three-dimensional geometric model was established based on the structure of a high-consistency hydraulic pulper used in the experimental production of lyocell fibers. The effective volume of the swelling machine was designed to be 0.4 m³. 3 The tank diameter D is 850 mm, the maximum outer diameter d of the impeller is 450 mm, the number of rotor blades is 4, and the impeller design speed is 80 rev / min. The rotor clearance c from the bottom is 5 mm, and there are 4 baffles evenly distributed on the inner wall of the tank. To reduce the computational load without affecting the overall accuracy of the flow field, the small local structures of the rotor are reasonably suppressed in the preprocessing stage, retaining only the geometric features that have a decisive influence on the macroscopic flow.

[0033] S102. The interaction between the wood pulp phase and the liquid phase is described by the Euler-Euler multiphase flow model, and a multiphase flow model is constructed.

[0034] The high-consistency pulper uses a solid-liquid two-phase system formed by mixing wood pulp and NMMO solution, which gradually evolves into a high-viscosity non-Newtonian pulp during the swelling process. The wood pulp particles have an equivalent diameter of approximately 0.001 cm and a density of 1550 kg / m³. 3 The NMMO solution concentration was 76%, and its dynamic viscosity was measured to be 4000 Pa·s, with a density of 1135.32 kg / m³. 3 The volume fraction of wood pulp is no more than 15%, which is considered a working condition with a high particle volume fraction and significant inter-particle interactions.

[0035] Furthermore, considering the aforementioned characteristics, an Eulerian-Eulerian multiphase flow model is employed to describe the solid-liquid two-phase system. This model treats both the wood pulp phase and the liquid phase as continuous media, characterizing particle inertia and interaction effects through interphase momentum exchange terms. Compared to the Eulerian-Lagrange model, it is more suitable for high-concentration systems and conditions with strong interactions.

[0036] The core of the Euler-Euler multiphase flow model is to treat each phase as a continuous phase and describe the phase distribution and interaction through interphase coupling equations. The following are the key formulas applicable to high-consistency hydraulic pulpers (wood pulp-NMMO solution-swollen slurry three-phase system), including conservation equations and interphase interaction models: Continuity equation (conservation of mass) For any phase k (k=1 is the solid phase of wood pulp, k=2 is the liquid phase of NMMO solution, and k=3 is the non-Newtonian phase of swollen pulp), the mass conservation equation is: In the formula, The volume fraction of phase k (satisfying) ); The density of phase k (wood pulp) , NMMO solution (The swelling slurry ρ3 is taken as a weighted average). Let k be the velocity vector of phase k; For phase k, the mass source term is 0 (in this study, there is no external mass input); The interphase mass transfer rate from phase j to phase k (corresponding to the phase transition from Newtonian fluid to non-Newtonian fluid, calculated by cumulative strain correlation).

[0037] Momentum equation (conservation of momentum) The momentum conservation equation for phase k considers inertial forces, pressure forces, viscous forces, and interphase interaction forces: In the formula, For pressure sharing between phases (uniform pressure in a three-phase system); The viscous stress tensor of phase k (Newtonian phase) (Non-Newtonian phases are dynamically corrected by the viscosity model). It is the vector of gravitational acceleration; The interphase forces between phase j and phase k (including drag force, lift force, etc., with drag force being the main focus in this study).

[0038] Interphase interaction force model (core coupling term) The drag force model (Gidaspow model, applicable to solid-liquid / liquid-liquid systems) The interphase drag force is the main interaction force, and its expression is: In the formula, The interval relaxation time; For phase j, the particle / droplet diameter (wood pulp particles) ); For the drag force correction factor (related to volume fraction), Time to take (Adapted to the working conditions of solid volume fraction ≤15% in this study), among which, dj is the density of phase j; dj is the particle diameter of phase j. Let k be the dynamic viscosity of phase k; For the drag force correction factor (related to volume fraction), Time to take (This is suitable for the working conditions where the solid volume fraction is ≤15% in this study).

[0039] Energy equation Considering the heat generated by stirring and shearing, the energy equation for phase k is: In the formula, The isobaric specific heat capacity of phase k; The temperature of phase k; Let k be the thermal conductivity of phase k; For phase k, the internal heat source (shear heat generation, ); The heat exchange rate between phases is denoted as .

[0040] S103. The SST k-ω model with rotational flow correction is used to describe the strong rotational and high shear characteristics of the flow field, and a turbulence model is constructed.

[0041] The flow in the pulper exhibits strong rotation, high shear, and significant adverse pressure gradient characteristics due to the high-speed rotating impeller. This study employs the SST k-ω turbulence model and introduces a rotational flow correction to improve the model's prediction accuracy in the rotating region. The SST k-ω model uses the k-ω form near the wall and gradually transitions to the k-ε form further away from the wall, effectively handling separated flows and complex rotating flows.

[0042] The transport equation for turbulent kinetic energy k is: In the formula, Density of the mixture ( ); It is turbulent kinetic energy; The j-direction component of the mixture velocity ( ); For the j-direction coordinates (corresponding to x / y / z); The molecular viscosity of the mixture ( ); The turbulent Prandtl number is the turbulent kinetic energy. ); For turbulent viscosity; For the generation of turbulent kinetic energy of the mixture ( , (The modulus of the strain rate tensor of the mixture). The turbulent kinetic energy dissipation related constant ( ); The specific dissipation rate.

[0043] Furthermore, the transport equation for the specific dissipation rate ω is: In the formula, The turbulent Prandtl number is the specific dissipation rate. ); For the proportionality constant of the generating term ( ); The dissipation correlation constant for specific dissipation rate ( ); It is a mixed function; Specific dissipation rate turbulent Prandtl number in the far-wall region ( ).

[0044] S104. The impeller rotation behavior is described using a motion mesh method, and the rotation speed is set for the rotation domain.

[0045] Data exchange between the rotating and stationary domains is achieved through an interface. The barrel wall, impeller surface, and other solid boundaries are all set as non-slip walls, while the free liquid surface is set as a symmetry boundary.

[0046] Furthermore, a first-order upwind scheme is used for the convection term to ensure computational stability, while a second-order central difference scheme is used for the diffusion term. The PRESTO! scheme is selected for the pressure term to improve the accuracy of the pressure field, and the SIMPLE method is used for the pressure-velocity coupling algorithm. The evolution characteristics of the flow field over time are captured through transient calculations.

[0047] S105. Using user-defined functions based on Ansys Fluent, the transition process from Newtonian fluid to non-Newtonian fluid is simulated.

[0048] To simulate the transition from Newtonian to non-Newtonian fluid during the swelling process of wood pulp and NMMO solution, a user-defined function (UDF) based on Ansys Fluent was introduced. The wood pulp fiber, NMMO solution, and swollen pulp were defined as a three-phase system, and the initial volume fractions of wood pulp fiber, NMMO solution, and swollen pulp were set.

[0049] The phase volume fraction is initialized using a custom function `init_phase_system`, and the initial cumulative strain value is set to zero for each computational unit. Within each time step, the cumulative strain is updated using the following formula: cum_strain = cum_strain + strain_rate × dt The phase transition process is described by introducing an exponential function based on cumulative strain: trans = 1 - exp(-cum_strain / CRITICAL_STRAIN) In the formula, CRITICAL_STRAIN is the critical strain threshold. When the cumulative strain reaches this threshold, the Newtonian fluid gradually transforms into a non-Newtonian swelling slurry, and the volume fraction of each phase is dynamically updated accordingly.

[0050] Furthermore, the rheological properties of the swollen slurry are described using a volume fraction-dependent viscosity model, and its dynamic viscosity is: In the formula, The initial viscosity, Viscosity in the fully swollen state. This represents the liquid component of the swelling slurry. This model can reflect the shear thinning behavior of the slurry during the swelling process.

[0051] The above methods enable a refined simulation of the internal flow field structure, turbulence characteristics, and rheological evolution process of a high-consistency pulper, providing a highly reliable data foundation for subsequent data-driven prediction and multi-objective optimization.

[0052] S2. Based on key parameters, experimental design and CFD simulation are used to construct a response surface model, and the parameter constraint space for subsequent optimization is defined according to the analysis results of the model.

[0053] In order to reveal the influence mechanism of key structural and operating parameters of high-concentration hydraulic pulpers on mixing performance and energy consumption while ensuring computational efficiency, and to provide a reasonable and physically interpretable search space for subsequent intelligent optimization algorithms, this invention introduces the Response Surface Methodology (RSM) to systematically construct the parameter constraint space based on CFD simulation.

[0054] Step S2 includes the following steps: S201. Select the rotor speed, the number of spoilers, and the number of rotor blades as core independent variables, and define the range of the rotor speed, the number of spoilers, and the number of rotor blades.

[0055] Among them, the rotor speed n, the number of baffles Nb, and the number of rotor blades Nv are selected as the core independent variables, and the reasonable range of variation of each parameter is defined under the premise of equipment structural safety and process operability.

[0056] S202. A three-factor, three-level, center composite test was used to construct the combination of rotor speed, number of spoilers, and number of rotor blades to obtain the test parameter set.

[0057] Among them, the use of a three-factor, three-level central composite experimental design (CCD) to construct representative parameter combinations can systematically characterize single-factor effects, factor interactions, and nonlinear secondary effects, avoiding computational redundancy caused by empirical selection or exhaustive enumeration of all parameters.

[0058] S203. For the test parameter group, a transient simulation was carried out based on the CFD flow field mechanism model of the high-consistency pulper to obtain the simulation results.

[0059] S204. Based on the simulation results, determine the stirring time and stirring energy consumption as response variables.

[0060] For each set of experimental parameters, transient simulations were conducted based on the CFD flow field mechanism model of a high-consistency pulper. The mixing uniformity criterion was that the difference in wood pulp volume fraction at each monitoring point was <3%. The mixing time t was extracted. The mixing energy consumption E was calculated by integrating the impeller output torque and rotational speed, and the two were used as response variables.

[0061] S205. Based on the experimental parameter set and its corresponding response variables, perform variance analysis to construct a quadratic response surface model for stirring time and stirring energy consumption.

[0062] Based on the obtained experimental parameter sets and their corresponding response variable data, a second-order polynomial mathematical model was established using multivariate nonlinear regression analysis to establish the relationship between the response variables (stirring time t and stirring energy consumption E) and the independent variables (rotor speed n, number of baffles Nb, and number of rotor blades Nv). The corresponding quadratic response surface regression equation is as follows: in, The predicted response variable is either the stirring time t or the stirring energy consumption E. , The independent variables for input are the horizontally encoded rotor speed n, the number of spoilers Nb, and the number of rotor blades Nv; This is the constant term (intercept) of the regression equation. The linear effect regression coefficient of the i-th independent variable reflects the basic influence of a single factor on the response variable; The regression coefficient of the second-order effect of the i-th independent variable reflects the nonlinear surface influence of a single factor. The regression coefficients represent the interaction between the i-th and j-th independent variables, reflecting the coupled influence of multiple parameters on the flow field and energy consumption. This is an unpredictable statistical random error term.

[0063] After determining the model structure, the Ordinary Least Squares (OLS) method was used to analyze the unknown coefficients in the above multiple quadratic regression equation. , , , The data is fitted and solved. After the solution is obtained, the coefficient of determination of the model is calculated. To evaluate the goodness of fit of the data and to construct an initial quadratic response surface model accordingly.

[0064] S206. Based on the analysis results of the quadratic response surface model, define the parameter constraint space.

[0065] The constructed quadratic response surface model, validated by ANOVA, is highly significant with no significant lack-of-fit terms, accurately reflecting the influence of parameters on the response. Rotational speed *n* plays a dominant role, while *Nb* and *Nv* influence the flow field structure and energy dissipation through interaction and quadratic terms. Based on this model, an explicit quadratic polynomial expression for the parameter-response relationship is obtained, enabling rapid quantitative prediction. This not only initially identifies the potential optimal region for multi-objective optimization but also reduces the training sample range for subsequent CNN-LSTM-Attention prediction models and the search space for the Multi-Objective Whale Optimization Algorithm (MOWOA), clarifying the boundaries of feasible engineering parameters and effectively connecting mechanistic modeling, statistical analysis, and intelligent optimization.

[0066] S3. Construct CFD data samples.

[0067] To address the problem that high-consistency pulper CFD simulations involve large computational loads and are difficult to directly apply to multi-objective optimization iterations, this invention proposes a physical constraint-based CFD data sample construction method for subsequent data-driven model training and prediction.

[0068] Step S3 includes: S301. Based on the parameter constraint space, the key parameters are sampled in the parameter space by combining the control variable method with Latin hypercube sampling.

[0069] First, given the determination of geometric parameters (number of rotor blades, number of spoilers, clearance from the bottom, etc.) and the range of physical property parameters, parameter space sampling is performed on key operating parameters such as rotational speed n, number of spoilers Nb, and number of rotor blades Nv by combining the controlled variable method with Latin hypercube sampling.

[0070] S302. Perform CFD simulation on each set of experimental parameters to obtain flow field evolution data throughout the swelling process.

[0071] For each set of parameters, transient CFD simulations were performed to obtain flow field evolution data throughout the swelling process, including but not limited to: velocity vector field, turbulent kinetic energy k, specific dissipation rate ω, shear rate field, volume fraction distribution of each phase, and energy consumption versus time curves. To characterize the spatial features of the flow field, the computational domain was discretized along the axial and radial axes, and statistics (mean, peak value, and standard deviation) were extracted from multiple characteristic sections.

[0072] S303. Introduce the time dimension into the sample construction process, sample key physical quantities at equal time steps, and obtain a high-dimensional sample structure of spatial features × time series, which serves as CFD data samples.

[0073] The time dimension is incorporated into the sample construction process, with key physical quantities sampled at equal time steps to form a high-dimensional sample structure of "spatial features × time series". Ultimately, each set of CFD condition samples is represented as: In the formula, Represents the set of spatial flow field characteristics. The time-series evolution characteristics are represented by t, where t is the stirring duration and E is the stirring energy consumption.

[0074] S4. Construct a CNN-LSTM-Attention model and train it using the CFD data samples to obtain a fast prediction model for flow field response.

[0075] To achieve rapid replacement prediction of CFD simulation results for high-consistency hydraulic pulpers and significantly reduce computational costs while maintaining prediction accuracy, this invention constructs a spatiotemporal feature joint prediction model that integrates a multi-scale convolutional neural network (CNN), a physically constrained long short-term memory network (LSTM), and a physically saliency-guided attention mechanism. This model is specifically designed to address the strong shear, multi-scale, highly nonlinear, and time-lag-dependent solid-liquid mixing flow field characteristics within the high-consistency pulper, enabling rapid prediction of mixing time and energy consumption under different operating conditions.

[0076] Step S4 includes the following steps: S401. Construct a multi-scale convolutional neural network subnetwork and extract the spatial distribution characteristics of the flow field in the CFD data samples to obtain microscale shearing characteristics, mesoscale energy transport characteristics and large-scale circulating flow characteristics.

[0077] In the model input layer, the CFD data samples constructed in step three are represented as a joint input vector X containing spatial and temporal features. The spatial features describe the velocity distribution, turbulent kinetic energy distribution, and shear rate characteristics at different locations, while the temporal features describe the dynamic evolution of the flow field over time. The model output is the predicted stirring time for the corresponding operating condition. The overall mapping relationship between the predicted stirring energy consumption value Ê and the overall mapping relationship can be expressed as: in, This represents a nonlinear mapping function composed of CNN, LSTM, and attention mechanisms, where θ is the set of model parameters.

[0078] The CNN subnetwork is used to automatically extract the spatial distribution features in CFD flow field data. Its core idea is to hierarchically represent the spatial non-uniformity of the complex flow field inside the pulper through local receptive field and weight sharing mechanism.

[0079] In this invention, the flow field results obtained from CFD simulation are mapped into a regularized multi-channel spatial feature tensor according to a predetermined sampling method. The channels include, but are not limited to, physical quantities such as velocity modulus, turbulent kinetic energy, and shear rate. Furthermore, to overcome the problem that traditional single-scale convolution cannot simultaneously characterize microscale shear structures, mesoscale energy accumulation structures, and large-scale circulating flows, this invention introduces a multi-scale spatial feature extraction structure into the CNN sub-network. Let the input be... Layer feature map is Then the first The calculation form of layer convolution features is as follows: in, This represents the convolution operation. For the first In a layer CNN, used for the first The convolution kernel weights of the scale branch, which are responsible for scalarizing the th scale branch. The input channel is mapped to the first... One output channel, For bias terms, It is a non-linear activation function.

[0080] Through the aforementioned multi-scale convolutional parallel operations, CNNs can simultaneously perceive flow field gradient changes and structural abrupt changes at different scales within a local spatial range. This allows for a more comprehensive identification of the spatial distribution boundaries between high-speed and low-speed regions, the scale characteristics of strong shearing regions, and the aggregation patterns of high-turbulent kinetic energy regions. These features are implicitly encoded in the convolutional feature map, avoiding reliance on human experience thresholds or rule-based criteria.

[0081] After processing by three convolutional layers and one pooling layer, the multidimensional feature map after pooling is converted into a one-dimensional vector form through a flattening layer. This high-level spatial feature representation is then reorganized into a time-series format, resulting in a spatial feature representation for time-series modeling. This representation is then passed as input to the subsequent LSTM sub-network to further model the dynamic behavior of the flow field features evolving with stirring time. The spatial features output by the CNN are reassembled according to time steps. For each time step t, the corresponding CFD flow field is encoded by the CNN to form a set of spatial feature vectors. Its form of expression is: In the formula, Let represent the flow field input data at time step t, CNN(·) represent the spatial feature extraction network, and Flatten(·) represent the feature flattening or dimensionality reduction operation.

[0082] Through the above processing, the original "space × time" high-dimensional flow field data is transformed into a feature sequence arranged in chronological order: This feature sequence serves as the input to the LSTM subnetwork, enabling a natural transition from "spatial structure representation" to "temporal dynamic modeling".

[0083] The LSTM subnetwork is used to model the temporal evolution of spatial feature sequences extracted by CNN. Its core lies in characterizing the cumulative effect and historical dependence of the flow field state with stirring time through a gating structure.

[0084] S402. The microscale shearing features, mesoscale energy transport features, and large-scale circulating flow features are reorganized into a feature sequence according to time steps. Physical constraints based on energy dissipation or shear intensity are introduced to correct the state of the memory unit, thereby constructing the evolution law of the flow field over time.

[0085] At time step t, the input to the LSTM is the spatial feature vector at the current time. And combined with the hidden state of the previous moment. With memory cell state ,in, and This represents the learnable weight matrix and bias terms. It is the Sigmoid activation function. This represents the hyperbolic tangent function, with an output range of (-1, 1).

[0086] The state update is accomplished through the following gating mechanism.

[0087] Forget Gate: In the formula, The input is the spatial feature vector at the current time. This is the hidden state from the previous moment. This is the output of the forget gate. It is a numerical vector between 0 and 1, used to control the proportion of the memory state information retained from the previous time step (0 indicates complete forgetting, 1 indicates complete retention). This is the Sigmoid activation function. This is the learnable weight matrix in the forget gate, which is continuously updated during model training to adjust the influence of input information on the forget gate output. This is a learnable bias term in the forget gate.

[0088] This is used to determine the extent to which historical flow field characteristics are retained at the current moment, thereby simulating the attenuation effect of early flow field characteristics on later states during the swelling process.

[0089] Input Gate: In the formula, The input is the spatial feature vector at the current time. This is the hidden state from the previous moment. The output coefficients of the input gate, with values ​​between (0, 1), are used to control the proportion of new spatial feature information written into the memory unit at the current time step. For the Sigmoid activation function, The learnable weight matrix in the input gate The learnable bias term in the input gate.

[0090] This is used to control the proportion of new spatial feature information written into memory cells at the current time step.

[0091] Candidate memory states: In the formula, The input is the spatial feature vector at the current time. This is the hidden state from the previous moment. Candidate memory states represent the potential state update amounts formed based on the current flow field characteristics. The hyperbolic tangent function is used to compress data and map it to the range (-1, 1), helping to regulate the data flow within a network. The learnable weight matrix is ​​calculated for candidate memory states. This is a learnable bias term used in the computation of candidate memory states.

[0092] This represents the potential state update based on the current flow field characteristics.

[0093] Based on this, a physical constraint correction term is introduced. The physical consistency correction of the memory cell state is performed, and its update form is as follows: in, This represents a physical constraint term constructed based on energy dissipation, shear strength, or flow continuity. For physical constraint weighting factors, Output Gate: Used to control the output ratio from the memory state to the current hidden state. The output coefficient of the output gate controls the proportion of output from the memorized state to the current hidden state. For the Sigmoid activation function, The weight matrix is ​​the learnable weight matrix in the output gate. This is the hidden state from the previous moment. The input is the spatial feature vector at the current time. This refers to the learnable bias term in the output gate.

[0094] After obtaining the output gate control coefficients, the hidden state at the current time step From the state of memory cells The result is obtained through nonlinear mapping and weighted output, and its calculation form is as follows: in, This represents element-wise multiplication. This represents the hidden state at the current time (step t). It comprehensively characterizes the overall evolutionary state of the system after incorporating the cumulative effects of historical flow fields and the driving force of current spatial characteristics. The control coefficients calculated for the output gate. This is the element-wise multiplication operation (Hadamard product). To calculate the current memory cell state Nonlinear mapping is performed using the hyperbolic tangent function, limiting its value to the range of (-1, 1).

[0095] Hidden state The comprehensive characterization at time step The system's overall evolution state, which integrates the cumulative effect of historical flow fields with the current spatial characteristics, contains both the temporal continuity information of the flow field structure and reflects the stability differences at different stages during stirring and swelling. This provides a highly reliable temporal characteristic basis for the subsequent attention mechanism to screen and weight key time stages.

[0096] Through the aforementioned CNN-LSTM coupling mechanism, this invention achieves a unified modeling of the "spatial structure-temporal evolution" characteristics of the flow field in a high-consistency pulper. Specifically, the CNN is responsible for extracting key flow field structural features at different time sections, while the LSTM, through a gated memory mechanism, depicts the dynamic laws governing the gradual accumulation, decay, and transformation of these structural features during the mixing process.

[0097] This modeling approach can effectively reflect the hysteresis and path dependence characteristics of the flow field caused by the transition from Newtonian to non-Newtonian fluid during the swelling process of wood pulp, providing a highly reliable dynamic feature representation for the subsequent key time step selection and multi-objective optimization of the attention mechanism.

[0098] S403. The hidden states at each time step output by the Long Short-Term Memory Network are weighted, and shear strength or energy dissipation level is introduced as a physical saliency indicator to guide the allocation of attention weights, thereby generating a global context vector that integrates features of key time periods.

[0099] After modeling the temporal characteristics of the flow field, the LSTM subnetwork outputs the corresponding hidden state vector at each time step t. This vector comprehensively represents the coupling state between the spatial structure of the flow field and its historical evolution information up to the current moment.

[0100] Arrange the hidden states of all time steps in chronological order to form a complete set of temporal features: This feature set not only contains the instantaneous feature information of the flow field at each moment, but also implies the potential differences in the contribution of different stages of the swelling process to the final stirring performance and energy consumption. It is the direct input for the Attention mechanism to perform feature weighting and filtering.

[0101] The Attention submodule uses the set of hidden states output by the LSTM. As input, a learnable weighting mechanism is introduced to assess the importance of the flow field state at different times. To further improve the engineering interpretability of the model, a physical saliency guiding factor is introduced into the attention mechanism.

[0102] For the t-th time step, its attention scoring function is defined as: In the formula, This indicates data-driven attention scoring. and For learnable weight parameters, For bias terms, This represents a physical significance index characterizing the shear intensity or energy dissipation level at that time step. For adjustment coefficients, This represents the unnormalized importance score for time step t.

[0103] To convert the above scores into comparable weighted coefficients, scores for all time steps are... Softmax normalization is performed to obtain the corresponding attention weights. : In the formula, And satisfy , for, is the attention weight at time step t, a value between 0 and 1, reflecting the model's attention to the flow field state at that time step when predicting mixing duration and energy consumption. exp(·) is the natural exponential function (based on $e$), used in the Softmax operation to highlight larger scores and ensure a positive output. This is the unnormalized importance score for time step t. It is calculated by combining the hidden states output by the LSTM with the physical significance index. The unnormalized importance score for the t-th time step is used as part of the denominator to calculate the exponential sum of the scores for all time steps.

[0104] This weighting coefficient reflects the degree of attention the model pays to the flow field state at time t when predicting the stirring duration and energy consumption. The larger the weight, the more significant the influence of the flow field structure and evolution characteristics within that time period on the final system response.

[0105] After obtaining the attention weights, a global context feature vector is constructed by weighted summation of the LSTM hidden states. : The context vector It integrates flow field feature information from key time periods throughout the stirring process, enabling the suppression of redundant or low-contribution time period features and the enhanced expression of features from key evolution stages.

[0106] S404. Input the global context vector into the fully connected layer to map the predicted stirring time and stirring energy consumption.

[0107] Among them, the context vector output by Attention The input is fed into a fully connected layer, and the system response prediction result is obtained through linear or nonlinear mapping: In the formula, and These represent the predicted stirring time and stirring energy consumption, respectively. and These are the learnable parameters of the output layer.

[0108] S405. The CNN-LSTM-Attention model is trained using the CFD data samples to obtain a trained fast prediction model for flow field response.

[0109] During the swelling and mixing process in a high-consistency pulper, the contribution of the flow field state at different time stages to the final mixing uniformity and energy consumption varies significantly. For example, the weight of the influence on system performance differs in the initial shear activation stage, the middle structural reconstruction stage, and the later stable flow stage.

[0110] By introducing an attention mechanism, this invention can automatically identify and enhance the focus on flow field features at key time stages, avoiding the information dilution problem caused by the traditional LSTM's "equal weighting" of all time steps, thereby significantly improving prediction accuracy and model engineering interpretability.

[0111] The trained CNN-LSTM-Attention model can output the prediction results of stirring time and energy consumption in milliseconds, providing an efficient and reliable fitness evaluation tool for subsequent multi-objective optimization.

[0112] S5. The multi-objective whale optimization algorithm is used to perform parameter co-optimization on the fast prediction model of the flow field response, and the Pareto optimal solution set is output as the optimal parameters of the fast prediction model of the flow field response, thus obtaining the optimized fast prediction model of the flow field response.

[0113] In addition to obtaining a rapid prediction model, this invention further introduces the Multi-Objective Whale Optimization Algorithm (MOWOA) to perform multi-objective collaborative optimization of the operating parameters of the high-concentration hydraulic pulper, so as to simultaneously minimize the mixing time and energy consumption.

[0114] Step S5 includes: S501. Using rotor speed, number of spoilers, and number of rotor blades as decision variables, and the parameter constraint space as the search boundary.

[0115] The key adjustable parameters of the pulper are used to construct the optimization decision vector: In the formula, n is the rotor speed, Nb is the number of baffles, and Nv is the number of rotor blades. The range of values ​​for each parameter is determined by the parameter constraint space constructed by the response surface method in step two.

[0116] S502. Using the fast prediction model of the flow field response as the fitness evaluation function, for any candidate solution X, the corresponding stirring time and stirring energy consumption are quickly calculated through the prediction model.

[0117] For any candidate solution The CNN-LSTM-Attention model serves as a non-linear prediction function. This is used to quickly calculate the corresponding objective function value. in, This indicates the predicted mixing time. This represents the predicted stirring energy consumption. This prediction result is directly used as input to MOWOA's multi-objective fitness function.

[0118] Furthermore, the multi-objective optimization problem solved by this invention can be expressed as: In the formula, To represent minimizing a multi-objective fitness function, the optimization objective is, under parameter constraints, to find the set of parameter combinations that simultaneously minimize the objective function value. To optimize the decision vector (candidate solution or parameter combination), it includes the key adjustable parameters of the pulper, specifically the rotor speed. Number of spoilers and the number of rotor blades constitute, This is the first sub-objective function in multi-objective optimization. This is the second sub-objective function in multi-objective optimization.

[0119] Further: In the formula, Given a combination of device parameters X, the stirring duration, quickly predicted by the preceding CNN-LSTM-Attention model, is the first objective that needs to be minimized. Given a combination of equipment parameters X, the energy consumption for stirring, which can be quickly predicted by the prediction model, is the second objective that needs to be minimized simultaneously. This is the first sub-objective function in multi-objective optimization. This is the second sub-objective function in multi-objective optimization. The optimization objective is to find a set of parameter combinations that minimize both stirring time and stirring energy consumption under parameter constraints. S503. Initialize the whale population by randomly generating multiple candidate solutions within the search boundary as the initial population.

[0120] In the algorithm iteration process, let the current optimal solution (or the guiding solution in the Pareto optimal solution set) be... any individual whale The location update method is defined as follows: In the formula, This is the position vector of the i-th individual whale in the current iteration (representing a specific set of candidate device parameters). This is the current optimal solution (or the guiding solution in the Pareto optimal solution set), which is equivalent to the "best prey location" found by the whale pod at this moment. This represents the updated position of the individual whale, i.e., the new parameter values ​​for this candidate solution in the next generation (t+1 iterations). Let be the relative distance vector between the individual whale and the current optimal solution. The oscillation coefficient vector, The convergence coefficient vector, and Let be a random number in the interval [0,1], and let a be the convergence factor.

[0121] in: In the formula, and for Random numbers within the interval This is a convergence factor that decreases linearly with the number of iterations, used to enhance global search capabilities in the early stages of the algorithm and strengthen local search capabilities in the later stages.

[0122] S504 simulates the encirclement and predation behavior of whales and the bubble net attack mechanism. In each iteration, the candidate solution is updated by randomly selecting either the shrinking encirclement mechanism or the spiral update mechanism.

[0123] To simulate the spiral approach behavior of whales around their prey, a logarithmic spiral model is introduced: In the formula, This represents the updated position of the individual whale, i.e., the new parameter values ​​for this candidate solution in the next generation (t+1 iterations). Let be the absolute straight-line distance between the current individual whale and the optimal solution, e be the base of the natural logarithm (a mathematical constant, approximately equal to 2.718), b be the logarithmic spiral shape constant, used to define the geometric shape of the whale's spiral ascent, and l be a random number in the interval [-1, 1], which determines the whale's specific position on the spiral path. This is the current optimal solution (or the guiding solution in the Pareto optimal solution set), which is equivalent to the "best prey location" found by the whale pod at this time.

[0124] in: The constant is the helical constant. for Random numbers within the interval. This update method is used to perform a fine search within the neighborhood of the optimal solution, improving the local convergence accuracy of the solution.

[0125] S505 introduces an adaptive weighting strategy to enhance global exploration capabilities in the early stages of iteration and local development capabilities in the later stages of iteration, thus avoiding getting trapped in local optima.

[0126] In each iteration, random probability is used. The decision of whether an individual whale will use an encirclement behavior or a bubble web attack mechanism: In the formula, p is the random probability in the interval [0,1]. When the random number is less than 0.5, the individual updates its position using a prey-surrounding mechanism (upper part of the formula); When the random number is greater than or equal to 0.5, the individual uses a spiral attack mechanism to update its position (lower part of the formula).

[0127] This mechanism enables the algorithm to adaptively switch between global exploration and local development.

[0128] S506. Construct a set of non-dominated solutions based on Pareto dominance relations, compare the dominance relations of candidate solutions in each generation of the population, select non-dominated solutions that are not dominated by any other solutions, and update the Pareto front.

[0129] In multi-objective scenarios, Pareto dominance is introduced to evaluate candidate solutions. For two solutions... and If the following conditions are met: In the formula, and These are two distinct candidate solutions in the parameter space. For the k-th objective function, specifically, k That is, f1 is the stirring time and f2 is the stirring energy consumption.

[0130] Then it is called Dominate The non-dominated solution set is constructed using this criterion, and the Pareto front is updated in each iteration.

[0131] S507. Once the maximum number of iterations or the convergence condition is reached, output the Pareto optimal solution set as the optimal parameters for the fast prediction model of the flow field response.

[0132] In this example, the fast prediction results provided by the CNN-LSTM-Attention model eliminate the need for CFD simulation in Pareto ranking and crowding calculation, thus significantly improving the efficiency of multi-objective optimization.

[0133] Furthermore, to improve the algorithm's search efficiency in complex nonlinear problems, this invention introduces an adaptive weighting strategy: in the early stages of iteration, the exploration factor weight is increased to enhance global search capabilities; in the later stages of iteration, the exploration weight is decreased to enhance exploration capabilities, achieving a fine-grained search of the neighborhood of the optimal solution, thereby avoiding getting trapped in local optima. The CNN-LSTM-Attention model, as a fast fitness function evaluation module for MOWOA, eliminates the need for repeated CFD simulations during the optimization process, significantly reducing computational costs. The algorithm ultimately outputs a Pareto optimal solution set that satisfies different trade-off requirements, providing various optional parameter combinations for engineering applications.

[0134] This example provides a method for optimizing a rapid prediction model of the flow field response of a high-consistency pulper. First, a CFD flow field mechanism model of the high-consistency pulper is constructed by introducing a user-defined function based on cumulative strain, which can realistically simulate the dynamic transition from Newtonian to non-Newtonian fluid. Then, a response surface model is constructed based on experimental design and CFD simulation of key parameters, and a parameter constraint space is defined according to the analysis results. Next, a high-dimensional CFD data sample containing spatial features and time series is constructed, and a CNN-LSTM-Attention model is designed to train the sample to obtain a rapid prediction model of the flow field response. Finally, the flow field is extracted using a multi-scale convolutional neural network. The microscale shear, mesoscale energy transport, and large-scale circulation characteristics are analyzed. A physically constrained long short-term memory network is used to modify memory units by introducing energy dissipation or shear intensity to accurately model the temporal evolution of the flow field. A physical saliency-guided attention mechanism strengthens focus on key time stages, thereby improving the accuracy of single-time flow field response assessment while maintaining prediction accuracy. Finally, a multi-objective whale optimization algorithm is employed, using the aforementioned fast prediction model as the fitness function for parameter collaborative optimization. An adaptive weighting strategy balances global search and local exploitation capabilities, outputting a Pareto optimal solution set that minimizes both stirring time and energy consumption. This method organically integrates high-fidelity CFD simulation with deep learning prediction models, overcoming the shortcomings of traditional pure physical modeling, such as high computational cost and difficulty in iterative optimization, while avoiding the prediction instability problem caused by the lack of physical constraints in pure data-driven models. It can efficiently obtain multi-objective optimal parameter combinations that balance stirring efficiency and energy consumption while significantly reducing computational resource consumption. This provides a highly reliable and efficient solution for optimizing the process parameters of high-consistency pulpers in lyocell fiber production, effectively improving the simulation accuracy and optimization efficiency of the swelling process, and reducing equipment operating energy consumption and R&D trial and error costs.

[0135] For those consistent with the above, please refer to Figure 3 , Figure 3 This application provides a schematic diagram of a system for optimizing a rapid prediction model of the flow field response of a high-consistency pulper. (See attached diagram.) Figure 3 As shown, the system includes: The first processing unit 1 is used to construct a CFD flow field mechanism model of a high-consistency pulper.

[0136] The second processing unit 2 is used for experimental design and CFD simulation based on key parameters, constructing a response surface model, and defining the parameter constraint space for subsequent optimization based on the analysis results of the model.

[0137] The third processing unit 3 is used to construct CFD data samples.

[0138] The fourth processing unit 4 is used to construct a CNN-LSTM-Attention model and train the CNN-LSTM-Attention model using the CFD data samples to obtain a fast prediction model for flow field response.

[0139] The fifth processing unit 5 is used to perform parameter co-optimization on the fast prediction model of the flow field response using the multi-objective whale optimization algorithm, and output the Pareto optimal solution set as the optimal parameters of the fast prediction model of the flow field response to obtain the optimized fast prediction model of the flow field response.

[0140] For examples consistent with the above embodiments, please refer to... Figure 4 , Figure 4 A schematic diagram of a terminal structure provided in an embodiment of this application is shown in the figure. It includes a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions. The program includes instructions for performing the following steps. A CFD flow field mechanism model of a high-consistency pulper was constructed.

[0141] Based on experimental design and CFD simulation using key parameters, a response surface model is constructed, and the parameter constraint space for subsequent optimization is defined according to the analysis results of the model.

[0142] Construct CFD data samples.

[0143] A CNN-LSTM-Attention model is constructed and trained using the CFD data samples to obtain a fast flow field response prediction model.

[0144] The multi-objective whale optimization algorithm is used to perform parameter co-optimization on the fast prediction model of the flow field response, and the Pareto optimal solution set is output as the optimal parameters of the fast prediction model of the flow field response, thus obtaining the optimized fast prediction model of the flow field response.

[0145] The above mainly describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the terminal includes the corresponding hardware structure and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0146] This application embodiment can divide the terminal into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0147] This application also provides a computer storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the fast prediction model optimization methods for high-consistency pulper flow field response described in the above method embodiments.

[0148] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program that causes a computer to perform some or all of the steps of any of the fast prediction model optimization methods for high-consistency pulper flow field response described in the above method embodiments.

[0149] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0150] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0151] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0152] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] Furthermore, the functional units in the various embodiments of the application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software program module.

[0154] If the integrated unit is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0155] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory, a random access memory, a magnetic disk, or an optical disk, etc.

[0156] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for optimizing a rapid prediction model of the flow field response of a high-consistency pulper, characterized in that, include: Constructing a CFD flow field mechanism model for a high-consistency pulper; Based on the experimental design and CFD simulation of key parameters, a response surface model is constructed, and the parameter constraint space for subsequent optimization is defined according to the analysis results of the model. Construct CFD data samples; A CNN-LSTM-Attention model is constructed and trained using the CFD data samples to obtain a fast flow field response prediction model. The multi-objective whale optimization algorithm is used to perform parameter co-optimization on the fast prediction model of the flow field response, and the Pareto optimal solution set is output as the optimal parameters of the fast prediction model of the flow field response, thus obtaining the optimized fast prediction model of the flow field response. The construction of the CFD flow field mechanism model for the high-consistency pulper includes: Based on the structure of the high-consistency hydraulic pulper used in the production of experimental lyocell fibers, a three-dimensional geometric model was constructed. The interaction between the wood pulp phase and the liquid phase was described using the Euler-Euler multiphase flow model, and a multiphase flow model was constructed. The SST k-ω model with rotational flow correction is used to describe the strong rotational and high shear characteristics of the flow field, and a turbulence model is constructed. The impeller rotation behavior is described using a motion mesh method, and the rotational speed is set for the rotation domain; The transition from Newtonian to non-Newtonian fluids is simulated using user-defined functions based on Ansys Fluent, including: The wood pulp fiber, NMMO solution and swelling slurry are defined as a three-phase system, and the initial volume fractions of wood pulp fiber, NMMO solution and swelling slurry are set. The phase volume fraction is initialized using a custom function `init_phase_system`, and the initial cumulative strain value is set to zero for each computational unit. Within each time step, the cumulative strain is updated using the following formula: cum_strain = cum_strain + strain_rate × dt The phase transition process is described by introducing an exponential function based on cumulative strain: trans = 1 - exp(-cum_strain / CRITICAL_STRAIN) In the formula, CRITICAL_STRAIN is the critical strain threshold. When the cumulative strain reaches this threshold, the Newtonian fluid gradually transforms into a non-Newtonian swelling slurry, and the volume fraction of each phase is dynamically updated accordingly. The rheological properties of the swollen slurry are described by a volume fraction-dependent viscosity model, and its dynamic viscosity is: In the formula, The initial viscosity, Viscosity in the fully swollen state. This represents the liquid fraction of the swelling slurry.

2. The method for optimizing the rapid prediction model of the flow field response of a high-consistency pulper according to claim 1, characterized in that, The experimental design and CFD simulation based on key parameters construct a response surface model, and delineate the parameter constraint space for subsequent optimization based on the analysis results of this model, including: Rotor speed, number of spoilers, and number of rotor blades are selected as core independent variables, and the ranges of the rotor speed, number of spoilers, and number of rotor blades are defined. A three-factor, three-level, center-based composite test was used to construct the combination of rotor speed, number of spoilers, and number of rotor blades to obtain the test parameter set. For the aforementioned set of test parameters, transient simulations were conducted based on a CFD flow field mechanism model of a high-consistency pulper, and simulation results were obtained. Based on the simulation results, the stirring time and stirring energy consumption are determined as response variables; Based on the experimental parameter set and its corresponding response variables, variance analysis was performed to construct a quadratic response surface model for stirring time and stirring energy consumption. Based on the analysis results of the quadratic response surface model, the parameter constraint space is defined.

3. The method for optimizing the rapid prediction model of the flow field response of a high-consistency pulper according to claim 2, characterized in that, The construction of CFD data samples includes: Based on the parameter constraint space, the key parameters are sampled in the parameter space by combining the control variable method with Latin hypercube sampling. CFD simulations were performed on each set of experimental parameters to obtain flow field evolution data throughout the swelling process; By introducing the time dimension into the sample construction process and sampling key physical quantities at equal time steps, a high-dimensional sample structure of spatial features × time series is obtained, which serves as CFD data samples.

4. The method for optimizing the rapid prediction model of the flow field response of a high-consistency pulper according to claim 3, characterized in that, The formula for expressing the CFD data sample is as follows: ， In the formula, Represents the set of spatial flow field characteristics. (t) represents the temporal evolution characteristics, where t is the stirring duration and E is the stirring energy consumption.

5. The method for optimizing the rapid prediction model of the flow field response of a high-consistency pulper according to claim 1, characterized in that, The construction of the CNN-LSTM-Attention model, and the training of the CNN-LSTM-Attention model using the CFD data samples to obtain a fast flow field response prediction model, includes: A multi-scale convolutional neural network subnetwork was constructed, and the spatial distribution characteristics of the flow field in the CFD data samples were extracted to obtain micro-scale shearing characteristics, meso-scale energy transport characteristics, and large-scale circulating flow characteristics. The microscale shearing features, mesoscale energy transport features, and large-scale circulating flow features are reorganized into a feature sequence according to time steps. Physical constraints based on energy dissipation or shear intensity are introduced to correct the state of the memory unit, thereby constructing the evolution law of the flow field over time. The hidden states at each time step of the output of the Long Short-Term Memory Network are weighted, and shear strength or energy dissipation level is introduced as a physical saliency index to guide the allocation of attention weights, thereby generating a global context vector that integrates features of key time periods. The global context vector is input into the fully connected layer to map the predicted stirring time and stirring energy consumption. The CNN-LSTM-Attention model is trained using the CFD data samples to obtain a trained fast prediction model for flow field response.

6. The method for optimizing the rapid prediction model of the flow field response of a high-consistency pulper according to claim 1, characterized in that, The multi-objective whale optimization algorithm is used to collaboratively optimize the parameters of the fast flow field response prediction model, and the Pareto optimal solution set is output as the optimal parameters of the fast flow field response prediction model, including: The rotor speed, the number of spoilers, and the number of rotor blades are used as decision variables, and the parameter constraint space is used as the search boundary. The fast prediction model of the flow field response is used as the fitness evaluation function. For any candidate solution X, the corresponding stirring time and stirring energy consumption are quickly calculated through the prediction model. Initialize the whale population by randomly generating multiple candidate solutions within the search boundary as the initial population; Simulates the encirclement and predation behavior of whales and the bubble net attack mechanism. In each iteration, the candidate solution is updated by randomly selecting either a shrinking encirclement mechanism or a spiral update mechanism. An adaptive weighting strategy is introduced to enhance global exploration capabilities in the early stages of iteration and local development capabilities in the later stages of iteration, thereby avoiding getting trapped in local optima. Based on the Pareto dominance relation, a non-dominated solution set is constructed. The dominance relation of candidate solutions in each generation of the population is compared, and non-dominated solutions that are not dominated by any other solutions are selected and the Pareto front is updated. Once the maximum number of iterations or the convergence condition is reached, the Pareto optimal solution set is output as the optimal parameters for the fast flow field response prediction model.

7. A rapid prediction model optimization system for the flow field response of a high-consistency pulper, characterized in that, include: The first processing unit is used to construct a CFD flow field mechanism model of a high-consistency pulper. The second processing unit is used for experimental design and CFD simulation based on key parameters, constructing a response surface model, and defining the parameter constraint space for subsequent optimization based on the analysis results of the model. The third processing unit is used to construct CFD data samples; The fourth processing unit is used to construct a CNN-LSTM-Attention model and train the CNN-LSTM-Attention model using the CFD data samples to obtain a fast flow field response prediction model. The fifth processing unit is used to perform parameter co-optimization on the fast prediction model of the flow field response using the multi-objective whale optimization algorithm, and output the Pareto optimal solution set as the optimal parameters of the fast prediction model of the flow field response, so as to obtain the optimized fast prediction model of the flow field response. The construction of the CFD flow field mechanism model for the high-consistency pulper includes: Based on the structure of the high-consistency hydraulic pulper used in the production of experimental lyocell fibers, a three-dimensional geometric model was constructed. The interaction between the wood pulp phase and the liquid phase was described using the Euler-Euler multiphase flow model, and a multiphase flow model was constructed. The SST k-ω model with rotational flow correction is used to describe the strong rotational and high shear characteristics of the flow field, and a turbulence model is constructed. The impeller rotation behavior is described using a motion mesh method, and the rotational speed is set for the rotation domain; The transition from Newtonian to non-Newtonian fluids is simulated using user-defined functions based on Ansys Fluent, including: The wood pulp fiber, NMMO solution and swelling slurry are defined as a three-phase system, and the initial volume fractions of wood pulp fiber, NMMO solution and swelling slurry are set. The phase volume fraction is initialized using a custom function `init_phase_system`, and the initial cumulative strain value is set to zero for each computational unit. Within each time step, the cumulative strain is updated using the following formula: cum_strain = cum_strain + strain_rate × dt The phase transition process is described by introducing an exponential function based on cumulative strain: trans = 1 - exp(-cum_strain / CRITICAL_STRAIN) In the formula, CRITICAL_STRAIN is the critical strain threshold. When the cumulative strain reaches this threshold, the Newtonian fluid gradually transforms into a non-Newtonian swelling slurry, and the volume fraction of each phase is dynamically updated accordingly. The rheological properties of the swollen slurry are described by a volume fraction-dependent viscosity model, and its dynamic viscosity is: In the formula, The initial viscosity, Viscosity in the fully swollen state. This represents the liquid fraction of the swelling slurry.

8. A terminal, characterized in that, The device includes a processor, an input device, an output device, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the fast prediction model optimization method for the flow field response of a high-consistency pulper as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the fast prediction model optimization method for the flow field response of a high-consistency pulper as described in any one of claims 1-6.

Images