Industrial large space temperature and humidity uniformity optimization method based on POD-mGPR proxy model

Abstract

The invention discloses an industrial large space temperature and humidity uniformity optimization method based on a POD-mGPR proxy model, and belongs to the field of industrial environment intelligent regulation and control, the method fuses computational fluid mechanics (CFD), intrinsic orthogonal decomposition (POD) and multivariate Gaussian process regression (mGPR), a high-dimensional data set is generated through CFD simulation, after a main mode is extracted through POD dimensionality reduction, the mGPR is used for constructing the proxy model, and the main mode is extracted through the mGPR; rapid and accurate prediction of the whole-space temperature and humidity field is realized; defining a non-uniformity coefficient, and establishing quantitative mapping of a control variable and a uniformity improvement effect; and finally, screening key variables through Sobol global sensitivity analysis, and solving an optimal control parameter combination in combination with a genetic algorithm. According to the method, the problems of insufficient perception, lack of quantitative feedback in control and low high-dimensional optimization efficiency of a traditional method are solved, the transformation from local experience control to global quantitative control is realized, and the temperature and humidity uniformity and the energy utilization efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for industrial environments, and in particular to a method for optimizing the uniformity of temperature and humidity in large industrial spaces based on the POD-mGPR surrogate model. Background Technology

[0002] Modern large-scale industrial workshops (such as aircraft manufacturing, data centers, and precision electronics) are not only typical multi-physics coupled dynamic systems, but their internal airflow organization is also affected by equipment heat dissipation, air conditioning suction, obstruction, and thermal pressure effects, forming a complex nonlinear flow field that includes vortices, jets, and stratification.

[0003] Traditional environmental control methods, firstly, rely on sparse point measurements at the perception level, characterizing the entire large spatial field through data from a few fixed-location sensors. This fails to capture crucial features such as horizontal and vertical gradients, dead zones in corners, and heat plumes above heat sources generated by equipment, leading to severe distortion in the control system's perception of the real environmental state. Secondly, at the control logic level, they primarily use simple PID or setpoint-based zone control. This approach is based on the fragile assumptions that "local measurement points can represent the whole" and "control response is linear and decoupled." In reality, adjusting the airflow of a particular air conditioner not only affects its local area but also disrupts the pressure balance and temperature distribution of the entire space through airflow disturbances, generating unpredictable chain reactions. Finally, in terms of optimization, they heavily depend on human experience, lacking the ability to perform closed-loop optimization based on overall performance goals. Faced with hundreds or thousands of control variables, relying solely on manual methods often fails to find the globally optimal or near-optimal solution.

[0004] To overcome the limitations of traditional methods, academia and industry have explored two main paths. The first path is a physical modeling method based on computational fluid dynamics simulation. This method has strict physical interpretability and can provide full-space, non-contact "virtual sensing" data, but it also suffers from high computational costs and poor real-time feedback control capabilities. The second path is a purely data-driven black-box model method, which uses historical operating data or experimental data to train machine learning models such as deep neural networks and random forests to directly establish a mapping relationship from control variables to the response of key measurement points. However, this method also suffers from poor physical interpretability and low accuracy.

[0005] In summary, current large-scale space environment control faces a core contradiction: high-fidelity physical models, represented by CFD, are accurate but slow, while purely data-driven models are fast but unreliable and have poor generalization. Neither can independently support a fully perceptive, closed-loop control intelligent control system. Therefore, the industry urgently needs an innovative technical framework that can organically integrate physical mechanisms and data intelligence. First, it must be able to provide high-precision full-space environmental field predictions at near real-time speeds, achieving a leap from "sparse point perception" to "full-field perception." Second, it must be able to utilize limited CFD simulation data to construct a high-dimensional, nonlinear, coupled mapping model between control variables and the full-field environmental response, and this model must have a physically consistent foundation. Third, it must explicitly represent global performance indicators as functions of control variables, thereby applying advanced optimization algorithms to solve for the optimal control command set that satisfies the constraints within seconds. Summary of the Invention

[0006] The purpose of this invention is to provide a method for optimizing the uniformity of temperature and humidity in large industrial spaces based on the POD-mGPR surrogate model. It addresses three core problems: insufficient perception of the spatial environment, lack of quantitative feedback in the control strategy, and low efficiency of high-dimensional optimization. The invention constructs an efficient surrogate model to achieve closed-loop control of the uniformity of temperature and humidity in large industrial spaces, thereby enabling precise control of the global large-space environment and quantifiable, closed-loop air conditioning control.

[0007] To achieve the above objectives, this invention provides a method for optimizing the temperature and humidity uniformity of large industrial spaces based on the POD-mGPR surrogate model, the steps of which are as follows: S1. Construct a large-space environment proxy model based on POD-mGPR, generate training data through CFD high-fidelity simulation, and achieve rapid and accurate prediction from control parameters to the temperature and humidity field of the entire space through POD dimensionality reduction and mGPR modeling. S2. Implement a control strategy based on the non-uniformity coefficient, define the temperature and humidity non-uniformity coefficient and the improvement amount, and establish a quantitative mapping relationship between the control variables and the improvement effect of temperature and humidity uniformity through variable operating condition sampling and Gaussian process regression fitting. S3. Implement an optimization method based on sensitivity analysis. Use Sobol global sensitivity analysis to screen key control variables, and use a genetic algorithm to search for the optimal parameter combination in the dimensionality-reduced variable space to achieve dimensionality reduction of high-dimensional control variables and solution of optimal parameter combination.

[0008] Preferably, S1 is executed according to the following sub-steps: S1.1 Physical Model and Computational Mesh Establishment: Based on the geometric drawings and equipment layout diagrams of the target industrial space, a geometric model including the building structure, process equipment, and air conditioning and ventilation system is constructed using 3D modeling software. The geometric model covers the location and size information of walls, columns, roof, industrial production equipment, air conditioning supply and return air vents, load-bearing walls, and windows. The geometric model is imported into mesh generation software, and a hybrid meshing technique combining unstructured meshes and boundary layer meshes is used to divide the mesh. Local densification is performed on key areas around the supply and return air vents and major heat source equipment. S1.2 Boundary condition parameterization and sampling design: The core control variables are determined as the air volume, air temperature, and air humidity of each air outlet. The Latin hypercube sampling method is used to generate dozens to hundreds of uniform and low-correlation control parameter combinations in the multidimensional control variable design space. Historical environmental parameters of the target space are collected, including the air conditioning system outlet parameters and the ambient temperature and humidity of each measuring point. The initial boundary conditions for computational fluid dynamics simulation are set. The initial boundary conditions include the heat transfer coefficient of the building envelope, the initial ambient temperature and humidity, the air conditioning outlet temperature and humidity, and the air volume. S1.3 High-fidelity CFD simulation and dataset generation: The mesh model from step S1.1 and the control parameters from step S1.2 are imported into the computational fluid dynamics software. The k-ε turbulence model is selected, and the energy equation and component transport equation are activated. Steady-state solutions are obtained using a set of control equations consisting of the continuity equation, momentum equation, energy equation, and component transport equation. After each set of simulations is completed, the full-field temperature and humidity data of the specified monitoring area are extracted. After all simulations are completed, a high-fidelity training dataset consisting of "input control parameters - output full-field data" is constructed. S1.4 Dimensionality reduction and feature extraction of spatial field data based on POD: The temperature field data and humidity field data obtained in step S1.3 are subjected to intrinsic orthogonal decomposition analysis respectively. Each high-dimensional spatial field data is represented as a linear combination of several characteristic modes. The characteristic modes are sorted from largest to smallest in terms of characterizing energy. The top-ranked characteristic modes are extracted. The high-dimensional field data of each sample is compressed into a low-dimensional amplitude vector that records the projection coefficients of the original field data on each principal mode. S1.5. Based on mGPR input-output mapping relationship learning, the control parameter combination in step S1.2 is used as the input feature, and the low-dimensional amplitude vector of temperature field and the low-dimensional amplitude vector of humidity field corresponding to step S1.4 are used as the output targets to train two multivariate Gaussian process regression models, forming a core proxy model for quickly predicting the amplitude of low-dimensional modal coefficients from the air outlet conditions of the air conditioning system. S1.6 Rapid reconstruction of full-field environmental information: For a new set of unsimulated control parameters, the corresponding temperature and humidity field modal coefficients are predicted using the core surrogate model trained in step S1.5. The modal coefficients are then linearly combined with the POD space principal modes extracted and stored in step S1.4 to reconstruct the full-space temperature distribution cloud map and humidity distribution cloud map.

[0009] Preferably, S2 is executed according to the following sub-steps: S2.1 Definition of Baseline Scenario and Non-uniformity Coefficient: Select a typical air conditioning operating condition as the baseline scenario. Reconstruct the overall temperature and humidity of the baseline scenario using the method in step S1.6, and calculate the non-uniformity coefficient according to the following formula:

[0010] in, For the temperature or humidity at each point in the space, Let n be the average temperature or average humidity at each point in the space, and n be the number of points in the space. S2.2 Variable operating condition sampling and non-uniformity coefficient calculation: Within the operating range of the control variables, the Latin hypercube sampling method is used to generate thousands of different variable air volume and variable temperature and humidity air supply scenarios. For each scenario, the fast reconstruction method in step S1.6 is called to obtain the predicted overall temperature and humidity, and the corresponding non-uniformity coefficient is calculated according to the formula in step S2.1 to form a "control variable combination - non-uniformity coefficient" dataset. S2.3 Calculation of the improvement amount of the non-uniformity coefficient, according to the formula Calculate the improvement amount of the non-uniformity coefficient for each group of variable operating condition scenarios in step S2.2, where The non-uniformity coefficient is the coefficient for the baseline scenario. For the first i The non-uniformity coefficient of the variable operating condition scenario; a positive value indicates that the uniformity of the scenario is better than that of the benchmark scenario. S2.4. Influence function fitting: Using the combination of control variables from step S2.2 as input and the improvement in the non-uniformity coefficient calculated in step S2.3 as output, Gaussian process regression is used to fit the continuous response function. ,in For the air outlet conditions, including the outlet temperature, humidity, and air volume of each air conditioning system, the fitting process provides the predicted mean and variance for uncertainty quantification.

[0011] Preferably, step S3 is performed as follows: S3.1 Global Sensitivity Analysis and Key Variable Screening: Based on the response function fitted in step S2.4, the Sobol global sensitivity analysis method is used to calculate the first-order sensitivity index of each control variable through Monte Carlo sampling. and total effect sensitivity index Select the overall effect sensitivity index The top k variables with the largest values ​​that are greater than a set threshold are selected as key optimization variables; the set threshold can be 0.05. S3.2 Intelligent optimization solution: With the objective of maximizing the improvement in the non-uniformity coefficient, the key optimization variables selected in step S3.1 are used as the optimization objects, and the response function fitted in step S2.4 is used as the objective function. Under the constraints of the air conditioning system satisfying the upper and lower limits of total air volume, the air volume range of a single air outlet, and the supply air temperature limit, a genetic algorithm is used to search for the optimal combination of control parameters in the dimensionality-reduced key variable space. All parameters (the six optimal air volumes + the remaining baseline parameters) are input into the surrogate model to verify the degree of improvement compared to the baseline state.

[0012] Preferably, in step S1.1, the building structure includes walls, columns, and a roof; the air conditioning and ventilation system explicitly includes the location and size parameters of the air supply and return vents; and the industrial production equipment includes its specific spatial layout information.

[0013] Preferably, in step S1.3, the computational fluid dynamics software is ANSYS Fluent, the solution method adopts the Coupled pressure-velocity coupling method, the momentum equation and turbulent kinetic energy adopt the first-order upwind scheme, the energy equation and turbulent dissipation rate adopt the second-order upwind scheme, and water vapor is used as an additional component to simulate the humidity field.

[0014] Preferably, in step S1.5, the multivariate Gaussian process regression model is trained using MATLAB's Statistics and Machine Learning Toolbox, the quadratic exponential kernel function is selected, and the training dataset is divided into a training set and a test set in a 4:1 ratio. The training set is used to train the model's hyperparameters, and the test set is used to verify the model's accuracy. The average relative error of the temperature field and humidity field reconstruction is less than 2%.

[0015] Preferably, in step S3.1, at least 10,000 samples are taken in the input space using the Saltelli sampling sequence, and the corresponding improvement in the non-uniformity coefficient is calculated using the GPR model trained in step S2.4, thereby accurately estimating the sensitivity index.

[0016] Preferably, in step S3.2, the parameters of the genetic algorithm (GA) are set as follows: population size 50, maximum number of iterations 100, crossover probability 0.8, mutation probability 0.1. After optimization, the optimal combination of control parameters is input into the core surrogate model constructed in S1 to verify the degree of uniformity improvement.

[0017] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) Achieve precise perception of the entire spatial environment and break through the limitations of traditional sparse point measurement. This invention uses the POD-mGPR proxy model to integrate the high-fidelity physical mechanism of CFD with the advantages of data-driven approach. It can quickly invert the temperature and humidity status of tens of thousands or even millions of locations in the entire space from the control parameters, accurately capture key features such as horizontal and vertical gradients, corner dead zones and heat plumes above heat sources, and completely solve the problem of the control system's distortion of the perception of the real environmental state, achieving a leapfrog improvement from "sparse point perception" to "full-field holographic perception".

[0018] (2) Establish a quantitative closed-loop control system to improve the scientific nature and accuracy of regulation. By defining the non-uniformity coefficient and the improvement amount, a quantitative mapping relationship between the control variable and the uniformity effect is constructed, so that "improving temperature and humidity uniformity" is transformed from a fuzzy target into a calculable and evaluable quantitative indicator. The system does not need to rely on human experience for "trial and error" adjustment, and can automatically iterate the optimal control strategy with the goal of "lowest non-uniformity coefficient", so that the air conditioning system is upgraded from a passive execution device to a continuously self-optimizing intelligent agent, significantly reducing regulation costs and improving control accuracy.

[0019] (3) Overcoming the challenges of high-dimensional optimization while balancing optimization efficiency and physical interpretability. By using Sobol global sensitivity analysis to identify key control variables, the high-dimensional optimization problem is effectively reduced in dimensionality. Combined with genetic algorithms, the optimal parameter combination is solved quickly, significantly improving optimization efficiency and feasibility. At the same time, sensitivity analysis can quantify the independent influence and interactive coupling effects of each variable, revealing the complex control mechanism of nonlinear systems. This makes the optimization process no longer a "black box operation," achieving both high efficiency and physical interpretability. Ultimately, it improves energy utilization efficiency while enhancing temperature and humidity uniformity.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating an embodiment of the method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR proxy model according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example like Figure 1 As shown, a method for optimizing the temperature and humidity uniformity of large industrial spaces based on the POD-mGPR surrogate model is presented, with the following steps: (1) Construction of a large space environment proxy model based on POD-mGPR.

[0026] S1.1 Physical Model and Computational Mesh Establishment. First, using 3D modeling software, a precise geometric model is established based on the spatial drawings, including the building envelope (walls, roof, columns), equipment outlines, and all supply and return air vents. Non-critical areas are appropriately simplified. This geometric model is then imported into mesh generation software, where local mesh refinement is performed on all supply and return air vents and the main heat-generating equipment areas, and the mesh quality is checked to ensure it meets standards.

[0027] S1.2 Boundary Condition Parameterization and Sampling Design. The control variables are the airflow, temperature, and humidity of each group of air vents (e.g., 12 groups). Using the Latin hypercube sampling (LHS) method, several groups (e.g., 1000 groups) of different control parameter combinations are generated within the aforementioned 36-dimensional design space. This ensures that the samples are uniformly distributed across each dimension and have minimal correlation between dimensions, forming the input sample set for CFD simulation. The boundary conditions are set as follows: the exterior walls and roof are set as Type III boundary conditions with fixed heat transfer coefficients; production equipment is converted to a fixed heat source based on its rated power, and lighting load is converted to a uniformly distributed heat source.

[0028] S1.3 High-fidelity CFD simulation and dataset generation. The mesh model and boundary conditions are imported one by one into the ANSYS Fluent solver. The energy equation and component transport equation are activated, with water vapor added as an additional component to simulate the humidity field. The coupled pressure-velocity method is used for the solution; the momentum equation and turbulent kinetic energy are solved using a first-order upwind scheme, while the energy equation and turbulent dissipation rate are solved using a second-order upwind scheme. Steady-state calculations are performed for each set of parameters. After convergence, the temperature and relative humidity values ​​of all mesh nodes in the entire computational domain are extracted to form a high-dimensional data vector. Finally, a dataset containing several samples (e.g., 2 million) is obtained, with each sample corresponding to a 36-dimensional input vector (control parameters) and a 6 million-dimensional output vector (full-field temperature and humidity data).

[0029] S1.4. Perform intrinsic orthogonal decomposition (POD) on the temperature and humidity field samples respectively. Taking the temperature field as an example, construct a matrix from the temperature field data. Calculate its spatial modes (POD modes) and the eigenvalues ​​corresponding to each mode using singular value decomposition (SVD). Extract the POD modes whose cumulative contribution rate (i.e., energy percentage) reaches more than 95% as the principal modes (such as the first 6 dimensions), which is sufficient to accurately characterize the spatial distribution characteristics of the original field. Each original high-dimensional temperature (humidity) field can be approximately represented as a linear combination of these principal modes, and its combination coefficients are the low-dimensional amplitude vectors.

[0030] S1.5 Input-Output Mapping Learning Based on mGPR. Two multivariate Gaussian process regression (mGPR) models are constructed. Taking 36-dimensional control parameters as input, the first 6 dimensions (temperature and humidity amplitudes) are outputs respectively. The quadratic exponential kernel function is selected using MATLAB's Statistics and Machine Learning Toolbox. 160 samples out of 200 are used as the training set to train the hyperparameters of the mGPR model; the remaining 40 samples are used as the test set to verify the model's accuracy. After training, if the average relative error of the reconstructed temperature and humidity fields on the test set is less than a threshold (e.g., 2%), the constructed surrogate model is considered to have good prediction accuracy.

[0031] S1.6 Rapid Reconstruction of Overall Environmental Information. For any new, unsimulated air supply parameters (36 dimensions), simply inputting them into the trained mGPR model will quickly yield the corresponding temperature and humidity amplitude prediction vectors. Then, by linearly superimposing this amplitude vector with the stored 6 temperature POD modes and 6 humidity POD modes, a temperature and humidity distribution cloud map of 2 million grid points in the entire workshop under this operating condition can be instantly reconstructed.

[0032] (2) Control strategy based on non-uniformity coefficient.

[0033] S2.1, Definition of Baseline Scenario and Non-uniformity Coefficient. The current normal operating conditions of the workshop are selected as the baseline scenario, and its control parameters are known. Using a surrogate model, the overall temperature and humidity under the baseline scenario are reconstructed. The non-uniformity coefficient is calculated using data from all grid points within the target area (a certain spatial plane). .

[0034] S2.2 Variable Operating Condition Sampling and Non-uniformity Coefficient Calculation. Within the allowable operating range of the control variables, the LHS method is used again to generate several sets (e.g., 5000 sets) of different supply air parameter combinations. Using a surrogate model, the overall temperature distribution under these 5000 scenarios is quickly predicted, and the temperature non-uniformity coefficient is calculated for each scenario. This results in a dataset containing 5,000 data points.

[0035] S2.3 Calculation of the improvement in non-uniformity coefficient. For each set of variable operating conditions i, calculate the improvement in its non-uniformity coefficient relative to the baseline scenario. When this value is greater than 0, it indicates that the temperature uniformity under this operating condition is better than that under the reference operating condition.

[0036] S2.4, Influence Function Fitting. To establish an explicit relationship between the control variables and the improvement effect, 5000 sets of 36-dimensional control parameters were used as input, and the corresponding... For output, a Gaussian process regression (GPR) model is trained to quantify the potential for uniformity improvement brought about by any air supply combination.

[0037] (3) Optimization method based on sensitivity analysis.

[0038] S3.1 Global Sensitivity Analysis and Key Variable Selection. Based on the aforementioned influence function, the Sobol global sensitivity analysis method is adopted. The input space is sampled several times (e.g., 10,000 times) using Saltelli sampling sequences. The corresponding unevenness coefficient improvement value is quickly calculated using a pre-trained GPR model, and then the first-order sensitivity index S for each control variable (36 variables) is estimated. i And the overall effect sensitivity index S Ti Analysis of the results revealed S. Ti The value is significantly higher than other variables (such as 6), and it is used as a "key optimization variable" to reduce the original 48-dimensional optimization problem to 6 dimensions.

[0039] S3.3. Intelligent Optimization Solution. The optimization problem is constructed as follows: the objective is to maximize the influence function, with the variables being the airflow of six key air vents, while the airflow of other vents and all supply air temperatures remain constant at baseline values. The constraint is the set airflow for each key air vent. A Genetic Algorithm (GA) is used for the solution. The population size is set to 50, the maximum number of iterations to 100, the crossover probability to 0.8, and the mutation probability to 0.1. Optimization is performed within a 6-dimensional space of key variables. Finally, an optimal airflow combination is obtained through optimization calculations. All parameters (the optimal six airflows + the remaining baseline parameters) are input into a surrogate model to verify the degree of improvement compared to the baseline state.

[0040] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art to meet different specific practical needs according to actual circumstances. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims. In the above description, numerous specific details have been set forth to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to implement the present invention. In other instances, to avoid obscuring the present invention, well-known techniques, such as specific construction details, operating conditions, and other technical conditions, have not been specifically described.

[0041] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model, characterized in that, The steps are as follows: S1. Construct a large-space environment proxy model based on POD-mGPR, generate training data through CFD high-fidelity simulation, and achieve rapid and accurate prediction from control parameters to the temperature and humidity field of the entire space through POD dimensionality reduction and mGPR modeling. S2. Implement a control strategy based on the non-uniformity coefficient, define the temperature and humidity non-uniformity coefficient and the improvement amount, and establish a quantitative mapping relationship between the control variables and the improvement effect of temperature and humidity uniformity through variable operating condition sampling and Gaussian process regression fitting. S3. Implement an optimization method based on sensitivity analysis. Use Sobol global sensitivity analysis to screen key control variables, and use a genetic algorithm to search for the optimal parameter combination in the dimensionality-reduced variable space to achieve dimensionality reduction of high-dimensional control variables and solution of optimal parameter combination.

2. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 1, characterized in that, S1 is executed according to the following sub-steps: S1.1 Physical Model and Computational Mesh Establishment: Based on the geometric drawings and equipment layout diagrams of the target industrial space, a geometric model including the building structure, process equipment, and air conditioning and ventilation system is constructed using 3D modeling software. The geometric model covers the location and size information of walls, columns, roof, industrial production equipment, air conditioning supply and return air vents, load-bearing walls, and windows. The geometric model is imported into mesh generation software, and a hybrid meshing technique combining unstructured meshes and boundary layer meshes is used to divide the mesh. Local densification is performed on key areas around the supply and return air vents and major heat source equipment. S1.2 Boundary condition parameterization and sampling design: The core control variables are determined as the air volume, air temperature, and air humidity of each air outlet. The Latin hypercube sampling method is used to generate dozens to hundreds of uniform and low-correlation control parameter combinations in the multidimensional control variable design space. Historical environmental parameters of the target space are collected, including the air conditioning system outlet parameters and the ambient temperature and humidity of each measuring point. The initial boundary conditions for computational fluid dynamics simulation are set. The initial boundary conditions include the heat transfer coefficient of the building envelope, the initial ambient temperature and humidity, the air conditioning outlet temperature and humidity, and the air volume. S1.3 High-fidelity CFD simulation and dataset generation: The mesh model from step S1.1 and the control parameters from step S1.2 are imported into the computational fluid dynamics software. The k-ε turbulence model is selected, the energy equation and component transport equation are activated, and the steady-state solution is performed using the control equation set composed of the continuity equation, momentum equation, energy equation and component transport equation. After each set of simulations is completed, the full-field temperature and humidity data of the specified monitoring area are extracted. After all simulations are completed, a high-fidelity training dataset consisting of "input control parameters - output full-field data" is constructed. S1.4 Dimensionality reduction and feature extraction of spatial field data based on POD: The temperature field data and humidity field data obtained in step S1.3 are subjected to intrinsic orthogonal decomposition analysis. Each high-dimensional spatial field data (dimension equal to the number of grid nodes) is represented as a linear combination of several feature modes (i.e., POD modes). The feature modes are sorted from largest to smallest in terms of representation energy. The top-ranked feature modes are extracted. The high-dimensional field data of each sample is compressed into a low-dimensional amplitude vector that records the projection coefficients of the original field data on each principal mode. S1.

5. Based on mGPR input-output mapping relationship learning, the control parameter combination in step S1.2 is used as the input feature, and the low-dimensional amplitude vector of temperature field and the low-dimensional amplitude vector of humidity field corresponding to step S1.4 are used as the output targets to train two multivariate Gaussian process regression models, forming a core proxy model for quickly predicting the amplitude of low-dimensional modal coefficients from the air outlet conditions of the air conditioning system. S1.6 Rapid reconstruction of full-field environmental information: For a new set of unsimulated control parameters, the corresponding temperature and humidity field modal coefficients are predicted using the core surrogate model trained in step S1.

5. The modal coefficients are then linearly combined with the POD space principal modes extracted and stored in step S1.4 to reconstruct the full-space temperature distribution cloud map and humidity distribution cloud map.

3. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 2, characterized in that, S2 is executed according to the following sub-steps: S2.1 Definition of Baseline Scenario and Non-uniformity Coefficient: Select a typical air conditioning operating condition as the baseline scenario. Reconstruct the overall temperature and humidity of the baseline scenario using the method in step S1.6, and calculate the non-uniformity coefficient according to the following formula: in, For the temperature or humidity at each point in the space, Let n be the average temperature or average humidity at each point in the space, and n be the number of points in the space. S2.2 Variable operating condition sampling and non-uniformity coefficient calculation: Within the operating range of the control variables, the Latin hypercube sampling method (LHS) is used to generate thousands of different variable air volume and variable temperature and humidity air supply scenarios. For each scenario, the fast reconstruction method in step S1.6 is called to obtain the predicted overall temperature and humidity, and the corresponding non-uniformity coefficient is calculated according to the formula in step S2.1 to form a "control variable combination - non-uniformity coefficient" dataset. S2.3 Calculation of the improvement amount of the non-uniformity coefficient, according to the formula Calculate the improvement amount of the non-uniformity coefficient for each group of variable operating condition scenarios in step S2.2, where The non-uniformity coefficient is the coefficient for the baseline scenario. For the first i The non-uniformity coefficient of the variable operating condition scenario; a positive value indicates that the uniformity of the scenario is better than that of the benchmark scenario. S2.

4. Influence function fitting: Using the combination of control variables from step S2.2 as input and the improvement in the non-uniformity coefficient calculated in step S2.3 as output, Gaussian process regression (GPR) is used to fit the continuous response function. ,in For the air outlet conditions, including the outlet temperature, humidity, and air volume of each air conditioning system, the fitting process provides the predicted mean and variance for uncertainty quantification.

4. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 3, characterized in that, Step S3 is specifically executed as follows: S3.1 Global Sensitivity Analysis and Key Variable Screening: Based on the response function fitted in step S2.4, the Sobol global sensitivity analysis method is used to calculate the first-order sensitivity index of each control variable through Monte Carlo sampling. and total effect sensitivity index Select the overall effect sensitivity index The top k variables with the largest values ​​that exceed a set threshold are selected as key optimization variables. S3.2 Intelligent optimization solution: With the goal of maximizing the improvement of the non-uniformity coefficient, the key optimization variables selected in step S3.1 are the optimization objects, and the response function fitted in step S2.4 is the objective function. Under the constraints of the air conditioning system that meet the upper and lower limits of total air volume, the range of air volume of a single air outlet, and the limit of supply air temperature, the genetic algorithm is used to search for the optimal combination of control parameters in the dimensionality-reduced key variable space.

5. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 4, characterized in that: In step S1.1, the building structure includes walls, columns, and roof; the air conditioning and ventilation system explicitly includes the location and size parameters of the air supply and return air vents; and the industrial production equipment includes its specific spatial layout information.

6. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 5, characterized in that: In step S1.3, the computational fluid dynamics software is ANSYS Fluent, the solution method is the Coupled pressure-velocity coupling method, the momentum equation and turbulent kinetic energy adopt the first-order upwind scheme, the energy equation and turbulent dissipation rate adopt the second-order upwind scheme, and water vapor is used as an additional component to simulate the humidity field.

7. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 6, characterized in that: In step S1.5, the multivariate Gaussian process regression model is trained using MATLAB's Statistics and Machine Learning Toolbox. The quadratic exponential kernel function is selected, and the training dataset is divided into a training set and a test set in a 4:1 ratio. The training set is used to train the model's hyperparameters, and the test set is used to verify the model's accuracy. The average relative error of the temperature field and humidity field reconstruction is less than 2%.

8. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 7, characterized in that: In step S3.1, at least 10,000 samples are taken in the input space using the Saltelli sampling sequence, and the corresponding improvement in the non-uniformity coefficient is calculated using the GPR model trained in step S2.4, thereby accurately estimating the sensitivity index.

9. The method for optimizing temperature and humidity uniformity in large industrial spaces based on the POD-mGPR surrogate model according to claim 8, characterized in that: In step S3.2, the parameters of the genetic algorithm are set as follows: population size 50, maximum number of iterations 100, crossover probability 0.8, mutation probability 0.

1. After optimization, the optimal combination of control parameters is input into the core surrogate model constructed in S1 to verify the degree of uniformity improvement.

Images