Rapid simulation prediction method and system for flowing heat exchange of heat exchanger based on deep network
By using a deep network-based full-process simulation of heat exchangers and a deep neural network collaborative architecture, the accuracy-efficiency problem in heat exchanger simulation is solved, enabling fast and accurate prediction of overall flow heat transfer, applicable to multi-structure and multi-working-medium scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI JIDING INFORMATION TECH CO LTD
- Filing Date
- 2026-01-14
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies suffer from the problem of balancing accuracy and efficiency in whole-machine flow heat transfer simulation of heat exchangers. Traditional physical simulation methods have long calculation cycles and large deviations in results, while artificial intelligence-assisted simulation models lack whole-machine process data and have insufficient adaptability, resulting in inaccurate prediction results.
A deep network-based simulation method is adopted. Training data is generated by constructing a full-process simulation of the heat exchanger. By combining a collaborative architecture of deep neural network and generative adversarial neural network, a nonlinear reduced-order model is constructed to achieve fast and accurate simulation.
While ensuring accuracy, simulation efficiency is greatly improved, reducing the simulation cycle from several days or weeks to seconds, ensuring that the prediction results are highly consistent with the actual performance, and adapting to multiple structure and multiple working medium scenarios.
Smart Images
Figure CN121936281A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of heat exchanger simulation and artificial intelligence, and in particular to a rapid simulation and prediction method and system for overall flow heat transfer in heat exchangers based on deep neural networks. Background Technology
[0002] Driven by the "dual carbon" goal, advanced energy systems such as supercritical CO2 Brayton cycle, vehicle-mounted nuclear power, and LNG gasification are developing rapidly. Heat exchangers such as printed circuit board heat exchangers (PCHEs) have become the core heat exchange equipment in these systems due to their high compactness, high temperature and high pressure resistance, and strong adaptability to working fluids with changing properties. The accuracy and efficiency of the whole-machine flow heat exchange simulation directly determine the reliability of the system design and the economy of operation.
[0003] Currently, the core contradiction in whole-machine flow heat transfer simulation technology for heat exchangers is the difficulty in balancing accuracy and efficiency, which manifests in the following two aspects:
[0004] I. Traditional physical simulation methods have inherent limitations. Full-scale, high-resolution computational fluid dynamics modeling requires reproducing the scale coupling between "meter-scale whole machine and millimeter-scale microchannels," "uneven flow distribution in manifold-microchannel systems," and "fluid-solid conjugate heat transfer," among other complex physical effects. This results in meshes numbering in the hundreds of millions and computation cycles lasting from days to weeks, making it extremely inefficient for design optimization and system-level simulations requiring rapid iteration. Oversimplified models, in an effort to improve computational efficiency, must ignore the aforementioned key coupling effects, leading to significant deviations in the prediction of overall heat exchange efficiency and pressure drop, failing to accurately reflect the actual operating state of the heat exchanger.
[0005] II. Significant shortcomings exist in AI-assisted heat exchanger simulation. First, the quality of training samples is insufficient. Existing research often uses simplified single-channel simulation models to generate training data, neglecting the key physical processes that determine the overall performance of the heat exchanger. This results in the trained AI models essentially only learning the physical laws of local units, leading to a fundamental and systematic deviation between the predicted results and the actual performance of the entire system. Second, the model architecture is limited. A single neural network cannot simultaneously capture the "macroscopic performance parameters" and "local flow field details" of the heat exchanger, making it unsuitable for complex scenarios involving multiple heat exchanger structures (straight channels, Z-shaped channels, airfoil channels) and multiple working fluids (supercritical CO2, methane, propane). Finally, the sample generation method is imperfect, failing to employ a full-process simulation method for the entire heat exchanger, resulting in training data that does not match actual engineering scenarios.
[0006] Therefore, the specific technical problem solved by the present invention is as follows:
[0007] I. To resolve the inherent contradiction between "accuracy and efficiency" in traditional heat exchanger whole-machine simulation, namely, to significantly improve simulation efficiency while ensuring the reproduction of complex physical effects such as "scale coupling between meter-level whole machine and millimeter-level microchannel", "uneven flow distribution in manifold-microchannel system" and "fluid-solid conjugate heat transfer", and at the same time avoid the accuracy defects of significant deviation in whole-machine heat exchange efficiency and pressure drop prediction caused by oversimplification of the model.
[0008] Second, address the issue of insufficient AI training sample quality by generating training data through a full-process simulation method for the entire heat exchanger. This compensates for the deficiencies in existing single-channel local simplified models, such as missing samples, "manifold flow deviation," and "global temperature field coupling," which are key physical processes of the entire machine. It also eliminates the root cause and systemic deviation between the prediction results of the AI model and the actual performance of the entire machine caused by the AI model only learning the physical laws of local units.
[0009] Third, improve the adaptability of AI model architecture, that is, design a deep neural network architecture that can simultaneously capture the macroscopic performance parameters of heat exchangers (heat exchange efficiency, total pressure drop, etc.) and local flow field details (microchannel vortex, near-wall temperature gradient, etc.) to solve the problem that a single neural network cannot adapt to complex scenarios with multiple heat exchanger structures (straight channel, Z-shaped channel, airfoil channel) and multiple working fluids (supercritical CO2, methane, propane).
[0010] Fourth, improve the sample generation method of AI. Through the whole process simulation of heat exchanger in the technical roadmap of "processing the working fluid properties of cold and hot channels → setting up physical models → calculating flow distribution → calculating the flow and heat transfer of the whole machine → iterative convergence", solve the problem that the existing sample generation method does not match the actual engineering scenario and ensure the engineering practicality of the training data. Summary of the Invention
[0011] This invention addresses the problems and shortcomings of existing technologies by providing a rapid simulation and prediction method and system for heat exchanger flow heat transfer based on deep networks.
[0012] The present invention solves the above-mentioned technical problems through the following technical solution:
[0013] This invention provides a rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks, characterized by comprising the following steps:
[0014] I. Constructing sample data reflecting the flow heat transfer phenomena in heat exchangers
[0015] S1. Set multiple sets of discretized boundary conditions for the heat exchanger. Each set of discretized boundary conditions includes the working fluid property parameters of the cold and hot channels of the heat exchanger and the parameters of the selected physical model.
[0016] S2. The heat exchanger is simplified using a flow resistance model. The flow distribution calculation is performed for each group of heat exchangers to obtain the corresponding flow distribution calculation results for each group.
[0017] S3. Using the flow distribution calculation results of each group as input, the multi-scale method is used to realize the flow heat transfer calculation of the entire heat exchanger for each group.
[0018] S4. If the overall flow heat transfer calculation results of the corresponding group do not converge, then use them to correct the working fluid properties and use the corrected parameters in S2 to perform iterative calculations until convergence, and obtain the actual flow heat transfer data of the heat exchanger corresponding to the group. The flow heat transfer data includes velocity field, pressure field and temperature field data. Multiple sets of discretized boundary conditions and corresponding actual flow heat transfer data constitute sample data.
[0019] II. Constructing the training set, validation set, and test set
[0020] S5. Perform matrixing, serialization, and normalization on each group of data in the sample data;
[0021] S6. The processed sample data is divided into training set, validation set and test set according to the proportion;
[0022] III. Constructing a Nonlinear Order Reduction Model
[0023] S7. Utilize the model training function of the training set, the model validation function of the validation set, and the model testing function of the test set to build and deploy a nonlinear reduced-order model based on a deep neural network.
[0024] IV. Rapid Simulation and Prediction of Heat Transfer in Heat Exchangers
[0025] S8. Input new discretized boundary conditions into the nonlinear reduced-order model for simulation prediction, and quickly output the simulation prediction results of heat exchanger flow heat transfer.
[0026] This invention also provides a rapid simulation and prediction system for heat exchanger flow heat transfer based on deep networks. Its features include a sample construction module for setting multiple sets of discretized boundary conditions for the heat exchanger. Each set of discretized boundary conditions includes the working fluid properties of the cold and hot channels of the heat exchanger and the parameters of the selected physical model. The heat exchanger is simplified using a flow resistance model. Flow distribution calculations are performed for each set to obtain the corresponding flow distribution calculation results. Using the flow distribution calculation results for each set as input, a multi-scale method is used to perform overall flow heat transfer calculations for each set of heat exchangers. If the overall flow heat transfer calculation results for that set do not converge, the working fluid properties are corrected, and the corrected parameters are used for iterative calculations until convergence, obtaining the actual overall flow heat transfer data for that set of heat exchangers. The flow heat transfer data includes velocity field, pressure field, and temperature field data. Multiple sets of discretized boundary conditions and corresponding actual flow heat transfer data constitute sample data.
[0027] The data matrix processing module is used to perform matrixing, serialization, and normalization on each group of data in the sample data. The processed sample data is divided into training set, validation set, and test set according to the proportion.
[0028] The reduced-order model building module is used to build and deploy a nonlinear reduced-order model based on a deep neural network by utilizing the model training function of the training set, the model validation function of the validation set, and the model testing function of the test set.
[0029] The rapid simulation prediction module is used to input new discretized boundary conditions into a nonlinear reduced-order model for simulation prediction, and quickly output the simulation prediction results of heat exchanger flow and heat transfer.
[0030] The present invention also provides an electronic device, characterized in that it includes:
[0031] processor;
[0032] Memory used to store processor-executable instructions;
[0033] The processor is configured to invoke instructions stored in memory to execute the rapid simulation and prediction method for heat exchanger flow heat transfer as described in any one of claims 1-8.
[0034] The present invention also provides a computer-readable storage medium storing computer program instructions thereon, characterized in that, when the computer program instructions are executed by a processor, they implement the rapid simulation and prediction method for heat exchanger flow heat transfer as described in any one of claims 1-8.
[0035] The positive and progressive effects of this invention are as follows:
[0036] 1. Fundamentally resolves the "accuracy-efficiency" contradiction: By combining whole-machine physical simulation of the heat exchanger with deep neural networks, while maintaining high accuracy (completely reproducing key physical effects such as manifold flow deviation, global temperature field coupling, and fluid-solid conjugate heat transfer), the nonlinear reduced-order model obtained from the trained deep neural network is used to quickly predict flow heat transfer data. This compresses the simulation cycle from the traditional CFD range of "days to weeks" to "seconds," achieving an order-of-magnitude improvement in efficiency and completely breaking through the efficiency bottleneck of traditional simulation methods in engineering design iterations.
[0037] 2. Ensuring the engineering reliability of AI predictions from the outset: An innovative full-process simulation method for the entire heat exchanger is used as the basis for generating AI training samples, which contain complete physical information of the entire system. This fundamentally eliminates the "systematic and root-cause prediction bias" caused by using locally simplified model samples, ensuring that the AI model's prediction results highly match the actual performance of the entire system, greatly improving its credibility in directly guiding engineering design.
[0038] 3. Architectural Innovation: The core computing layer adopts a collaborative architecture that integrates deep neural networks (Deep NNs) with deep feedforward neural networks (DFNN) and generative adversarial neural networks (GAN) to construct a nonlinear reduced-order model (ROM) for rapid inference of new discretized boundary conditions and output simulation prediction results of heat exchanger flow heat transfer. This significantly reduces computational complexity and improves simulation efficiency to the second level. Attached Figure Description
[0039] Figure 1 and Figure 2 This is a flowchart of a preferred embodiment of the present invention, which describes a rapid simulation and prediction method for heat exchanger flow heat transfer based on deep networks.
[0040] Figure 3 This is a block diagram of a rapid simulation and prediction system for heat exchanger flow heat transfer based on deep networks, which is a preferred embodiment of the present invention.
[0041] Figure 4 This is a schematic diagram of a rectangular channel according to a preferred embodiment of the present invention.
[0042] Figure 5 This is a schematic diagram of the mesh division result of a preferred embodiment of the present invention.
[0043] Figure 6 This is a boundary condition setting diagram for a preferred embodiment of the present invention.
[0044] Figure 7 The convergence criterion reference diagram is a preferred embodiment of the present invention.
[0045] Figure 8 This is a residual curve conformity report graph of a preferred embodiment of the present invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0047] For ease of description, only the parts relevant to the present invention are shown in the accompanying drawings. The terms "first," "second," etc., used in this invention are merely for the convenience of describing the technical solutions of the invention and do not have a specific limiting effect; they are all general references and do not constitute a limitation on the technical solutions of the present invention.
[0048] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a rapid simulation and prediction method for heat exchanger flow heat transfer based on deep networks. Through a full-link process of "constructing high-quality training data → training and inference," it achieves rapid and accurate simulation of the overall flow heat transfer of the heat exchanger. The technical framework follows a logical closed-loop framework of "sample extraction—data preparation—model training—rapid prediction." Specifically, it includes the following steps:
[0049] I. Constructing sample data reflecting the flow heat transfer phenomena in heat exchangers
[0050] To achieve accurate predictions from the AI model, the primary task is to generate training data that fully reflects the complex physical phenomena of the entire heat exchanger. This embodiment employs a rigorously designed full-process simulation method for the entire heat exchanger: "processing of working fluid properties and setting of physical models in cold and hot channels → flow distribution calculation → overall flow heat transfer calculation → iterative convergence"—a complete simulation logic.
[0051] Step 101: Set multiple sets of discretized boundary conditions for the heat exchanger. Each set of discretized boundary conditions (DBCs) includes the working fluid property parameters of the cold and hot channels of the heat exchanger and the parameters of the selected physical model.
[0052] Step 102: The heat exchanger is simplified using a flow resistance model. The flow distribution calculation is performed for each group of heat exchangers to obtain the corresponding flow distribution calculation results for each group.
[0053] In this step, the heat exchanger is simplified using a flow resistance model: the cold and hot channels of the heat exchanger are divided into cold and hot microchannels in the middle section and cold and hot flow channel systems on both sides. Cold and hot microchannel mesh models are constructed separately for the cold and hot microchannels, and cold and hot flow channel system mesh models are constructed separately for the cold and hot flow channel systems. For the cold and hot microchannel mesh models, they are divided along the length direction to obtain multiple one-to-one corresponding cold and hot treatment units. One corresponding cold treatment unit and one hot treatment unit constitute a heat exchange unit.
[0054] For each group, the flow distribution calculation of the heat exchanger is performed to obtain the corresponding flow distribution calculation results: For each group, the latest cold and hot channel working fluid property parameters, physical model parameters and flow resistance model are used to simulate the cold and hot microchannel mesh model respectively. At the same time, the latest cold and hot channel working fluid property parameters and physical model parameters are used to simulate the cold and hot flow channel system mesh model to obtain the flow distribution calculation results of the cold and hot microchannels of the heat exchanger for each group.
[0055] Flow resistance model: Establish a channel mesh model representing cold and hot microchannels, set operating conditions, and construct a flow resistance model based on the relationship between the channel's flow resistance characteristics and the inlet flow rate.
[0056] ;in, This indicates the pressure drop in the channel. This represents the quadratic coefficient of the resistance term. This represents the first-order coefficient of the resistance term. This indicates the density of the fluid in the channel. This indicates the velocity of the fluid in the channel; it is obtained through multiple experiments on the channel. and The value of .
[0057] Step 103: Using the flow distribution calculation results corresponding to each group as input, the multi-scale method is used to realize the flow heat transfer calculation of the entire heat exchanger for each group.
[0058] In this step, the multi-scale method includes a system-scale model and a three-dimensional local model. The system-scale model is used to solve the overall macroscopic flow and heat transfer parameters of the heat exchanger, including a flow resistance model for macroscopic flow and a one-dimensional heat transfer model for axial temperature distribution. The three-dimensional local model is used to solve the detailed three-dimensional flow heat transfer field at the cold and hot microchannels in key regions.
[0059] Furthermore, a mesh-like three-dimensional local model with cross-sectional features is constructed for each heat exchange unit, and a one-dimensional heat transfer model of the axial temperature distribution of the heat exchanger applicable to each heat exchange unit is constructed. Based on the inlet temperatures of the cold and hot microchannels, a temperature condition table containing the inlet temperatures of each cold and hot treatment unit is established. Based on the flow distribution calculation results of the cold and hot microchannels and the temperature condition table, simulation calculations are performed on each three-dimensional local model to obtain the outlet temperatures and heat transfer of each heat exchange unit's cold and hot treatment units. Then, the heat transfer of each heat exchange unit is substituted into the one-dimensional heat transfer model for solution to obtain the axial temperature distribution of the heat exchanger as temperature field data.
[0060] One-dimensional heat transfer model:
[0061]
[0062] In the formula, and These represent the specific heat capacities of the fluids in the heat and cold treatment units, respectively. and These represent the temperature changes in the heat and cold processing units, respectively. The heat exchange capacity of each heat exchange unit is input as known quantities in tabular form, and the result is obtained by solving the above formula. and This allows us to obtain the axial temperature distribution of the heat exchanger as temperature field data.
[0063] Step 104: If the overall flow heat transfer calculation results for the corresponding group do not converge, then use them to correct the working fluid properties and use the corrected parameters in step 102 to perform iterative calculations until convergence, thereby obtaining the actual flow heat transfer data of the heat exchanger for the corresponding group. The flow heat transfer data includes velocity field, pressure field and temperature field data. Multiple sets of discretized boundary conditions and corresponding actual flow heat transfer data constitute sample data.
[0064] II. Constructing the training set, validation set, and test set
[0065] Step 105: Perform matrixing, serialization, and normalization on each group of data in the sample data.
[0066] Step 106: The processed sample data is divided into training set, validation set and test set according to the proportion.
[0067] III. Constructing a Nonlinear Order Reduction Model (ROM)
[0068] Step 107: Utilize the model training function of the training set, the model validation function of the validation set, and the model testing function of the test set to construct and deploy a physical constraint-enhanced nonlinear reduced-order model based on a deep neural network.
[0069] Step 107 specifically includes:
[0070] Step 1071: Construct a Deep Feedforward Neural Network (DFNN) as a macroscopic global mapping network and a Generative Adversarial Neural Network (GAN) as a local detail enhancement network. The DFNN and GAN form a collaborative architecture. The DFNN is used to establish the main nonlinear relationship from the discretized boundary conditions to the basic distribution of the overall macroscopic performance parameters and the physical field of the whole machine, and outputs flow heat transfer data with global rationality, providing a stable foundation for local detail enhancement. The GAN is used to take the flow heat transfer data output by the DFNN and the discretized boundary conditions as input conditions, and focuses on reproducing local complex physical phenomena such as manifold flow deviation, microchannel vortex, and near-wall temperature gradient, making up for the DFNN's deficiency in capturing microscopic details.
[0071] A deep feedforward neural network consists of an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer corresponds to the dimension of the discretized boundary conditions, and the number of nodes in the output layer corresponds to the total dimension of the discretized physical quantities (including velocity, pressure, and temperature). The number of hidden layers and the number of neurons in each layer are adjustable hyperparameters.
[0072] The generative adversarial neural network (GAN) consists of a generator G and a discriminator D. The generator G adopts a deconvolutional neural network structure, and its input is a low-dimensional random vector + flow heat transfer data output by a deep feedforward neural network + discretized boundary conditions. The output is high-resolution local velocity field, pressure field, and temperature field details. The discriminator D adopts a convolutional neural network structure, and its input is velocity field, pressure field, and temperature field data + discretized boundary conditions. The velocity field, pressure field, and temperature field data are real flow heat transfer data or data generated by the generator G. The output is a scalar value of 0-1, where 1 indicates that it is judged as real data and 0 indicates that it is judged as generated data. At the same time, a physical feature extraction layer is embedded to verify the physical rationality of the velocity field, pressure field, and temperature field.
[0073] Step 1072: Train the collaborative architecture of the deep feedforward neural network and the generative adversarial neural network using the training set, validate the collaborative architecture using the validation set, and test the collaborative architecture using the test set until the prediction accuracy of the collaborative architecture on the test set meets the preset convergence criterion. Then the deep feedforward neural network and the generative adversarial neural network are trained and solidified into an independent and stable nonlinear reduced-order model.
[0074] The deep feedforward neural network uses the backpropagation algorithm and the Adam optimizer to perform iterative training until convergence, thereby determining the optimal weights and bias parameters of the network.
[0075] By using an adversarial training mechanism, the generator G and the discriminator D are trained alternately and iteratively until convergence, and it is determined that the generator G can output high-fidelity local velocity field, pressure field and temperature field details.
[0076] IV. Rapid Simulation and Prediction of Heat Transfer in Heat Exchangers
[0077] Step 108: Input the new discretized boundary conditions into the nonlinear reduced-order model for simulation prediction, and quickly output the simulation prediction results of heat exchanger flow heat transfer.
[0078] The nonlinear order reduction model can be encapsulated in a fast simulation prediction module, and a dedicated model calling interface can be provided in the software's graphical user interface. After the user inputs new discretized boundary conditions (New DBCs) through the interface, the module automatically calls the nonlinear order reduction model to perform calculations. The nonlinear order reduction model completes the calculations in seconds and outputs the predicted results of velocity, pressure and temperature across the entire field.
[0079] like Figure 3 As shown, this embodiment of the invention also provides a rapid simulation and prediction system for heat exchanger flow heat transfer based on deep networks, including a sample construction module 1, a data matrix processing module 2, a reduced-order model construction module 3, and a rapid simulation and prediction module 4.
[0080] Sample construction module 1 is used to set multiple sets of discretized boundary conditions for the heat exchanger. Each set of discretized boundary conditions includes the working fluid property parameters of the cold and hot channels of the heat exchanger and the parameters of the selected physical model. The heat exchanger is simplified using a flow resistance model. For each set, the flow distribution calculation of the heat exchanger is performed to obtain the corresponding flow distribution calculation results. Using the flow distribution calculation results of each set as input, a multi-scale method is used to realize the overall flow heat transfer calculation of the corresponding heat exchanger. If the overall flow heat transfer calculation results of the corresponding set do not converge, the working fluid property parameters are corrected and the corrected parameters are used for iterative calculation until convergence, to obtain the actual flow heat transfer data of the corresponding heat exchanger. The flow heat transfer data includes velocity field, pressure field and temperature field data. Multiple sets of discretized boundary conditions and corresponding actual flow heat transfer data constitute the sample data.
[0081] The data matrix processing module 2 is used to perform matrixing, serialization and normalization on each group of data in the sample data. The processed sample data is divided into training set, validation set and test set according to the proportion.
[0082] The reduced-order model building module 3 is used to build and deploy a physically constrained, nonlinear reduced-order model based on a deep neural network by utilizing the model training function of the training set, the model validation function of the validation set, and the model testing function of the test set.
[0083] The rapid simulation prediction module 4 is used to input new discretized boundary conditions into the nonlinear reduced-order model for simulation prediction and quickly output the simulation prediction results of heat exchanger flow heat transfer.
[0084] This invention also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the aforementioned method.
[0085] This invention also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the aforementioned method.
[0086] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.
[0087] In this invention, a high-quality AI training sample generation system is constructed with the whole process simulation of the heat exchanger as the core. The simulation steps of "processing the working fluid properties and setting the physical model in the cold and hot channels → calculating the flow distribution → calculating the flow heat transfer of the whole machine → iterative convergence" are adopted to ensure that the AI training samples come from the real flow heat transfer of the whole heat exchanger and eliminate systematic bias caused by sample defects.
[0088] In this invention, a design is based on a deep neural network (Deep NNs) that integrates a deep feedforward neural network (DFNN) and a generative adversarial neural network (GAN). This design is used to train the model on data and boundary conditions obtained from the simulation of the entire heat exchanger, resulting in a nonlinear reduced-order model (ROM). This model can simultaneously and accurately capture the macroscopic performance parameters and local flow field details of the heat exchanger, and has the ability to adapt to multiple structures and working fluid scenarios.
[0089] In this invention, through efficient computation, the simulation cycle is compressed from several days to seconds while ensuring the reduction of the complex physical effects of the heat exchanger as a whole. This completely breaks through the efficiency bottleneck of traditional simulation and meets the high-efficiency requirements of engineering design iteration and system-level simulation.
[0090] The key point of this invention is:
[0091] 1. Ensure the authenticity of the generated samples: The training samples are obtained by using a full-process simulation method of the heat exchanger. This ensures that the training sample data includes key physical processes of the heat exchanger, such as "manifold flow deviation, global temperature field coupling, and fluid-solid conjugate heat transfer", thus solving the root cause of the deviation caused by local simplification of samples in the existing technology.
[0092] 2. Architectural Innovation: The core computing layer adopts a collaborative architecture that integrates deep neural networks (Deep NNs) with deep feedforward neural networks (DFNN) and generative adversarial neural networks (GAN) to construct a nonlinear reduced-order model (ROM) for rapid inference of new discretized boundary conditions and output simulation prediction results of heat exchanger flow heat transfer. This significantly reduces computational complexity and improves simulation efficiency to the second level.
[0093] 3. Multi-scenario adaptation and efficiency breakthrough: Through the combination of architecture design and ROM, it can not only adapt to complex scenarios of heat exchangers with multiple structures (straight channel, Z-shaped channel, airfoil channel) and multiple working fluids (supercritical CO2, methane, propane), but also completely solve the efficiency pain point of traditional CFD (computational fluid dynamics) of "hundreds of millions of grids and several days of calculation cycle".
[0094] This embodiment uses a typical rectangular channel PCHE as an example, and the specific implementation steps are as follows:
[0095] (a) Definition of rectangular channel parameters
[0096] 1. Geometric parameters and model establishment
[0097] The rectangular channel dimensions are: length 1000mm, width 3mm, thickness 60mm (see...) Figure 4 and Figure 5 ).
[0098] 2. Physical Model
[0099] The main content of this calculation is the flow heat transfer and conjugate heat transfer process in the heat exchange channel. The physical model to be considered is as follows:
[0100] Pressure-velocity coupling algorithm: The SIMPLEC algorithm is adopted.
[0101] Turbulence model: Turbulence model.
[0102] 3. Working fluid physical properties
[0103] The physical properties of water at 15 MPa can be obtained by consulting the NIST property database:
[0104] Table 1-15 MPa water properties
[0105]
[0106] The NIST query results are shown below:
[0107] Table 2 - NIST Database (Partial)
[0108]
[0109] Since its physical properties change little, this embodiment uses normal physical properties for processing.
[0110] 4. Boundary conditions
[0111] This embodiment sets up a total of 5 verification conditions.
[0112] Fluid pressure: 15MPa, inlet temperature 180℃, inlet mass flow rate 2000kg / m² / s, 1500kg / m² / s, heating power 40kW, 20kW. See the table below for specific operating conditions:
[0113] Table 3 - Boundary Conditions under Different Operating Conditions
[0114]
[0115] The inlet is set as a velocity inlet, the outlet as a pressure outlet, the upper and lower walls as heat flux walls, and the remaining boundaries as adiabatic walls. See [link / details]. Figure 6 Boundary condition settings.
[0116] (ii) Training sample generation
[0117] 1. CFD simulation and data extraction: Complete steady-state simulation of five operating conditions of the PCHE machine, and then extract velocity field and temperature field data using ReadVTM.py and ReadVTR.py tools; extract sample number sequence using idxList.py and sample parameter sequence using paralnList.py.
[0118] 2. Data matrix processing: The PSP_VTK2HDF.py tool reads VTK format flow heat transfer data (including velocity field, pressure field, and temperature field data) from 5 operating conditions, converts them into HDF5 format matrices, and generates a complete sample database containing the "boundary condition-matrix" mapping relationship.
[0119] (III) Deep Neural Network Training
[0120] 1. Data preprocessing: The 5 sets of samples were divided into training set (3 sets), validation set (1 set), and test set (1 set), and the data scale was normalized.
[0121] 2. DFNN Training: Construct an input layer (3 nodes: inlet temperature, flow rate, and heating heat flux), 3 hidden layers (using the ReLU activation function), and an output layer (corresponding to the temperature and velocity values of discrete nodes); use the Adam optimizer and iteratively train until the loss on the test set converges (convergence criteria are described in the report). Figure 7 "Fast simulation training convergence residuals" means that the loss value tends to be stable and has no obvious fluctuations.
[0122] 3. GAN Training: The generator uses a deconvolutional network (TCNN), and the discriminator uses a convolutional network (CNN). Training is performed alternately until the generated flow field (velocity field, pressure field) is highly consistent with the statistical distribution of real CFD data (judgment criterion: residual curve conformance report). Figure 8 (Convergence trend of "CFD calculated residual curve").
[0123] 4. Model solidification: The trained DFNN+GAN architecture and weight parameters are serialized to generate a ROM file, which is then packaged and named "Fast Simulation Prediction Module".
[0124] (iv) Simulation verification and error analysis
[0125] 1. Fast simulation call: By inputting 5 sets of working condition boundary conditions, the ROM model is called to complete the fast calculation. The calculation time for each set of working conditions is ≤1 second.
[0126] 2. Result Comparison
[0127] The results of the rapid simulation were compared with those of traditional 3D CFD, and the specific discrepancies are as follows:
[0128]
[0129] It should be noted that the CFD sample data used in this embodiment to train the nonlinear reduced-order model (ROM) all contain complete flow and heat transfer data (i.e., including velocity, pressure, and temperature fields). Therefore, the trained ROM model can directly output complete flow and heat transfer data under new operating conditions. The table above only lists the two most representative physical quantities, velocity and temperature, for quantitative comparison. Since the pressure and velocity fields are tightly coupled through the fluid control equations, the high accuracy of the velocity field prediction (maximum deviation -7.01%) indicates from a physical mechanism perspective that the model's prediction of the pressure field also has reliable engineering accuracy. This ROM model can be directly used to output key macroscopic performance parameters such as overall system pressure drop.
[0130] 3. Test Conclusion:
[0131] Accuracy requirements: The maximum deviations of temperature and speed in the five operating conditions were 1.55% and -7.01%, respectively, both meeting the accuracy requirement of "not exceeding 10%". Furthermore, based on the predicted accuracy of the velocity field and the completeness of the model training, it can be confirmed that this method also possesses practical engineering accuracy for predicting the pressure field and overall pressure drop.
[0132] Efficiency requirements: Fast simulation performs forward propagation calculations based on a trained nonlinear reduced-order model (ROM), eliminating the need to repeatedly solve the three-dimensional Navier-Stokes equations and complex turbulence models. Traditional CFD requires time-consuming steps such as mesh discretization and iterative equation solving, while fast simulation only requires milliseconds to seconds of computation, significantly improving computational efficiency compared to traditional three-dimensional CFD.
[0133] It should be noted that the above embodiments primarily describe preferred implementations of the collaborative operation of deep feedforward neural networks (DFNNs) and generative adversarial neural networks (GANs). In practical applications, the construction of the nonlinear order reduction model of this invention is not limited to the above collaborative architecture. Depending on different simulation accuracy, efficiency, and complexity requirements, any of the following implementation methods can also be adopted:
[0134] (I) Implementation of Independent Deep Feedforward Neural Network: In step S7, only the deep feedforward neural network is constructed. The network is trained using the training set defined in step S6, with the aforementioned discretized boundary conditions as input and the corresponding real flow heat transfer data as output labels. Hyperparameter tuning and model selection are performed using the validation set defined in step S6. The model accuracy is evaluated using the test set defined in step S6. Once the model's prediction accuracy meets the preset convergence criterion, the trained deep feedforward neural network is solidified into an independent nonlinear reduced-order model. This implementation method is simple in structure, stable in training, and suitable for rapid prediction of the macroscopic flow heat transfer performance of heat exchangers.
[0135] (II) Independent Generative Adversarial Neural Network Implementation: In step S7, only the generative adversarial neural network is constructed. Using the dataset divided in step S6, the network learns the mapping from discretized boundary conditions to high-fidelity flow heat transfer data through adversarial training between its internal generator and discriminator; and finally, its generator is solidified into an independent nonlinear reduced-order model. This implementation is suitable for scenarios with extremely high requirements for the reproduction of local flow heat transfer details.
[0136] (III) Collaborative architecture implementation: As described above in the preferred embodiment, local detail enhancement is achieved by working collaboratively with deep feedforward neural networks and generative adversarial neural networks while ensuring global physical consistency.
[0137] All of the above embodiments are based on the sample data construction, processing and partitioning process described in this invention (steps S1-S6), and achieve rapid simulation prediction by calling the constructed nonlinear reduced-order model (step S8).
[0138] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of the present invention is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, and all such changes and modifications fall within the scope of protection of the present invention.
Claims
1. A rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks, characterized in that, It includes the following steps: I. Constructing sample data reflecting the flow heat transfer phenomena in heat exchangers S1. Set multiple sets of discretized boundary conditions for the heat exchanger. Each set of discretized boundary conditions includes the working fluid property parameters of the cold and hot channels of the heat exchanger and the parameters of the selected physical model. S2. The heat exchanger is simplified using a flow resistance model. The flow distribution calculation is performed for each group of heat exchangers to obtain the corresponding flow distribution calculation results for each group. S3. Using the flow distribution calculation results of each group as input, the multi-scale method is used to realize the flow heat transfer calculation of the entire heat exchanger for each group. S4. If the overall flow heat transfer calculation results of the corresponding group do not converge, then use them to correct the working fluid properties and use the corrected parameters in S2 to perform iterative calculations until convergence, and obtain the actual flow heat transfer data of the heat exchanger corresponding to the group. The flow heat transfer data includes velocity field, pressure field and temperature field data. Multiple sets of discretized boundary conditions and corresponding actual flow heat transfer data constitute sample data. II. Constructing the training set, validation set, and test set S5. Perform matrixing, serialization, and normalization on each group of data in the sample data; S6. The processed sample data is divided into training set, validation set and test set according to the proportion; III. Constructing a Nonlinear Order Reduction Model S7. Utilize the model training function of the training set, the model validation function of the validation set, and the model testing function of the test set to build and deploy a nonlinear reduced-order model based on a deep neural network. IV. Rapid Simulation and Prediction of Heat Transfer Flow in Heat Exchangers S8. Input new discretized boundary conditions into the nonlinear reduced-order model for simulation prediction, and quickly output the simulation prediction results of heat exchanger flow heat transfer.
2. The rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks as described in claim 1, characterized in that, S7 specifically includes: S71. Construct a deep feedforward neural network as a macro-global mapping network and a generative adversarial neural network as a local detail enhancement network. The deep feedforward neural network and the generative adversarial neural network form a collaborative architecture. The deep feedforward neural network is used to establish the main nonlinear relationship from the discretized boundary conditions to the basic distribution of the overall macroscopic performance parameters and the physical field of the whole field, and outputs flow heat transfer data with global rationality. The generative adversarial neural network is used to take the flow heat transfer data output by the deep feedforward neural network and the discretized boundary conditions as input conditions, and focuses on reproducing local complex physical phenomena. S72. Train the collaborative architecture of the deep feedforward neural network and the generative adversarial neural network using the training set, validate the collaborative architecture using the validation set, and test the collaborative architecture using the test set until the prediction accuracy of the collaborative architecture on the test set meets the preset convergence criterion. Then the deep feedforward neural network and the generative adversarial neural network are trained and solidified into an independent and stable nonlinear reduced-order model.
3. The rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks as described in claim 2, characterized in that, In S71, the deep feedforward neural network contains an input layer, multiple hidden layers and an output layer. The number of nodes in the input layer corresponds to the dimension of the discretized boundary conditions, the number of nodes in the output layer corresponds to the total dimension of the discretized physical quantities of the entire field, and the number of hidden layers and the number of neurons in each layer are adjustable hyperparameters. In S72, the deep feedforward neural network uses the backpropagation algorithm and the Adam optimizer to perform iterative training until convergence, thereby determining the optimal weights and bias parameters of the network.
4. The rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks as described in claim 2, characterized in that, In S71, the generative adversarial neural network consists of a generator G and a discriminator D. The generator G adopts a deconvolutional neural network structure. The input is a low-dimensional random vector + flow and heat transfer data output by a deep feedforward neural network + discretized boundary conditions. The output is high-resolution local velocity field, pressure field and temperature field details. The discriminator D adopts a convolutional neural network structure. The input is velocity field, pressure field and temperature field data + discretized boundary conditions. The velocity field, pressure field and temperature field data are real flow heat transfer data or data generated by generator G. The output is a scalar value of 0-1, where 1 indicates that it is judged as real data and 0 indicates that it is judged as generated data. At the same time, a physical feature extraction layer is embedded to verify the physical rationality of the velocity field, pressure field and temperature field. In S72, the generator G and discriminator D are trained alternately and iteratively until convergence using an adversarial training mechanism, which determines that the generator G can output high-fidelity local velocity field, pressure field and temperature field details.
5. The rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks as described in claim 1, characterized in that, In S2, the heat exchanger is simplified using a flow resistance model: the cold and hot channels are divided into cold and hot microchannels in the middle section and cold and hot flow channel systems on both sides. Cold and hot microchannel mesh models are constructed separately for the cold and hot microchannels, and cold and hot flow channel system mesh models are constructed separately for the cold and hot flow channel systems. For the cold and hot microchannel mesh models, they are divided along the length direction to obtain multiple one-to-one corresponding cold and hot treatment units. One corresponding cold treatment unit and one hot treatment unit constitute a heat exchange unit. For each group, the flow distribution calculation of the heat exchanger is performed to obtain the corresponding flow distribution calculation results: For each group, the latest cold and hot channel working fluid property parameters, physical model parameters and flow resistance model are used to simulate the cold and hot microchannel mesh model respectively. At the same time, the latest cold and hot channel working fluid property parameters and physical model parameters are used to simulate the cold and hot flow channel system mesh model to obtain the flow distribution calculation results of the cold and hot microchannels of the heat exchanger for each group. In S3, the flow distribution calculation results of each group are used as input, and the multi-scale method is used to realize the flow heat transfer calculation of the entire heat exchanger for each group: the multi-scale method includes a system scale model and a three-dimensional local model. The system scale model is used to solve the macroscopic flow and heat transfer parameters of the heat exchanger as a whole. It includes a flow resistance model for macroscopic flow and a one-dimensional heat transfer model for axial temperature distribution. The three-dimensional local model is used to solve the detailed three-dimensional flow heat transfer field at the cold and hot microchannels in the key areas. S4. If the overall flow heat transfer calculation results for the corresponding group do not converge, the overall flow heat transfer calculation results are substituted into S2 as initial field conditions to correct the working fluid properties and physical model parameters. Iterative calculations are performed until convergence is achieved, and the actual flow heat transfer data of the heat exchanger corresponding to the group is obtained. Multiple sets of discretized boundary conditions and corresponding actual flow heat transfer data constitute sample data.
6. The rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks as described in claim 5, characterized in that, Flow resistance model: Establish a channel mesh model representing cold and hot microchannels, set operating conditions, and construct a flow resistance model based on the relationship between the channel's flow resistance characteristics and the inlet flow rate. ； in, This indicates the pressure drop in the channel. This represents the quadratic coefficient of the resistance term. This represents the first-order coefficient of the resistance term. This indicates the density of the fluid in the channel. Indicates the velocity of the fluid in the channel; Through multiple experiments on the channel, we obtained and The value of .
7. The rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks as described in claim 5, characterized in that, In S3, a mesh-like three-dimensional local model with cross-sectional features is constructed for each heat exchange unit, and a one-dimensional heat transfer model of the axial temperature distribution of the heat exchanger applicable to each heat exchange unit is constructed. Based on the inlet temperatures of the cold and hot microchannels, a temperature condition table containing the inlet temperatures of each cold and hot treatment unit is established. Based on the flow distribution calculation results of the cold and hot microchannels and the temperature condition table, simulation calculations are performed on each three-dimensional local model to obtain the outlet temperatures and heat transfer of each heat exchange unit's cold and hot treatment units. Then, the heat transfer of each heat exchange unit is substituted into the one-dimensional heat transfer model for solution to obtain the axial temperature distribution of the heat exchanger as temperature field data.
8. The rapid simulation and prediction method for heat transfer flow in heat exchangers based on deep networks as described in claim 7, characterized in that, One-dimensional heat transfer model: ； In the formula, and These represent the specific heat capacities of the fluids in the heat and cold treatment units, respectively. and These represent the temperature changes in the heat and cold processing units, respectively. The heat exchange capacity of each heat exchange unit is input as known quantities in tabular form, and the result is obtained by solving the above formula. and This allows us to obtain the axial temperature distribution of the heat exchanger as temperature field data.
9. A rapid simulation and prediction system for heat transfer flow in heat exchangers based on deep networks, characterized in that, The sample construction module is used to set multiple sets of discretized boundary conditions for the heat exchanger. Each set of discretized boundary conditions includes the working fluid properties of the cold and hot channels of the heat exchanger and the parameters of the selected physical model. The heat exchanger is simplified using a flow resistance model. For each set, the flow distribution of the heat exchanger is calculated to obtain the corresponding flow distribution calculation results. Using the flow distribution calculation results of each set as input, a multi-scale method is used to calculate the overall flow heat transfer of the corresponding heat exchanger. If the overall flow heat transfer calculation results of the corresponding set do not converge, the working fluid properties are corrected, and the corrected parameters are used for iterative calculation until convergence, obtaining the actual flow heat transfer data of the corresponding heat exchanger. The flow heat transfer data includes velocity field, pressure field, and temperature field data. The multiple sets of discretized boundary conditions and the corresponding actual flow heat transfer data constitute the sample data. The data matrix processing module is used to perform matrixing, serialization, and normalization on each group of data in the sample data. The processed sample data is divided into training set, validation set, and test set according to the proportion. The reduced-order model building module is used to build and deploy a nonlinear reduced-order model based on a deep neural network by utilizing the model training function of the training set, the model validation function of the validation set, and the model testing function of the test set. The rapid simulation prediction module is used to input new discretized boundary conditions into a nonlinear reduced-order model for simulation prediction, and quickly output the simulation prediction results of heat exchanger flow and heat transfer.
10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in memory to execute the rapid simulation and prediction method for heat exchanger flow heat transfer as described in any one of claims 1-8.
11. A computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the rapid simulation and prediction method for heat exchanger flow heat transfer as described in any one of claims 1-8.