Condenser performance prediction method combining data driving and physical constraint
By combining data-driven and physical constraint methods, an adaptive deep operator network model is built and dynamic weight adjustment is performed, which solves the problems of poor generalization and high computational cost in condenser performance prediction, and achieves efficient and accurate performance prediction.
Patent Information
- Application Number
- CN202510409681.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-02
- Publication Date
- 2025-07-18
AI Technical Summary
The prior art has problems in the prediction of condenser performance, which has strong dependence on training data and lack of physical constraints, resulting in poor generalization. The CFD simulation calculation is large and costly, making it difficult to meet the needs of real-time optimization and intelligent control.
Combining the condenser performance prediction method of data-driven and physical constraints, the fitted model is constructed through the fusion training of adaptive dynamically adjusted deep operator network model and physical constraint model, and the dynamic weight adjustment strategy balance model training is used to meet the physical laws and data-driven needs.
The prediction accuracy, generalization ability and computing efficiency of the model are improved, and efficient real-time prediction of condenser performance is achieved.
Smart Images
Figure CN120337993A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of condenser performance prediction, and in particular to a condenser performance prediction method combining data-driven and physical constraints. Background Art
[0002] A condenser is a key component in a refrigeration system, and its performance directly affects the energy efficiency ratio (COP) and operating cost of the refrigeration system. Physical modeling of the condenser can often lead to improvements in process design, control strategies, etc., thereby further enhancing the energy efficiency ratio of the refrigerator. Currently, the following methods are mainly used for modeling the condenser:
[0003] 1. Fitting of experimental data: Training a statistical model, such as multiple regression or neural network, through a large amount of experimental data for condenser performance prediction. However, this method is highly dependent on training data and lacks physical constraints, resulting in large errors in the model test results within the range of non-training data, with poor generalization ability and difficulty in making correct responses when dealing with unknown working conditions.
[0004] 2. Computational Fluid Dynamics (CFD) simulation: Simulating the thermodynamic and fluid dynamic behaviors inside the condenser through CFD to provide high-precision modeling and prediction. However, the computational amount is extremely large, and a large amount of experimental data is required for parameter calibration, resulting in high time cost, high difficulty, and difficulty in meeting the requirements of real-time optimization and intelligent control, and it is not applicable to the real-time control of real systems. Summary of the Invention
[0005] In order to solve the above technical problems existing in the prior art, the present invention proposes a condenser performance prediction method combining data-driven and physical constraints, and its specific technical solutions are as follows:
[0006] A condenser performance prediction method combining data-driven and physical constraints includes the following steps:
[0007] Step 1: Obtain data of the condenser under different working conditions and preprocess the data;
[0008] Step 2: Construct a deep operator network model with adaptive dynamic adjustment;
[0009] Step 3: Construct a physical constraint model to make the deep operator network model constructed in Step 2 conform to the laws of thermodynamics and fluid mechanics equations during training;
[0010] Step 4: Use the preprocessed data to perform fusion training on the physical constraint model and the deep operator network model with adaptive dynamic adjustment to obtain a final fitting model;
[0011] Step 5: Use the fitting model to predict the condenser performance index.
[0012] Further, in Step 1, through the whole-machine experiment under specific working conditions or the data of the fine-tuned simulation model, the key input parameters x and the target output variables y of the condenser under different working conditions are obtained, and the data set is obtained, so as to construct the mapping relationship of y = f(x);
[0013] Among them, the input variable x includes the refrigerant temperature T r , the refrigerant pressure P r , the ambient temperature T env , the wind speed v w and the condenser structure parameters: fin pitch and pipe diameter;
[0014] The output variable y includes the heat transfer efficiency η, the condensation temperature T c , the thermal resistance R th .
[0015] Further, in Step 1, the preprocessing of the data is carried out by normalizing the data.
[0016] Further, in Step 2, the deep operator network model is improved: for the branch network in the deep operator network model, the dynamic depth d B (x) is used to automatically control the number of layers of the branch network, and the global feature expression extracted by the branch network is Among them, the calculation method of the dynamic depth d B (x) is d B (x) = d min +|α·C(x)|, where d min is the minimum network depth, α is the scaling factor used to control the sensitivity of depth change; C(x) is the variability of the data input gradient, and the calculation method is: The number of layers of the branch network can be adjusted according to the size of C(x);
[0017] At the same time, for the backbone network in the deep operator network model, the local distribution feature expression extracted is Among them, the calculation method of the dynamic depth d T (s) is d T (s) = d min +|β·C T (s)|, β is the scaling factor used to control the sensitivity of depth change, C T (s) is used to measure the complexity of spatial features, and the calculation method is: The number of layers of the backbone network can be adjusted according to the size of C T (s).
[0018] Further, in Step 3, the construction of the physical constraint model is specifically as follows:
[0019] First, construct the energy conservation equation based on the first law of thermodynamics:
[0020]
[0021] where ρ is the fluid density, c p is the specific heat capacity, T is the temperature, u and v are the velocity components, k is the thermal conductivity, and q is the external heat source term;
[0022] Secondly, construct the mass conservation equation:
[0023]
[0024] where u = (u, v, m) is the flow velocity vector;
[0025] Finally, construct the momentum conservation equation:
[0026]
[0027] where p is the pressure, μ is the dynamic viscosity, and F is the external force.
[0028] Furthermore, in step four, for the fusion training, a fusion training strategy with dynamic weight adjustment is adopted, specifically:
[0029] When the training step is t, the weights λ1 of the physical loss of the physical constraint model and λ2 of the data loss of the deep operator network model are adjusted according to the following strategy:
[0030]
[0031] λ2(t) = 1 - λ1(t),
[0032] where α controls the weight change speed, giving priority to data-driven in the initial stage and strengthening physical constraints in the later stage.
[0033] Furthermore, new operating conditions including refrigerant temperature, pressure, and ambient temperature are input to the fitting model to test the prediction performance.
[0034] Advantages of the present invention: The present invention fuses and improves the deep operator network and the physics-informed neural network to model the condenser. Among them, the traditional deep operator network is improved, and a fusion training method based on dynamic weight adjustment is designed to balance the problem of difficult training control brought by the fusion model, thereby effectively improving the prediction accuracy, generalization ability, and calculation efficiency of the model. Description of the Drawings
[0035] Figure 1It is a flowchart of a condenser performance prediction method combining physical constraints and data-driven in an embodiment of the present invention;
[0036] Figure 2 It is a flowchart of a method for improving and modeling a deep operator network in an embodiment of the present invention. Detailed implementation manners
[0037] In order to make the objectives, technical solutions and technical effects of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings of the specification and embodiments.
[0038] As Figure 1 shown, this embodiment discloses a condenser performance prediction method combining data-driven and physical constraints, including the following steps:
[0039] Step 1: Data acquisition and preprocessing.
[0040] Through the whole-machine experiment under specific working conditions or the data of the fine-tuned simulation model, the key input parameters x and the target output variable y of the condenser under different working conditions are obtained, and the data set is obtained, so as to construct the mapping relationship of y = f(x).
[0041] Among them, the input variable x includes the refrigerant temperature T r , the refrigerant pressure P r , the ambient temperature T env , the wind speed v w and the condenser structure parameters, including the fin pitch and the pipe diameter. The output variable y includes the heat transfer efficiency η, the condensation temperature T c , the thermal resistance R th .
[0042] In order to improve the stability of network training, the data is processed in a normalized manner.
[0043] Step 2: As Figure 2 shown, the improvement and modeling of the deep operator network.
[0044] The traditional deep operator network includes two parts, a branch network and a backbone network. The branch network is used to learn the feature mapping of the global input parameters of the condenser, as follows:
[0045] B(x) = {b1, b2,..., b k},
[0046] Among them, B(x) represents the extracted global features, and k is the dimension of the hidden layer.
[0047] The backbone network is used to learn the spatial distribution features of the output variables, as follows:
[0048] T(s) = {t1, t2, ..., t k},
[0049] where s represents the spatial coordinates inside the condenser, such as the temperature and velocity fields in the heat exchange area, and T(s) represents the extracted local distribution features.
[0050] The present invention improves the traditional deep operator network and proposes an adaptive dynamic adjustment deep operator network model, which is as follows.
[0051] First, use the dynamic depth d B (x) to automatically control the number of layers of the branch network. That is, use to replace the traditional The calculation method of the dynamic depth is d B (x) = d min + |α·C(x)|, where d min is the minimum network depth. α is a scaling factor used to control the sensitivity of depth change. C(x) is the variability of the data input gradient, and its calculation method is: After that, the number of layers of the branch network can be adjusted according to its size to achieve the effect of adaptive dynamic adjustment.
[0052] At the same time, in the backbone network, an adaptive adjustment strategy is also adopted. That is, use to replace the traditional The calculation method of the dynamic depth is d T (s) = d min + |β·C T (s)|. β is a scaling factor used to control the sensitivity of depth change. C T (s) is a measure of the complexity of spatial features, and its calculation method is: After that, the number of layers of the backbone network can be adjusted according to its size to achieve the effect of adaptive dynamic adjustment.
[0053] Step 3: Physical constraint construction.
[0054] To ensure that the prediction results satisfy physical laws, the present invention adopts a physical constraint model to make the network model constructed in Step 2 conform to the laws of thermodynamics and fluid mechanics equations during training.
[0055] First, construct the energy conservation equation based on the first law of thermodynamics:
[0056]
[0057] where ρ is the fluid density, c p is the specific heat capacity, T is the temperature, u and v are the velocity components respectively, k is the thermal conductivity, and q is the external heat source term.
[0058] Secondly, construct the mass conservation equation:
[0059]
[0060] where \(u=(u, v, m)\) is the flow velocity vector.
[0061] Finally, construct the momentum conservation equation:
[0062]
[0063] where \(p\) is the pressure, \(\mu\) is the dynamic viscosity, and \(F\) is the external force.
[0064] Step 4: A fusion training method based on dynamic weight adjustment.
[0065] When training by fusing the improved deep operator network and the physics-informed neural network, it may be difficult to control the balance between physical constraints and data-driven. Therefore, the present invention proposes a fusion training strategy with dynamic weight adjustment.
[0066] At the training step \(t\), the weights \(\lambda_1\) and \(\lambda_2\) of the physical loss and the data loss are adjusted according to the following strategy:
[0067]
[0068] \(\lambda_2(t)=1 - \lambda_1(t)\),
[0069] where \(\alpha\) controls the weight change speed, making it mainly data-driven in the initial stage and strengthening physical constraints in the later stage.
[0070] After training is completed, the final fitted model is output, and this model can be used to predict key indicators such as the heat transfer efficiency and condensation temperature of the condenser in real time.
[0071] Step 5: Test the modeling performance of the condenser.
[0072] To verify the accuracy of the established model, new operating conditions, including refrigerant temperature, pressure, ambient temperature, etc., can be input to the model, and the condenser performance indicators can be predicted through the output.
[0073] The above is only the preferred embodiment of the present invention and does not impose any formal restrictions on the present invention. Although the implementation process of the present invention has been described in detail above, for those familiar with the art, they can still modify the technical solutions recorded in the foregoing examples or make equivalent replacements for some of the technical features. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.
Claims
1. A condenser performance prediction method combining data-driven and physical constraints, characterized in that, It includes the following steps: Step 1: Obtain the data of the condenser under different working conditions and preprocess the data; Step 2: Construct a deep operator network model with adaptive dynamic adjustment; Step 3: Construct a physical constraint model to make the deep operator network model constructed in Step 2 comply with the laws of thermodynamics and the equations of fluid mechanics during training; Step 4: Use the preprocessed data to conduct fusion training on the physical constraint model and the deep operator network model with adaptive dynamic adjustment to obtain the final fitting model; Step 5: Use the fitting model to predict the performance indicators of the condenser.
2. The condenser performance prediction method according to claim 1, wherein In Step 1, through the full-machine experiment under specific working conditions or the simulation model data after fine-tuning, the key input parameters x and the target output variable y of the condenser under different working conditions are obtained, and the data set is obtained, so as to construct the mapping relationship of y = f(x); Among them, the input variable x includes the refrigerant temperature T r , the refrigerant pressure P r , the ambient temperature T env , the wind speed v w and the condenser structure parameters: fin pitch and tube diameter; The output variable y includes the heat transfer efficiency η, the condensation temperature T c , and the thermal resistance R th .
3. The condenser performance prediction method according to claim 2, wherein In Step 1, the preprocessing of the data is to process the data in a normalized manner.
4. The condenser performance prediction method according to claim 1, wherein In step two, the deep operator network model is improved: for the branch network in the deep operator network model, the dynamic depth d B (x) is used to automatically control the number of layers of the branch network, and the global feature expression extracted by the branch network is where the calculation method of the dynamic depth d B (x) is d B (x) = d min + |α·C(x)|, where d min is the minimum network depth, α is a scaling factor used to control the sensitivity of depth change; C(x) is the variability of the data input gradient, and its calculation method is: The number of layers of the branch network can be adjusted according to the magnitude of C(x); Meanwhile, for the backbone network in the deep operator network model, the extracted local distribution feature expression is where the dynamic depth d T (s) is calculated as d T (s) = d min + |β · C T (s)|, β is a scaling factor used to control the sensitivity of depth change, and C T (s) measures the spatial feature complexity and is calculated as: The number of layers of the backbone network can be adjusted according to the magnitude of C T (s).
5. The condenser performance prediction method according to claim 1, wherein, In Step 3, the construction of the physical constraint model is specifically as follows: First, construct the energy conservation equation based on the first law of thermodynamics: where ρ is the fluid density, c p is the specific heat capacity, T is the temperature, u and v are the velocity components, k is the thermal conductivity, and q is the external heat source term; Secondly, construct the mass conservation equation: where u=(u, v, m) is the velocity vector; Finally, construct the momentum conservation equation: where p is the pressure, μ is the dynamic viscosity, and F is the external force.
6. The condenser performance prediction method according to claim 1, wherein, In Step 4, the fusion training adopts a fusion training strategy with dynamic weight adjustment, specifically: When at the training step t, the weight λ1 of the physical loss of the physical constraint model and the weight λ2 of the data loss of the deep operator network model are adjusted according to the following strategy: λ2(t)=1 - λ1(t), where α controls the weight change speed, making it mainly data-driven in the initial stage and strengthening physical constraints in the later stage.
7. The condenser performance prediction method according to claim 1, wherein, Input the new operating conditions including refrigerant stability, pressure, and ambient temperature into the fitting model to test the prediction performance.