ORC working medium thermophysical property prediction method based on graph neural network

Abstract

The invention relates to the technical field of thermal physical properties of working media, and discloses an ORC working medium thermal physical property prediction method based on a graph neural network, and the method comprises the following steps: S1, working medium molecular graph construction: for a target ORC working medium, obtaining molecular structure information of the target ORC working medium; if the working medium is a single working medium, constructing a single molecular graph based on atoms and chemical bonds; and if the working medium is the mixed working medium, constructing a mixed working medium diagram based on atoms and chemical bonds of each component molecule and a preset intermolecular interaction relationship. According to the ORC working medium thermophysical property prediction method based on the graph neural network, the graph neural network is applied to end-to-end prediction of the ORC working medium thermophysical property, and working medium molecules or mixture components are abstracted into a graph structure, so that a model can learn a complex mapping relation between a molecular structure and macroscopic thermophysical property from an atomic level; the limitation that a traditional method depends on experimental data or empirical equations is broken through, and the model is endowed with strong molecular representation learning ability and generalization prediction potential for a new working medium.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of working medium thermophysical properties, and particularly relates to an ORC working medium thermophysical property prediction method based on a graph neural network. BACKGROUND

[0002] The organic Rankine cycle is a key technology for efficient utilization of medium and low temperature heat sources, and the system design and performance optimization thereof seriously depend on the accuracy of working medium thermophysical properties such as density, specific heat capacity at constant pressure, thermal conductivity and dynamic viscosity. The ORC working medium is often an organic mixture, and the properties thereof change in a wide temperature and pressure range during system operation, and the property change presents strong nonlinearity and complexity, especially near the near-critical region or the phase transition region, and there is a “property mutation” feature similar to that of a pure working medium.

[0003] According to the search, a cross-critical working medium thermophysical property intelligent prediction method suitable for the property mutation feature (CN120706239A) is provided, and an intelligent segmented polynomial method for the property mutation of a pure working medium is proposed, but the essence is still based on pure numerical fitting, and is not associated with the physical essence (molecular structure) of the working medium, and cannot realize prediction of new working media.

[0004] At present, the ORC working medium thermophysical property is mainly obtained by relying on a traditional machine learning fitting method: a neural network, a support vector machine or the like is used to globally fit experimental data, however, such a method has the following significant limitations: 1) Most of them are “black box” models, and have poor physical interpretability; 2) The working medium is regarded as a whole, and the molecular structure information thereof cannot be effectively utilized, and the generalization ability for working media or mixtures not appearing in the training set is weak; 3) A single model is used to process the whole working condition range, and it is difficult to simultaneously consider the modeling accuracy of the smooth change region and the mutation region of the property, and oscillation is easily generated in the mutation region or overfitting is easily generated in the smooth region. SUMMARY

[0005] In view of the above problems that the essence of the existing method is still based on pure numerical fitting, is not associated with the physical essence (molecular structure) of the working medium, and cannot realize prediction of new working media, the present application is proposed.

[0006] Therefore, the present application aims to provide an ORC working medium thermophysical property prediction method based on a graph neural network, and the purpose is to: apply the graph neural network to end-to-end prediction of the ORC working medium thermophysical property, abstract the working medium molecules or mixture components into a graph structure, enable the model to learn the complex mapping relationship between the molecular structure and the macroscopic thermophysical property from the atomic level, break the limitation of the traditional method depending on experimental data or empirical equations, and endow the model with strong molecular representation learning ability and generalization prediction potential for new working media.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for predicting the thermophysical properties of ORC working fluids based on graph neural networks, comprising the following steps: S1: Construction of working fluid molecular diagram: For the target ORC working fluid, obtain its molecular structure information; if it is a single working fluid, construct a single molecular diagram based on its atoms and chemical bonds; if it is a mixed working fluid, construct a mixed working fluid diagram based on the atoms, chemical bonds and preset intermolecular interactions of each component molecule; among them, the node features of the molecular diagram include atom type, charge, and hybridization state, and the edge features include chemical bond type and bond length; S2: Operating condition coding: Obtain the temperature and pressure values ​​of the target prediction point and perform standardized coding; S3: Thermophysical property prediction: The working fluid molecular diagram constructed in step S1 and the operating conditions encoded in step S2 are input together into a pre-trained thermophysical property prediction model to obtain the predicted value of at least one thermophysical property parameter of the ORC working fluid at the temperature and pressure values; the thermophysical property prediction model is a graph neural network model, which is trained to learn the mapping relationship from the molecular diagram and operating conditions to the thermophysical property parameters, and contains a dynamic calculation module that adapts to nonlinear changes in physical properties.

[0008] As a preferred embodiment of the ORC working fluid thermophysical property prediction method based on graph neural network described in this invention, in step S1, for the mixed working fluid, the construction of the mixed working fluid graph specifically involves: constructing molecular subgraphs for each component molecule respectively; and adding virtual edges representing intermolecular interactions between each molecular subgraph according to the component molar ratio and a preset intermolecular interaction model, thereby connecting each molecular subgraph to form a unified mixed working fluid graph.

[0009] As a preferred embodiment of the ORC working fluid thermophysical property prediction method based on graph neural network described in this invention, the thermophysical property prediction model adopts a multi-task learning architecture, and its output layer simultaneously predicts multiple thermophysical parameters, including density, specific heat capacity at constant pressure, thermal conductivity and dynamic viscosity.

[0010] As a preferred embodiment of the ORC working fluid thermophysical property prediction method based on graph neural networks described in this invention, the dynamic calculation module is an expert hybrid module selected based on operating conditions; this module includes K parallel expert networks {E_1,E_2,...,E_K} and a routing network R; the routing network R takes the operating condition vector c encoded in step S2 as input and calculates a set of routing weights. [ 1, 2,..., For the nodes or graph-level features output by the l-th layer of a graph neural network. The modulated features output by the dynamic calculation module Calculate using the following formula: =

[0011] Among them, routing weight It is generated by the routing network R in the following way: g=Softmax(R(c)) The Softmax ( The function ensures and K Through training, the routing network R learns to map different operating conditions, such as normal zones and near-critical zones, to different routing weight distributions, thereby dynamically selecting or combining the most relevant expert network pairs for features. Transformation is performed to adaptively handle the nonlinear changes in the thermophysical properties of the working fluid under different operating conditions.

[0012] As a preferred embodiment of the ORC working fluid thermophysical property prediction method based on graph neural networks described in this invention, the training steps of the thermophysical property prediction model include: Constructing a training dataset: Collect thermal property data of various ORC working fluids at different temperatures and pressures from standard thermal property databases or experimental data, and construct corresponding molecular diagram-working condition-thermal property label data pairs according to steps S1 and S2. Model training: Using the data pairs as input, the thermal property prediction model is trained end-to-end with the goal of minimizing the loss between the predicted thermal properties and the actual thermal properties; the loss function includes a weighted term for the prediction error in the region of abrupt change in properties.

[0013] As a preferred embodiment of the ORC working fluid thermophysical property prediction method based on graph neural network described in this invention, the weighting term for the prediction error of the property change region is implemented by automatically identifying the change region according to the gradient of the property parameters with temperature change in the training data, and increasing the loss weight of the data points in the region when calculating the loss function.

[0014] As a preferred embodiment of the ORC working fluid thermophysical property prediction method based on graph neural networks described in this invention, the method further includes, after obtaining the predicted value in step S3: Uncertainty quantification: Using the integration or Bayesian inference method of the thermal property prediction model, output the uncertainty range of the predicted values ​​of the thermal property parameters.

[0015] As a preferred embodiment of the present invention, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the ORC working fluid thermophysical property prediction method based on graph neural networks.

[0016] As a preferred embodiment of the present invention, a computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the ORC working fluid thermophysical property prediction method based on graph neural networks.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: 1. This invention applies graph neural networks to end-to-end prediction of the thermophysical properties of ORC working fluids. By abstracting the working fluid molecules or mixture components into graph structures, the model can learn the complex mapping relationship between molecular structure and macroscopic thermophysical properties at the atomic level. This breaks the limitations of traditional methods that rely on experimental data or empirical equations, and endows the model with powerful molecular representation learning capabilities and the potential for generalized prediction of new working fluids.

[0018] 2. This invention introduces an adaptive gating mechanism or conditional calculation module into the output layer or intermediate layer of a graph neural network, enabling the model to dynamically adjust its internal feature representation or calculation path according to the input temperature and pressure conditions. This allows the model to automatically and smoothly adapt to the modeling needs of different physical property regions, including smooth and abrupt regions, avoiding the compromise of accuracy of a single global model and the subjectivity of manual segmentation.

[0019] 3. By designing a special graph construction method, this invention constructs a multi-connected graph of the mixture or processes the graphs of each component separately and then merges them. This method can handle the prediction of the thermal properties of single working fluids and mixed working fluids within a unified framework. The model can learn the influence of the interactions between different molecules on the overall properties of the mixture.

[0020] 4. Compared to black-box global neural networks, the prediction of graph neural networks is based on molecular graph structures. Its attention mechanism can be used to analyze which atoms or chemical bonds contribute more to specific physical properties, thus providing a certain degree of physical attribution for property prediction, enhancing the credibility and practicality of the model. It can also quickly and accurately predict the complete set of thermophysical properties of any candidate ORC working fluid under specified operating conditions. It can be seamlessly integrated into ORC system modeling, process simulation and optimization algorithms, greatly accelerating the working fluid selection and system design iteration process, and has important engineering application value. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the steps of the ORC working fluid thermophysical property prediction method based on graph neural networks of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Example Reference Figure 1 As an embodiment of the present invention, a method for predicting the thermal properties of ORC working fluids based on graph neural networks is provided. This method for predicting the thermal properties of ORC working fluids based on graph neural networks includes the following steps: S1: Molecular Graph Construction of Working Proteome: For the target ORC working proteome, obtain its SMILES expression or three-dimensional molecular structure file as molecular structure information; if it is a single working proteome, construct a single molecular graph based on its atoms and chemical bonds; if it is a mixed working proteome, construct a unified mixed working proteome graph based on the atoms, chemical bonds, and preset intermolecular interactions of each component molecule; wherein, the node feature vector of the molecular graph includes at least the atom type, formal charge, and one-hot encoding of hybridization state, and the edge feature vector includes at least the chemical bond type and the reference bond length obtained through molecular mechanics calculations or a preset database; S2: Working condition coding: Obtain the absolute temperature value T and absolute pressure value P of the target prediction point, and perform standardization or normalization coding based on the statistics of the training dataset to obtain a one-dimensional or multi-dimensional working condition feature representation. S3: Thermophysical property prediction: The structural data, node features, and edge features of the working fluid molecular graph constructed in step S1, as well as the operating condition features encoded in step S2, are input into a pre-trained thermophysical property prediction model to obtain the predicted values ​​of one or more thermophysical property parameters of the ORC working fluid under the operating conditions. The thermophysical property prediction model is a graph neural network model based on a message passing mechanism. It is trained end-to-end to learn the complex nonlinear mapping relationship from molecular structure and operating conditions to thermophysical property parameters, and it contains a feature modulation module that can dynamically adjust its internal calculation path or parameters according to the operating condition features. Step S4: Assessment of prediction uncertainty; specifically, after obtaining the point prediction value in step S3, the uncertainty of the prediction result is quantified using one or more of the following methods: Model ensemble method: A committee is formed by using multiple thermal property prediction models with the same structure but different initializations or different subsets of training data to predict the same input. The statistical distribution (such as mean and standard deviation) of the prediction results of multiple models is used as the final prediction and its uncertainty measure. Monte Carlo Dropout method: During the forward propagation of a trained thermal property prediction model, the Dropout layer is activated multiple times to obtain multiple random forward propagation results under the same input, and the statistical distribution of these results represents the prediction uncertainty. Bayesian neural network method: Treats model weights as a probability distribution, approximates the posterior distribution through variational inference or Markov chain Monte Carlo method, and obtains the probability distribution of predicted values ​​accordingly. In step S1, for mixed working fluids, constructing a unified mixed working fluid diagram specifically includes the following sub-steps: Molecular subgraphs are constructed for each pure component molecule, and each subgraph contains the atomic nodes, chemical bond edges and their features inside the component molecule; Between molecular subgraphs of different components, virtual edges representing intermolecular interactions are added based on the mole fraction ratio of each component and considering the possible types of intermolecular interactions; wherein the initial weight or features of the virtual edges are initialized by learnable parameters containing component concentration information or a distance function based on molecular descriptors. The virtual edges connect all component molecular subgraphs to form a globally connected heterogeneous graph, which serves as the mixed working fluid graph and is used to uniformly characterize the microstructure of the mixture system.

[0024] The thermal property prediction model adopts a multi-task learning architecture, which includes a shared graph feature extraction backbone network and multiple parallel property prediction sub-networks. The shared backbone network is used to extract deep-level graph-level global feature representations from the input molecular graph. The multiple property prediction sub-networks take the fusion result of the global feature representation and the operating condition features as input and independently output different predicted values ​​of thermal property parameters. The thermal property parameters include at least density, specific heat capacity at constant pressure, thermal conductivity, and dynamic viscosity.

[0025] The dynamic calculation module is an expert hybrid module based on operating condition selection; this module includes K parallel expert networks {E_1,E_2,...,E_K} and a routing network R; the routing network R takes the operating condition vector c encoded in step S2 as input and calculates a set of routing weights. [ 1, 2,..., For the nodes or graph-level features output by the l-th layer of a graph neural network. The modulated features output by the dynamic calculation module Calculate using the following formula: =

[0026] Among them, routing weight It is generated by the routing network R in the following way: g=Softmax(R(c)) The Softmax ( The function ensures and K Through training, the routing network R learns to map different operating conditions, such as normal zones and near-critical zones, to different routing weight distributions, thereby dynamically selecting or combining the most relevant expert network pairs for features. Transformation is performed to adaptively handle the nonlinear changes in the thermophysical properties of the working fluid under different operating conditions.

[0027] The training process of the thermal property prediction model specifically includes the following steps: Training dataset construction: Collect thermal property data samples covering a variety of ORC candidate working fluids and in a wide temperature and pressure range from standard thermal property databases, public literature or experimental measurements; for each sample, according to its working fluid composition, temperature and pressure, construct the corresponding molecular diagram data object and working condition encoding vector offline according to steps S1 and S2, and use the real property values ​​as supervision labels to form a complete training sample set; End-to-end model training: Stochastic gradient descent or its variants are used as the optimization algorithm. The model parameters are updated iteratively by calculating the predicted values ​​through forward propagation and the gradients through backpropagation. The loss function used consists of a basic regression loss term and at least one regularization or enhancement term, where the basic regression loss term is used to measure the overall deviation between the predicted values ​​and the true labels.

[0028] The loss function includes a region-adaptive weighting term, which is implemented as follows: before or during training, for each sample i in the training dataset, calculate the numerical gradient or approximate gradient of its physical property label value y_i with respect to temperature T_i. y / T|_i; Based on the magnitude of this gradient, assign a dynamic loss weight w_i to sample i, where w_i is | y / T|_i is a monotonically non-decreasing function; when calculating batch loss, the basic regression loss of each sample is multiplied by the corresponding weight w_i and then summed and averaged, so that the model focuses more on learning the data patterns in areas of drastic changes in physical properties during training.

[0029] This invention achieves: Accurate prediction: By directly learning the mapping from molecular structure to physical properties through graph neural networks, high-precision prediction of the thermophysical properties of single and mixed ORC working fluids over a wide range of operating conditions is achieved.

[0030] Adaptive mutation handling: The innovative dynamic computing module (expert hybrid) and training strategy (gradient-weighted loss) enable the model to automatically and smoothly adapt to regions of smooth physical properties and regions of mutation, overcoming the limitations of traditional single models or manual segmentation methods.

[0031] Strong generalization and interpretability: The model makes predictions based on molecular graphs, which has the potential to generalize to unknown working fluids. The visualization of graph attention mechanism and routing weights can provide a certain physical explanation for the prediction results.

[0032] Highly practical for engineering applications: It provides a complete implementation plan from data preparation, model training to prediction and uncertainty quantification, and can be easily integrated into ORC working fluid screening and system design software, greatly improving design efficiency and reliability.

[0033] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the thermophysical properties of ORC working fluids based on graph neural networks, characterized in that, Includes the following steps: S1: Construction of working fluid molecular diagram: For the target ORC working fluid, obtain its molecular structure information; if it is a single working fluid, construct a single molecular diagram based on its atoms and chemical bonds; if it is a mixed working fluid, construct a mixed working fluid diagram based on the atoms, chemical bonds and preset intermolecular interactions of each component molecule; among them, the node features of the molecular diagram include atom type, charge, and hybridization state, and the edge features include chemical bond type and bond length; S2: Operating condition coding: Obtain the temperature and pressure values ​​of the target prediction point and perform standardized coding; S3: Thermophysical property prediction: The working fluid molecular diagram constructed in step S1 and the operating conditions encoded in step S2 are input together into a pre-trained thermophysical property prediction model to obtain the predicted value of at least one thermophysical property parameter of the ORC working fluid at the temperature and pressure values; the thermophysical property prediction model is a graph neural network model, which is trained to learn the mapping relationship from the molecular diagram and operating conditions to the thermophysical property parameters, and contains a dynamic calculation module that adapts to nonlinear changes in physical properties.

2. The ORC working fluid thermophysical property prediction method based on graph neural networks according to claim 1, characterized in that: In step S1, for the mixed working fluid, the construction of the mixed working fluid diagram specifically involves: constructing molecular sub-diagrams for each component molecule; and adding virtual edges representing intermolecular interactions between each molecular sub-diagram according to the component molar ratio and a preset intermolecular interaction model, thereby connecting each molecular sub-diagram to form a unified mixed working fluid diagram.

3. The ORC working fluid thermophysical property prediction method based on graph neural networks according to claim 1, characterized in that: The thermal property prediction model adopts a multi-task learning architecture, and its output layer simultaneously predicts multiple thermal property parameters, including density, specific heat capacity at constant pressure, thermal conductivity, and dynamic viscosity.

4. The ORC working fluid thermophysical property prediction method based on graph neural networks according to claim 1, characterized in that: The dynamic calculation module is an expert hybrid module based on operating conditions selection; this module includes K parallel expert networks {E_1,E_2,...,E_K} and a routing network R; the routing network R takes the operating condition vector c encoded in step S2 as input and calculates a set of routing weights. [ 1, 2,..., For the nodes or graph-level features output by the l-th layer of a graph neural network. The modulated features output by the dynamic calculation module Calculate using the following formula: = Among them, routing weight It is generated by the routing network R in the following way: g=Softmax(R(c)) The Softmax ( The function ensures and K Through training, the routing network R learns to map different operating conditions, such as normal zones and near-critical zones, to different routing weight distributions, thereby dynamically selecting or combining the most relevant expert network pairs for features. Transformation is performed to adaptively handle the nonlinear changes in the thermophysical properties of the working fluid under different operating conditions.

5. The ORC working fluid thermophysical property prediction method based on graph neural networks according to claim 1, characterized in that: The training steps for the thermal property prediction model include: Constructing a training dataset: Collect thermal property data of various ORC working fluids at different temperatures and pressures from standard thermal property databases or experimental data, and construct corresponding molecular diagram-working condition-thermal property label data pairs according to steps S1 and S2. Model training: Using the data pairs as input, the thermal property prediction model is trained end-to-end with the goal of minimizing the loss between the predicted thermal properties and the actual thermal properties; the loss function includes a weighted term for the prediction error in the region of abrupt change in properties.

6. The ORC working fluid thermophysical property prediction method based on graph neural networks according to claim 5, characterized in that: The weighting term for the prediction error of the abrupt change region is implemented as follows: the abrupt change region is automatically identified based on the gradient of the change of physical property parameters with temperature in the training data, and the loss weight of the data points in that region is increased when calculating the loss function.

7. The ORC working fluid thermophysical property prediction method based on graph neural networks according to claim 6, characterized in that: After obtaining the predicted value in step S3, the process further includes: Uncertainty quantification: Using the integration or Bayesian inference method of the thermal property prediction model, output the uncertainty range of the predicted values ​​of the thermal property parameters.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images