A graph physical information network construction method for power system state analysis

Abstract

This invention discloses a method for constructing a graph physical information network for power system state analysis, comprising: based on the time-varying topology of the power network and historical measurement data of nodes, generating high-dimensional embedded features of nodes that integrate spatiotemporal dependencies through a spatiotemporal graph attention module and inputting them into a hierarchical neural network predictor module, outputting at least one of the following: predicted values ​​of the electrical state of nodes, predicted values ​​of the internal physical fields of equipment, and predicted values ​​of the health and aging state of equipment; constructing a physical constraint loss term, which together with a data fidelity loss term constitutes a multi-physics deep coupling loss function; dynamically adjusting the gradients of each loss term in the multi-physics deep coupling loss function and resolving conflicts, and updating the network parameters based on the adjusted gradients until the model converges, thereby generating a graph physical information network for power system state analysis.

Description

Technical Field

[0001] This invention relates to the field of power system analysis and artificial intelligence, and more specifically, to a method for constructing a graph physical information network for power system state analysis. Background Technology

[0002] With the large-scale grid integration of renewable energy and the increasing complexity of power loads, the operation of power transmission and transformation systems faces unprecedented challenges. Equipment aging, extreme weather, and potential operational errors can all trigger cascading failures, threatening system safety. Therefore, accurate prediction of the health status of critical equipment and real-time, precise understanding of the power flow distribution across the entire network are crucial for preventative maintenance, optimized scheduling, and emergency response.

[0003] Traditional power system analysis methods primarily rely on accurate mathematical models and numerical solution techniques. These methods perform well when system parameters are known and operating conditions are stable, but they have limitations in handling parameter uncertainties, errors introduced by model simplification, and real-time computation of large-scale systems. In recent years, data-driven methods, represented by deep learning, have shown great potential in power system prediction tasks. However, these methods typically treat the system as a "black box," lacking physical interpretability, and may produce predictions that violate physical laws in areas where training data is incomplete, thus casting doubt on their reliability in critical decision-making.

[0004] The emergence of Physics-Informed Neural Networks (PINNs) offers a new approach to solving the aforementioned problems. PINNs incorporate the residuals of partial differential equations (PDEs) as a penalty term into the loss function of the neural network, enabling the network to learn data patterns while its output satisfies specific physical constraints. PINNs have achieved success in several fields, including fluid mechanics and heat conduction. In the field of power systems, initial applications of PINNs have also demonstrated their potential in state estimation and dynamic simulation.

[0005] However, the application of existing PINNs in power systems still faces some challenges: (1) Complex topology and dynamics: Power networks have complex graph structures, and the topology may change dynamically due to switching operations or faults. Traditional PINNs are usually based on fixed spatial domains, making it difficult to directly capture this topological dependence and dynamics.

[0006] (2) Multi-physics coupling: The state of power equipment is not only affected by its own electrical parameters, but also closely related to thermodynamic factors such as the surrounding environment and heat dissipation conditions. Power flow distribution itself is also a manifestation of electromagnetic field laws. How to couple multiple physical laws within a unified framework is a difficult problem.

[0007] (3) Balance between data and physics: In practical applications, high-quality measurement data may be sparse or noisy. How to achieve the best balance between data-driven and physical constraints, and ensure the generalization ability and physical consistency of the model, is a problem that needs to be solved. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method for constructing a graph-based physical information network for power system state analysis.

[0009] According to one aspect of the present invention, a method for constructing a graph physical information network for power system state analysis is provided, comprising: Based on the time-varying topology of the power network and historical measurement data of nodes, a high-dimensional embedding feature of nodes that integrates spatiotemporal dependencies is generated through a spatiotemporal graph attention module. The high-dimensional embedded features of the nodes are input into the hierarchical neural network predictor module, which outputs at least one of the following: the predicted electrical state of the nodes in the power system, the predicted physical field of the equipment, and the predicted health and aging state of the equipment. Based on the predicted values ​​of node electrical state, the predicted values ​​of equipment internal physical field, and the predicted values ​​of equipment health aging state, a physical constraint loss term is constructed, which includes at least one of the following: node power balance residual, equipment heat conduction partial differential equation residual, and equipment aging dynamics ordinary differential equation residual. This term, together with the data fidelity loss term, constitutes a multi-physics deep coupling loss function. Through the gradient dynamic balancing and conflict resolution module, the gradients of each loss term in the multi-physics deep coupling loss function are dynamically weighted and conflict resolved. The network parameters are updated based on the adjusted gradients until the model converges, generating a graphical physical information network for power system state analysis.

[0010] Optionally, based on the time-varying topology of the power network and historical measurement data of nodes, a high-dimensional embedding feature of nodes incorporating spatiotemporal dependencies is generated through a spatiotemporal graph attention module, including: Based on the time-varying adjacency matrix or Laplace matrix of the power network, spatial topological features are extracted using graph convolution operation. The graph convolution operation approximates the spectral domain filter with Chebyshev polynomial to generate node spatial features. Input the historical time series corresponding to the node spatial features into the temporal self-attention layer or Transformer encoder layer to capture the temporal dependencies within the historical time window and generate node temporal features. By fusing node spatial features and node temporal features, high-dimensional embedding features of nodes are generated.

[0011] Optionally, the hierarchical neural network predictor module includes: The node electrical state predictor is used to output the voltage amplitude and voltage phase angle of the node as the predicted value of the node electrical state based on the node's high-dimensional embedding features and time coding information. The device internal physics field predictor is used to output the three-dimensional temperature field distribution inside the device as the predicted value of the device internal physics field based on the node high-dimensional embedding features, spatial coordinate encoding and device loss power information. The equipment health and aging status predictor is connected to the equipment's internal physical field predictor. It is used to output the equipment's cumulative aging degree as the predicted value of the equipment's health and aging status based on the hot spot temperature and historical heat load accumulation information extracted from the predicted value of the equipment's internal physical field.

[0012] Optionally, the physical constraint loss term specifically includes: Based on the voltage amplitude and voltage phase angle in the predicted electrical state of the nodes, the balance residuals of injected active power and reactive power are calculated, and physical constraints for AC power flow are constructed. Based on the three-dimensional temperature field distribution in the predicted physical field of the equipment, the residual between it and the partial differential equation of heat conduction is calculated, and the physical constraint term of heat conduction is constructed. Based on the cumulative aging degree in the predicted value of equipment health aging state, the residual between it and the Arrhenius aging ordinary differential equation is calculated, and the physical constraint term of aging kinetics is constructed.

[0013] Optionally, it also includes: further superimposing the residual of the equipment heat conduction boundary conditions and the residual of the initial temperature distribution and aging state into the physical constraint loss term as boundary and initial condition constraint terms.

[0014] Optionally, a gradient dynamic balancing and conflict resolution module is used to dynamically adjust the weights and resolve conflicts of the gradients of each loss term in the multiphysics deep coupling loss function, and to update the network parameters based on the adjusted gradients, including: Calculate the gradient of each loss term in the multiphysics deep coupling loss function with respect to the network parameters to obtain the gradient vector of each loss term; Calculate the cosine similarity between each pair of gradient vectors, construct the gradient conflict matrix, and determine the gradient conflict degree of each loss term based on the gradient conflict matrix; The weights of each loss term in the multiphysics deep coupling loss function are dynamically adjusted based on the gradient conflict degree, so that the loss term with a larger conflict degree receives a smaller weight coefficient, in order to seek the Pareto optimal joint descent direction. The joint gradient direction is obtained by weighted combination of the gradients of each loss term after weight adjustment, and the network parameters are updated accordingly.

[0015] Optionally, the weights can be dynamically adjusted. The calculation formula is: Among them, the exponential moving average of the data loss gradient norm The calculation expression is: In the formula, Indicates the number of training iterations for the model; Represents the gradient of neural network parameters; Indicates the first The L2 norm of the data loss gradient during step training; This represents the exponential moving average decay coefficient. Each physical loss item Maximum gradient component The calculation expression is: In the formula, The expression for calculating the gradient conflict parameter is as follows: In the formula, Here are the hyperparameters; the conflict degree is... , This represents the cosine similarity between physical gradients.

[0016] According to another aspect of the present invention, a graph physical information network construction device for power system state analysis is provided, comprising: The first generation module is used to generate high-dimensional embedding features of nodes that integrate spatiotemporal dependencies based on the time-varying topology of the power network and historical measurement data of nodes through the spatiotemporal graph attention module. The output module is used to input the high-dimensional embedded features of the nodes into the hierarchical neural network predictor module, and output at least one of the following: the predicted electrical state of the nodes of the power system, the predicted physical field of the equipment, and the predicted health and aging state of the equipment. The module is used to construct a physical constraint loss term based on the predicted values ​​of node electrical state, the predicted values ​​of equipment internal physical field, and the predicted values ​​of equipment health aging state. This term includes at least one type of physical constraint loss term, which is composed of the residual of node power balance, the residual of equipment heat conduction partial differential equation, and the residual of equipment aging dynamics ordinary differential equation. Together with the data fidelity loss term, it constitutes a multi-physics deep coupling loss function. The second generation module is used to dynamically adjust the weights and resolve conflicts of the gradients of each loss term in the deep coupling loss function of multiphysics through the gradient dynamic balancing and conflict resolution module, and update the network parameters based on the adjusted gradients until the model converges, thereby generating a graphical physical information network for power system state analysis.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing a computer program for performing the methods described in any of the above aspects of the present invention.

[0018] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the method described in any of the preceding aspects of the present invention.

[0019] Therefore, this invention proposes a novel high-dimensional dynamically coupled graph-physical information network for power system state analysis. This network handles the topology and dynamic changes of the power network by introducing a dynamic graph embedding module, embeds AC power flow equations and equipment thermodynamic equations into the loss function using a multi-physics coupling module, and optimizes the learning process through an adaptive weighted loss mechanism. It has the following beneficial effects: (1) Unified high-fidelity prediction framework: An end-to-end framework was constructed, consisting of a collaborative spatiotemporal graph attention module, a hierarchical neural network predictor module, a multi-physics field deep coupling loss function, and a gradient dynamic balance and conflict resolution loss function, which enabled comprehensive and high-precision prediction of the multi-scale state and long-term aging trend of the power system.

[0020] (2) Spatiotemporal graph attention module: A new mechanism is proposed that integrates dynamic graph theory and practical attention to efficiently capture the complex high-dimensional spatiotemporal dependencies in the evolution of power networks.

[0021] (3) Hierarchical neural network predictor: It adopts a modular and hierarchical prediction architecture to achieve accurate decoupling and collaborative prediction of multiple types and scales of physical quantities such as electrical state, internal physical field, and equipment aging, effectively improving the prediction accuracy and generalization ability of different physical quantities. At the same time, it utilizes hierarchical dependency relationships to reuse intermediate prediction results and improve computational efficiency.

[0022] (4) Multiphysics deep coupling loss function: Within the unified PINN framework, the deep coupling and collaborative prediction of enhanced electrical dynamics, three-dimensional high-fidelity thermal field inside the equipment and physical aging model are realized for the first time.

[0023] (5) Gradient dynamic balancing and conflict resolution loss function: Design an advanced adaptive loss strategy to ensure the robustness and physical consistency of model training by dynamically balancing the magnitude of multi-source gradients and actively resolving gradient conflicts between physical constraints. Attached Figure Description

[0024] Exemplary embodiments of the present invention can be more fully understood by referring to the following figures: Figure 1 This is a flowchart illustrating a method for constructing a graph physical information network for power system state analysis, provided by an exemplary embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a graph physical information network construction device for power system state analysis provided in an exemplary embodiment of the present invention; Figure 3 This is the structure of an electronic device provided in an exemplary embodiment of the present invention. Detailed Implementation

[0025] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0026] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0027] Those skilled in the art will understand that the terms "first," "second," etc., in the embodiments of the present invention are only used to distinguish different steps, devices, or modules, and do not represent any specific technical meaning, nor do they indicate a necessary logical order between them.

[0028] It should also be understood that in the embodiments of the present invention, "multiple" can refer to two or more, and "at least one" can refer to one, two or more.

[0029] It should also be understood that any component, data or structure mentioned in the embodiments of the present invention can generally be understood as one or more unless explicitly defined or given contrary instructions in the context.

[0030] Furthermore, the term "and / or" in this invention is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this invention generally indicates that the preceding and following related objects have an "or" relationship.

[0031] It should also be understood that the description of the various embodiments in this invention emphasizes the differences between the various embodiments, and the similarities or similarities can be referred to each other. For the sake of brevity, they will not be described in detail.

[0032] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0033] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0034] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0035] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0036] The embodiments of this invention can be applied to electronic devices such as terminal devices, computer systems, and servers, and can operate together with a wide range of other general-purpose or special-purpose computing system environments or configurations. Well-known examples of terminal devices, computing systems, environments, and / or configurations suitable for use with electronic devices such as terminal devices, computer systems, and servers include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments including any of the above systems, etc.

[0037] Electronic devices such as terminal devices, computer systems, and servers can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. Computer systems / servers can be implemented in distributed cloud computing environments, where tasks are executed by remote processing devices linked through communication networks. In distributed cloud computing environments, program modules can reside on local or remote computing system storage media, including storage devices.

[0038] Exemplary methods Figure 1 This is a flowchart illustrating a method for constructing a graph-physical information network for power system state analysis, provided by an exemplary embodiment of the present invention. This embodiment can be applied to electronic devices, such as… Figure 1 As shown, the method 100 for constructing a graph-physical information network for power system state analysis includes the following steps: Step 101: Based on the time-varying topology of the power network and historical measurement data of nodes, generate high-dimensional embedding features of nodes that integrate spatiotemporal dependencies through the spatiotemporal graph attention module. Step 102: Input the high-dimensional embedding features of the nodes into the hierarchical neural network predictor module, and output at least one of the following: the predicted electrical state of the nodes of the power system, the predicted physical field of the equipment, and the predicted health and aging state of the equipment. Step 103: Based on the predicted values ​​of node electrical state, the predicted values ​​of equipment internal physical field, and the predicted values ​​of equipment health aging state, construct a physical constraint loss term that includes at least one of the following: node power balance residual, equipment heat conduction partial differential equation residual, and equipment aging dynamics ordinary differential equation residual. This term, together with the data fidelity loss term, constitutes a multi-physics deep coupling loss function. Step 104: Through the gradient dynamic balancing and conflict resolution module, the gradients of each loss term in the multi-physics deep coupling loss function are dynamically weighted and conflict resolved. The network parameters are updated based on the adjusted gradients until the model converges, generating a graphical physical information network for power system state analysis.

[0039] Specifically, addressing the technical problems existing in the background art, this invention aims to design a novel high-dimensional dynamically coupled graph physical information network for power system state analysis. This network handles the topology and dynamic changes of the power network by introducing a dynamic graph embedding module, embeds AC power flow equations and equipment thermodynamic equations into the loss function using a multi-physics coupling module, and optimizes the learning process through an adaptive weighted loss mechanism.

[0040] The complete technical solution provided by this invention I. Problem Definition and Symbols Consider power grid ,in It is a set of nodes. It is a moment edge set, It is a node feature matrix (such as voltage) Phase angle Injected power , It is the edge feature matrix, where F N This represents the number of feature dimensions of a node. F E This represents the feature dimension of the edge. The goal is to predict a period of time in the future. Inside: 1. High-dimensional equipment status: For example, the three-dimensional temperature field inside transformer windings / cores. ,in Spatial coordinates; cumulative aging of key components .

[0041] 2. Dynamic power flow and system status: Bus voltage ,in This represents the predicted value of the bus voltage amplitude. This represents the predicted value of the bus voltage phase angle; branch power flow. ,in This represents the predicted active power value. This represents the predicted reactive power value, and even the system frequency. Or related dynamic variables.

[0042] II. High-Dimensional Dynamically Coupled Graph Physical Information Network Framework (1) Input layer: historical high-frequency measurement data, network topology sequence (Time-varying adjacency matrix or Laplace matrix), equipment geometric model, material parameters, boundary condition functions, and external disturbance sequence.

[0043] (2) Spatiotemporal Graph Attention Module: The core is the fusion of dynamic graph convolutional network and temporal self-attention mechanism. It learns the high-dimensional embedding representations of nodes and edges in time and space. ,in D h represents the dimension of the implicit feature vectors of nodes and edges.

[0044] (3) Hierarchical Neural Network Predictor Module: A set of interconnected MLPs or CNN-MLP hybrid structures for field prediction. These MLPs receive embeddings... (node (embedding), and time Spatial coordinates External control signals The output consists of various physical quantities: voltage. Phase angle Spatial coordinates Cumulative aging .

[0045] (4) Gradient dynamic equilibrium and conflict resolution loss function: includes data fidelity terms and multiple coupled physical constraint terms (PDEs / ODEs residuals).

[0046] III. Spatiotemporal Graph Attention Module This spatiotemporal graph attention module aims to autonomously learn and extract deep, high-dimensional embedded features from complex power network data with time-varying topology, node states, and edge attributes by leveraging the synergistic effect of dynamic convolution and multidimensional attention mechanisms based on graph theory. These features comprehensively reflect the multi-scale, nonlinear spatiotemporal dynamic correlations and potential evolutionary patterns of the system.

[0047] Define the graph Laplace operator Or normalized form ,in It is a degree matrix. It is an adjacency matrix. It is the identity matrix. Graph convolution operations can be defined in the Fourier domain: in It is the node feature matrix. , yes eigendecomposition, It is an eigenvalue diagonal matrix. It is the eigenvector matrix. It represents the Hadamardi (or Hadama) stack. The parameter is A learnable spectral filter. Due to the large computational cost of eigenvalue decomposition, this invention employs Chebyshev polynomial approximation. : in Denotes the order of the Chebyshev polynomial expansion. Indicates the first One weight parameter to be learned, express Chebyshev polynomial of the first kind, This represents the diagonal matrix of eigenvalues ​​of the normalized graph Laplacian operator. To capture temporal dynamics, the output sequence of the spatiotemporal graph attention module is used. ( (The historical time window) is input to a temporal self-attention layer or a Transformer encoder layer: in This represents the attention mechanism. This represents the query matrix in the Transformer. V represents the key matrix in the Transformer, and V represents the value matrix in the Transformer. Represents the feature dimension. The final output node. At any moment The context embedding is This embedding integrates historical information and current graph structure information.

[0048] IV. Hierarchical Neural Network Predictor The hierarchical neural network predictor module is the core output unit of the high-dimensional dynamically coupled graph-physical information network framework. It is responsible for embedding the deep, high-dimensional node features learned by the spatiotemporal graph attention module. In addition to other relevant external inputs, these are mapped to accurate predictions of multi-scale, multi-physical field variables in the power system. The "layered" characteristic of the layered neural network predictor is reflected in its modular design and potential hierarchical dependencies, meaning that the prediction of some physical quantities may depend on the preliminary prediction results of other physical quantities, forming a computational order or information flow. The layered neural network predictor consists of a set of sub-neural networks with specific functions. Each sub-network (usually a multilayer perceptron MLP, or a CNN-MLP hybrid structure for field prediction, Fourier neural operator FNO, etc.) is optimized to predict a specific physical quantity or a set of related physical quantities.

[0049] (1) Node electrical state predictor The hierarchical neural network predictor module contains a set of core subnetworks dedicated to predicting key electrical state variables of each node in the power network, primarily voltage magnitude and phase angle. Specifically, for any node in the network... Its voltage amplitude Consists of a dedicated voltage predictor network Output. The input to this network is a fusion of nodes generated by the spatiotemporal graph attention module. High-dimensional spatiotemporal embedding features Time information encoded by periodic or Fourier features and possible external control signals Ultimately, through the learned network parameters The predicted voltage amplitude is obtained by mapping: Similarly, the voltage phase angle of the node By another phase angle predictor network Generate, its input structure is the same as Maintain consistency, that is in, Indicates the amount of external control signal input. This represents the parameters of the neural network. Special attention must be paid to its inherent characteristics when processing phase angle prediction. The periodicity of these predictions may necessitate specific techniques in the network output layer design or post-processing stages to ensure the physical meaning and numerical stability of the prediction results, and to avoid unnecessary phase transitions. These electrical state predictors form the basis for system power flow and dynamic analysis.

[0050] (2) Internal physical field predictor of the equipment With transformer Internal three-dimensional temperature field For example (where) (Intra-device spatial coordinates) Here It can be an embedding of bus nodes connected to the transformer, or a graph embedding specifically learned for the transformer equipment itself. Spatial coordinates High-dimensional Fourier feature encoding is crucial for helping standard MLPs learn high-frequency functions. This represents time information encoded by periodic or Fourier features. This represents the power loss of the device, which can be the output of another sub-network, reflecting the hierarchical dependencies within the hierarchical neural network predictor. Indicates the first Parameters of the internal three-dimensional temperature field prediction subnetwork of the device. The structure can use small convolutional networks or graph neural networks to process spatial information.

[0051] (3) Equipment health and aging status predictor For transformers Cumulative aging metrics such as equivalent aging based on degree of aggregation (DP) The calculation formula is as follows: in, From detailed temperature field The key hotspot temperatures extracted from them Indicates the first The cumulative aging of the equipment is predicted by sub-network parameters. This can be achieved by... The output is obtained through spatial pooling or specific point querying. This represents the cumulative effect of historical heat load, which is an important input for some aging models. This represents the initial aging state. This subnetwork essentially learns to solve or approximate the integral form of the Arrhenius equation, the aging model.

[0052] Through the meticulously designed modular structure, feature fusion mechanism, and output layer adjustments tailored to specific physical quantities, the hierarchical neural network predictor module can effectively transform the abstract spatiotemporal features learned from dynamic graphs into comprehensive and accurate predictions of the multifaceted, multi-physical state of the power system. This hierarchical and specialized design is key to the high-dimensional dynamic coupled graph physical information network's ability to handle complex coupled systems and generate high-fidelity predictions.

[0053] V. Loss Function for Deep Coupling of Multiphysics Fields The core driving force of the high-dimensional dynamic coupling graph physical information network model comes from its carefully designed total loss function for deep multi-physics coupling. This total loss function... It not only includes data fidelity items that measure the degree of agreement between model predictions and actual observed data. More importantly, it uses a series of physical constraints. This involves embedding profound principles of physics into the training process of neural networks. These physical constraints stem from the partial and ordinary differential equations that must be followed in the operation of power systems and the evolution of equipment states. (Total loss function) It can be represented as: in , , and These are the weighting coefficients for each loss, dynamically adjusted by the GDB-CR mechanism. It refers to the number of types of core physical constraints.

[0054] (1) Data fidelity loss item If available historical or real-time measurement data (e.g., voltage provided by SCADA, PMU) Phase angle Power or the temperature of the device measured by the sensor. (etc.), the data fidelity loss term penalizes the deviation between the model's predicted values ​​and these true observations. It is typically expressed as mean squared error (MSE) or a variant thereof: in It is the number of data points. Iterate through all available measurement types. The model is for the first The data point of the th data point Predicted output of physical quantities It corresponds to the actual value. These are the weights of different types of measurement data.

[0055] (2) Physical constraints 1. Dynamic AC power flow and line physical constraints Node power balance residuals. For each bus in the network. and any configurable time point The active power it injects and reactive power The power flow equations must be satisfied. The residual is defined as: in , These are the voltage amplitude and phase angle predicted by the HNNP module. These are elements of the admittance matrix of the time-varying network. It is a known power generation or load.

[0056] Residual of dynamic voltage-current relationship in transmission lines. For connection nodes. and The transmission lines (parameters are) The relationship between its voltage drop and current can be approximately described by the following differential equation: in It is the predicted current flowing through the line, which can be derived from the predicted power and voltage or output by a dedicated current prediction subnetwork, and its time derivative. Obtained through automatic differentiation. Therefore, this portion of the physical loss is: in It is the total number of configuration points used to evaluate electrical dynamic residuals.

[0057] 2. Physical constraints on three-dimensional heat conduction of the equipment The key component of the transformer is the winding. For example, its internal temperature field (in The evolution of (in spatial coordinates) follows the partial differential equation of heat conduction: in , , These are density, specific heat capacity, and thermal conductivity, respectively; these material properties can depend on spatial location. and predicted temperature Volumetric heat generation rate Mainly determined by the predicted current density and resistivity (or its reciprocal conductivity) as a function of temperature )Decide: The residual of the PDE is defined as follows: The heat conduction loss is then: The integral is obtained through the spatial domain. and time point Monte Carlo approximation is performed using upsampled configuration points. It refers to the number of volume configuration points.

[0058] 3. Kinetic and physical constraints of equipment insulation aging Based on the Arrhenius model, the device Aging indicators such as equivalent cumulative aging or loss of polymerization degree of insulating material. The dynamic evolution is described by the following ordinary differential equation: in It is from the predicted detailed temperature field The key hotspot temperatures extracted from them It refers to the pre-factor. It is activation energy. This is the universal gas constant. The derivative of this ODE is calculated using automatic differentiation. Its residual is defined as: The aging kinetic loss is then: in It is the number of time allocation points. It is a collection of equipment used for aging assessment.

[0059] (3) Boundary conditions and initial conditions constraints For the PDEs and ODEs mentioned above, the uniqueness and physical meaning of their solutions also depend on the corresponding boundary conditions and initial conditions.

[0060] 1. Thermal conduction boundary conditions for It is necessary to have it at the component boundary. Apply conditions such as Dirichlet, Neumann, or Robin conditions to the boundary, such as convective heat transfer boundary conditions. The residual is: in It is the outer normal vector of the boundary. It is a temperature-dependent convective heat transfer coefficient. It is the temperature of the surrounding fluid medium.

[0061] 2. Initial conditions The device at the initial moment Temperature distribution and aging state Must be with known initial values and Consistent. The residuals are calculated as follows: Overall It is a weighted average of the sum of squares of the residuals of all these boundary and initial conditions, evaluated at the corresponding boundary and initial time points: VI. Gradient Dynamic Equilibrium and Conflict Resolution Loss Function This module aims to address the issues of inconsistent convergence speeds and conflicting gradient directions among loss terms in multiphysics and multi-objective optimization. This invention employs a gradient game theory approach based on Nash equilibrium, or a variant of the multi-objective gradient descent algorithm. Let the total loss function in the multiphysics deep coupling loss function be... ,in The rest are physical constraints. The objective solution is to find a set of weights. and network parameters .

[0062] 1. Gradient normalization and benchmarking In each training batch Calculate the gradient of each loss term. Calculate the moving average L2 norm of the gradient. The initial weights can be set to... .

[0063] 2. Gradient conflict detection and Pareto optimization For a set of tasks, if there exists a common descent direction... Makes all , If so, these tasks can be improved simultaneously; otherwise, conflicts exist. MGDA aims to find a Pareto-optimal direction for updating shared parameters. This can be achieved by solving the following quadratic programming problem (for all physical constraint gradients). Data gradient (Handle separately or give dominant position): Solution A combined gradient direction is given. .These Can be used as dynamic weights . The weights of the data items can be set to 1 or a large constant, or their gradients can be made equal to the gradients of the data items. Maximize the inner product.

[0064] 3. Dynamic weight adjustment mechanism In each training step Calculate each loss term For parameters gradient Calculate the exponential moving average of the gradient norm of the data loss. For each physical loss item Calculate its maximum gradient component Dynamically adjust weights : Here Used to handle gradient conflicts. The specific calculation steps are as follows: Calculate the cosine similarity matrix between all pairs of physical gradients. Define the degree of conflict: but It can be obtained from the following formula: in This is a hyperparameter that controls the penalty for gradient conflicts. This means that if the gradient of one physical loss term conflicts significantly with the gradient directions of other physical loss terms, its weight will be appropriately reduced in order to seek the optimal overall optimization path.

[0065] Therefore, this invention proposes a novel high-dimensional dynamically coupled graph-physical information network for power system state analysis. This network handles the topology and dynamic changes of the power network by introducing a dynamic graph embedding module, embeds AC power flow equations and equipment thermodynamic equations into the loss function using a multi-physics coupling module, and optimizes the learning process through an adaptive weighted loss mechanism. It has the following beneficial effects: (1) Unified high-fidelity prediction framework: An end-to-end framework was constructed, consisting of a collaborative spatiotemporal graph attention module, a hierarchical neural network predictor module, a multi-physics field deep coupling loss function, and a gradient dynamic balance and conflict resolution loss function, which enabled comprehensive and high-precision prediction of the multi-scale state and long-term aging trend of the power system.

[0066] (2) Spatiotemporal graph attention module: A new mechanism is proposed that integrates dynamic graph theory and practical attention to efficiently capture the complex high-dimensional spatiotemporal dependencies in the evolution of power networks.

[0067] (3) Hierarchical neural network predictor: It adopts a modular and hierarchical prediction architecture to achieve accurate decoupling and collaborative prediction of multiple types and scales of physical quantities such as electrical state, internal physical field, and equipment aging, effectively improving the prediction accuracy and generalization ability of different physical quantities. At the same time, it utilizes hierarchical dependency relationships to reuse intermediate prediction results and improve computational efficiency.

[0068] (4) Multiphysics deep coupling loss function: Within the unified PINN framework, the deep coupling and collaborative prediction of enhanced electrical dynamics, three-dimensional high-fidelity thermal field inside the equipment and physical aging model are realized for the first time.

[0069] (5) Gradient dynamic balancing and conflict resolution loss function: Design an advanced adaptive loss strategy to ensure the robustness and physical consistency of model training by dynamically balancing the magnitude of multi-source gradients and actively resolving gradient conflicts between physical constraints.

[0070] Exemplary device Figure 2 This is a schematic diagram of a graph-physical information network construction device for power system state analysis provided in an exemplary embodiment of the present invention. Figure 2 As shown, the device 200 includes: The first generation module 210 is used to generate high-dimensional embedding features of nodes that integrate spatiotemporal dependencies based on the time-varying topology of the power network and historical measurement data of nodes through the spatiotemporal graph attention module. The output module 220 is used to input the high-dimensional embedded features of the nodes into the hierarchical neural network predictor module, and output at least one of the following: the predicted electrical state of the nodes of the power system, the predicted physical field of the equipment, and the predicted health and aging state of the equipment. Module 230 is used to construct a physical constraint loss term based on the predicted values ​​of node electrical state, the predicted values ​​of equipment internal physical field, and the predicted values ​​of equipment health aging state. This term includes at least one type of physical constraint loss term, which is composed of the residual of node power balance, the residual of equipment heat conduction partial differential equation, and the residual of equipment aging dynamics ordinary differential equation. Together with the data fidelity loss term, it constitutes a multi-physics deep coupling loss function. The second generation module 240 is used to dynamically adjust the weights and resolve conflicts of the gradients of each loss term in the deep coupling loss function of multiphysics through the gradient dynamic balancing and conflict resolution module, and update the network parameters based on the adjusted gradients until the model converges, thereby generating a graphical physical information network for power system state analysis.

[0071] Optionally, the first generation module 210 includes: Based on the time-varying adjacency matrix or Laplace matrix of the power network, spatial topological features are extracted using graph convolution operation. The graph convolution operation approximates the spectral domain filter with Chebyshev polynomial to generate node spatial features. Input the historical time series corresponding to the node spatial features into the temporal self-attention layer or Transformer encoder layer to capture the temporal dependencies within the historical time window and generate node temporal features. By fusing node spatial features and node temporal features, high-dimensional embedding features of nodes are generated.

[0072] Optionally, the hierarchical neural network predictor module includes: The node electrical state predictor is used to output the voltage amplitude and voltage phase angle of the node as the predicted value of the node electrical state based on the node's high-dimensional embedding features and time coding information. The device internal physics field predictor is used to output the three-dimensional temperature field distribution inside the device as the predicted value of the device internal physics field based on the node high-dimensional embedding features, spatial coordinate encoding and device loss power information. The equipment health and aging status predictor is connected to the equipment's internal physical field predictor. It is used to output the equipment's cumulative aging degree as the predicted value of the equipment's health and aging status based on the hot spot temperature and historical heat load accumulation information extracted from the predicted value of the equipment's internal physical field.

[0073] Optionally, the physical constraint loss term specifically includes: Based on the voltage amplitude and voltage phase angle in the predicted electrical state of the nodes, the balance residuals of injected active power and reactive power are calculated, and physical constraints for AC power flow are constructed. Based on the three-dimensional temperature field distribution in the predicted physical field of the equipment, the residual between it and the partial differential equation of heat conduction is calculated, and the physical constraint term of heat conduction is constructed. Based on the cumulative aging degree in the predicted value of equipment health aging state, the residual between it and the Arrhenius aging ordinary differential equation is calculated, and the physical constraint term of aging kinetics is constructed.

[0074] Optionally, the device 200 further includes: a superposition module, used to further superimpose the residual of the equipment heat conduction boundary conditions and the residual of the initial temperature distribution and aging state into the physical constraint loss term, as boundary and initial condition constraint terms.

[0075] Optionally, the second generation module 240 uses a gradient dynamic balancing and conflict resolution module to dynamically adjust the weights and resolve conflicts of the gradients of each loss term in the multiphysics deep coupling loss function, and updates the network parameters based on the adjusted gradients, including: Calculate the gradient of each loss term in the multiphysics deep coupling loss function with respect to the network parameters, and obtain the gradient vector of each loss term; Calculate the cosine similarity between each pair of gradient vectors, construct the gradient conflict matrix, and determine the gradient conflict degree of each loss term based on the gradient conflict matrix; The weights of each loss term in the multiphysics deep coupling loss function are dynamically adjusted based on the gradient conflict degree, so that the loss term with a larger conflict degree receives a smaller weight coefficient, in order to seek the Pareto optimal joint descent direction. The joint gradient direction is obtained by weighted combination of the gradients of each loss term after weight adjustment, and the network parameters are updated accordingly.

[0076] Optionally, the weights can be dynamically adjusted. The calculation formula is: Among them, the exponential moving average of the data loss gradient norm The calculation expression is: In the formula, Indicates the number of training iterations for the model; Represents the gradient of neural network parameters; Indicates the first The L2 norm of the data loss gradient during step training; This represents the exponential moving average decay coefficient. Each physical loss item Maximum gradient component The calculation expression is: In the formula, The expression for calculating the gradient conflict parameter is as follows: In the formula, Here are the hyperparameters; the conflict degree is... , This represents the cosine similarity between physical gradients.

[0077] Exemplary electronic devices Figure 3 This is the structure of an electronic device provided in an exemplary embodiment of the present invention. For example... Figure 3 As shown, the electronic device 30 includes one or more processors 31 and memory 32.

[0078] The processor 31 may be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions.

[0079] The memory 32 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 31 may execute the program instructions to implement the methods of the software programs of the various embodiments of the present invention described above, and / or other desired functions. In one example, the electronic device may also include an input device 33 and an output device 34, these components being interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0080] In addition, the input device 33 may also include, for example, a keyboard, a mouse, etc.

[0081] The output device 34 can output various information to the outside. The output device 34 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0082] Of course, for the sake of simplicity, Figure 3 Only some of the components of this electronic device relevant to the present invention are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device may include any other suitable components depending on the specific application.

[0083] Exemplary computer program products and computer-readable storage media In addition to the methods and apparatus described above, embodiments of the present invention may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of the present invention described in the "Exemplary Methods" section above.

[0084] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of the present invention. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0085] Furthermore, embodiments of the present invention may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps of the methods according to various embodiments of the present invention described in the "Exemplary Methods" section above.

[0086] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0087] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the aforementioned specific details.

[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0089] The block diagrams of devices, systems, devices, and systems involved in this invention are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, systems, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0090] The methods and systems of the present invention may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of the present invention are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, the present invention may also be implemented as a program recorded on a recording medium, the program comprising machine-readable instructions for implementing the methods according to the present invention. Thus, the present invention also covers recording media storing programs for performing the methods according to the present invention.

[0091] It should also be noted that in the systems, apparatus, and methods of the present invention, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered equivalents of the present invention. The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the invention. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the invention. Therefore, the invention is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0092] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the invention to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for constructing a graph-based physical information network for power system state analysis, characterized in that, include: Based on the time-varying topology of the power network and historical measurement data of nodes, a high-dimensional embedding feature of nodes that integrates spatiotemporal dependencies is generated through a spatiotemporal graph attention module. The high-dimensional embedded features of the nodes are input into the hierarchical neural network predictor module, which outputs at least one of the following: the predicted electrical state of the nodes in the power system, the predicted physical field of the equipment, and the predicted health and aging state of the equipment. Based on the predicted values ​​of the node electrical state, the predicted values ​​of the internal physical field of the equipment, and the predicted values ​​of the health and aging state of the equipment, a physical constraint loss term is constructed, which includes at least one of the following: node power balance residual, equipment heat conduction partial differential equation residual, and equipment aging dynamics ordinary differential equation residual. This term, together with the data fidelity loss term, constitutes a multi-physics deep coupling loss function. The gradient dynamic balancing and conflict resolution module dynamically adjusts the gradients of each loss term in the deep coupling loss function of the multiphysics field and resolves conflicts. The network parameters are then updated based on the adjusted gradients until the model converges, generating a graphical physical information network for power system state analysis.

2. The method according to claim 1, characterized in that, Based on the time-varying topology of power networks and historical node measurement data, a spatiotemporal graph attention module is used to generate high-dimensional node embedding features that fuse spatiotemporal dependencies, including: Based on the time-varying adjacency matrix or Laplace matrix of the power network, spatial topological features are extracted using graph convolution operation. The graph convolution operation approximates the spectral domain filter with Chebyshev polynomial to generate node spatial features. The historical time sequence corresponding to the node spatial features is input into the temporal self-attention layer or the Transformer encoder layer to capture the temporal dependencies within the historical time window and generate node temporal features. The node spatial features and node temporal features are fused to generate the node high-dimensional embedding features.

3. The method according to claim 1, characterized in that, The hierarchical neural network predictor module includes: A node electrical state predictor is used to output the voltage amplitude and voltage phase angle of the node as the predicted value of the node electrical state based on the node's high-dimensional embedding features and time coding information. The device internal physics field predictor is used to output the three-dimensional temperature field distribution inside the device as the predicted value of the device internal physics field based on the node's high-dimensional embedding features, spatial coordinate encoding, and device loss power information. The equipment health aging state predictor is connected to the equipment internal physical field predictor. It is used to output the cumulative aging degree of the equipment as the predicted value of the equipment health aging state based on the hot spot temperature and historical heat load accumulation information extracted from the predicted value of the equipment internal physical field.

4. The method according to claim 1, characterized in that, The physical constraint loss term specifically includes: Based on the voltage amplitude and voltage phase angle in the predicted electrical state of the node, the balance residual between the injected active power and reactive power of the node is calculated, and the physical constraint term of AC power flow is constructed. Based on the three-dimensional temperature field distribution in the predicted physical field values ​​inside the device, the residual between the temperature field distribution and the partial differential equation of heat conduction is calculated, and a physical constraint term for heat conduction is constructed. Based on the cumulative aging degree in the predicted health aging state of the equipment, the residual between it and the Arrhenius aging ordinary differential equation is calculated, and the aging kinetic physical constraint term is constructed.

5. The method according to claim 4, characterized in that, Also includes: In the physical constraint loss term, the residual of the equipment heat conduction boundary condition and the residual of the initial temperature distribution and aging state are further superimposed as the boundary and initial condition constraint term.

6. The method according to claim 1, characterized in that, The gradient dynamic balancing and conflict resolution module dynamically adjusts the gradients of each loss term in the multiphysics deep coupling loss function and resolves conflicts, and updates the network parameters based on the adjusted gradients, including: Calculate the gradient of each loss term in the multiphysics deep coupling loss function with respect to the network parameters to obtain the gradient vector of each loss term; Calculate the cosine similarity between each pair of gradient vectors, construct a gradient conflict matrix, and determine the gradient conflict degree of each loss term based on the gradient conflict matrix; The weights of each loss term in the multiphysics deep coupling loss function are dynamically adjusted based on the gradient conflict degree, so that the loss term with a larger conflict degree receives a smaller weight coefficient, in order to seek the Pareto optimal joint descent direction. The joint gradient direction is obtained by weighted combination of the gradients of each loss term after weight adjustment, and the network parameters are updated accordingly.

7. The method according to claim 6, characterized in that, Dynamically adjusted weights The calculation formula is: Among them, the exponential moving average of the data loss gradient norm The calculation expression is: In the formula, Indicates the number of training iterations for the model; Represents the gradient of neural network parameters; Indicates the first The L2 norm of the data loss gradient during step training; This represents the exponential moving average decay coefficient. Each physical loss item Maximum gradient component The calculation expression is: In the formula, The expression for calculating the gradient conflict parameter is as follows: In the formula, Here are the hyperparameters; the conflict degree is... , This represents the cosine similarity between physical gradients.

8. A graph-physical information network construction device for power system state analysis, characterized in that, include: The first generation module is used to generate high-dimensional embedding features of nodes that integrate spatiotemporal dependencies based on the time-varying topology of the power network and historical measurement data of nodes through the spatiotemporal graph attention module. The output module is used to input the high-dimensional embedded features of the node into the hierarchical neural network predictor module, and output at least one of the following: the predicted electrical state of the node in the power system, the predicted physical field of the equipment, and the predicted health and aging state of the equipment. The construction module is used to construct, based on the predicted values ​​of the node electrical state, the predicted values ​​of the internal physical field of the equipment, and the predicted values ​​of the health and aging state of the equipment, at least one type of physical constraint loss term is constructed, which includes the node power balance residual, the equipment heat conduction partial differential equation residual, and the equipment aging dynamics ordinary differential equation residual, and together with the data fidelity loss term, constitutes a multi-physics deep coupling loss function. The second generation module is used to dynamically adjust the weights and resolve conflicts of the gradients of each loss term in the deep coupling loss function of the multiphysics field through the gradient dynamic balancing and conflict resolution module, and update the network parameters based on the adjusted gradients until the model converges, thereby generating a graphical physical information network for power system state analysis.

9. The apparatus according to claim 8, characterized in that, The first generation module includes: Based on the time-varying adjacency matrix or Laplace matrix of the power network, spatial topological features are extracted using graph convolution operation. The graph convolution operation approximates the spectral domain filter with Chebyshev polynomial to generate node spatial features. The historical time sequence corresponding to the node spatial features is input into the temporal self-attention layer or the Transformer encoder layer to capture the temporal dependencies within the historical time window and generate node temporal features. The node spatial features and node temporal features are fused to generate the node high-dimensional embedding features.

10. The apparatus according to claim 8, characterized in that, The hierarchical neural network predictor module includes: A node electrical state predictor is used to output the voltage amplitude and voltage phase angle of the node as the predicted value of the node electrical state based on the node's high-dimensional embedding features and time coding information. The device internal physics field predictor is used to output the three-dimensional temperature field distribution inside the device as the predicted value of the device internal physics field based on the node's high-dimensional embedding features, spatial coordinate encoding, and device loss power information. The equipment health aging state predictor is connected to the equipment internal physical field predictor. It is used to output the cumulative aging degree of the equipment as the predicted value of the equipment health aging state based on the hot spot temperature and historical heat load accumulation information extracted from the predicted value of the equipment internal physical field.

11. The apparatus according to claim 8, characterized in that, The physical constraint loss term specifically includes: Based on the voltage amplitude and voltage phase angle in the predicted electrical state of the node, the balance residual between the injected active power and reactive power of the node is calculated, and the physical constraint term of AC power flow is constructed. Based on the three-dimensional temperature field distribution in the predicted physical field values ​​inside the device, the residual between the temperature field distribution and the partial differential equation of heat conduction is calculated, and a physical constraint term for heat conduction is constructed. Based on the cumulative aging degree in the predicted health aging state of the equipment, the residual between it and the Arrhenius aging ordinary differential equation is calculated, and the aging kinetic physical constraint term is constructed.

12. The apparatus according to claim 11, characterized in that, Also includes: The superposition module is used to further superimpose the residual of the equipment heat conduction boundary conditions and the residual of the initial temperature distribution and aging state into the physical constraint loss term, as boundary and initial condition constraint terms.

13. The apparatus according to claim 8, characterized in that, The second generation module uses a gradient dynamic balancing and conflict resolution module to dynamically adjust the weights and resolve conflicts of the gradients of each loss term in the multiphysics deep coupling loss function, and updates the network parameters based on the adjusted gradients, including: Calculate the gradient of each loss term in the multiphysics deep coupling loss function with respect to the network parameters to obtain the gradient vector of each loss term; Calculate the cosine similarity between each pair of gradient vectors, construct a gradient conflict matrix, and determine the gradient conflict degree of each loss term based on the gradient conflict matrix; The weights of each loss term in the multiphysics deep coupling loss function are dynamically adjusted based on the gradient conflict degree, so that the loss term with a larger conflict degree receives a smaller weight coefficient, in order to seek the Pareto optimal joint descent direction. The joint gradient direction is obtained by weighted combination of the gradients of each loss term after weight adjustment, and the network parameters are updated accordingly.

14. The apparatus according to claim 13, characterized in that, Dynamically adjusted weights The calculation formula is: Among them, the exponential moving average of the data loss gradient norm The calculation expression is: In the formula, Indicates the number of training iterations for the model; Represents the gradient of neural network parameters; Indicates the first The L2 norm of the data loss gradient during step training; This represents the exponential moving average decay coefficient. Each physical loss item Maximum gradient component The calculation expression is: In the formula, The expression for calculating the gradient conflict parameter is as follows: In the formula, Here are the hyperparameters; the conflict degree is... , This represents the cosine similarity between physical gradients.

15. A computer-readable storage medium, characterized in that, The storage medium stores a computer program for performing the method described in any one of claims 1-7.

16. An electronic device, characterized in that, The electronic device includes: processor; Memory used to store the processor's executable instructions; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the method described in any one of claims 1-7.

Images