Power system modeling method and system based on conservation information embedded graph learning, medium and equipment
The method addresses scalability and interpretability issues in spatiotemporal dynamical systems by using conservation information embedded graph learning to predict nonlinear partial differential equations, achieving improved accuracy and efficiency in complex scenarios.
Patent Information
- Application Number
- CN202410958839.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-12-06
AI Technical Summary
Traditional physical techniques struggle with complex real-world problems due to poor scalability, sensitivity to parameters, high computational cost, and limited interpretability in spatiotemporal dynamical systems, while data-centric approaches like neural networks require extensive data and lack interpretability.
A dynamic system modeling method using conservation information embedded graph learning, which discretizes the physical field, embeds symmetry into graph neural networks, and uses message passing mechanisms to learn continuous equations, incorporating initial/boundary conditions and physical prior knowledge to predict nonlinear partial differential equations.
The method reduces the need for extensive training data, enhances interpretability, and improves scalability and accuracy in predicting complex spatiotemporal dynamics, particularly in irregular domains.
Smart Images

Figure CN119089632A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power system modeling, and in particular to a power system modeling method, system, medium and equipment based on conservation information embedding graph learning. Background Art
[0002] Accurate modeling and simulation of spatiotemporal dynamical systems play an important role in many scientific and engineering applications, and can be used to better understand, design, optimize and control target systems. However, the complexity of real-world problems poses huge challenges that are difficult to solve with traditional physical techniques. In particular, this problem is particularly prominent when the prior physical knowledge is incomplete or repeated forward analysis is required to assimilate the data.
[0003] Many classical methods assume that the problem scenario is in a good physical condition (under certain initial and boundary conditions) and try to solve such system problems by solving partial differential equations (PDEs). In general, the solution process may be achieved by selecting an appropriate numerical solver. For example, existing climate models are numerically solved on a spatiotemporal grid in the context of data sampling from the perspectives of physical, biological, and chemical principles, thus laying an important foundation for predicting climate change and extreme weather events. However, this type of method has poor scalability to the problem dimension, is sensitive to parameters, and has a very large search space, which usually requires a lot of computational cost for the solution process.
[0004] The advent of the era of big data on the Internet has opened up a new path to address these challenges: using the power of machine learning (e.g., convolutional neural networks, recurrent neural networks, graph neural networks, variational autoencoders, pre-trained large models, etc.) in a data-centric approach to model and solve such problems. However, in most scientific and engineering problems, it is expensive to obtain these data sets from experiments or simulations. Since these methods are highly dependent on the rich information implicit in a large number of data sets, the accuracy of the model's predictions or solutions is severely limited. In addition, due to the over-parameterized black box nature of the model, the interpretability of such methods is relatively vague, and it is impossible to interpret specific physical phenomena from a deep or mechanistic level; in addition, the extrapolation and generalization capabilities of the model are generally poor.
[0005] In recent years, the method of embedding known prior knowledge into physics-informed neural networks (PINNs) has shown the potential to alleviate these basic problems. However, since the simulation mechanism of the network for physical fields is based on point-by-point global approximation, in the case of complex geometry and high-dimensional computational domains, this method requires randomly extracting a large number of configuration points to cover the entire space-time computational domain and ensure model accuracy. This learning strategy leads to low training efficiency, poor convergence, and the need for explicit control equations.
[0006] Neural operators provide an alternative approach to learn the latent Green's function of PDEs. The method does not require explicit prior knowledge of the PDE formula. Neural operator learning methods, including DeepONet, Fourier Neural Operator (FNO) and their variants, have made significant leaps in approximating solutions to PDE systems. However, the generalizability of these models also depends heavily on the abundance of labeled training data. When the complexity of the studied system remains high (e.g., multidimensional spatiotemporal dynamics in irregular domain geometry), such methods may perform poorly and show insufficient scalability. Summary of the invention
[0007] In view of the above problems, the purpose of the present invention is to provide a power system modeling method, system, medium and equipment based on conservation information embedding graph learning, which can effectively reduce the amount of training data and effectively identify potential power mechanisms.
[0008] To achieve the above-mentioned objectives, in the first aspect, the technical solution adopted by the present invention is: a dynamic system modeling method based on conservation information embedded graph learning, which includes: discretizing the physical field in time and space, embedding symmetry into the underlying architecture of the graph neural network to learn continuous equations that are universal in time and space; the initial / boundary conditions in the forward problem are "strongly encoded" as part of the model input and output; when the explicit PDE formula is completely unknown or partially missing, the physical prior knowledge in a macroscopic sense is embedded in the neural network structure through different encoding methods to serve the simulation of different physical systems; based on the topological structure relationship that naturally has strong interpretability, a message passing mechanism is used to learn the high-dimensional feature representation of different information in the physical field, which is used to learn and predict the long-term behavior of nonlinear partial differential equations, so as to obtain a conservation information embedded graph neural network for learning dynamics in complex scenarios.
[0009] Furthermore, conservation information is embedded into the graph neural network, including: The spatial module, with the help of the multi-grid concept, generates a network space structure with information of different scales and converts the physical state into a low-dimensional graph representation; The encoder elevates the physical state represented by the low-dimensional graph into latent features represented in a high-dimensional space; The SymMPNN module processes the spatial and channel information of the local neighborhood of each node on the graph through aggregation operations to handle high-dimensional latent features; The temporal module regards the multi-layer message passing mechanism of the SymMPNN module as an implicit transmission of the concept of time, and further updates the high-dimensional latent features through a learnable time integration network; The decoder maps the updated high-dimensional latent features back to the original physical domain.
[0010] Furthermore, a multi-grid graph is defined, including: Using the mesh generation algorithm, the initial mesh is constructed. represents a graph where represents a node set, Represents an edge set; definition is the initial graph, where ; Random Drop Generate a new graph with nodes of preset proportions in as the subsequent initial graph; and repeat the operation and aggregate all graphs into ; The generated graph Organize by layers, with the graph with the most nodes as the top layer and the graph with the least nodes as the bottom layer; The entire picture Conceptualized as the union of all edges and nodes, the new graph The node is always A subset of nodes is constructed, and the edge reconstruction process can capture information across extended distances, resulting in a multi-grid distribution.
[0011] Furthermore, the physical state represented by the low-dimensional graph is promoted to the potential features represented in the high-dimensional space, including: Encoded initial node features Contains The physical state of the node , the corresponding position features and node type; where represents the dimension of the latent space; Encoded initial edge features Capture connected nodes and nodes The relative position information between them; Initial edge features are encoded for various variables to enhance the expressiveness of edge feature representation.
[0012] Furthermore, the SymMPNN module includes multiple SymMPNN layers, which form physical constraints by constructing implicit asymmetric matrices and symmetric matrices to model conservation laws and inequality principles on the graph, and process high-dimensional latent features through layer-by-layer iteration.
[0013] Furthermore, at each SymMPNN layer, each node exchanges information with itself and its neighbors; and residual connection operations are added between adjacent SymMPNN layers to allow direct message passing at different scales; Given a graph with available node and edge features, the message passing mechanism of each SymMPNN layer is:
[0014]
[0015]
[0016] in, and Respectively Layer SymMPNN Node and High-dimensional node features; For Node and High-dimensional edge features between and The corresponding middle edge features are inferred to represent symmetric and asymmetric information transfer; It is the edge update function; for The layers have trainable weights that are shared across the entire graph; For Node The index set of neighboring nodes; Represents the Hadamard product operation; is the state update function of the node feature; is the flux update function for symmetric message passing; Indicates asymmetry, Indicates symmetry.
[0017] Furthermore, the multi-layer message passing mechanism of the SymMPNN module is regarded as a cascade transmission in time, including: Assume that a series of fixed or non-fixed sub-time intervals are implicitly generated ; Each SymMPNN layer is considered as the evolution of information within a time interval; The time integration strategy adopted at time t is:
[0018]
[0019] in, is the learnable coefficient, is the initial feature of the node, for The learnable functions of the layer SymMPNN module, is the total number of layers of the SymMPNN module; Represents the output of the final SymMPNN layer, which is the solution Order approximation.
[0020] In the second aspect, the technical solution adopted by the present invention is: a dynamic system modeling system based on conservation information embedded graph learning, which includes: a first processing module, which discretizes the physical field in time and space, and embeds symmetry into the underlying architecture of the graph neural network to learn continuous equations that are universal in time and space; the initial / boundary conditions in the forward problem are "strongly encoded" as part of the model input and output; the second processing module, when the explicit PDE formula is completely unknown or partially missing, the physical prior knowledge in the macroscopic sense is embedded in the neural network structure through different encoding methods to serve the simulation of different physical systems; the construction module, based on the topological structure relationship that naturally has strong interpretability, adopts a message passing mechanism to learn the high-dimensional feature representation of different information in the physical field, which is used to learn and predict the long-term behavior of nonlinear partial differential equations, so as to obtain a conservation information embedded graph neural network for learning dynamics in complex scenarios.
[0021] In a third aspect, the technical solution adopted by the present invention is: a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device executes any one of the above methods.
[0022] In a fourth aspect, the technical solution adopted by the present invention is: a computing device, comprising: one or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the above methods.
[0023] The present invention adopts the above technical solution, which has the following advantages: 1. The present invention uses an end-to-end graph-based deep learning model (i.e., CiGNN) to predict the evolution of complex spatiotemporal dynamics. The symmetric and asymmetric information transmitted on the edge by the SymMPNN module is used for graph representation learning through the general differential form of the conservation law based on the basic principle of symmetry. Embedding this prior knowledge into the network architecture design is equivalent to adding inductive bias to the learning process, thereby alleviating the need for a large amount of training data.
[0024] 2. The multi-grid spatial strategy and implicit time integrator of the present invention further improve the performance of CiGNN. The multi-grid strategy refines the computational domain by superimposing multiple grid scales, enabling the network to capture local and non-local information across spatial scales. The time integrator aims to improve the long-term rollout prediction accuracy and stability of the model by learning the correlation of latent features of the sequence in the SymMPNN layer. The resulting graph learning model is interpretable, and the trained model can be generalized to predict complex spatiotemporal dynamics on different ICs, BCs, and geometric grids. 3. The CiGNN model of the present invention can well identify the underlying dynamical mechanisms and show remarkable accuracy. It is suitable for learning and predicting various spatiotemporal dynamics in spatial domains with arbitrary geometric shapes. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 It is an overall flow chart of a power system modeling method based on conservation information embedding graph learning in an embodiment of the present invention; Figure 2a Schematic diagram of the network structure of the Conservation Informed Graph Neural Network (CiGNN) model in an embodiment of the present invention; Figure 2b It is a spatial block in the CiGNN model in the embodiment of the present invention used to define a multi-grid graph; Figure 2c It is a SymMPNN block diagram used for learning high-dimensional graph representation in the CiGNN model in an embodiment of the present invention; Figure 2d It is a Flux update component that transmits symmetric and asymmetric information on the edge of the graph in the CiGNN model in the embodiment of the present invention; Figure 2e It is the time block for advancing the high-dimensional sequence features in the CiGNN model in the embodiment of the present invention; Figure 2f is a schematic diagram of a potential time integrator in a CiGNN model according to an embodiment of the present invention; Figure 3a It is a performance comparison of the CiGNN model in the embodiment of the present invention and other benchmark models on a two-dimensional viscous Burgers equation test set, as well as a comparison chart between the CiGNN predicted value and the true value; Figure 3bIt is a performance comparison of the CiGNN model in the embodiment of the present invention and other benchmark models on the three-dimensional GS equation test set, as well as a comparison chart of the CiGNN predicted value and the true value.
[0026] Figure 3c is a schematic diagram of a CF sample configuration in which fluid flows in from left (inflow) to right (outflow) in an embodiment of the present invention; Figure 3d It is a performance comparison of the CiGNN model in an embodiment of the present invention and other benchmark models on a two-dimensional CF test set, as well as a comparison chart of the CiGNN predicted value and the true value; Figure 3e : is a comparison diagram of the multi-step training effects of CiGNN based on the Burgers dataset in an embodiment of the present invention and baselines (e.g., FNO, MGN and MP-PDE); Figure 4a is the geographic and sensor information of the BS data set in the embodiment of the present invention; Figure 4b This is an example of 180 days of observation data at the position of the five-pointed star icon in A at 12 m below sea level in the embodiment of the present invention; Figure 4c are the Pearson correlation curves and snapshots of the predicted velocity fields on the 1st day, the 69th day, and the 137th day in the embodiment of the present invention; Figure 4d 1 and 2 are Pearson correlation curves and snapshots of the predicted water temperature on the 1st, 69th and 137th days in an embodiment of the present invention. DETAILED DESCRIPTION
[0027] Physically-informed neural networks (PINNs) often suffer from the limitation of low-dimensional parameterization and scalability issues, requiring the use of fully connected neural networks and gridless point-wise training. While domain decomposition training and discrete learning schemes can alleviate these issues, the PINNs family of methods faces several other challenges. For example, the need for explicit governing equations may not be met, model generalization cannot be guaranteed, and solution accuracy depends heavily on the correct choice of weight hyperparameters of the loss function terms.
[0028] The neural operator uses a neural network to learn a gridless, infinite-dimensional operator, which to a certain extent gets rid of the limitation of grid resolution. However, the evaluation cost of the integral operator is relatively high. To this end, under the premise of finite dimensions, FNO alleviates the amount of calculation through fast Fourier transform and effectively solves the above-mentioned extrapolation / generalization problem. However, due to the excessive reliance on a large number of rich training samples in the parameter space, it is difficult to learn accurate neural operators if the training data samples are too sparse. Although the method of adding prior physical knowledge to the neural operator network as a constraint effectively solves the model's demand for data volume, similar to PINNs, neural operators still have problems such as poor scalability, low training efficiency, poor convergence, and poor interpretability to a certain extent.
[0029] The present invention solves the above problems by adopting discrete learning based on physical information. First, the physical field is discretized in space and time, and the basic concepts of traditional models are combined with neural networks to add initial / boundary conditions to the network in the form of "strong encoding" to solve the model scalability and extrapolation / generalization problems, while enhancing the interpretability of the model to a certain extent. Subsequently, when the explicit PDE formula is completely unknown or partially missing, the physical prior knowledge in a macroscopic sense is embedded in the neural network structure through different encoding methods to serve the simulation of different physical systems. By retaining general physical laws in the architecture, the present invention can learn complex spatiotemporal dynamics more strongly, showing significant advantages in breaking the trade-off between accuracy and efficiency (such as model size), while the previous state-of-the-art models are subject to such limitations. Finally, in order to deal with the problem that most models cannot handle complex irregular geometric computational domains and boundaries, the present invention uses graph neural networks to represent physical features, so as to learn the long-term behavior of nonlinear partial differential equations (for example, simulating particle ensemble dynamics, continuous media and rigid bodies, and predicting global weather) with the help of graphs, which are natural structural relationships with strong interpretability (such as spatial proximity).
[0030] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.
[0031] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.
[0032] In one embodiment of the present invention, a method for modeling a power system based on conservation information embedding graph learning is provided. In this embodiment, Figure 1 As shown, the method comprises the following steps: 1) Discretize the physical field in time and space, embed symmetry into the underlying architecture of the graph neural network to learn universal continuous equations in time and space; the initial / boundary conditions involved in the forward problem are "strongly encoded" as part of the model input and output; 2) When the explicit PDE formula is completely unknown or partially missing, the macroscopic physical prior knowledge is embedded into the neural network structure through different encoding methods to serve the simulation of different physical systems; 3) Based on topological structural relationships (e.g., graphs) that are naturally highly interpretable, a message passing mechanism is used to learn high-dimensional feature representations of different information in the physical field, which is used to learn and predict the long-term behavior of nonlinear partial differential equations, in order to obtain a conservation-informed graph neural network (CiGNN) for learning dynamics in complex scenarios.
[0033] The CiGNN of the present invention uses GNNs and symmetry principles to learn dynamics in complex scenarios, such as irregular sparse structures, various initial conditions (ICs) and boundary conditions (BCs). Based on this new neural network architecture, the present invention can learn the required physical laws.
[0034] The goal of this invention is to build a learning model from the Euler perspective to approximate the solution of a specific system given limited measurement data. This invention adopts an end-to-end learning framework of "encoder-processor-decoder". In particular, it focuses on learning partial differential equations on irregular grids, which enhances the versatility of real-world scenarios.
[0035] In the above step 3), if Figure 2a to Figure 2f As shown in Figure 2, conservation information is embedded in the graph neural network, including: The spatial module, with the help of the multi-grid concept, generates a network space structure with information of different scales and converts the physical state into a low-dimensional graph representation; The encoder elevates the physical state represented by the low-dimensional graph into latent features represented in a high-dimensional space; The SymMPNN module processes the spatial and channel information of the local neighborhood of each node on the graph through aggregation operations to handle high-dimensional latent features; The temporal module regards the multi-layer message passing mechanism of the SymMPNN module as an implicit transmission of the concept of time, and further updates the high-dimensional latent features through a learnable time integration network; The decoder maps the updated high-dimensional latent features back to the original physical domain.
[0036] exist Figure 2a to Figure 2f middle, Indicates time The input state, Indicates time The predicted state of are the high-dimensional latent embeddings of different SymMPNN layers. and Represent high-dimensional graph node and edge features. and The edges and Through edge functions The intermediate buffer result is: and represents the node index, Represents the network layer. , , They are the Conserved Information Embedding Graph Neural Network (CiGNN) proposed in this paper, which uses GNNs and symmetry principles to learn dynamics in complex scenarios, such as irregular sparse structures, various initial conditions (ICs) and boundary conditions (BCs).
[0037] In the above embodiment, in the space module, a multi-grid strategy is adopted to conveniently obtain key structural information; this strategy uses grids of different densities in different regions of the structure.
[0038] Specifically, defining a multi-grid graph includes the following steps: 3.1.1) Use a mesh generation algorithm (such as the Delaunay triangulation algorithm) to construct an initial mesh. represents a graph where represents a node set, represents an edge set; 3.1.2) Definition is the initial graph, where ; 3.1.3) Random Drop Generate a new graph with nodes of preset proportions in as the subsequent initial graph; and repeat the operation and aggregate all graphs into ; 3.1.4) The generated graph Organize by layers, with the graph with the most nodes as the top layer and the graph with the least nodes as the bottom layer; That is, as k increases ( ), the generated graph structure Contains fewer nodes.
[0039] 3.1.5) The entire graph Conceptualized as the union of all edges and nodes, the new graph The node is always A subset of nodes is constructed, and the edge reconstruction process can capture information across extended distances, resulting in a multi-grid distribution.
[0040] In this embodiment, a refinement ratio is also defined: to describe the relationship between the number of nodes in the context graph. For example, if have nodes, then the subsequent graph there will be In this embodiment, it is preferred to refine the ratio Set to 0.1. The entire graph can be conceptualized as the union of all edges and nodes as follows:
[0041] It should be noted that The node is always The edge reconstruction process can capture information across extended distances by using a subset of nodes. This results in a multi-scale grid distribution, enabling the network to capture local and non-local relationships and learn complex details. It is worth noting that the multi-grid technique only performs edge reconstruction during preprocessing (not yet training), following the random masking of nodes at each level. Figure 2b As shown, the mask operation generates a large number of subgraphs, which can alleviate the over-reliance on local information and enhance node-level feature extraction.
[0042] In the above embodiment, in the encoder, the physical state represented by the low-dimensional graph is promoted to a potential feature represented in the high-dimensional space, including the following steps: 3.2.1) Encoded initial node features Contains The physical state of the node , the corresponding position features and node type; where Indicates the channel size (that is, the dimension of the latent space); position It is defined in the global coordinate system.
[0043] 3.2.2) Encoded initial edge features Capture connected nodes and nodes The relative position information between them; 3.2.3) Encode various variables (e.g., angles) to enhance the expressiveness of edge feature representation.
[0044] Specifically, the corresponding expression is given by the following formula:
[0045]
[0046] in, and Represents a learnable function (such as MLP); Is a reflection node and nodes The vector features of the relative position information between them; Represents relative distance in physical space; For Node Type; is the angle information of the edge in different directions. In addition, Represents a join operation.
[0047] In the above embodiment, the SymMPNN module includes multiple SymMPNN layers, which form physical constraints by constructing implicit asymmetric matrices and symmetric matrices to model conservation laws and inequality principles on the graph, and process high-dimensional potential features through layer-by-layer iteration, such as Figure 2a shown.
[0048] In this embodiment, in each SymMPNN layer, each node exchanges information with itself and its neighbors; and residual connection operations are added between adjacent SymMPNN layers to allow direct message transmission at different scales.
[0049] Given a graph with available node and edge features, the message passing mechanism of each SymMPNN layer is:
[0050]
[0051]
[0052] in, and Respectively Layer SymMPNN Node and High-dimensional node features; For Node and High-dimensional edge features between and The corresponding intermediate edge features (e.g., bidirectional) are inferred to represent symmetric and asymmetric information transfer; It is the edge update function; for The layers have trainable weights that are shared across the entire graph; For Node The index set of neighboring nodes; Represents the Hadamard product operation; is the state update function of the node features (such as MLP); is the flux update function for symmetric message passing (like MLP); Indicates asymmetry, It is symmetrical. It should be noted that when When , the CiGNN model of the present invention achieves more reliable performance.
[0053] In the above embodiment, in the time module, the multi-layer message transmission mechanism of the SymMPNN module is regarded as a cascade transmission in time, including the following steps: 3.3.1) Assume that a series of fixed or non-fixed sub-time intervals are implicitly generated ; For example, ,in and Represents the physical time step.
[0054] 3.3.2) Each SymMPNN layer is regarded as the evolution of information within a time interval, which is equivalent to the sub-stepping process commonly used in numerical integration methods; Furthermore, the Taylor series expansion shows that if and only if the coefficients When the sum of is equal to 1, the method is consistent.
[0055] Therefore, the time integration strategy adopted at time t is:
[0056]
[0057] in, is the learnable coefficient, is the initial feature of the node, for The learnable functions of the layer SymMPNN module, is the total number of layers of the SymMPNN module; Represents the output of the final SymMPNN layer, which is the solution Order approximation.
[0058] Each increment in this embodiment can be regarded as a coefficient And each function The product of the estimated slopes specified is the sum of the weighted increments. To this end, a trainable 1D convolutional filter is used to approximate the coefficients in the equation. Figure 2d As shown, this temporal integration strategy does not add too much additional cost to network training and inference.
[0059] In the above embodiment, in the decoder, the updated high-dimensional latent features are mapped back to the original physical domain. The physical domain receives the input from the encoder and generates a predicted increment for the last time step as output. This increment is then added to the initial state as a residual value to determine the new state. The learning process of the decoder is given by the following formula:
[0060] in, is a differentiable function (e.g., MLP), is the total number of layers of SymMPNN. For all MLP hidden layers, GELU activation function is used, while for the output layer, linear activation is considered. In addition, layer normalization is applied after each MLP to improve training convergence.
[0061] In summary, the CiGNN of the present invention mainly includes a SymMPNN module, a spatial module and a temporal module. In the SymMPNN module, in order to model the conservation laws and inequality principles on the graph, the present invention constructs implicit asymmetric matrices and symmetric matrices to form physical constraints. In addition, the present invention uses MPNNs as the basic core module to learn low-frequency information. By encoding prior knowledge into MPNNs, the SymMPNN module of the present invention is more interpretable in solving control partial differential equations and real-world problems. At the same time, in the spatial module, the present invention adopts a multi-grid space strategy to propagate information from different ranges or scales. This enables the network to capture local and non-local relationships, facilitating the acquisition of more detailed information. In the temporal module, the present invention no longer regards multi-layer MPNNs as the gradual acquisition of high-order information in space, but regards the multi-layer message passing mechanism of multi-layer MPNNs as cascade transmission in time. This means that implicit time intervals, whether fixed or non-fixed, can be generated between layers. Therefore, the present invention adopts a simple but effective temporal strategy based on numerical physics methods to reduce the computational errors of long-term rollout predictions, thereby allowing deeper neural operators.
[0062] Embodiment, through the experimental setting in this embodiment, several different situations are mainly explored, including the convection-diffusion equation, the reaction-diffusion system, the Navier-Stokes equation, the cylinder flow (CF) dataset and the real-world Red Sea (BS) dataset, to solve the challenges of time series prediction. Figures 3a to 4d As shown, the effectiveness of the model of the present invention on various PDEs and real-world datasets is demonstrated. In order to evaluate the performance of SymGNN, the present invention compares it with four baseline models: Deep Operator Network (DeepONet), Fourier Neural Operator (FNO), MeshGraphNet (MGN) and MP-PDE solver. It should be noted that the model of the present invention is trained using paired input-output data in all scenarios and adapts to various input-output modes. In order to evaluate stability and generalization, the present invention predicts all next time steps based only on a single initial input. Unless otherwise stated, the root mean square error (Rooted Mean Squared Error, RMSE for short) is used as the main evaluation indicator in the experiments of the present invention. Due to the spatial characteristics of the graph, some models (such as FNO) may not be well suited for these structures. Therefore, the present invention avoids using FNO for experiments on CF and BS datasets.
[0063] like Figure 3aAs shown in FIG. 1 , the performance comparison of the present invention and other benchmark models on the two-dimensional viscous Burgers equation test set and the comparison diagram between the CiGNN predicted value and the true value are shown; Figure 3b As shown in the figure, the performance comparison of the present invention and other benchmark models on the three-dimensional GS equation test set and the comparison chart between the CiGNN predicted value and the true value. Figure 3c As shown in FIG. 1 , a schematic diagram of the CF sample setting of the fluid flowing from left (inflow) to right (outflow) in the present invention. When different data sets are generated, the position and size of the cylinder are different, while the inflow speed remains unchanged. The yellow circle represents the position of the cylinder in the training data set, and the white circle represents the test data set. Figure 3d As shown in FIG. 1 , the performance comparison of the present invention and other benchmark models on a two-dimensional CF test set and the comparison between the CiGNN prediction value and the true value are shown. Figure 3e As shown, it is a comparison of the multi-step training effects of CiGNN based on the Burgers dataset in the present invention and the baseline (e.g., FNO, MGN and MP-PDE). It should be noted that FNO is not suitable for processing irregular grids. For comparison, irregular grids are omitted. The coordinates and radius of each white circle are (0.30 m, 0.30 m, 0.08 m), (0.15 m, 0.10 m, 0.07 m), (0.15 m, 0.25 m, 0.04 m), (0.10 m, 0.10 m, 0.04 m), (0.30 m, 0.20 m, 0.08 m). Taking the test dataset as an example, the generalization ability of the training model is evaluated under different changes in the position bc of the cylinder. s lists the snapshots of the generalization test case 3 and case 5 predicted by cignn.
[0064] like Figure 4a As shown in , it is the geographic and sensor information of the BS dataset. Figure 4b The following is an example of 180 days of observation data at the five-pointed star icon position in A at 12 m below sea level, including the eastward velocity , northward speed , water temperature .like Figure 4c As shown in Figure 1, the Pearson correlation curves and snapshots of the predicted velocity field on the 1st day, the 69th day, and the 137th day. Figure 4d The following are the Pearson correlation curves and snapshots of the predicted water temperature on the 1st day, the 69th day, and the 137th day. The above three time points represent the sampling moments in the early, middle, and late stages of the prediction, respectively. Figure 4c and Figure 4d The curves and snapshots in the figure are from 2021. The blue line shows the average of the correlations, and the light blue areas show deviations (e.g. ).
[0065] Based on artificial and field observation datasets, the present invention evaluates the effectiveness of the CiGNN model in predicting the dynamics of various spatiotemporal systems, such as convection-diffusion systems, reaction-diffusion systems, flow through a cylinder, and the hydrodynamics of the Black Sea. With limited training data, the present invention demonstrates that the CiGNN model can well identify the underlying dynamical mechanisms and exhibits remarkable accuracy. The results show that the CiGNN model is suitable for learning to predict various spatiotemporal dynamics in spatial domains with arbitrary geometric shapes.
[0066] In one embodiment of the present invention, a power system modeling system based on conservation information embedding graph learning is provided, which includes: The first processing module discretizes the physical field in time and space, embeds symmetry into the underlying architecture of the graph neural network to learn universal continuous equations in time and space; the initial / boundary conditions involved in the forward problem are "strongly encoded" as part of the model input and output; The second processing module, when the explicit PDE formula is completely unknown or partially missing, embeds the macroscopic physical prior knowledge into the neural network structure through different encoding methods to serve the simulation of different physical systems; The building module, based on the topological structure relationship that naturally has strong interpretability, adopts the message passing mechanism to learn the high-dimensional feature representation of different information in the physical field, which is used to learn and predict the long-term behavior of nonlinear partial differential equations, so as to obtain the conservation information embedded graph neural network for learning dynamics in complex scenarios.
[0067] In the above embodiment, the conservation information is embedded in the graph neural network, including: The spatial module, with the help of the multi-grid concept, generates a network space structure with information of different scales and converts the physical state into a low-dimensional graph representation; The encoder elevates the physical state represented by the low-dimensional graph into latent features represented in a high-dimensional space; The SymMPNN module processes the spatial and channel information of the local neighborhood of each node on the graph through aggregation operations to handle high-dimensional latent features; The temporal module regards the multi-layer message passing mechanism of the SymMPNN module as an implicit transmission of the concept of time, and further updates the high-dimensional latent features through a learnable time integration network; The decoder maps the updated high-dimensional latent features back to the original physical domain.
[0068] In the above embodiment, defining a multi-grid graph includes: Using the mesh generation algorithm, the initial mesh is constructed. represents a graph where represents a node set, represents an edge set; definition is the initial graph, where ; Random Drop Generate a new graph with nodes of preset proportions in as the subsequent initial graph; and repeat the operation and aggregate all graphs into ; The generated graph Organize by layers, with the graph with the most nodes as the top layer and the graph with the least nodes as the bottom layer; The entire picture Conceptualized as the union of all edges and nodes, the new graph The node is always A subset of nodes is constructed, and the edge reconstruction process can capture information across extended distances, resulting in a multi-grid distribution.
[0069] In the above embodiment, the physical state represented by the low-dimensional graph is promoted to a potential feature represented in a high-dimensional space, including: Encoded initial node features Contains The physical state of the node , the corresponding position features and node type; where represents the dimension of the latent space; Encoded initial edge features Capture connected nodes and nodes The relative position information between them; Initial edge features are encoded for various variables to enhance the expressiveness of edge feature representation.
[0070] In the above embodiment, the SymMPNN module includes multiple SymMPNN layers, which form physical constraints by constructing implicit asymmetric matrices and symmetric matrices to model conservation laws and inequality principles on the graph, and process high-dimensional potential features through layer-by-layer iteration.
[0071] In the above embodiment, in each SymMPNN layer, each node exchanges information with itself and its neighbors; and residual connection operations are added between adjacent SymMPNN layers to allow direct message transmission at different scales; Given a graph with available node and edge features, the message passing mechanism of each SymMPNN layer is:
[0072]
[0073]
[0074] in, and Respectively Layer SymMPNN Node and High-dimensional node features; For Node and High-dimensional edge features between and The corresponding middle edge features are inferred to represent symmetric and asymmetric information transfer; It is the edge update function; for The layers have trainable weights that are shared across the entire graph; For Node The index set of neighboring nodes; Represents the Hadamard product operation; is the state update function of the node feature; is the flux update function for symmetric message passing; Indicates asymmetry, Indicates symmetry.
[0075] In the above embodiment, the multi-layer message passing mechanism of the SymMPNN module is regarded as a cascade transmission in time, including: Assume that a series of fixed or non-fixed sub-time intervals are implicitly generated ; Each SymMPNN layer is considered as the evolution of information within a time interval; The time integration strategy adopted at time t is:
[0076]
[0077] in, is the learnable coefficient, is the initial feature of the node, for The learnable functions of the layer SymMPNN module, is the total number of layers of the SymMPNN module; Represents the output of the final SymMPNN layer, which is the solution Order approximation.
[0078] The system provided in this embodiment is used to execute the above-mentioned method embodiments. Please refer to the above-mentioned embodiments for specific processes and detailed contents, which will not be repeated here.
[0079] A computing device provided in one embodiment of the present invention may be a terminal, which may include: a processor, a communications interface, a memory, a display screen, and an input device. The processor, the communications interface, and the memory communicate with each other through a communication bus. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory, wherein the non-volatile storage medium stores an operating system and a computer program, and the computer program is executed by the processor to implement the methods in the above embodiments; the internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be implemented through WIFI, a management network, NFC (near field communication) or other technologies. The display screen may be a liquid crystal display screen or an electronic ink display screen, and the input device may be a touch layer covered on the display screen, or a key, trackball or touchpad provided on the housing of the computing device, or an external keyboard, touchpad or mouse, etc. The processor may call the logic instructions in the memory.
[0080] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0081] In one embodiment of the present invention, a computer program product is provided, wherein the computer program product includes a computer program stored on a non-transitory computer-readable storage medium, wherein the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the methods provided by the above-mentioned method embodiments.
[0082] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores server instructions, and the computer instructions enable a computer to execute the methods provided in the above embodiments.
[0083] The above embodiment provides a computer-readable storage medium, whose implementation principle and technical effect are similar to those of the above method embodiment, and will not be repeated here.
[0084] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0085] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0086] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A dynamic system modeling method based on conservation information embedding graph learning, characterized in that: include: Discretize the physical field in space and time, embed symmetry into the underlying architecture of the graph neural network to learn universal continuous equations in space and time; The initial / boundary conditions in the forward problem are "strongly encoded" as part of the model input and output; When the explicit PDE formula is completely unknown or partially missing, the macroscopic physical prior knowledge is embedded into the neural network structure through different encoding methods to serve the simulation of different physical systems. Based on the topological structure relationship that is naturally highly interpretable, a message passing mechanism is used to learn the high-dimensional feature representation of different information in the physical field, which is used to learn and predict the long-term behavior of nonlinear partial differential equations, so as to obtain the conservation information embedded graph neural network for learning dynamics in complex scenarios.
2. The method for modeling a dynamic system based on conservation information embedding graph learning as claimed in claim 1, characterized in that: Conservation information is embedded in graph neural networks, including: The spatial module, with the help of the multi-grid concept, generates a network space structure with information of different scales and converts the physical state into a low-dimensional graph representation; The encoder elevates the physical state represented by the low-dimensional graph into latent features represented in a high-dimensional space; The SymMPNN module processes the spatial and channel information of the local neighborhood of each node on the graph through aggregation operations to handle high-dimensional latent features; The temporal module regards the multi-layer message passing mechanism of the SymMPNN module as an implicit transmission of the concept of time, and further updates the high-dimensional latent features through a learnable time integration network; The decoder maps the updated high-dimensional latent features back to the original physical domain.
3. The method for modeling a dynamic system based on conservation information embedding graph learning as claimed in claim 2, characterized in that: Define a multi-grid graph, including: Using the mesh generation algorithm, the initial mesh is constructed. represents a graph where represents a node set, Represents an edge set; definition is the initial graph, where ; Random Drop Generate a new graph with nodes of preset proportions in as the subsequent initial graph; and repeat the operation and aggregate all graphs into ; The generated graph Organize by layers, with the graph with the most nodes as the top layer and the graph with the least nodes as the bottom layer; The entire picture Conceptualized as the union of all edges and nodes, the new graph The node is always A subset of nodes is constructed, and the edge reconstruction process can capture information across extended distances, resulting in a multi-grid distribution.
4. The method for modeling a dynamic system based on conservation information embedding graph learning as claimed in claim 2, characterized in that: The physical state represented by the low-dimensional graph is promoted to the latent features represented in the high-dimensional space, including: Encoded initial node features Contains The physical state of the node , the corresponding position features and node type; where represents the dimension of the latent space; Encoded initial edge features Capture connected nodes and nodes The relative position information between them; Initial edge features are encoded for various variables to enhance the expressiveness of edge feature representation.
5. The method for modeling a dynamic system based on conservation information embedding graph learning as claimed in claim 2, characterized in that: The SymMPNN module consists of multiple SymMPNN layers, which form physical constraints by constructing implicit asymmetric and symmetric matrices to model conservation laws and inequality principles on the graph, and process high-dimensional latent features through layer-by-layer iteration.
6. The method for modeling a dynamic system based on conservation information embedding graph learning as claimed in claim 5, characterized in that: In each SymMPNN layer, each node exchanges information with itself and its neighbors; And add residual connection operations between adjacent SymMPNN layers to allow direct message passing at different scales; Given a graph with available node and edge features, the message passing mechanism of each SymMPNN layer is: in, and Respectively Layer SymMPNN Node and High-dimensional node features; For Node and High-dimensional edge features between and The corresponding middle edge features are inferred to represent symmetric and asymmetric information transfer; It is the edge update function; for The layers have trainable weights that are shared across the entire graph; For Node The index set of neighboring nodes; Represents the Hadamard product operation; is the state update function of the node feature; is the flux update function for symmetric message passing; Indicates asymmetry, Indicates symmetry.
7. The method for modeling a dynamic system based on conservation information embedding graph learning as claimed in claim 2, characterized in that: The multi-layer message passing mechanism of the SymMPNN module is regarded as a cascade transmission in time, including: Assume that a series of fixed or non-fixed sub-time intervals are implicitly generated ; Each SymMPNN layer is considered as the evolution of information within a time interval; The time integration strategy adopted at time t is: in, is the learnable coefficient, is the initial feature of the node, for The learnable functions of the layer SymMPNN module, is the total number of layers of the SymMPNN module; Represents the output of the final SymMPNN layer, which is the solution Order approximation.
8. A dynamic system modeling system based on conservation information embedding graph learning, characterized in that: include: The first processing module discretizes the physical field in space and time, embeds the symmetry into the underlying architecture of the graph neural network to learn universal continuous equations in space and time; The initial / boundary conditions in the forward problem are "strongly encoded" as part of the model input and output; The second processing module, when the explicit PDE formula is completely unknown or partially missing, embeds the macroscopic physical prior knowledge into the neural network structure through different encoding methods to serve the simulation of different physical systems; The building module, based on the topological structure relationship that naturally has strong interpretability, adopts the message passing mechanism to learn the high-dimensional feature representation of different information in the physical field, which is used to learn and predict the long-term behavior of nonlinear partial differential equations, so as to obtain the conservation information embedded graph neural network for learning dynamics in complex scenarios.
9. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any one of the methods of claims 1 to 7.
10. A computing device, characterized in that: include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any one of the methods described in claims 1 to 7.
