Parameter and data processing correction method and system for power grid energy management

Abstract

The application provides a parameter and data processing correction method and system for power grid energy management, comprising constructing a double-layer graph data model of the power grid, including an operating state view taking a bus as a node and an asset parameter view taking a power transmission branch as a node, and a topological correlation matrix connecting the two views; inputting the double-layer graph data model into an edge node switching convolutional encoder, performing alternating bidirectional feature updating between the operating state view and the asset parameter view through the topological correlation matrix, and generating a latent feature representation; through decoding reconstruction and feature projection, jointly performing physical consistency constraint checking and feature decoupling contrast learning, for mapping the latent feature representation to mutually independent parameter and measurement subspaces, and distinguishing the contribution degree of the anomaly in the two subspaces through contrast learning, to separate and identify measurement device faults and branch parameter errors; and outputting the distinguishing result of the anomaly type and the corresponding operation and maintenance decision according to the checking and contrast learning results.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and power system energy management technology, specifically relating to a parameter and data processing and correction method and system for power grid energy management. Background Technology

[0002] The Energy Management System (EMS) of a power system is the core of modern power grid dispatch automation and operation management. It undertakes the functions of real-time monitoring, safety early warning, and optimized dispatch of the large power grid. Based on the monitoring and prediction of real-time power flow distribution of the power grid by measurement terminals, it provides data support for the dispatch center's decision-making, fault handling, and economic settlement. However, in actual power grid management, data quality directly restricts the effectiveness of supervision: on the one hand, measurement instrument failures, communication noise, or network attacks lead to mixed and poor data in measurement data, interfering with the accurate monitoring of the state; on the other hand, aging transmission lines, meteorological influences, or delayed record updates cause deviations between the physical parameters such as resistance and reactance in the system topology model and the actual situation (i.e., parameter errors). Such uncertainties not only cause state estimation errors but also lead to power metering errors, misleading relay protection settings, and even erroneous dispatch instructions, threatening the safe and economical operation of the power grid.

[0003] To address the aforementioned anomaly identification problem, existing solutions mainly fall into two categories: First, residual analysis techniques based on statistical principles (such as residual sensitivity methods and state augmentation methods). However, these rely heavily on idealized assumptions about the statistical distribution of measurement residuals, and are prone to residual overload effects in scenarios where poor data and parameter errors coexist. This can lead to the operation and maintenance system being unable to identify the source of errors, resulting in misjudgments or missed detections. Furthermore, they lack generalization ability when facing the unknown parameter drift of large-scale power grids, making it difficult to balance efficiency and accuracy. Second, with the evolution of artificial intelligence technology, data-driven methods based on deep learning, especially graph neural networks (GNNs), have emerged. While the introduction of graph convolutional networks has improved feature extraction capabilities, significant shortcomings remain: the low-pass filtering characteristics of general graph convolutional networks smooth out differences in node features, masking high-frequency abnormal signals reflecting faults or data tampering, and reducing sensitivity to subtle risks. More importantly, existing models are mostly black-box architectures that do not explicitly incorporate inherent physical constraints of power systems, such as Kirchhoff's laws (KCL) and Ohm's law, into the decision-making logic. They lack physical interpretability, cannot decouple conflicts from heterogeneous data sources, and struggle to provide reliable auxiliary decision-making basis, thus limiting their practical application in high-reliability monitoring and forecasting of power grids. Summary of the Invention

[0004] To address the shortcomings and deficiencies of existing technologies, this invention provides a parameter and data processing and correction method and system for power grid energy management, aiming to solve the technical problem of the coupling between poor measurement data and erroneous branch parameters in the power grid, making it difficult to accurately identify and attribute the causes.

[0005] The core of this invention lies in constructing a unified two-layer graph data model, mapping the physical entities of the power grid into an operational status view with buses as monitoring units and an asset parameter view with transmission branches as management units, and defining the connection relationship between the two views through a topological association matrix. Based on this, an edge-node switching convolutional encoder is designed, utilizing the topological association matrix to achieve bidirectional, alternating feature updates and cross-validation of operational status and asset parameter information, generating a latent feature representation that integrates topological association logic. Based on this representation, this invention innovatively combines physical consistency constraint verification and feature decoupling comparative learning: the former applies hard constraint correction and soft constraint penalty to electrical quantities based on Kirchhoff's current law and Ohm's law, ensuring that the output results conform to the physical mechanism; the latter maps latent features to independent parameter subspaces and measurement subspaces, and utilizes weighted hard negative sample comparative learning to effectively decouple mixed abnormal signals, thereby accurately distinguishing the contribution of anomalies and pinpointing whether the source is a measurement equipment failure or a branch parameter error. Finally, the system can output targeted equipment calibration instructions or parameter correction suggestions based on the verification and attribution results, achieving collaborative correction and closed-loop management of parameters and data. This invention also integrates mechanisms such as graded risk response and human-machine collaborative incremental learning, which improves the engineering practicality and adaptability of the method.

[0006] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0007] A parameter and data processing correction method for power grid energy management includes:

[0008] A two-layer graph data model of the power grid is constructed. The model includes an operation status view with bus as the node and an asset parameter view with transmission branch as the node, as well as a topological correlation matrix connecting the two views. The node features of the operation status view include voltage measurement values ​​and corresponding measurement reliability indicators, and the node features of the asset parameter view include branch impedance parameters, current measurement values ​​and corresponding parameter consistency indicators.

[0009] The two-layer graph data model is input into the edge node switching convolutional encoder, and alternating bidirectional feature updates are performed between the running status view and the asset parameter view through the topological association matrix to generate a potential feature representation that integrates running and asset information.

[0010] Based on the latent feature representation, through decoding reconstruction and feature projection, physical consistency constraint verification and feature decoupling comparison learning are jointly performed; wherein, the physical consistency constraint verification is used to verify the reconstructed electrical quantities according to at least one of Kirchhoff's current law and Ohm's law;

[0011] The feature decoupling contrastive learning is used to map the latent feature representation to mutually independent parameter subspaces and measurement subspaces, and to distinguish the contribution of anomalies in the two subspaces through contrastive learning, so as to separate and identify measurement equipment faults and branch parameter errors.

[0012] Based on the verification and comparison learning results, output the distinction results of anomaly types and the corresponding operation and maintenance decisions.

[0013] Furthermore, the measurement reliability index is the standardized residual, and the parameter consistency index is the standardized Lagrange multiplier; the features of the operating status view also include the real and imaginary parts of the voltage phasor acquired by the synchronous phasor measurement unit, and the features of the asset parameter view also include the standardized residuals corresponding to the branch resistance, reactance parameters, and current measurement components.

[0014] Furthermore, the edge-node switching convolutional encoder is composed of multiple stacked edge-node switching convolutional layers. The alternating feature update of the two views includes: first, fusing the topological aggregation features inside the running state view with the modulation features of the asset parameter view through the topological association matrix mapping element by element to update the running state features; then, fusing the modulation features of the updated running state features through the topological association matrix transpose mapping with the topological aggregation features inside the asset parameter view element by element to update the asset parameter features.

[0015] Furthermore, the physical consistency constraint verification includes: projecting the reconstructed predicted branch current vector onto a feasible solution space that satisfies node current conservation by using the pseudo-inverse of the node-branch correlation matrix; and calculating the difference between the projected current and the theoretical current as a soft constraint loss based on the reconstructed node voltage and branch impedance.

[0016] Furthermore, the feature decoupling contrastive learning is projected through two multilayer perceptrons with non-shared parameters to construct positive and negative sample pairs: the parameter subspace embedding and measurement subspace embedding of the same branch are positive sample pairs, and the parameter subspace embedding of a certain branch and the measurement subspace embedding of other branches are negative sample pairs; the consistency of embedding is measured by cosine similarity, and high similarity negative samples are given higher weights, minimizing the loss function to bring physically related embedding pairs closer and push away mismatched embedding pairs.

[0017] Furthermore, it also includes a graded risk response step: calculating a risk intensity coefficient based on the ratio of the real-time anomaly score to a preset threshold, and triggering different levels of response actions according to the risk intensity coefficient; wherein, the preset threshold is determined based on the maximum anomaly score of the historical anomaly-free verification dataset; the real-time anomaly score is obtained by calculating the Euclidean distance between the original input features and their reconstructed features of the running status view and the asset parameter view.

[0018] Furthermore, the distinction of anomaly contribution is achieved by calculating a fault attribution tendency index, which is calculated based on the embedding bias between the measurement subspace and the parameter subspace, as well as the default residual ratio of Kirchhoff's current law and Ohm's law.

[0019] Furthermore, it also includes a human-machine collaborative closed-loop management mechanism: the handling results of front-line operation and maintenance are injected into the historical case library as high-weight samples, and incremental fine-tuning of the side node switching convolutional encoder and feature decoupling projection mechanism is carried out during the model iteration cycle.

[0020] Furthermore, the method also includes a data fidelity reconstruction task: based on the latent feature representation, the original features of the running state view and the asset parameter view are reconstructed respectively through two independent decoders; and a view-level attention mechanism is introduced to dynamically weight and calculate the reconstruction error of the two views as the loss.

[0021] And, a parameter and data processing and correction system for grid energy management, comprising:

[0022] The model building module is used to build a two-layer graph data model of the power grid. The two-layer graph data model is based on graph theory and includes an operation status view with the bus as the monitoring unit and an asset parameter view with the transmission branch as the management unit. The connection logic between the two views is established through a topological association matrix, and real-time power grid measurement data, offline asset parameters, measurement reliability indicators and parameter consistency indicators are integrated.

[0023] The feature extraction module employs an edge-node switching convolutional encoder, which enables bidirectional information interaction and feature cross-modulation between the running status view and the asset parameter view through the topological association matrix. It alternately updates the feature matrices of the two types of views to generate a low-dimensional latent representation that integrates the topological association logic.

[0024] The decoding and reconstruction module is used to reconstruct the original features of the running state view and the asset parameter view based on the low-dimensional latent representation through two independent decoders, providing a basis for subsequent verification and attribution;

[0025] The physical consistency constraint verification module is used to perform physical consistency constraint verification based on the output of the decoding and reconstruction module. It forces the output to conform to the physical operation mechanism of the power grid through a hard constraint correction mechanism based on Kirchhoff's current law and a soft constraint penalty mechanism based on Ohm's law.

[0026] The fault attribution module is configured with a feature projection unit and a contrastive learning unit. The feature projection unit maps the low-dimensional latent representation to independent parameter subspaces and measurement subspaces. The contrastive learning unit with weighted hard negative samples decouples the mixed abnormal signals and locks the source of the abnormality.

[0027] The decision output module is used to output targeted equipment calibration instructions or asset parameter correction suggestions based on the joint results of the physical consistency constraint verification module and the fault attribution module, thereby completing the collaborative correction of parameters and data.

[0028] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0029] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0030] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0031] It effectively overcomes the residual flooding effect of traditional statistical methods in scenarios where bad data and parameter errors coexist. At the same time, it overcomes the oversmoothing limitation of general graph neural networks. Through the collaborative design of a two-layer graph data model and edge node switching convolution, it realizes bidirectional verification of operating data and asset parameters, significantly improving the accuracy and sensitivity of abnormal signal identification in complex power grid environments, and can effectively detect even weak or hidden anomalies.

[0032] By explicitly incorporating inherent physical constraints of the power grid, such as Kirchhoff's Current Law and Ohm's Law, into the model's decision-making process, the AI-driven anomaly identification process is given physical interpretability. This avoids non-physical false predictions that may arise from purely data-driven models, significantly improving the credibility and compliance of the decision-making results and providing a solid and reliable basis for auxiliary decision-making for the dispatch center.

[0033] By leveraging the feature decoupling and comparative learning mechanism, we have successfully achieved accurate separation and attribution of measurement equipment failures and branch parameter errors. This has completely solved the pain point that existing technologies struggle to clarify the source of errors, effectively reduced ineffective inspection dispatches and wasted human resources, and provided clear guidance for operations and maintenance departments to develop targeted asset maintenance plans and data governance strategies.

[0034] By establishing a tiered risk response process and a human-machine collaborative closed-loop management mechanism, a full lifecycle management chain has been constructed, from anomaly detection and accurate attribution to instruction generation and model iteration. This not only improves the agility and precision of power grid operation and maintenance, but also continuously optimizes model performance as real-world application scenarios accumulate, forming a virtuous cycle and providing strong support for the safe, stable operation and economically efficient dispatch of the power grid. Attached Figure Description

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0036] Figure 1This is a flowchart illustrating the closed-loop decision-making process for power system operation monitoring and asset maintenance based on multi-task collaborative supervision, as described in this embodiment of the invention.

[0037] Figure 2 The figures show a comparison of the PR curves of different models in the mixed anomaly scenario of the IEEE 118-node system according to embodiments of the present invention. The left figure is the PR curve for scenario one and the right figure is the PR curve for scenario two.

[0038] Figure 3 The diagram shows the identification results of node voltage data under a typical mixed abnormal scenario in an embodiment of the present invention. The left diagram shows the identification results of the real part of the voltage, and the right diagram shows the identification results of the imaginary part of the voltage. Detailed Implementation

[0039] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0040] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] To address the shortcomings of existing technologies, this invention provides a parameter and data processing and correction method and system for power grid energy management. By integrating graph neural network technology with power system physical constraints, it achieves accurate identification, correction, and operation and maintenance decision support for power grid operation data and asset parameters. This solution focuses on solving the anomaly identification challenge caused by the coexistence of malfunctioning data and incorrect parameters. It constructs a comprehensive technical system covering data modeling, feature extraction, compliance verification, fault attribution, and closed-loop management. The technical solution of this invention is described in detail below with specific implementation details:

[0043] This invention first constructs a unified management information model for panoramic power grid monitoring. Based on graph theory, it overcomes the limitations of traditional single-dataflow modeling by mapping the physical entities of the power system into a two-layer architecture: a bus operation view and a branch asset view. It constructs a unified graph data structure including node features, edge features, and correlation matrices. This structure integrates real-time operational indicators such as voltage and current measurements, and deeply fuses offline physical parameters such as resistance and reactance, as well as asset health indicators such as Lagrange multiplier statistics. It explicitly defines the topological connection logic between managed objects, providing a standardized data foundation for subsequent cross-level information interaction and anomaly tracing in complex power grids.

[0044] To achieve collaborative monitoring of operational data and asset parameters, this invention designs a feature extraction mechanism based on Side-Node Switching Convolution (SGCN). Addressing the interactive dependency between the two, an alternating information transmission path is constructed using an association matrix, completing bidirectional verification between the operational view and the asset view during the feature encoding stage. Specifically, bus node monitoring information can dynamically modulate the feature representation of branch parameters, and the equipment status at the branch level can also be fed back to the node level. This collaborative processing mode effectively overcomes the smoothing effect of traditional low-pass filtering algorithms on high-frequency abnormal signals, ensuring the system's perception of weak fault symptoms and generating potential state representations containing topological correlation logic, providing comprehensive information support for management decisions.

[0045] To ensure the compliance and reliability of decision-making, this invention embeds a physical rule verification module based on Kirchhoff's Current Law (KCL) and Ohm's Law in the decoding and reconstruction stage. It employs random feature masks and hard-constraint projection techniques to force the output results to remain within the feasible region that satisfies physical conservation laws. By penalizing reconstruction results that violate electrical constitutive relations, false alarms caused by spurious data associations are eliminated at the source, ensuring the physical interpretability of monitoring signals, reducing the incidence of invalid equipment maintenance dispatches, and improving the economic efficiency of operation and maintenance management.

[0046] To address the challenge of distinguishing between measurement equipment failures and line parameter drift, this invention constructs a causal analysis framework with in-graph feature decoupling. It designs a contrastive learning task based on subspace projection and uses weighted hard negative sample discrimination logic to enable the system to learn a causal mechanism where "parameters determine state," dividing the latent feature space into independent parameter and measurement subspaces. This mechanism achieves clear decoupling of mixed abnormal signals, pinpointing the true source of anomalies by maximizing mutual information. This allows the system to output clearly targeted maintenance recommendations, distinguishing whether anomalies originate from measurement noise at the information layer or equipment aging at the physical layer, thus supporting the development of targeted asset maintenance plans and data governance strategies.

[0047] This invention also constructs a human-machine collaborative hierarchical risk response and closed-loop management mechanism. Based on the risk quantification assessment results, the system initiates a dynamic response process from the early warning level to the handling level: when the risk is at the verification level, an AI-assisted analysis report is generated; when the risk reaches the handling level, a calibration work order for metering equipment or a correction suggestion for the asset ledger is automatically generated. Simultaneously, an expert confidence feedback mechanism is introduced to collect the handling results reported by frontline maintenance personnel, incrementally fine-tuning the system's core algorithm modules to achieve closed-loop management throughout the entire lifecycle, from anomaly detection and fault attribution to administrative instruction generation and technological iteration.

[0048] The implementation of the embodiments of the present invention will be shown and described in more detail below:

[0049] 1. Construct a power system management data model for panoramic supervision.

[0050] To achieve comprehensive digital monitoring of power system operating status and asset parameters, this embodiment first constructs a unified management data model that includes a bus operation view (node ​​graph), a branch asset view (edge ​​graph), and topology association logic (association matrix). This model maps physical power grid entities to a node graph with buses as monitoring units. And the edge graph with transmission branches as management units .in, This represents the set of buses under controlled monitoring. This represents the total number of busbars. This is the bus adjacency matrix, used to describe the connection relationships between monitoring points; This is the characteristic matrix of the busbar operating state, and each row vector... Characterizing the first Each monitoring unit Dimensional state data. Correspondingly, This represents a collection of branch road assets that require maintenance and management. The total number of branch roads; Given an adjacency matrix of an edge graph, if the branches and If connected to the same bus, then the corresponding element Set to 1, otherwise set to 0; This is a branch asset feature matrix used to characterize the physical profile parameters and real-time load status of transmission components. It is used for the bus operation view (node ​​diagram). To fully integrate real-time monitoring data and its reliability assessment information, the first... Feature vector of each bus monitoring unit The structure is as follows:

[0051]

[0052] In the formula, and These represent the data collected by the synchronous phasor measurement unit (PMU) at the [number]th ... The real and imaginary parts of the bus voltage phasor are measured as the basic data for real-time operation monitoring. and These represent the standardized residuals (NR) corresponding to the real and imaginary parts of the voltage measurement, respectively. The standardized residual here is a reliability index of the measurement equipment calculated based on weighted least squares (WLS) preprocessing. Its value directly reflects whether there is a possibility of failure or data deviation in the front-end acquisition device, and is a key basis for the system to generate equipment calibration instructions.

[0053] The specific steps for extracting indicators are as follows:

[0054] First, perform WLS state estimation by minimizing the objective function. The system state estimate is obtained. and measurement residual vector ,in For measurement vectors, For the measurement equation, Let be the covariance matrix of the measurement error.

[0055] Subsequently, the covariance matrix of the residuals is calculated. ,in It is a sensitivity matrix. To measure the Jacobian matrix.

[0056] Finally, the first Equipment reliability index (i.e., standardized residual) at each measurement point Through its original residual It is calculated by dividing by its standard deviation, i.e. .

[0057] For branch asset views (side graph) To achieve joint investigation of errors in asset parameter files and abnormal operational data, the first... Feature vector of branch assets The structure is as follows:

[0058]

[0059] In the formula, This refers to the offline physical parameters stored in the branch asset ledger, specifically including the branch resistance. and reactance ; The Normalized Lagrange Multiplier (NLM) is used as an asset parameter consistency warning indicator to quantitatively assess the risk of deviation between the current ledger parameters and the actual physical power grid characteristics. This represents the complex current measurement vector of the branch, which contains four components: the real and imaginary parts of the current at the beginning and end of the branch obtained from the PMU. It is used to monitor the line load. This represents the standardized residual vector that corresponds one-to-one with the aforementioned current measurement components; the symbol || indicates the vector concatenation operation.

[0060] The steps for extracting the Non-Monitoring Model (NLM) for Asset Parameter Consistency are as follows:

[0061] First, the residual vector after convergence is estimated based on WLS. Using the parameter Jacobian matrix Calculate the original Lagrange multiplier vectors This step projects the deviations at the measurement level into the asset parameter management space.

[0062] Next, calculate covariance matrix .

[0063] Finally, the Standardized Lagrange multipliers with 1 parameter By using the original multipliers Divide by its standard deviation (i.e., the square root of the diagonal elements of the covariance matrix) to obtain, i.e. .

[0064] This feature construction method introduces the NR indicator, reflecting the status of monitored equipment, and the NLM indicator, reflecting the quality of asset archives, at the data input layer, providing physical prior data for the subsequent management system to accurately decouple the source of faults. To establish a clear mapping between the operational view and the asset view, this invention defines an association matrix. Serving as a topological bridge connecting the two. Elements in the matrix. The definition is as follows:

[0065]

[0066] In the formula, This represents the total number of busbar nodes. The total number of branch nodes is represented by the correlation matrix T. This matrix describes the physical connection topology between power system management objects. It plays a crucial role in information transmission and data aggregation in subsequent edge node switching convolutional networks, ensuring that the supervisory model can strictly adhere to the actual physical architecture of the power grid for cross-view feature interaction and fault tracing.

[0067] 2. An operation and asset collaborative supervision encoder based on edge-node switching convolution (SGCN)

[0068] To implement the running view With Asset View To achieve deep fusion and bidirectional interaction of heterogeneous features, this invention designs a Switching Graph Convolutional Network (SGCN) encoder to realize runtime-asset collaborative supervision. This encoder consists of... It is composed of stacked convolutional layers with layer switching, and each layer utilizes the correlation matrix. As a topological bridge for information transmission, it achieves cross-modulation of features while preserving the structural characteristics of each graph. Its core purpose is to use the physical parameters and load status of branch assets to assist in verifying the authenticity of bus voltage data, and vice versa, thereby completing the preliminary logical self-consistency review during the feature extraction stage.

[0069] Assumption and They represent the first The bus operation status matrix and branch asset characteristic matrix of the layer (when At that time, it is the initial input feature. and ).

[0070] In the In the feature propagation process of layer-based collaborative supervision, the bus operating status is updated first. This update mechanism is a hybrid aggregation verification process: on the one hand, it aggregates the voltage features of adjacent monitoring points through graph convolution to assess the consistency of the region's operating status; on the other hand, it receives parameter and current information from connected branch assets as a supervision modulation signal. The specific update formula is defined as follows:

[0071]

[0072] In the formula, Indicates the updated number Layer bus node feature matrix; It is a non-linear activation function (the ReLU function is preferred in this embodiment); This represents the Hadamard Product, used to implement element-wise feature-gated modulation. The right side of the formula mainly consists of two parts: the first part... Standard graph convolutions representing the interior of the busbar operating status view are used to achieve region operating status aggregation, where... It is the learnable weight matrix of the feature transformation of the nodes in this layer. It is the normalized node graph Laplacian matrix, which is calculated as follows: , here This is the original bus adjacency matrix. It is the identity matrix. This is the degree matrix. Part Two. The modulation signal from the asset view constitutes the asset-monitored modulation, where It is a learnable projection matrix that maps edge features to the modulation space; It is a diagonalization operator used to transform eigenvectors into diagonal matrices for matrix operations; For the aforementioned correlation matrix, Transpose it. The physical meaning of this term is that the information flow first passes through... Operations are aggregated at the branch dimension, and then through... and The mapping accurately transmits and accumulates the asset health status of each branch to the two busbars it connects to, realizing physical mechanism modeling for verifying operational data using asset parameters.

[0073] Immediately afterwards, within the same convolutional layer, based on the just-updated bus operating state... This design ensures the causal dependence of information flow, meaning that the assessment of branch asset status should be aware of the latest bus voltage status at the current moment. The formula for updating asset characteristics is as follows:

[0074]

[0075] In the formula, For the updated number Layer-branch roadside feature matrix. Similar to the busbar operating status update, the right side of this formula also includes two parts: internal aggregation and external modulation. This indicates that the graph convolution within the branch asset feature matrix, based on topological adjacency relationships, achieves asset association aggregation. The learnable weight matrix for edge feature transformation. Let be the normalized adjacency matrix of the edge graph. This constitutes the modulated signal from the running view, where These are the projection weights from node features to the modulation space. This is achieved through the transpose of the correlation matrix. Left multiplication and By multiplying by the right, the updated bus operating status is effectively broadcast and mapped back to the branches it is connected to, realizing the information flow of "reverse verification of asset archives using real-time operating data".

[0076] go through After the alternating convolutional operations described above, the model finally outputs a low-dimensional latent representation that integrates the correlation between the entire network's asset information and operational measurement data, denoted as... and These two potential representation matrices are not merely data compression, but also high-confidence features after multiple rounds of "operation-asset" bidirectional verification, which will serve as the shared data foundation for subsequent multi-task decision-making modules (fault attribution and parameter identification).

[0077] 3. Multi-task learning and decoding (AI-assisted decision-making module)

[0078] 3.1 Data fidelity reconstruction and consistency verification based on view-level attention mechanism

[0079] The data fidelity reconstruction task is a fundamental component of the AI-assisted supervision framework of this invention, with its core management objective being to establish a digital standard baseline for the normal operation of the power system. This module maps high-confidence decision benchmarks (i.e., latent representations) in the latent space to the original management data space using a decoder, and learns the standard distribution patterns of normal operation and healthy assets by minimizing reconstruction errors. If real-time data cannot be effectively reconstructed, it indicates that the current state deviates from the benchmark, thereby triggering an anomaly alarm.

[0080] Considering the heterogeneity between the bus operation view and the branch asset view in terms of feature dimensions and management logic, this invention designs two parameter-independent graph convolutional network (GCN) inference units: an operation status verifier (a decoder based on a node graph) and an asset parameter verifier (a decoder based on an edge graph), which are used to recover the original monitoring data and archive parameters from the bus decision benchmark and the branch decision benchmark, respectively. The mathematical expression of the specific consistency verification and reconstruction process is as follows:

[0081]

[0082] In the formula, and These represent the standard busbar operation characteristic matrix and the standard branch asset characteristic matrix obtained from the reconstruction, respectively. and Each of these is its own adjacency matrix, used to aggregate topological association information during the decoding process; and This is the latent feature representation of the encoder output after collaborative supervision and verification.

[0083] Given that abnormal risks in power systems often exhibit localized focus—that is, faults may preferentially manifest as abnormal operating indicators (e.g., voltage exceeding limits, concentrated in the node view) or abnormal asset profiles / lines (e.g., parameter drift, concentrated in the side view)—this invention introduces a view-level attention mechanism to enable the monitoring system to adaptively focus on views containing richer risk information. This mechanism first calculates the global importance score of each view using a shared attention neural network to determine the current focus of the monitoring system. Taking the bus operation view as an example, its importance score... The calculation formula is:

[0084]

[0085] In the formula, This indicates that a global average pooling operation is performed on the node graph features to obtain a panoramic operational status vector of the view. This represents the total number of busbar nodes. and These are the learnable weight matrix and bias vector in the attention network, used for feature transformation. This is the hyperbolic tangent activation function, used to introduce nonlinearity; A learnable attention vector used to map the transformed features to a scalar score. Similarly, the importance score of the side view can be calculated. Subsequently, the scores are normalized using the Softmax function to obtain the final supervision weight coefficients: where, and Let the weight coefficients of the node view and the edge view be respectively, in the loss calculation, satisfying... .

[0086] Based on the aforementioned reconstructed features and attention weights, a data fidelity loss function is defined. The weighted sum of the reconstruction errors of the two views:

[0087]

[0088] In the formula, and These are the feature matrices of the bus nodes and the feature matrices of the branches, respectively, from the original input. This represents the Frobenius Norm, used to measure the overall numerical difference between matrices. Minimizing this loss function... The AI ​​model is prompted to learn the compliant operation logic under normal operating conditions of the power system. At the same time, the attention mechanism enables the system to dynamically adjust the degree of attention to different management dimensions during the training process, thereby amplifying the residual signal of the abnormal view in a targeted manner when an anomaly occurs, providing a highly sensitive judgment basis for subsequent accurate fault location and decision-making.

[0089] 3.2 Compliance Verification and Logic Gating Mechanism Based on Hybrid Physical Rule Constraints

[0090] To address the issue that traditional, purely data-driven unsupervised models are prone to degenerating into identity mappings during training, leading to the system's inability to effectively identify errors in asset parameter files, this embodiment introduces a hybrid physical compliance verification mechanism at the decision decoding end, integrating hard-constraint topology correction and soft-constraint parameter guidance. A logic gating mechanism is introduced into this intelligent supervision system, the core purpose of which is to utilize the inviolable physical laws of the power system (Kirchhoff's laws and Ohm's law) to forcibly restrict the reconstruction results of the AI ​​model within a physically interpretable compliant solution space, thereby preventing the system from outputting false decision signals that violate the power grid's operating mechanism.

[0091] First, a hard-constrained projection layer based on Kirchhoff's Current Law (KCL) is introduced at the decoder output as a topology conservation check based on KCL hard constraints. This layer utilizes the node-branch correlation matrix of the power system. The initial branch current monitoring vector generated by the model Forced projection onto the physically feasible solution space that satisfies node current conservation, resulting in a compliant current vector after topological logic correction. The calculation formula is as follows:

[0092]

[0093] In the formula, The node-branch incidence matrix represents the power system and is used to accurately describe the directed topology of the power grid. This represents the total number of monitoring nodes in the system. Total number of controlled branch assets. Matrix The element construction rules follow the definition of graph theory: if a node branch road The starting point (i.e., where the current flows out of the node) is the element. If node branch road The termination point (i.e., where current flows into the node) then If node and branch road If not directly connected, then Therefore, item Physically, it represents the net outflow current vector of each node predicted by the model.

[0094] A current vector is injected into the node as a boundary condition for physical conservation; its value is determined by the node complex power measured by the system. and node voltage phasors Through formula The calculation yielded, where This indicates the conjugate operation. Representation matrix The Moore-Penrose pseudo-inverse matrix is ​​used to implement residual inverse correction in the least squares sense.

[0095] Because the topology of a power system typically makes the correlation matrix... Since the matrix is ​​not square (the number of branches is usually greater than the number of nodes), it cannot be directly inverted. Therefore, this invention uses the Singular Value Decomposition (SVD) method to pre-calculate the matrix. Let the matrix... The SVD decomposition form is ,in and They are orthogonal matrices, For a diagonal matrix containing singular values, the pseudo-inverse matrix is ​​calculated as follows: Here, Yes The matrix is ​​obtained by taking the reciprocals of all non-zero singular values ​​and leaving the zero value unchanged. Since the power grid topology remains unchanged over a certain system operating period... This projection operation can be calculated in one go during system initialization and used as a constant matrix in forward propagation. KCL default residuals at the node level ( The error is propagated back to the branch asset level, and the correction is subtracted from the original forecast, thus ensuring the output monitoring signal. It strictly satisfies the law of conservation of node current, eliminating non-physical data fluctuations caused by algorithm randomness.

[0096] To obtain the current that satisfies the topological constraints Subsequently, the system further combines the node voltages reconstructed by the operating state verifier. Branch impedance reconstructed by asset parameter validator Construct a physical consistency loss function based on Ohm's law. This loss function aims to penalize reconstructions that violate the constitutive relations of electrical branches, and its mathematical expression is defined as:

[0097]

[0098] In the formula, This represents the set of all controlled branch assets in the power network. Total assets; For the first The complex current value of each branch after projection correction (i.e., a high-confidence current reference). and The nodes obtained by reconstructing the running state verifier are respectively and nodes Complex voltage phasors; The first result obtained by reconstructing the asset parameter validator The complex impedance parameters of the branch (derived from the reconstructed resistor) and reactance constitute).

[0099] By minimizing this loss function, when biases in the input asset profile parameters cause a violation of Ohm's law, the model will use topologically compliant current as a logical anchor point to generate a strong gradient signal to forcibly correct the impedance parameters. The inference results are obtained through this mechanism. Essentially, this mechanism establishes a physical logic-based supervisory mechanism within the algorithm, fundamentally suppressing false reconstructions that violate physical laws, significantly improving the management system's sensitivity to identifying erroneous asset parameters, and ensuring that the final generated maintenance instructions have a solid physical basis.

[0100] 3.3 Fault Attribution Judgment and Differentiated Operation and Maintenance Decision-Making Mechanism Based on Intra-Graph Feature Decoupling

[0101] Although the aforementioned edge-node switching convolutional cooperative supervised encoder (SGCN) has effectively mapped branch physical asset parameters and real-time current measurements to a unified high-dimensional latent representation space, in actual operation and maintenance management decision-making processes, this implicit feature fusion lacks explicit supervision signals and makes it difficult to distinguish the inherent influence of two types of heterogeneous information: archival data and monitoring data.

[0102] To address the pain points of this technology and avoid ambiguous maintenance instructions output by the system, this embodiment introduces a fault attribution mechanism based on in-graph feature decoupling (i.e., a contrastive learning task). The core idea of ​​this mechanism is to use AI algorithms to construct a strict logical mapping relationship between asset parameters and operational monitoring data under normal operating conditions in the potential decision space. By maximizing the mutual information between the two, it ensures that any unilateral anomaly on either side (whether it is parameter changes caused by equipment aging or data jumps caused by sensor failure) will disrupt this consistency, thereby generating a significant decoupled alarm signal and providing accurate fault attribution support for maintenance personnel.

[0103] To avoid feature competition between the fault attribution and determination task and the core data fidelity reconstruction task in the shared representation space, the system first introduces a projection head mechanism to perform subspace mapping of decision features.

[0104] Specifically, an independent multilayer perceptron (MLP) with two parameters that are not shared is defined as the projection function of the decision dimension. and For the first output of the shared encoder Potential representation of branch road assets The asset embedding representation is obtained by projecting it onto the asset attribute subspace and the monitoring state subspace respectively. With monitoring embedded representation Based on this, positive and negative sample pairs are constructed: samples from the same branch are paired. Asset representation and monitoring representation Defined as a positive sample pair, its physical meaning lies in the fact that a given set of physical asset parameters and operating conditions necessarily correspond to a unique electrical operating state, which is the benchmark for the system to determine equipment health; the branch... The asset parameters are represented by those of other branches within the batch. ( The monitoring indicates Defined as negative sample pairs, which represent a mismatch between assets and their operational status, these are used to train the system's ability to identify logical errors.

[0105] To overcome the false negative problem that may arise from random negative sample sampling in traditional contrastive learning, and to improve the management system's sensitivity to minor asset parameter deviations (such as early insulation aging), this embodiment employs the InfoNCE loss function based on difficult negative sample sampling. For any branch... Its in-plot contrast loss function The definition is as follows:

[0106]

[0107] In the formula, This represents the cosine similarity function, used to measure the similarity between two vectors in the feature space. This is a temperature hyperparameter used to adjust the model's smoothness of the similarity distribution and its attention to difficult samples. The denominator contains... This is a hard negative sample penalty term, calculated by weighted summation of the similarities of all negative sample pairs. This term introduces a concentration parameter. To control the distribution of weights: The higher the value, the greater the weight the model assigns to negative samples with high similarity (i.e., difficult to distinguish). By minimizing this loss function, the AI-assisted decision-making model is prompted to bring physically matching "asset-monitoring" feature pairs closer together in the latent space, while pushing away physically mismatched feature combinations. This mechanism ensures that during the online supervision phase, regardless of whether the error is due to asset profile errors... Is the offset caused by a malfunction in the measuring equipment? Offsets significantly increase the feature distance between the two, effectively decoupling the mixed reconstruction errors into separable feature conflict signals. This enables the system to output clearly directional, differentiated operation and maintenance suggestions. When the system indicates a shift in dominance, it suggests updating the logbook or inspecting the line; when When offset is dominant, the system recommends checking the sensors or troubleshooting the communication.

[0108] Combining the above three designs (data reconstruction supervision, physical compliance verification, and fault attribution judgment), this invention constructs a joint optimization function oriented towards comprehensive management objectives. The learning objectives of data fidelity, physical compliance, and logical attribution are balanced by a weighted summation method.

[0109]

[0110] In the formula, The aforementioned data fidelity reconstruction loss is used to ensure the accuracy of panoramic supervision; The physical consistency constraint loss is used to ensure that decision-making results comply with the physical operation rules of the power grid (to avoid unauthorized command). The collection of all controlled branch assets; and These are the balancing hyperparameters for the physical constraint term and the contrastive learning term, respectively, used to adjust the regularization strength of the auxiliary task on the main reconstruction task.

[0111] After model training is complete, the system enters the online real-time monitoring and decision support phase. For any given power system snapshot data, it is first input into the trained SGCN intelligent management model to obtain the reconstructed standard operating state node features. and standard asset parameter characteristics Subsequently, by calculating the difference between the original input features and the reconstructed features, the difference is quantified into bus monitoring units. Abnormal index and branch roads Asset Health Index The specific calculation formula is as follows: and ,in The Euclidean norm of a vector ( (norm). Finally, a threshold-based strategy based on statistical maximum values ​​is used for risk identification: the threshold is set as the maximum value of all anomaly scores generated by the model on a historical validation dataset without anomalies. If the anomaly score of a monitoring node or asset branch calculated in real time exceeds this threshold, it is determined that the component has distorted operating data or incorrect asset parameters, and the corresponding alarm and work order generation process is immediately triggered.

[0112] 4. Intelligent operation and maintenance decision generation and closed-loop management based on fault attribution

[0113] The system provided in this embodiment also constructs a proactive defense and operation and maintenance decision-making system with cognitive intelligence. After obtaining the aforementioned high-confidence anomaly score based on physical consistency constraints, the system no longer relies on a single hard-line disconnection logic, but instead drives a dynamic hierarchical management response process based on the risk quantification assessment results. This process aims to achieve closed-loop management of the entire lifecycle of abnormal events in the power grid through deep collaboration between AI algorithms and operation and maintenance personnel.

[0114] First, to transform abstract mathematical feature differences into actionable administrative management criteria, the system defines a risk grading function based on statistical boundaries. Let the set of anomaly scores calculated by the model on a historical validation set without anomalies be... Its maximum value is defined as the baseline threshold. The current anomaly score is calculated for any unit (bus node or branch asset) in real-time monitoring. (i.e., the aforementioned) or The system introduces a risk intensity coefficient. This quantifies the degree to which the current state deviates from the normal baseline. Based on this, the following piecewise risk response decision function is constructed. :

[0115]

[0116] In the formula, and The management tolerance over-parameter is set according to the power grid safety management specifications (the recommended setting in this embodiment is...). These correspond to the boundaries for determining suspected risks and actual faults, respectively.

[0117] When the system determines the risk to be at Level I (early warning level), given that abnormal signals may originate from transient disturbances or system edge conditions, the management strategy focuses on enhancing awareness rather than blind intervention. The system automatically triggers a data governance subroutine to implement a high-frequency sampling strategy for the monitored area, increasing the data archiving density from the usual "minute level" to the "second level," and pushes a yellow early warning signal to the dispatch monitoring screen, prompting on-duty personnel to maintain logical attention to the area. This process requires no manual intervention and aims to accumulate high-precision time-series evidence for subsequent analysis.

[0118] As the risk level escalates to Level II (verification level), the system enters AI-assisted analysis mode. At this point, a single data deviation becomes statistically significant. The system automatically retrieves the historical load curve of the abnormal unit, topological correlation data of neighboring nodes, and the discrimination results from the aforementioned in-graph feature decoupling module, generating an intelligent analysis report of the abnormal event that includes a fault probability distribution and preliminary attribution suggestions. This report is pushed to management personnel in the relevant area via mobile maintenance terminals, where experts conduct cross-verification based on actual on-site conditions. This mechanism effectively combines the computing power of AI with human experience and wisdom, avoiding false alarms that might occur with purely automated systems and ensuring the prudence of management decisions.

[0119] Once the risk intensity exceeds Level III (handling level), indicating significant parameter errors or measurement failures in the physical power grid, the system immediately initiates an automated administrative handling process. Without manual approval, the system directly generates standardized physical maintenance work orders or asset correction recommendations based on the fault attribution results (determined to equipment failure or file error), and directly dispatches them to the nearest grid-based emergency repair team or asset management department via the enterprise ERP interface. This flat management model, with data directly reaching the work teams, significantly shortens the administrative workflow time from fault occurrence to on-site handling, achieving agility and precision in power grid operation and maintenance management.

[0120] After triggering the aforementioned verification or handling-level response mechanisms, simple alarm signals are no longer sufficient to support the needs of refined power grid management. To avoid blind dispatching and resource waste caused by unclear fault sources in traditional operation and maintenance models, this system further initiates a differentiated operation and maintenance decision generation process based on in-graph feature decoupling. The core logic of this process lies in mapping heterogeneous feature deviations in the potential space to specific administrative execution instructions, thereby achieving precise separation of information layer faults (distorted measurement data) and physical layer hidden dangers (incorrect asset records) at the management level. To this end, the system first defines a fault attribution tendency index. This is used to quantify the competitive relationship between the monitoring subspace and the asset subspace in terms of feature drift.

[0121]

[0122] In the formula, and These are the real-time monitoring embedded representation and the asset parameter embedded representation output by the aforementioned projection head, respectively. For the corresponding normal reference vector, To prevent the use of tiny constants with a denominator of zero, the latter part of the exponential term introduces the ratio of the physical constraint residuals, where... Reflects the degree of violation of topological conservation. It reflects the degree of breach of electrical constitutive relations.

[0123] When the system calculates The value is significantly greater than the preset attribution threshold (e.g. When the measurement anomaly threshold is reached, it indicates that the anomalies in the feature space are mainly driven by non-physical jumps in the monitoring data, accompanied by high KCL residuals. Based on this, the system determines the fault attribute as an anomaly in the reliability of the measurement equipment. In this scenario, the AI ​​decision engine will automatically bypass the maintenance process for the primary physical line and instead generate a field calibration work order for metering and communication equipment. This work order not only includes the ID and geographical location of the abnormal equipment, but also embeds the theoretical confidence interval derived from the model, guiding the information and communication department (ICT) to perform accuracy verification, communication error troubleshooting, or secondary circuit detection for specific PMU terminals or SCADA acquisition modules. This attribution-based decision-making mechanism effectively prevents maintenance personnel from mistakenly inspecting normal transmission lines when data errors occur, greatly reducing ineffective manpower and vehicle dispatch costs, and demonstrating the optimization of technical management efficiency under data-driven approaches.

[0124] Conversely, when The value is significantly smaller than the attribution threshold (i.e. When the asset parameters exceed the abnormal threshold, it means that the monitoring data is topologically consistent, but the asset parameters exhibit significant outliers in the feature space and violate Ohm's law. Based on this, the system determines the fault attribute as a discrepancy between the asset ledger and the physical assets. At this point, the system shifts its decision-making focus to the asset management dimension, automatically generating an asset ledger verification and correction recommendation and pushing it to relevant personnel. This recommendation not only indicates that the resistance and reactance parameters of the current branch have aging drift or input errors, but more importantly, the system utilizes the decoder's reconstruction capabilities to directly provide recommended parameter correction values ​​that conform to the current physical operating conditions within the recommendation. Based on this decision recommendation, administrators can quickly update and correct outdated or erroneous ledgers in the ERP system. This allows for the cleaning and maintenance of digital assets without interrupting power grid operations, ensuring the accuracy and timeliness of basic power grid management data.

[0125] Furthermore, as a preferred option, to achieve a continuously evolving management closed loop, the system provided in this embodiment can also consider introducing an expert-in-the-loop mechanism. After the frontline maintenance team completes the on-site handling of the aforementioned work orders (such as replacing aging sensors or correcting incorrect line parameters), the handling results, backed up via the mobile work terminal, will be automatically captured by the system and injected into the historical case library as high-weight "strong-label" samples. In subsequent model iteration cycles, the system uses these manually verified real management cases to incrementally fine-tune the SGCN encoder and fault attribution projector. This mechanism ensures that the monitoring system can continuously improve its decision-making accuracy as the power grid architecture changes and management experience accumulates.

[0126] Based on the design scheme provided in this embodiment, this invention uses Graph Neural Networks (GNNs) as the core technology carrier, integrates physical constraints and contrastive learning mechanisms, and constructs a parameter and data processing correction method and system for power grid energy management, such as... Figure 1 As shown, its overall operation monitoring and management process is as follows:

[0127] First, heterogeneous information is collected through a multi-dimensional data access layer, simultaneously acquiring core information from the asset view (including branch archive parameters and load detection data) and the node view (bus operation status data), providing comprehensive data support for subsequent processing. Then, the operation-asset collaborative supervisory encoder (SGCN) maps the two types of heterogeneous data to a high-confidence decision benchmark space through an edge node logical interaction verification mechanism, generating asset health benchmark features and operation status benchmark features.

[0128] The aforementioned benchmark features are simultaneously transferred to two core processing modules: on the one hand, they enter the data fidelity verification and decoding module, which generates data fidelity supervision indicators through the asset parameter verifier and the operating status verifier, and then outputs physical violation penalty indicators by integrating Kirchhoff's Current Law (KCL) topological conservation and Ohm's Law physical rule compliance gating; on the other hand, they flow into the fault attribution and logical tracing module, which constructs a logical mapping between assets and monitoring data through decision subspace projection, and generates attribution judgment supervision indicators by combining logical consistency comparison and hard sample mining.

[0129] Two types of monitoring indicators are input into the comprehensive management optimization objective function (integrating reconstruction loss, physical constraint loss, and attribution loss). After the model optimization is completed, it enters the risk threshold assessment and operation and maintenance decision generation stage. Based on the multi-level risk quantification assessment results, the corresponding hierarchical response strategy is matched: Level I executes enhanced monitoring, Level II initiates AI-assisted assessment and human-machine review, and Level III triggers automated handling.

[0130] Finally, the anomaly type is determined through fault attribution calculation (ψ value determination): if it is determined to be an operational measurement anomaly, the system automatically generates a calibration work order for the metering equipment and forwards it to the communications / ICT department; if it is determined to be an asset parameter anomaly, an asset ledger correction suggestion is generated and pushed to the equipment management department and ERP system. After the on-site maintenance personnel complete the handling, they fill the results back into the system, and incremental fine-tuning of the model is achieved through feedback from high-confidence samples, forming a closed-loop management system covering the entire lifecycle of "data acquisition - feature processing - decision generation - execution feedback - model iteration".

[0131] Compared to existing residual sensitivity analysis methods and general data-driven anomaly detection techniques, this invention's identification method based on edge node switching convolution and multi-task learning offers significant technical advantages in power system operation monitoring and equipment maintenance management. Addressing the problem that existing technologies, in scenarios where erroneous data and parameter errors coexist, suffer from residual flooding effects due to the inability to distinguish error sources, leading to scheduling monitoring failures or misjudgments in maintenance decisions, this invention constructs a physical consistency constraint regularization mechanism. This mechanism integrates Kirchhoff's laws and Ohm's laws of power systems into the optimization objective of the deep learning model, effectively suppressing the blind fitting of the monitoring model to erroneous parameters. This fundamentally solves the false alarm problem caused by high-leverage point contamination, significantly improving the management system's anti-interference capability and decision-making reliability in highly disruptive environments.

[0132] Based on the in-graph feature decoupling and comparison learning mechanism, this invention establishes a strong causal relationship between asset parameters and real-time measurements in the latent feature space through AI algorithms, achieving accurate decoupling and classification of mixed abnormal signals. This enables the system to clearly distinguish between line parameter drift and measurement equipment noise, providing the operation and maintenance department with accurate fault attribution basis. The differentiated execution instructions generated accordingly can clarify the direction of fault handling, reducing trial and error costs and human resource consumption.

[0133] This invention not only achieves accurate identification of single anomalies but also constructs a closed-loop management system encompassing data computation, administrative execution, and knowledge accumulation by introducing a tiered risk response and expert feedback mechanism. Multi-scenario simulation experiments based on the IEEE 118-bus system demonstrate that this invention significantly outperforms traditional statistical methods and advanced baseline models in key monitoring metrics such as the area under the precision-recall curve (PRAUC) and classification accuracy. Furthermore, it exhibits low sensitivity to hyperparameter changes and excellent robustness, effectively supporting real-time online monitoring, risk prediction, and refined asset management of the power grid, demonstrating significant engineering practical value and administrative benefits.

[0134] The following is a simulation example:

[0135] 1. Example Setup

[0136] To verify the effectiveness and robustness of the system design proposed in this invention in actual power grid administration, this embodiment constructs a simulation environment based on the IEEE 118-bus standard system. The basic operating data of this environment is generated using MATPOWER simulation, and baseline values ​​are applied to the system's baseline load demand and generator active power dispatch plan. Random disturbances within the specified range generated 7,000 normal operation profiles that conform to power flow convergence constraints. This step aims to simulate the random fluctuation characteristics exhibited in the daily dispatching of the power system, thereby providing the intelligent system with a normal baseline training set covering diverse operating conditions. This allows the system to learn and master the physical operation logic and asset association patterns of the power grid under compliant conditions. For each generated operation management snapshot, the system automatically executes the aforementioned data access process, constructs a unified management data model including bus operation views, branch asset views, and topology association logic, and performs Z-score standardization on all monitoring indicators to eliminate dimensional differences between heterogeneous management data (such as voltage amplitude and branch impedance), ensuring the standardization of the input to the decision model.

[0137] To evaluate the intelligent system's decision-making generalization ability in the face of complex challenges and multiple concurrent failures, this embodiment designs a hybrid anomaly test set that includes both measurement equipment failures and asset record errors. This test set is not simply a superposition of noise, but aims to simulate the real "operation-asset" dual risk scenario in power grid operation and maintenance, covering a gradient of risk intensity configurations. Specifically, for the simulation of measurement equipment failures, two dimensions are defined: deviation amplitude (set as 10, 15, and 20 times the standard deviation of measurement noise) and failure penetration rate (3%, 5%, and 10%), used to test the system's signal discrimination capability under different signal-to-noise ratios. For the simulation of asset record errors, parameter drift intensity is divided into three management levels: low (deviation 1%-5%), medium (deviation 5%-10%), and high (deviation 10%-15%), aiming to examine the system's differentiated response levels to minor insulation aging (early warning) and serious parameter entry errors (emergency correction). In addition, to test the system's ability to resist the residual flooding effect that leads to incorrect decisions under extreme conditions, this embodiment also specifically constructs an adversarial confusion scenario, that is, injecting current errors with large amplitude and opposite direction into adjacent branches of the common bus to induce strong false topology associations, thereby verifying whether the system can adhere to the principle of physical compliance under strong interference and avoid issuing incorrect scheduling instructions.

[0138] The decision-making core of this intelligent supervision system is built on the PyTorch deep learning framework and deployed on a computing server equipped with an Intel i7-10700 CPU and an NVIDIA RTX 3060 Ti GPU, simulating the cloud-edge collaborative intelligent computing environment of an actual dispatch center. The proposed model follows an unsupervised learning paradigm, using only normal historical management data for benchmark training, which aligns with the current situation in power grids where complete fault labels are lacking. The encoder network is configured as a three-layer stacked edge-node switching convolutional layer, mapping the originally discrete monitoring data to a 32-dimensional high-confidence decision benchmark space. During model training, the Adam optimizer is used to minimize the multi-task joint management objective function, which includes data fidelity, physical compliance, and attribution discrimination. Dropout technology and an early shutdown strategy are employed to prevent the model from overfitting to specific operating conditions, ensuring its robustness under unknown management scenarios.

[0139] Regarding the management effectiveness evaluation system, this embodiment abandons a single numerical fitting index and instead adopts a comprehensive evaluation framework that reflects the quality of administrative supervision. The area under the precision-recall curve (PRAUC) is used as the core indicator to measure the system's "threshold-free comprehensive supervision quality," while recall, precision, and accuracy are introduced as specific indicators to quantify operational decision-making performance. The calculation logic for each management effectiveness indicator is as follows: In terms of the evaluation system, this embodiment uses the area under the precision-recall curve (PRAUC) as a threshold-independent comprehensive evaluation index, while recall, precision, and accuracy are used as classification performance indicators. The calculation formulas for each classification indicator are as follows:

[0140]

[0141] In the formula, (True Positive) represents the number of real risks in the system that are correctly identified and alerted by the system, reflecting the sensitivity of the alert. (False Positive) represents the number of false positives that are falsely reported as abnormal during normal operation. This indicator is directly related to the generation rate of invalid work orders and reflects the system's anti-interference capability. (False Negative) indicates a false negative example, representing the number of real risks that are missed by the system, reflecting blind spots in system security supervision; (True Negative) represents the number of normal states that are correctly recognized as compliant by the system.

[0142] PRAUC, as a core indicator for measuring the quality of anomaly scoring without a preset threshold, is calculated as follows: by dynamically adjusting the risk response threshold... To obtain a series of corresponding accuracy rates and recall rate By comparing the points, the PR curve is plotted. The PRAUC value is the area under the PR curve, and its mathematical expression is:

[0143]

[0144] The closer the RAUC index is to 1, the more it indicates that the intelligent system can maintain a high risk recall rate (no missed potential risks) while maintaining an extremely low false alarm rate. This proves that it can provide high-quality and highly reliable decision support for power grid operation and maintenance, and is particularly suitable for actual power supervision scenarios where the ratio of positive to negative samples is seriously unbalanced (i.e., faults are scarce but the consequences are severe).

[0145] 2. Result Comparison and Analysis

[0146] To systematically verify the comprehensive performance of the system constructed in this invention under complex power grid management environments, this embodiment sets up multiple mixed risk scenarios on the IEEE 118-node standard management model and compares this system (based on the SGCN-MTL architecture) with existing representative auxiliary tools. The comparison objects cover traditional statistical monitoring tools (maximum standardized residual method NR, standardized Lagrange multiplier method NLM), general data-driven monitoring models (autoencoder AE), and cutting-edge graph neural network algorithms (GCN-AE, Sub-CR).

[0147] First, the area under the precision-recall curve (PRAUC) metric was used to evaluate the overall risk ranking and discrimination capabilities of each system without the need for a preset threshold. Experimental results are as follows: Figure 2As shown, in a complex scenario where moderate-intensity asset profile errors coexist with measurement faults of varying intensities, the intelligent system proposed in this invention exhibits significant decision-making advantages, with its PRAUC values ​​significantly outperforming all baseline methods. Specifically, the regulatory effectiveness of the basic GCN-AE model is even lower than that of the traditional AE model without utilizing topological information, and it suffers severe degradation in low-intensity hidden risk scenarios. This confirms that the low-pass filtering characteristics of standard graph convolutional networks easily smooth out high-frequency feature signals reflecting equipment faults during the aggregation process, leading to a decrease in the sensitivity of the regulatory system to subtle risks. In contrast, although the Sub-CR method achieves some results by incorporating a contrastive learning mechanism, it struggles to effectively distinguish between measurement bias and parameter drift under the interference of asset parameter errors due to the lack of explicit physical compliance gating. This system, by introducing physical consistency constraints and feature masking mechanisms, forces the decision-making model to strictly adhere to the physical constitutive relationship of the power grid in the latent space, thereby effectively overcoming the oversmoothing problem of the algorithm and maintaining a significant performance advantage in scenarios dealing with hidden risks.

[0148] Furthermore, to evaluate the application value of the system in actual operation and maintenance dispatching, after setting a unified risk response threshold, quantitative evaluation was conducted using recall rate (corresponding to security supervision coverage), precision rate (corresponding to dispatching accuracy), and accuracy rate indicators. As shown in Table 1, the method proposed in this invention exhibits extremely high management robustness when dealing with mixed anomalies of "operation-asset," especially overcoming the drawback of baseline models sacrificing accuracy for recall rate in terms of precision rate. In the scenario of hidden anomalies with low signal-to-noise ratio, although the advanced baseline Sub-CR can obtain a reasonable recall rate through data fitting, its precision rate is low due to environmental noise interference. This means that a large number of invalid verification alarms will be generated in the actual production environment; while this system effectively filters out non-physical false fluctuations by relying on physical consistency constraints, significantly improving the precision rate of operation and maintenance decisions. Especially in scenarios with strong asset parameter interference, purely data-driven models are prone to confusing parameter deviations with measurement errors, leading to increased false alarm rates and consequently, incorrect maintenance dispatches. This invention, however, benefits from an in-graph feature decoupling mechanism, clearly distinguishing the source of violations of Ohm's law as "asset records" rather than "real-time measurements," thus maintaining a consistently high decision-making accuracy. This fully demonstrates the fault attribution capability of this multi-task collaborative framework under complex coupled interference. In contrast, while traditional NR methods have higher recall rates, their accuracy is extremely low (due to numerous false alarms), leading to a surge in maintenance costs and lacking practical engineering application value.

[0149] Table 1. Data identification performance of different models in four mixed anomaly scenarios.

[0150]

[0151] To more intuitively demonstrate the qualitative differences in decision support capabilities among different systems, the example selected a typical mixed anomaly snapshot for waveform analysis. The results are shown below. Figure 3 In this scenario, both asset parameter errors and Gaussian noise were simultaneously injected into the system. Visualization results show that the system successfully detected all real anomaly risk points, demonstrating an extremely high regulatory recall rate. Particularly in voltage imaginary part monitoring, the offset of some anomalies did not significantly exceed the background noise band (i.e., within three standard deviations), leading to missed detections by statistical tools relying on residual amplitudes (such as NLM and NR) and some deep learning models, creating blind spots in safety supervision. However, this system, leveraging physical topological constraints, successfully identified these hidden risks, achieving early warning. Furthermore, for large voltage surges, the system maintained the same high sensitivity as statistical methods. This indicates that the method of this invention retains the ability to capture significant risks while effectively solving the problem of separating weak signals under complex interference, achieving accurate joint identification and differentiated management of power system operation data distortion and asset parameter errors.

[0152] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0153] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer 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 device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0154] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0155] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0156] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive various other forms of parameter and data processing correction methods and systems for grid energy management. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A parameter and data processing and correction method for power grid energy management, characterized in that, include: A two-layer graph data model of the power grid is constructed. The model includes an operation status view with bus as the node and an asset parameter view with transmission branch as the node, as well as a topological correlation matrix connecting the two views. The node features of the operation status view include voltage measurement values ​​and corresponding measurement reliability indicators, and the node features of the asset parameter view include branch impedance parameters, current measurement values ​​and corresponding parameter consistency indicators. The two-layer graph data model is input into the edge node switching convolutional encoder, and alternating bidirectional feature updates are performed between the running status view and the asset parameter view through the topological association matrix to generate a potential feature representation that integrates running and asset information. Based on the latent feature representation, through decoding reconstruction and feature projection, physical consistency constraint verification and feature decoupling comparative learning are jointly performed; wherein... The physical consistency constraint verification is used to verify the reconstructed electrical quantities based on at least one of Kirchhoff's current law and Ohm's law. The feature decoupling contrastive learning is used to map the latent feature representation to mutually independent parameter subspaces and measurement subspaces, and to distinguish the contribution of anomalies in the two subspaces through contrastive learning, so as to separate and identify measurement equipment faults and branch parameter errors. Based on the verification and comparison learning results, output the distinction results of anomaly types and the corresponding operation and maintenance decisions; The edge-node switching convolutional encoder is composed of multiple stacked edge-node switching convolutional layers. The alternating feature update of the two views includes: first, the topological aggregation features inside the running state view and the modulation features of the asset parameter view mapped by the topological association matrix are fused element by element to update the running state features; then, the modulation features of the updated running state features mapped by the topological association matrix are multiplied and fused element by element with the topological aggregation features inside the asset parameter view to update the asset parameter features. The physical consistency constraint verification includes: projecting the reconstructed predicted branch current vector to the feasible solution space that satisfies node current conservation by using the pseudo-inverse of the node-branch correlation matrix; and calculating the difference between the projected current and the theoretical current as a soft constraint loss based on the reconstructed node voltage and branch impedance. The feature decoupling contrastive learning achieves projection through two multilayer perceptrons with non-shared parameters, constructing positive and negative sample pairs: the parameter subspace embedding and measurement subspace embedding of the same branch are positive sample pairs, and the parameter subspace embedding of a certain branch and the measurement subspace embedding of other branches are negative sample pairs; the consistency of embedding is measured by cosine similarity, and higher weights are given to negative samples with high similarity to minimize the loss function.

2. The parameter and data processing and correction method for power grid energy management according to claim 1, characterized in that: The measurement reliability index is the standardized residual, and the parameter consistency index is the standardized Lagrange multiplier; the features of the operating status view also include the real and imaginary parts of the voltage phasors collected by the synchronous phasor measurement unit, and the features of the asset parameter view also include the standardized residuals corresponding to the branch resistance, reactance parameters and current measurement components.

3. The parameter and data processing and correction method for power grid energy management according to claim 1, characterized in that: It also includes a graded risk response step: calculating a risk intensity coefficient based on the ratio of the real-time anomaly score to a preset threshold, and triggering different levels of response actions according to the risk intensity coefficient; wherein, the preset threshold is determined based on the maximum anomaly score of the historical anomaly-free verification dataset; the real-time anomaly score is obtained by calculating the Euclidean distance between the original input features and their reconstructed features of the running status view and the asset parameter view.

4. The parameter and data processing and correction method for power grid energy management according to claim 1, characterized in that: The distinction contribution of anomalies is achieved by calculating a fault attribution tendency index, which is calculated based on the embedding bias between the measurement subspace and the parameter subspace, as well as the default residual ratio of Kirchhoff's current law and Ohm's law.

5. The parameter and data processing and correction method for power grid energy management according to claim 1, characterized in that: It also includes a human-machine collaborative closed-loop management mechanism: the handling results of front-line operation and maintenance are injected into the historical case library as high-weight samples, and the edge node switching convolutional encoder and feature decoupling projection mechanism are incrementally fine-tuned during the model iteration cycle.

6. The parameter and data processing and correction method for grid energy management according to claim 1, characterized in that: The method also includes a data fidelity reconstruction task: based on the latent feature representation, the original features of the running status view and the asset parameter view are reconstructed respectively through two independent decoders; A view-level attention mechanism is introduced to dynamically weight and calculate the reconstruction error between the two views as the loss.

7. A parameter and data processing and correction system for grid energy management, used to implement the method as described in claim 1, characterized in that, include: The model building module is used to build a two-layer graph data model of the power grid. The two-layer graph data model is based on graph theory and includes an operation status view with the bus as the monitoring unit and an asset parameter view with the transmission branch as the management unit. The connection logic between the two views is established through a topological association matrix, and real-time power grid measurement data, offline asset parameters, measurement reliability indicators and parameter consistency indicators are integrated. The feature extraction module employs an edge-node switching convolutional encoder, which enables bidirectional information interaction and feature cross-modulation between the running status view and the asset parameter view through the topological association matrix. It alternately updates the feature matrices of the two types of views to generate a low-dimensional latent representation that integrates the topological association logic. The decoding and reconstruction module is used to reconstruct the original features of the running state view and the asset parameter view based on the low-dimensional latent representation through two independent decoders, providing a basis for subsequent verification and attribution; The physical consistency constraint verification module is used to perform physical consistency constraint verification based on the output of the decoding and reconstruction module. It forces the output to conform to the physical operation mechanism of the power grid through a hard constraint correction mechanism based on Kirchhoff's current law and a soft constraint penalty mechanism based on Ohm's law. The fault attribution module is configured with a feature projection unit and a contrastive learning unit. The feature projection unit maps the low-dimensional latent representation to independent parameter subspaces and measurement subspaces. The contrastive learning unit with weighted hard negative samples decouples the mixed abnormal signals and locks the source of the abnormality. The decision output module is used to output targeted equipment calibration instructions or asset parameter correction suggestions based on the joint results of the physical consistency constraint verification module and the fault attribution module, thereby completing the collaborative correction of parameters and data.

Images