Power distribution network operation state supervision method and system based on artificial intelligence

Abstract

The application discloses a power distribution network operation state supervision method and system based on artificial intelligence, and the contents include data preparation, multi-modal time sequence feature extraction, time sequence graph abnormal diffusion modeling and abnormality discrimination. The application relates to the technical field of power grid intelligent supervision, and particularly discloses a power distribution network operation state supervision method and system based on artificial intelligence. According to the scheme, electrical and environmental modalities are separated, a double-flow Transformer is used to generate dynamic gate element-by-element modulation electrical characteristics from the environmental flow, the sensitivity and interpretability to environment-induced abnormalities are improved, a learnable time sequence edge is added on the physical topology, distance-sensitive graph attention and a self-recurrence discrete diffusion equation are combined to simulate the diffusion of abnormalities along the physical path, the reconstruction error, diffusion abnormality degree and speed are fused and adaptively set thresholds, gradient sensitivity analysis and attention path tracking are combined to accurately locate and output power grid operation state structured alarms, false alarms are reduced, and fault response is accelerated.

Description

Technical Field

[0001] This invention relates to the field of intelligent power grid monitoring technology, specifically to a method and system for monitoring the operation status of distribution networks based on artificial intelligence. Background Technology

[0002] With the large-scale deployment of distributed sensor networks and high-precision measurement devices, power distribution networks can now obtain multimodal time-series data at the second or even millisecond level, spurring a surge in research on data-driven operational status perception and anomaly detection. Artificial intelligence, especially deep learning methods, is rapidly penetrating the power system field, such as multi-head self-attention... The use of graph neural networks for temporal modeling, graph neural networks for topology sensing, and dynamic graph models based on diffusion or propagation processes for characterizing fault propagation all provide new approaches to capturing complex spatiotemporal dependencies. Meanwhile, the development of statistical extremum theory, interpretability methods, and edge computing provides feasible paths for constructing adaptive thresholds, improving localization interpretability, and achieving online deployment.

[0003] However, existing methods often rely on single-modal or simply pieced-together temporal representations, making it difficult to characterize the nonlinear modulation and conditional dependence of environmental factors on electrical quantities. Traditional graph models are mostly based on static topology, neglecting the impact of instantaneous coupling between nodes and physical and characteristic distances on anomaly propagation, thus failing to realistically simulate the temporal diffusion process of cascading faults. Anomaly detection generally relies on single indicators or manually set thresholds, resulting in insufficient noise resistance and adaptability, and the location methods often depend on heuristic rules, leading to low accuracy and confidence. In addition, the problems of data cleaning, synchronization and missing value handling brought about by high-resolution data, the scarcity of labels and insufficient model interpretability, as well as the computational and deployment costs in large-scale distribution networks, all restrict the reliable application of existing technologies in actual power grids. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for monitoring the operation status of distribution networks based on artificial intelligence. Addressing the limitations of existing single-current or simply spliced ​​time-series representation methods that cannot effectively characterize the nonlinear modulation of electrical quantities by environmental factors and struggle to simultaneously preserve long-range time-series dependencies and inter-modal conditional interactions, this solution employs environmental modulation dual-current... The architecture separates and embeds electrical and environmental flows, each capturing long-range dependencies. Dynamic gating of the environmental flow generates element-wise nonlinear modulation of the electrical flow, enabling selective amplification and suppression of electrical features by the environment. This significantly improves the sensitivity and interpretability of environmentally induced anomalies while maintaining high-quality temporal representations for subsequent anomaly detection. Addressing the shortcomings of existing graph models—such as lack of learnable temporal coupling, neglect of the impact of physical and feature distances on anomaly propagation, and inability to simulate the cascading propagation of anomalies along physical and temporal paths—this scheme introduces learnable temporal edges into the physical topology and parameterizes the adjacency matrix. It then calculates propagation weights using a graph attention mechanism with distance and feature penalties, and finally simulates multi-step propagation using a discrete iterative diffusion equation with autoregressive retained terms and controllable perturbations. The diffusion process can accurately reproduce the diffusion amplification or attenuation behavior of anomalies along physical and temporal paths, thereby achieving highly sensitive modeling and robust prediction of early propagation patterns of cascading faults. Addressing the shortcomings of existing anomaly detection methods based solely on single indicators or fixed thresholds, which are prone to false alarms and low location accuracy, this solution integrates reconstruction error, diffusion anomaly degree, and diffusion velocity to construct a comprehensive anomaly score. It then uses extreme value theory to adaptively estimate node thresholds for robust judgment. Furthermore, it combines gradient sensitivity analysis, attention-based path tracing, and node removal experiments to comprehensively calculate the location score and path confidence, ultimately outputting structured alarms. This makes anomaly detection more adaptive, interpretable, and accurate in location, effectively reducing false alarm rates and improving fault response efficiency.

[0005] The technical solution adopted in this invention is as follows: a method for monitoring the operation status of a power distribution network based on artificial intelligence, the method comprising the following steps:

[0006] Step S1: Data preparation. Real-time multimodal high-resolution time series data are collected from sensors at each node of the distribution network, including voltage, current, active and reactive power, frequency, and ambient temperature and humidity. The raw data is subjected to sliding window local statistical detection and significant outliers are replaced by local linear interpolation. Then, each parameter is normalized according to the independent minimum and maximum intervals and the mapping coefficients are saved.

[0007] Step S2: Multimodal temporal feature extraction. The electrical parameter flow and the environmental parameter flow are separated and mapped to a unified latent space through linear embedding and position encoding respectively. Then, the long-range temporal dependencies are captured by a multi-layer multi-head self-attention encoder. Subsequently, the electrical flow is subjected to element-level nonlinear modulation using the dynamic gating signal generated by the environmental flow, and the fused temporal feature representation is output.

[0008] Step S3: Temporal graph anomaly diffusion modeling. Learnable temporal coupling edges are introduced and the adjacency matrix is ​​parameterized. The propagation weight is calculated by combining a graph attention mechanism that considers physical distance and feature distance. The multi-step diffusion propagation of anomaly features is simulated on the graph using a discrete iterative diffusion equation with autoregressive retention terms and controllable perturbation is added to obtain a diffusion-enhanced node representation. At the same time, the reconstruction error, diffusion anomaly degree, and diffusion speed are calculated for anomaly characterization. The reconstruction accuracy, diffusion consistency, and temporal smoothness are constrained by a composite loss function.

[0009] Step S4: Anomaly detection. Based on reconstruction error, diffusion anomaly degree, and diffusion speed, a node-level comprehensive anomaly score is constructed. The peak overthreshold theory is used to fit the generalized Pareto distribution to determine the node adaptive threshold for anomaly detection. For node combinations that trigger anomalies, the location score and path confidence are calculated through gradient sensitivity analysis, attention-based path tracing, and node removal experiments. Finally, the detection and location results are transformed into a structured alarm containing information such as time node anomaly score and location path.

[0010] Furthermore, in step S1, the data preparation specifically includes the following steps:

[0011] Step S11: Data acquisition. Multimodal time series data, including voltage, current, active power, reactive power, frequency, and ambient temperature and humidity, are collected in real time from sensors at power grid nodes. The data is acquired through a distributed sensor network to form a high-resolution time series.

[0012] Step S12: Clean the raw data. First, use a sliding window to calculate local statistics to detect outliers. If the outlier exceeds the threshold, it is marked as an outlier and replaced with local linear interpolation.

[0013] Step S13: Normalization. Normalize the cleaned data for each parameter, mapping each parameter sequence to the [0,1] interval. Simultaneously, calculate and record the normalization parameters to support the interpretation of inverse normalization.

[0014] Furthermore, in step S2, the multimodal temporal feature extraction proposes an environment-modulated dual-stream approach. The architecture separates and processes electrical and environmental parameters, and then uses a dynamic gating signal generated from the environmental flow to nonlinearly modulate the electrical flow characteristics. Specifically, this includes the following steps:

[0015] Step S21: Dual-stream embedding, which splits the standardized data into two independent streams, electrical mode and environmental mode, and maps them to the latent space of the same dimension through a linear embedding layer, and adds sinusoidal position coding to preserve the dynamic characteristics of each mode;

[0016] Step S22: Merge, apply independent 3-layer methods to the two streams respectively. The encoder captures long-range dependencies and then innovatively introduces an environmental modulation attention module: the environmental flow output is used as a query to actively query electrical flow features to generate a modulation vector, which is then dynamically... Gating modulates electrical characteristics at the element level to achieve parameterized modeling of the amplification effect of environmental factors on electrical fluctuations; finally, it outputs a fused sequence to capture environmentally induced electrical risk patterns.

[0017] Further, in step S3, the time-series graph anomaly propagation modeling involves designing a dynamic propagation graph attention network. This is achieved by explicitly adding learnable weighted temporal edges to the graph topology and introducing feature distance-sensitive attention coefficients. This allows anomaly features to preferentially propagate along physical and temporal paths, enabling accurate simulation and high-sensitivity detection of early propagation patterns of cascading faults. Specifically, this includes the following steps:

[0018] Step S31: Construct a time-series graph, treating the distribution network as a graph at each sampling time, where the set of nodes is the power grid measurement point, the initial static edge set comes from the physical topology, and an additional learnable time-series edge set represents the instantaneous coupling between nodes in time. Generate the adjacency matrix of the graph at each time step and make it differentiable parameterized.

[0019] Step S32: Introduce graph attention. When propagating on the graph, the adjacency weights are modulated with attention so that the attention not only depends on feature similarity but is also sensitive to physical distance and feature distance. This allows abnormal features to spread preferentially on paths with physical nearest neighbors and similar features. Multi-head attention with distance penalty is used to output the propagation weight matrix between nodes.

[0020] Step S33: Discrete graph diffusion. The diffusion equation of discrete iteration is used to simulate the propagation of anomalous features on the graph. By combining attention weights and autoregressive retention terms, the node representation after multi-step diffusion is obtained, which can simulate the dynamics of anomalous amplification or attenuation along the path. The diffusion process also introduces controllable noise terms.

[0021] Step S34: Anomaly alert quantification. Based on the changes in node representation before and after diffusion and the neighborhood prediction deviation, two metrics, diffusion anomaly degree and diffusion rate, are defined for subsequent anomaly discrimination.

[0022] Step S35: Training objective: Define a composite loss function so that the model can both reconstruct normal samples and enhance the discrimination of anomalies while ensuring the consistency of the diffusion path and the smoothness of time. Use mini-batch time window training.

[0023] Furthermore, in step S4, the anomaly detection specifically includes the following steps:

[0024] Step S41: Anomaly scoring fusion, based on The obtained diffusion anomaly degree and diffusion rate are used to construct a node-level comprehensive anomaly score, and anomalies are judged by statistical extreme value theory.

[0025] Step S42: Anomaly source localization. This involves locating the set of nodes triggering the anomaly and tracing the causal path back to output the most likely initial fault node and cascading propagation path. A combination of gradient influence, attention path tracking, and neighborhood permutation experiments is used to obtain the localization score and path confidence. Specifically, this includes the following steps:

[0026] Step S421: Define the effect of gradient sensitivity on the score;

[0027] Step S422: Attention path tracking. Perform attention path tracking and find the maximum product path from the high-anomaly node to other nodes according to the attention weight, transforming it into a shortest path problem; select the path with the highest confidence as the candidate cascaded link.

[0028] Step S423: Calculate the final positioning score and the change in the overall abnormal score of the system after deleting the node; finally, normalize the three types of indicators and fuse them according to their weights to obtain the final positioning score.

[0029] Step S43: Anomaly Alarm. The judgment and location results are converted into structured alarm messages and output. Each alarm message includes a timestamp, node and line identifier, anomaly score, location score, and propagation path.

[0030] The distribution network operation status monitoring system based on artificial intelligence provided by the present invention includes a data preparation module, a multimodal time series feature extraction module, a time series diagram anomaly propagation modeling module, and an anomaly discrimination module;

[0031] The data preparation module collects multimodal high-resolution time series data in real time from sensors at each node of the distribution network, including voltage, current, active and reactive power, frequency, and ambient temperature and humidity. It performs sliding window local statistical detection on the raw data and replaces significant outliers with local linear interpolation. Then, it normalizes each parameter according to the independent minimum and maximum intervals and saves the mapping coefficients. Finally, it sends the data to the multimodal time series feature extraction module.

[0032] The multimodal time series feature extraction module receives data sent by the data preparation module, separates the electrical parameter stream and the environmental parameter stream, and maps them to a unified latent space through linear embedding and position encoding respectively. Then, each of them is captured by a multi-layer multi-head self-attention encoder to capture long-range time series dependencies. Subsequently, the electrical stream is subjected to element-level nonlinear modulation using the dynamic gating signal generated by the environmental stream, and the fused time series feature representation is output. The data is then sent to the time series graph anomaly diffusion modeling module.

[0033] The time-series graph anomaly diffusion modeling module receives data sent by the multimodal time-series feature extraction module, introduces learnable temporal coupling edges and parameterizes the adjacency matrix, calculates propagation weights by combining a graph attention mechanism that considers physical distance and feature distance, simulates the multi-step diffusion propagation of anomaly features on the graph using a discrete iterative diffusion equation with autoregressive retention terms and adds controllable perturbations to obtain a diffusion-enhanced node representation. At the same time, it calculates reconstruction error, diffusion anomaly degree and diffusion speed for anomaly characterization, and constrains reconstruction accuracy, diffusion consistency and temporal smoothness with a composite loss function, and sends the data to the anomaly discrimination module.

[0034] The anomaly discrimination module receives data sent by the time-series graph anomaly diffusion modeling module, constructs a node-level comprehensive anomaly score based on reconstruction error, diffusion anomaly degree, and diffusion speed, and uses peak overthreshold theory to fit a generalized Pareto distribution to determine the node adaptive threshold for anomaly judgment. For node combinations that trigger anomalies, the module calculates the location score and path confidence through gradient sensitivity analysis, attention-based path tracing, and node removal experiments. Finally, the discrimination and location results are converted into a structured alarm containing information such as time node anomaly score and location path.

[0035] The beneficial effects achieved by the present invention using the above solution are as follows:

[0036] (1) In view of the shortcomings of existing single-stream or simple splicing timing representation methods, which cannot effectively characterize the nonlinear modulation of electrical quantities by environmental factors and are difficult to retain long-range timing dependence and inter-modal conditional interaction at the same time, this scheme modulates dual streams through environmental modulation. The architecture separates and embeds electrical flow and environmental flow, each capturing long-range dependencies. The environmental flow generates dynamic gating to nonlinearly modulate the electrical flow element by element, enabling selective amplification and suppression of electrical features by the environment. This significantly improves the sensitivity and interpretability of environmentally induced anomalies, while maintaining high-quality temporal representation to facilitate subsequent anomaly detection.

[0037] (2) In view of the problems that existing graph models lack learnable temporal coupling, ignore the influence of physical distance and feature distance on anomaly propagation, and cannot simulate the cascading propagation of anomalies along physical and temporal paths, this scheme introduces learnable temporal edges on the physical topology and parameterizes the adjacency matrix. It calculates the propagation weight by combining a graph attention mechanism with distance and feature penalty, and then simulates the multi-step diffusion process with a discrete iterative diffusion equation with autoregressive retained terms and controllable perturbation. This scheme can accurately reproduce the diffusion amplification or attenuation behavior of anomalies along physical and temporal paths, thereby achieving high-sensitivity modeling and robust prediction of the early propagation mode of cascading faults.

[0038] (3) In view of the shortcomings of existing anomaly detection based on a single indicator or fixed threshold, which are prone to false alarms and low positioning accuracy, this solution constructs a comprehensive anomaly score by integrating reconstruction error, diffusion anomaly degree and diffusion speed, and uses extreme value theory to adaptively estimate node thresholds to achieve robust judgment. At the same time, it combines gradient sensitivity analysis, attention-based path tracking and node removal experiment to comprehensively calculate positioning score and path confidence, and finally outputs structured alarms, making anomaly detection more adaptive, interpretable and accurate, effectively reducing false alarm rate and improving fault response efficiency. Attached Figure Description

[0039] Figure 1 A schematic diagram of the distribution network operation status monitoring method based on artificial intelligence provided by the present invention;

[0040] Figure 2 A schematic diagram of the distribution network operation status monitoring system based on artificial intelligence provided by the present invention;

[0041] Figure 3 This is a schematic diagram of step S1;

[0042] Figure 4 This is a schematic diagram of step S2;

[0043] Figure 5 This is a schematic diagram of step S3;

[0044] Figure 6 This is a schematic diagram of step S4.

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0047] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0048] Example 1, see Figure 1 The present invention provides an artificial intelligence-based method for monitoring the operation status of a power distribution network, which includes the following steps:

[0049] Step S1: Data preparation. Real-time multimodal high-resolution time series data are collected from sensors at each node of the distribution network, including voltage, current, active and reactive power, frequency, and ambient temperature and humidity. The raw data is subjected to sliding window local statistical detection and significant outliers are replaced by local linear interpolation. Then, each parameter is normalized according to the independent minimum and maximum intervals and the mapping coefficients are saved.

[0050] Step S2: Multimodal temporal feature extraction. The electrical parameter flow and the environmental parameter flow are separated and mapped to a unified latent space through linear embedding and position encoding respectively. Then, the long-range temporal dependencies are captured by a multi-layer multi-head self-attention encoder. Subsequently, the electrical flow is subjected to element-level nonlinear modulation using the dynamic gating signal generated by the environmental flow, and the fused temporal feature representation is output.

[0051] Step S3: Temporal graph anomaly diffusion modeling. Learnable temporal coupling edges are introduced and the adjacency matrix is ​​parameterized. The propagation weight is calculated by combining a graph attention mechanism that considers physical distance and feature distance. The multi-step diffusion propagation of anomaly features is simulated on the graph using a discrete iterative diffusion equation with autoregressive retention terms and controllable perturbation is added to obtain a diffusion-enhanced node representation. At the same time, the reconstruction error, diffusion anomaly degree, and diffusion speed are calculated for anomaly characterization. The reconstruction accuracy, diffusion consistency, and temporal smoothness are constrained by a composite loss function.

[0052] Step S4: Anomaly detection. Based on reconstruction error, diffusion anomaly degree, and diffusion speed, a node-level comprehensive anomaly score is constructed. The peak overthreshold theory is used to fit the generalized Pareto distribution to determine the node adaptive threshold for anomaly detection. For node combinations that trigger anomalies, the location score and path confidence are calculated through gradient sensitivity analysis, attention-based path tracing, and node removal experiments. Finally, the detection and location results are transformed into a structured alarm containing information such as time node anomaly score and location path.

[0053] Example 2, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S1, the data preparation specifically includes the following steps:

[0054] Step S11: Data Acquisition. Multimodal time-series data, including voltage, current, active power, reactive power, frequency, and ambient temperature and humidity, is collected in real-time from sensors at power grid nodes. The data is acquired through a distributed sensor network with a sampling frequency of 1Hz, forming a high-resolution time series, as shown below:

[0055] ;

[0056] in, It is the original time series dataset, containing One sampling point; It is the i-th timestamp, in seconds, incrementing from 0; It is the voltage amplitude, measured in kilovolts. This is the current amplitude, measured in amperes (A). It is active power, measured in megawatts; It is reactive power, measured in megavars. It is frequency, measured in Hertz; This is the ambient temperature, expressed in degrees Celsius. It refers to relative humidity; This indicates the total number of sampling points; data is collected separately for each node.

[0057] Step S12: Clean the raw data. First, use a sliding window to calculate local statistics to detect outliers. If an outlier exceeds a threshold, it is marked as an anomaly and replaced with local linear interpolation, as shown below: ;if, ;but ;in, This represents the i-th value in any parameter sequence; and The center is Window size The local mean and standard deviation; It is a standardized score used to identify points that significantly deviate from the local trend; It is a fixed threshold, based on rule; These are the replaced values; the interpolation formula uses the two nearest normal points before and after; it is applied independently to all parameters and all nodes to form a clean dataset. ;

[0058] Step S13: Normalization. Perform min-max normalization on the cleaned data for independent parameters, mapping each parameter sequence to the [0,1] interval. Simultaneously, calculate and record the normalization parameters to support the interpretation of inverse normalization, as shown below:

[0059] ;

[0060] in, For the normalized formula, The formula for inverse normalization, It is the normalized value; and The parameter is in the entire The minimum and maximum values ​​are used within a fixed range for frequency. Highlight deviations; use actual extreme values ​​for voltage, current, and power, and use [specific values] for environmental parameters. and Scope of experience; forming standardized datasets Save each parameter and Used for explanation.

[0061] Example 3, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S2, the multimodal temporal feature extraction proposes an environment-modulated dual-stream approach. The architecture separates and processes electrical and environmental parameters, and then uses a dynamic gating signal generated from the environmental flow to nonlinearly modulate the electrical flow characteristics. Specifically, this includes the following steps:

[0062] Step S21: Dual-stream embedding. The standardized data is split into two independent streams: electrical mode and environmental mode. Each stream is mapped to a latent space of the same dimension through a linear embedding layer, and sinusoidal position coding is added to preserve the dynamic characteristics of each mode. The formulas for dual-stream embedding and position coding are as follows:

[0063] ;

[0064] in, , , , , , and These represent the normalized voltage amplitude, current amplitude, active power, reactive power, frequency, ambient temperature, and relative humidity, respectively. It is the first Each electrical flow embedding vector has dimensions. , Indicates electrical embedding weights; Indicates electrical embedded bias; It is the first Each environmental flow embedding has a dimension of 256; Indicates the context embedding weights. Indicates the environment embedding bias; It is a standard sinusoidal position code. , , Values ​​range from 0 to 127 to ensure timing sequence information; all weights are passed through... Initialization; forming a sequence, applying layer normalization after each layer. ;

[0065] Step S22: Merge, apply independent 3-layer methods to the two streams respectively. The encoder captures long-range dependencies and then innovatively introduces an environmental modulation attention module: the environmental flow output is used as a query to actively query electrical flow features to generate a modulation vector, which is then dynamically... Gating modulates electrical characteristics at the element level, enabling parameterized modeling of the amplification effect of environmental factors on electrical fluctuations. The electrical flow processing is represented as follows: Environmental flow processing is represented as: Modulation fusion is represented as: Each Transformer encoder layer contains Multi-head self-attention, 32-head dimension, feedforward network, residual connection and The feedforward network has an intermediate dimension of 512, which is composed of... Activation; attention calculation is For layer input projection, Indicates the transpose symbol; Indicates the fusion weight. Indicates fusion bias; yes Function, generation The gate vector; It is element-wise multiplication; the modulation term allows environmental features to selectively amplify specific dimensions of electrical features; the final output is a fused sequence. Dimension 256, capturing environmentally induced electrical risk patterns.

[0066] By performing the above operations, this solution addresses the shortcomings of existing single-stream or simple splicing timing representation methods, which cannot effectively characterize the nonlinear modulation of electrical quantities by environmental factors and are difficult to simultaneously preserve long-range timing dependencies and inter-modal conditional interactions. Instead, it employs environmental modulation of dual-stream systems. The architecture separates and embeds electrical flow and environmental flow, each capturing long-range dependencies. The environmental flow generates dynamic gating to nonlinearly modulate the electrical flow element by element, enabling selective amplification and suppression of electrical features by the environment. This significantly improves the sensitivity and interpretability of environmentally induced anomalies, while maintaining high-quality temporal representation to facilitate subsequent anomaly detection.

[0067] Example 4, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S3, the time-series graph anomaly propagation modeling involves designing a dynamic diffusion graph attention network. By explicitly adding learnable weighted temporal edges to the graph topology and introducing feature distance-sensitive attention coefficients, the abnormal features are preferentially propagated along physical and temporal paths. This enables accurate simulation and high-sensitivity detection of early propagation modes of cascading faults. Specifically, it includes the following steps:

[0068] Step S31: Construct a time sequence diagram, treating the distribution network as a graph at each sampling time. , where the set of nodes For power grid measurement points, the initial static edge set Derived from physical topology, with an additional set of learnable temporal edges. This represents the instantaneous coupling between nodes over time, generating the adjacency matrix for each timemap. And it is parameterized to be differentiable, as follows:

[0069] ;

[0070] in, Let the adjacency matrix at time t be denoted as . The physical topology adjacency matrix is ​​explicitly given by the wiring diagram. If the nodes and If physically connected, the corresponding element is 1; otherwise, it is 0. Represents a learnable temporally coupled scoring matrix; through a perceptron Output, representing time Additional coupling strength between nodes; This represents the activation function. Let e ​​denote the natural logarithm, and e denote the natural constant. Indicates time The initial feature vector of each node; For node feature dimensions; Temporal differences representing node features; A two-layer structure is adopted: first, each node is embedded into a low-dimensional space. Then through outer product Generate a rating matrix, add a learnable bias, and crop the diagonal to 0;

[0071] Step S32: Introduce graph attention. During propagation on the graph, attention modulation is applied to the adjacency weights, making the attention not only dependent on feature similarity but also sensitive to physical distance and feature distance. This ensures that anomalous features are preferentially propagated along paths with physical neighbors and similar features. Multi-head attention with distance penalty is used, and the propagation weight matrix between nodes is output. , means as follows:

[0072] ;

[0073] in, Represents a node For neighboring nodes Attention score Indicates time node The input features can be The row vector, and This represents the linear projection matrix of the query and the key. Dimension representing the attention of each individual; Represents a node and The physical distance between them and Let represent the learnable distance and the feature penalty weights, respectively. Represents the square of the Euclidean norm. Indicates section The set of candidate neighbors is defined as physical neighbors; Represents a node To the neighbors Normalized attention weights Indicates time Attention-weighted adjacency matrix;

[0074] Step S33: Discrete graph diffusion. The propagation of anomalous features on the graph is simulated using a discrete iterative diffusion equation, combined with attention weights. By combining the autoregressive retained terms, we obtain the node representation after multi-step diffusion, which enables the simulation of the dynamics of anomalies cascading amplification or attenuation along the path. The diffusion process also introduces a controllable noise term, as shown below:

[0075] ;

[0076] in, Indicates the diffusion process Step, moment The node representation matrix, and the nodes are node vectors; This represents the initial characterization of the diffusion process, taking steps... Fusion output And perform linear mapping to match dimensions ; The degree matrix is ​​defined as follows: ; yes The inverse matrix; It represents the diffusion coefficient, which controls the balance between propagation and autoregressive retention; Indicates the diffusion perturbation term; Indicates the number of diffusion iterations. This represents the node characterization after diffusion is complete.

[0077] Step S34: Anomaly Quantification. Based on the changes in node representation before and after diffusion and the neighborhood prediction bias, two metrics, diffusion anomaly degree and diffusion rate, are defined for subsequent anomaly detection. First, the node reconstruction error is calculated using the following formula: ;in, The formula for diffusion anomaly is: The diffusion rate formula is: ;in, Indicate steps Nodes after cleaning and normalization At any moment The original observation vector, This represents the observation vector reconstructed by the decoder, and... Same dimension, Indicate steps Output nodes Post-diffusion characterization, decoder It is a two-layer perceptron structure, Map back to the observation space and output the reconstruction; Indicates reconstruction error, Indicates the degree of diffusion anomaly. Indicates the diffusion rate. For time step, Number of neighbors;

[0078] Step S35: Training objective: Define a composite loss function so that the model can both reconstruct normal samples and enhance the discrimination of anomalies while ensuring the consistency of the diffusion path and the smoothness of time. Mini-batch time window training is used, as shown below:

[0079] ;

[0080] in, This represents the total loss, which needs to be minimized. Indicates the reconstruction loss. To train a set of consecutive time windows contained in a mini-batch, ; Indicates diffusion loss, Indicates smoothing loss. ; Represents the regularization term. This represents the set of all trainable parameters of the model; Indicates removal The subsequent set of all consecutive time windows; , , and Representing the loss weights, the optimizer uses The initial learning rate is The batch time window length is set to 1. Second.

[0081] By performing the above operations, this scheme addresses the problems of existing graph models lacking learnable temporal coupling, ignoring the impact of physical and feature distances on anomaly propagation, and being unable to simulate the cascading propagation of anomalies along physical and temporal paths. It introduces learnable temporal edges into the physical topology and parameterizes the adjacency matrix. Combined with a graph attention mechanism with distance and feature penalties, it calculates propagation weights and then uses a discrete iterative diffusion equation with autoregressive retained terms and controllable perturbations to simulate the multi-step diffusion process. This can accurately reproduce the diffusion amplification or attenuation behavior of anomalies along physical and temporal paths, thereby achieving highly sensitive modeling and robust prediction of early propagation patterns of cascading faults.

[0082] Example 5, see Figure 1 and Figure 6 This embodiment is based on the above embodiment. In step S4, the anomaly detection specifically includes the following steps:

[0083] Step S41: Anomaly scoring fusion, based on The obtained reconstruction error diffusion anomaly and diffusion rate Construct a node-level comprehensive anomaly score And through statistical extreme value theory An anomaly is indicated as follows:

[0084] ;

[0085] in, Represents a node At any moment The overall abnormal score; if Then determine the node At any moment This is an anomaly, threshold. Through statistical extreme value theory Estimate: For historical normal windows The extreme value exceeding the threshold portion fits the generalized Pareto distribution. And select significance level Obtain the threshold; , and Indicates the scoring weight. and This represents the sample mean and standard deviation of the reconstruction error within the historical window. and The sample mean and standard deviation represent the degree of diffusion anomalies within a historical window. and The sample mean and standard deviation represent the diffusion rate within the historical window; Representing node-level thresholds, fitting details: Let Percentile is the high threshold For more than peak Fitting Accounting in significance quantiles below and take ;

[0086] Step S42: Anomaly source localization. This involves locating the set of nodes triggering the anomaly and tracing the causal path back to output the most likely initial fault node and cascading propagation path. A combination of gradient influence, attention path tracking, and neighborhood permutation experiments is used to obtain the localization score and path confidence. Specifically, this includes the following steps:

[0087] Step S421: Define the gradient sensitivity effect score as follows:

[0088] ;

[0089] in, Represents a node The gradient influence represents the sensitivity to changes in overall outlier scores; Indicate steps Initial node characteristics;

[0090] Step S422: Attention path tracking, based on Attention path tracing involves finding the maximum product path from high-anomaly nodes to other nodes based on attention weights, transforming it into a shortest path problem: given a set of starting points... , for path Define path confidence The path with the highest confidence level is selected as the candidate cascade link; among which, Indicate steps Middle node right Attention weights;

[0091] Step S423: Calculate the final positioning score and calculate the deleted node. The system's overall abnormal score changes after input; finally, the three types of indicators are normalized and fused according to their weights to obtain the final positioning score. , means as follows:

[0092] ;

[0093] in, This represents an influence metric based on attention aggregation, for nodes. Sum of outgoing edge weights; This indicates the overall abnormal differences after node removal. , and Indicates the fusion weights; Indicates the final positioning score;

[0094] Step S43: Anomaly Alarm. The judgment and location results are converted into structured alarm messages and output. Each alarm message includes a timestamp, node and line identifier, anomaly score, location score, and propagation path.

[0095] By performing the above operations, this solution addresses the shortcomings of existing anomaly detection methods that rely solely on a single indicator or fixed threshold, which are prone to false alarms and low positioning accuracy. It constructs a comprehensive anomaly score by fusing reconstruction error, diffusion anomaly degree, and diffusion speed, and uses extreme value theory to adaptively estimate node thresholds for robust judgment. Furthermore, it combines gradient sensitivity analysis, attention-based path tracing, and node removal experiments to comprehensively calculate positioning scores and path confidence, ultimately outputting structured alarms. This makes anomaly detection more adaptive, interpretable, and accurate in positioning, effectively reducing false alarm rates and improving fault response efficiency.

[0096] Example 6, see Figure 2 Based on the above embodiments, the artificial intelligence-based power distribution network operation status monitoring system provided by the present invention includes a data preparation module, a multimodal time series feature extraction module, a time series diagram anomaly propagation modeling module, and an anomaly discrimination module.

[0097] The data preparation module collects multimodal high-resolution time series data in real time from sensors at each node of the distribution network, including voltage, current, active and reactive power, frequency, and ambient temperature and humidity. It performs sliding window local statistical detection on the raw data and replaces significant outliers with local linear interpolation. Then, it normalizes each parameter according to the independent minimum and maximum intervals and saves the mapping coefficients. Finally, it sends the data to the multimodal time series feature extraction module.

[0098] The multimodal time series feature extraction module receives data sent by the data preparation module, separates the electrical parameter stream and the environmental parameter stream, and maps them to a unified latent space through linear embedding and position encoding respectively. Then, each of them is captured by a multi-layer multi-head self-attention encoder to capture long-range time series dependencies. Subsequently, the electrical stream is subjected to element-level nonlinear modulation using the dynamic gating signal generated by the environmental stream, and the fused time series feature representation is output. The data is then sent to the time series graph anomaly diffusion modeling module.

[0099] The time-series graph anomaly diffusion modeling module receives data sent by the multimodal time-series feature extraction module, introduces learnable temporal coupling edges and parameterizes the adjacency matrix, calculates propagation weights by combining a graph attention mechanism that considers physical distance and feature distance, simulates the multi-step diffusion propagation of anomaly features on the graph using a discrete iterative diffusion equation with autoregressive retention terms and adds controllable perturbations to obtain a diffusion-enhanced node representation. At the same time, it calculates reconstruction error, diffusion anomaly degree and diffusion speed for anomaly characterization, and constrains reconstruction accuracy, diffusion consistency and temporal smoothness with a composite loss function, and sends the data to the anomaly discrimination module.

[0100] The anomaly discrimination module receives data sent by the time-series graph anomaly diffusion modeling module, constructs a node-level comprehensive anomaly score based on reconstruction error, diffusion anomaly degree, and diffusion speed, and uses peak overthreshold theory to fit a generalized Pareto distribution to determine the node adaptive threshold for anomaly judgment. For node combinations that trigger anomalies, the module calculates the location score and path confidence through gradient sensitivity analysis, attention-based path tracing, and node removal experiments. Finally, the discrimination and location results are converted into a structured alarm containing information such as time node anomaly score and location path.

[0101] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0103] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for monitoring the operation status of a power distribution network based on artificial intelligence, characterized in that: The method includes the following steps: Step S1: Data preparation. Real-time multimodal high-resolution time series data are collected from sensors at each node of the distribution network, including voltage, current, active and reactive power, frequency, and ambient temperature and humidity. The raw data is subjected to sliding window local statistical detection and significant outliers are replaced by local linear interpolation. Then, each parameter is normalized according to the independent minimum and maximum intervals and the mapping coefficients are saved. Step S2: Multimodal temporal feature extraction. The electrical parameter flow and the environmental parameter flow are separated and mapped to a unified latent space through linear embedding and position encoding respectively. Then, the long-range temporal dependencies are captured by a multi-layer multi-head self-attention encoder. Subsequently, the electrical flow is subjected to element-level nonlinear modulation using the dynamic gating signal generated by the environmental flow, and the fused temporal feature representation is output. Step S3: Temporal graph anomaly diffusion modeling. Learnable temporal coupling edges are introduced and the adjacency matrix is ​​parameterized. The propagation weight is calculated by combining a graph attention mechanism that considers physical distance and feature distance. The multi-step diffusion propagation of anomaly features is simulated on the graph using a discrete iterative diffusion equation with autoregressive retention terms and controllable perturbation is added to obtain a diffusion-enhanced node representation. At the same time, the reconstruction error, diffusion anomaly degree, and diffusion speed are calculated for anomaly characterization. The reconstruction accuracy, diffusion consistency, and temporal smoothness are constrained by a composite loss function. Step S4: Anomaly detection. Based on reconstruction error, diffusion anomaly degree, and diffusion speed, a node-level comprehensive anomaly score is constructed. The peak overthreshold theory is used to fit the generalized Pareto distribution to determine the node adaptive threshold for anomaly detection. For node combinations that trigger anomalies, the location score and path confidence are calculated through gradient sensitivity analysis, attention-based path tracing, and node removal experiments. Finally, the detection and location results are transformed into a structured alarm containing information such as time node anomaly score and location path.

2. The method for monitoring the operation status of a power distribution network based on artificial intelligence according to claim 1, characterized in that: In step S2, the multimodal temporal feature extraction proposes an environment-modulated dual-stream approach. The architecture separates and processes electrical and environmental parameters, and then uses a dynamic gating signal generated from the environmental flow to nonlinearly modulate the electrical flow characteristics. Specifically, this includes the following steps: Step S21: Dual-stream embedding, which splits the standardized data into two independent streams, electrical mode and environmental mode, and maps them to the latent space of the same dimension through a linear embedding layer, and adds sinusoidal position coding to preserve the dynamic characteristics of each mode; Step S22: Merge, apply independent 3-layer methods to the two streams respectively. The encoder captures long-range dependencies and then innovatively introduces an environmental modulation attention module: the environmental flow output is used as a query to actively query electrical flow features to generate a modulation vector, which is then dynamically... Gating modulates electrical characteristics at the element level to achieve parameterized modeling of the amplification effect of environmental factors on electrical fluctuations; finally, it outputs a fused sequence to capture environmentally induced electrical risk patterns.

3. The method for monitoring the operation status of a power distribution network based on artificial intelligence according to claim 1, characterized in that: In step S3, the time-series graph anomaly propagation modeling involves designing a dynamic propagation graph attention network. This is achieved by explicitly adding learnable weighted temporal edges to the graph topology and introducing feature distance-sensitive attention coefficients. This allows anomaly features to preferentially propagate along physical and temporal paths, enabling accurate simulation and high-sensitivity detection of early propagation patterns of cascading faults. Specifically, this includes the following steps: Step S31: Construct a time-series graph, treating the distribution network as a graph at each sampling time, where the set of nodes is the power grid measurement point, the initial static edge set comes from the physical topology, and an additional learnable time-series edge set represents the instantaneous coupling between nodes in time. Generate the adjacency matrix of the graph at each time step and make it differentiable parameterized. Step S32: Introduce graph attention. When propagating on the graph, the adjacency weights are modulated with attention so that the attention not only depends on feature similarity but is also sensitive to physical distance and feature distance. This allows abnormal features to spread preferentially on paths with physical nearest neighbors and similar features. Multi-head attention with distance penalty is used to output the propagation weight matrix between nodes. Step S33: Discrete graph diffusion. The diffusion equation of discrete iteration is used to simulate the propagation of anomalous features on the graph. By combining attention weights and autoregressive retention terms, the node representation after multi-step diffusion is obtained, which can simulate the dynamics of anomalous amplification or attenuation along the path. The diffusion process also introduces controllable noise terms. Step S34: Anomaly alert quantification. Based on the changes in node representation before and after diffusion and the neighborhood prediction deviation, two metrics, diffusion anomaly degree and diffusion rate, are defined for subsequent anomaly discrimination. Step S35: Training objective: Define a composite loss function so that the model can both reconstruct normal samples and enhance the discrimination of anomalies while ensuring the consistency of the diffusion path and the smoothness of time. Use mini-batch time window training.

4. The method for monitoring the operation status of a power distribution network based on artificial intelligence according to claim 1, characterized in that: In step S4, the anomaly detection specifically includes the following steps: Step S41: Anomaly scoring fusion, based on The obtained diffusion anomaly degree and diffusion rate are used to construct a node-level comprehensive anomaly score, and anomalies are judged by statistical extreme value theory. Step S42: Anomaly source localization. This involves locating the set of nodes triggering the anomaly and tracing the causal path back to output the most likely initial fault node and cascading propagation path. A combination of gradient influence, attention path tracking, and neighborhood permutation experiments is used to obtain the localization score and path confidence. Specifically, this includes the following steps: Step S421: Define the effect of gradient sensitivity on the score; Step S422: Attention path tracking. Perform attention path tracking and find the maximum product path from the high-anomaly node to other nodes according to the attention weight, transforming it into a shortest path problem; select the path with the highest confidence as the candidate cascaded link. Step S423: Calculate the final positioning score and the change in the overall abnormal score of the system after deleting the node; finally, normalize the three types of indicators and fuse them according to their weights to obtain the final positioning score. Step S43: Anomaly Alarm. The judgment and location results are converted into structured alarm messages and output. Each alarm message includes a timestamp, node and line identifier, anomaly score, location score, and propagation path.

5. The method for monitoring the operation status of a power distribution network based on artificial intelligence according to claim 1, characterized in that: In step S1, the data preparation specifically includes the following steps: Step S11: Data acquisition. Multimodal time series data, including voltage, current, active power, reactive power, frequency, and ambient temperature and humidity, are collected in real time from sensors at power grid nodes. The data is acquired through a distributed sensor network to form a high-resolution time series. Step S12: Clean the raw data. First, use a sliding window to calculate local statistics to detect outliers. If the outlier exceeds the threshold, it is marked as an outlier and replaced with local linear interpolation. Step S13: Normalization. Normalize the cleaned data for each parameter, mapping each parameter sequence to the [0,1] interval. Simultaneously, calculate and record the normalization parameters to support the interpretation of inverse normalization.

6. An artificial intelligence-based distribution network operation status monitoring system, used to implement the artificial intelligence-based distribution network operation status monitoring method as described in any one of claims 1-5, characterized in that: It includes a data preparation module, a multimodal time series feature extraction module, a time series graph anomaly diffusion modeling module, and an anomaly detection module.

7. The distribution network operation status monitoring system based on artificial intelligence according to claim 6, characterized in that: The data preparation module collects multimodal high-resolution time series data in real time from sensors at each node of the distribution network, including voltage, current, active and reactive power, frequency, and ambient temperature and humidity. It performs sliding window local statistical detection on the raw data and replaces significant outliers with local linear interpolation. Then, it normalizes each parameter according to the independent minimum and maximum intervals and saves the mapping coefficients. Finally, it sends the data to the multimodal time series feature extraction module. The multimodal time series feature extraction module receives data sent by the data preparation module, separates the electrical parameter stream and the environmental parameter stream, and maps them to a unified latent space through linear embedding and position encoding respectively. Then, each of them is captured by a multi-layer multi-head self-attention encoder to capture long-range time series dependencies. Subsequently, the electrical stream is subjected to element-level nonlinear modulation using the dynamic gating signal generated by the environmental stream, and the fused time series feature representation is output. The data is then sent to the time series graph anomaly diffusion modeling module. The time-series graph anomaly diffusion modeling module receives data sent by the multimodal time-series feature extraction module, introduces learnable temporal coupling edges and parameterizes the adjacency matrix, calculates propagation weights by combining a graph attention mechanism that considers physical distance and feature distance, simulates the multi-step diffusion propagation of anomaly features on the graph using a discrete iterative diffusion equation with autoregressive retention terms and adds controllable perturbations to obtain a diffusion-enhanced node representation. At the same time, it calculates reconstruction error, diffusion anomaly degree and diffusion speed for anomaly characterization, and constrains reconstruction accuracy, diffusion consistency and temporal smoothness with a composite loss function, and sends the data to the anomaly discrimination module. The anomaly discrimination module receives data sent by the time-series graph anomaly diffusion modeling module, constructs a node-level comprehensive anomaly score based on reconstruction error, diffusion anomaly degree, and diffusion speed, and uses peak overthreshold theory to fit a generalized Pareto distribution to determine the node adaptive threshold for anomaly judgment. For node combinations that trigger anomalies, the module calculates the location score and path confidence through gradient sensitivity analysis, attention-based path tracing, and node removal experiments. Finally, the discrimination and location results are converted into a structured alarm containing information such as time node anomaly score and location path.

Images