Method and system for monitoring load state of power distribution area based on stereo portrait analysis

Abstract

The invention discloses a power distribution area load state monitoring method and system based on stereo portrait analysis, relates to the technical field of urban power grids, and effectively improves the positioning precision and response speed of a load fluctuation area by introducing a three-dimensional image reconstruction and dynamic partitioning mechanism; a graph neural network and space-time long-short-term memory network modeling method is combined, so that the trend perception capability of the load state evolution process is enhanced; further, multi-source fusion recognition of abnormal behaviors is realized through association matching of the image change track and the load feature sequence, and the monitoring accuracy and sensitivity are effectively improved; based on risk index assessment and a map visualization expression mode, an abnormal area positioning result is visual and transparent, and intelligent early warning and decision assistance are supported; the method has the advantages of high data fusion capability, high risk identification precision, high visual interactivity and the like, and the intelligent and practical level of load state monitoring of the power distribution area is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of urban power grid technology, and in particular to a method and system for monitoring the load status of distribution substations based on three-dimensional image analysis. Background Technology

[0002] With the rapid development of urban power grids, distribution substations, as crucial nodes in the final transmission of electricity, directly impact the safety and reliability of users' electricity supply. Current technologies primarily rely on the remote acquisition and analysis of electrical parameters such as current and voltage for load status monitoring in distribution substations, supplemented by some image monitoring methods. However, these methods generally suffer from a specific problem: a lack of three-dimensional correlation analysis capabilities between the spatial structure and operating status of power distribution equipment, particularly limiting their ability to identify the causal relationship between abnormal load fluctuations and the on-site physical environment. Traditional two-dimensional images often lead to errors in identifying key components during equipment recognition due to angular obstruction or changes in lighting conditions, failing to accurately locate the equipment corresponding to load anomalies, thus affecting the accuracy and timeliness of fault warnings. Therefore, designing a distribution substation load status monitoring method based on three-dimensional image analysis to improve the accuracy of anomaly identification is essential. Summary of the Invention

[0003] This invention provides a method and system for monitoring the load status of distribution substations based on three-dimensional image analysis, which solves the technical problem of how to improve the accuracy of identifying abnormal load status in distribution substations.

[0004] The first aspect of this invention provides a method for monitoring the load status of a distribution substation based on three-dimensional image analysis, comprising:

[0005] A three-dimensional image model is constructed using multi-view image data of key equipment in the distribution area, and a spatial equipment feature set is generated based on the three-dimensional image model.

[0006] The feature set of the space equipment is subjected to multimodal fusion encoding to obtain a coupled feature vector;

[0007] A load state feature model is constructed using the coupled feature vectors, and a dynamic partitioning mechanism is introduced to mark load fluctuation areas in real time.

[0008] Based on the labeling results, the load trend index sequence is dynamically monitored, and the coupling determination result between the load trend index sequence and the image change trajectory feature sequence of the load fluctuation area is determined.

[0009] Anomaly identification is performed on the load fluctuation regions where the coupling determination result indicates high coupling, and multi-dimensional feature inversion is performed on potential anomaly regions to generate load anomaly response maps;

[0010] Based on the load anomaly response map, the potential anomaly areas are standardized and evaluated to generate a distribution area load status monitoring and analysis report.

[0011] Optionally, the step of constructing a three-dimensional image model using multi-view image data of key equipment within the acquired distribution area, and generating a spatial equipment feature set based on the three-dimensional image model, includes:

[0012] Acquire multi-view image data of key equipment in the distribution radio area, and perform image preprocessing on the multi-view image data;

[0013] Feature point matching is performed on the preprocessed results to obtain feature matching point pairs;

[0014] The feature matching point pairs are image registered and transformed, and coordinates are calculated in conjunction with a preset spatial projection model to determine the spatial three-dimensional coordinates of the feature matching points.

[0015] The sparse reconstruction method is used to integrate all the three-dimensional spatial coordinates to generate an initial point cloud.

[0016] The initial point cloud is density-enhanced, and a spatial feature model is constructed using the enhanced initial point cloud;

[0017] The spatial feature model is optimized to obtain a three-dimensional image model;

[0018] Based on the three-dimensional image model, the contour information of the key equipment is extracted, and the contour information is segmented and semantically recognized to obtain equipment instances with semantic tags;

[0019] The device instances are vector-encoded, and the encoding results are summarized to generate spatial device feature sets for each key device.

[0020] Optionally, the step of performing multimodal fusion encoding on the feature set of the space device to obtain a coupled feature vector includes:

[0021] The space device feature set is structured and encoded to obtain a structured dataset of device features;

[0022] Electrical parameters are collected based on the structured dataset of device features to obtain a time series dataset of electrical parameters;

[0023] A device image feature dataset is established based on the aforementioned three-dimensional image model;

[0024] The device image feature dataset is mapped to the electrical parameter time series dataset to generate an initial multimodal dataset;

[0025] The initial multimodal dataset is aligned and standardized to obtain a standard multimodal dataset;

[0026] A multimodal neural network with a pre-configured fusion attention mechanism is used to jointly model the standard multimodal dataset to obtain the coupling feature vectors of each key device.

[0027] Optionally, the step of constructing a load state feature model using the coupled feature vector and introducing a dynamic partitioning mechanism to mark load fluctuation areas in real time includes:

[0028] The coupled feature vectors are input into the preset graph neural network architecture model to construct an initial device topology graph;

[0029] The actual physical connection relationships between each key device are embedded based on the initial device topology diagram, and weights are defined to represent the coupling strength to obtain the target device topology diagram.

[0030] The local neighborhood features of the target device topology graph are iteratively aggregated through graph convolution operations to generate a load state feature model.

[0031] The load state characteristic model is divided into grids to obtain multiple spatial sub-regions of the distribution substation;

[0032] Based on each preset time step, the rate of change of coupling characteristics of multiple spatial sub-regions is calculated in real time.

[0033] When the rate of change of the coupling feature is greater than a preset change threshold, the spatial sub-region is marked as a load fluctuation region.

[0034] Optionally, the step of dynamically monitoring the load trend index sequence based on the labeling results and determining the coupling result between the load trend index sequence and the image change trajectory feature sequence of the load fluctuation area includes:

[0035] The load status time-series output sequence of the load fluctuation region is obtained by parsing the labeling results.

[0036] The load state time series output sequence is used as input to a pre-set spatial-temporal long short-term memory network model to construct a load change time series model;

[0037] By monitoring the first trend index of the load change time series model output, a load trend index sequence is obtained;

[0038] Extract the image change trajectory feature sequence of each key device within the load fluctuation area;

[0039] Normalize the load trend index sequence and the image change trajectory feature sequence;

[0040] The similarity index between the normalized load trend index sequence and the normalized image change trajectory feature sequence is calculated using the dynamic time warping matching algorithm.

[0041] When the similarity index is greater than or equal to a preset index threshold, it is determined that the image changes of the key equipment are highly coupled with the abnormal load fluctuations in the load fluctuation area.

[0042] When the similarity index is less than the preset index threshold, it is determined that the image changes of the key equipment are not highly coupled with the abnormal load fluctuations in the load fluctuation area.

[0043] Optionally, the step of identifying anomalies in the load fluctuation regions where the coupling determination result indicates high coupling, and performing multi-dimensional feature inversion on potential anomaly regions to generate a load anomaly response map, includes:

[0044] Calculate a second trend index for the load fluctuation region where high coupling exists;

[0045] The second trend indicator is used as input to a preset multi-parameter stability assessment model to perform a stability assessment and obtain a stability score.

[0046] When the stability score is greater than or equal to a preset score threshold, and the second trend index meets the preset stability characteristic condition, the load fluctuation area is marked as a stable change area.

[0047] When the stability score is less than the preset score threshold, and the second trend indicator meets the preset abnormal characteristic conditions, the load fluctuation area is marked as a potential abnormal area.

[0048] Extract multimodal features from the potential anomaly regions to generate a joint feature vector;

[0049] Calculate the regional risk index score of the joint feature vector;

[0050] Layer mapping is performed on the regional risk index scores, and the visualization results are overlaid on the three-dimensional image model to generate a load anomaly response map.

[0051] Optionally, the step of standardizing the assessment of the potential anomaly area based on the load anomaly response map and generating a distribution substation load status monitoring and analysis report includes:

[0052] Analyze the image features, electrical parameter change features, and time-series evolution data of each potential abnormal region in the load anomaly response spectrum;

[0053] Risk assessment is performed by inputting the image features of the potential anomaly region, the electrical parameter change features, and the time-series evolution data into a pre-set expert rule base.

[0054] The risk assessment results are standardized and formatted to generate a load status monitoring and analysis report for the distribution area.

[0055] The second aspect of this invention provides a distribution area load status monitoring system based on stereoscopic image analysis, comprising:

[0056] The construction module is used to construct a three-dimensional image model using multi-view image data of key equipment in the acquired distribution area, and to generate a spatial equipment feature set based on the three-dimensional image model.

[0057] The encoding module is used to perform multimodal fusion encoding on the feature set of the space device to obtain a coupled feature vector;

[0058] The marking module is used to construct a load state feature model using the coupled feature vectors and introduce a dynamic partitioning mechanism to mark load fluctuation areas in real time.

[0059] The monitoring module is used to dynamically monitor the load trend index sequence based on the labeling results, and determine the coupling judgment result between the load trend index sequence and the image change trajectory feature sequence of the load fluctuation area;

[0060] The identification module is used to identify anomalies in the load fluctuation areas where the coupling determination result indicates high coupling, and to perform multi-dimensional feature inversion on potential anomaly areas to generate a load anomaly response map.

[0061] The evaluation module is used to perform standardized evaluation of the potential abnormal areas based on the load anomaly response map and generate a distribution area load status monitoring and analysis report.

[0062] The third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the distribution area load status monitoring method based on stereoscopic image analysis as described in any of the preceding claims.

[0063] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the distribution area load status monitoring method based on stereoscopic image analysis as described in any of the preceding claims.

[0064] As can be seen from the above technical solutions, the present invention has the following advantages:

[0065] This invention effectively improves the positioning accuracy and response speed of load fluctuation areas by introducing a three-dimensional image reconstruction and dynamic zoning mechanism. It enhances the trend perception capability of load state evolution by combining graph neural networks and spatial-temporal long short-term memory network modeling methods. Furthermore, through the correlation and matching of image change trajectories and load feature sequences, it achieves multi-source fusion identification of abnormal behavior, effectively improving the accuracy and sensitivity of monitoring. Based on risk index assessment and map visualization, the abnormal area positioning results are intuitive and transparent, supporting intelligent early warning and decision support. This invention has advantages such as strong data fusion capability, high risk identification accuracy, and strong visualization interactivity, significantly improving the intelligence and practicality of distribution transformer area load state monitoring, and has broad application prospects and engineering promotion value. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 A flowchart illustrating the steps of a method for monitoring the load status of a distribution substation based on stereoscopic image analysis, provided in an embodiment of the present invention;

[0068] Figure 2 A structural block diagram of a distribution substation load status monitoring system based on stereoscopic image analysis provided in an embodiment of the present invention;

[0069] Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0070] This invention provides a method and system for monitoring the load status of distribution substations based on three-dimensional image analysis, which addresses the technical problem of improving the accuracy of identifying load anomalies in distribution substations.

[0071] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0072] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a method for monitoring the load status of a distribution substation based on stereoscopic image analysis, as provided in this embodiment of the invention.

[0073] This invention provides a method for monitoring the load status of distribution substations based on three-dimensional image analysis, comprising:

[0074] Step 101: Construct a three-dimensional image model using multi-view image data of key equipment in the distribution area, and generate a spatial equipment feature set based on the three-dimensional image model.

[0075] Further, step 101 may include the following sub-steps:

[0076] S11. Acquire multi-view image data of key equipment in the distribution area and perform image preprocessing on the multi-view image data.

[0077] Key equipment, in the context of a power distribution substation, refers to various electrical devices that play a crucial role in the normal operation of the power distribution system and the stability and security of power supply. These devices typically include, but are not limited to, transformers, which are used to change voltage levels and achieve the rational transmission and distribution of electrical energy; switchgear, which controls, protects, and monitors circuits, enabling the connection and disconnection of circuits; cables, as important carriers of electrical energy, responsible for transmitting power from one point to another; and circuit breakers, fuses, and other equipment that directly affects the reliable operation of the power system.

[0078] Multi-view image data refers to a collection of images of key equipment taken from different angles and positions within the distribution area.

[0079] Image preprocessing refers to operations performed on raw multi-view images, including grayscale conversion, format unification, and resolution adjustment.

[0080] In this embodiment of the invention, grayscale preprocessing operations such as grayscale conversion are performed on the acquired multi-view images to unify the image format and resolution.

[0081] S12. Perform feature point matching on the preprocessing results to obtain feature matching point pairs.

[0082] Feature point matching refers to the process of identifying and establishing the relationship between corresponding feature points in images from different perspectives or at different times through algorithms.

[0083] A feature matching point pair refers to a set of two corresponding feature points from different images, determined through a feature point matching process. One feature point comes from a reference image, and the other comes from the image to be matched. The two feature points are consistent in spatial location or feature attributes.

[0084] Key points refer to unique and identifiable local feature points in an image.

[0085] A descriptor is a vectorized description of information such as the image gradient around a key point. It can reflect the features of the key point, and the features extracted by the scale-invariant feature transformation algorithm have scale invariance and rotation invariance, which can adapt to the differences in device shape under different viewpoints.

[0086] Nearest neighbor matching refers to finding the most similar match for each descriptor in a set of descriptors from different images.

[0087] The random sampling consensus algorithm is an algorithm that iteratively selects random samples, fits a model, and verifies sample consistency to eliminate outliers.

[0088] In this embodiment of the invention, for each preprocessed image, the Scale-Invariant Feature Transform (SIFT) algorithm is first used to extract key points and corresponding descriptors. Then, based on the descriptors of key points between images, nearest neighbor matching is performed to filter out possible corresponding feature points between different images, forming a candidate corresponding point set. Finally, the Random Sample Consensus (RANSAC) algorithm is used to verify the matching relationship in the candidate corresponding point set, eliminating mismatched points that do not conform to geometric constraints, and retaining stable and reliable feature matching point pairs.

[0089] S13. Perform image registration and transformation on the feature matching point pairs, and calculate the coordinates in conjunction with the preset spatial projection model to determine the spatial three-dimensional coordinates of the feature matching points.

[0090] Image registration and transformation refer to the process of adjusting images from different viewpoints, times, or sensors to the same coordinate system by calculating the geometric transformation relationship between different images. "Registration" is the operation of establishing the correspondence between images, while "transformation" is the operation of unifying the spatial position of images through geometric transformations such as affine and projection, with the aim of eliminating positional deviations between images.

[0091] A pre-defined spatial projection model refers to a mathematical model that describes the projection of three-dimensional spatial points onto a two-dimensional image plane through an imaging device. It includes the geometric laws and intrinsic and extrinsic parameters of camera imaging and is used to establish the mapping relationship between two-dimensional image points and three-dimensional spatial points.

[0092] Three-dimensional spatial coordinates refer to the X, Y, and Z coordinate parameters used in a pre-defined world coordinate system to describe the specific location of a feature point in three-dimensional space. These coordinates uniquely determine the orientation and position of a feature point in space and are the core data for constructing a three-dimensional image model.

[0093] The homography matrix is ​​used to describe the projection mapping relationship of points in the same plane between different images.

[0094] The fundamental matrix refers to the epipolar geometric constraints between two camera views.

[0095] Camera intrinsic parameters refer to the imaging characteristics of the camera itself, such as focal length and principal point coordinates.

[0096] Camera extrinsic parameters refer to the position and orientation of the camera in the world coordinate system.

[0097] Triangulation is a method that uses observations of the same point from different camera angles to determine the spatial location of that point through geometric calculations.

[0098] In this embodiment of the invention, the homography matrix or fundamental matrix between images is first calculated based on the feature matching point pairs to achieve affine or projection transformation of the images, and finally all viewpoint images are registered to the same reference coordinate system. Then, by combining the preset spatial projection model and the camera's intrinsic and extrinsic parameters, the projection relationship between the registered images of each viewpoint and the world coordinate system is established. Triangulation is performed on the registered feature matching point pairs to obtain the spatial three-dimensional coordinates of each feature matching point in the world coordinate system.

[0099] S14. Use the sparse reconstruction method to integrate all spatial three-dimensional coordinates to generate an initial point cloud.

[0100] Sparse reconstruction methods refer to techniques for constructing preliminary point clouds from the three-dimensional coordinates of a limited set of feature points.

[0101] Coordinate integration refers to the process of merging, calibrating, and associating multiple discrete 3D coordinates obtained from different perspectives and calculated through feature matching point pairs, according to a unified world coordinate system. The purpose is to eliminate possible deviations in coordinate calculations from different perspectives and integrate scattered 3D coordinate data into a coherent and consistent overall dataset.

[0102] The initial point cloud refers to a set of point data formed by the spatial three-dimensional coordinates of key feature points, which can initially represent the spatial distribution of equipment.

[0103] In this embodiment of the invention, a sparse reconstruction method is used to integrate all these spatial three-dimensional coordinates to generate an initial point cloud. The spatial boundary positions of each key device in the distribution area are then calibrated using this initial point cloud, thus initially presenting the spatial distribution pattern of the devices.

[0104] S15. Density enhancement is performed on the initial point cloud, and a spatial feature model is constructed using the enhanced initial point cloud.

[0105] Density enhancement refers to increasing the number of points in a point cloud through technical means, thereby increasing the density of the point cloud and enabling it to present the shape of objects in greater detail.

[0106] Structured light scanning technology refers to the use of structured light projection and acquisition to obtain three-dimensional information of an object's surface to enhance point clouds.

[0107] Deep neural network reconstruction technology refers to the use of neural networks to learn the features of point clouds to generate denser point clouds.

[0108] Spatial feature model refers to a preliminary three-dimensional image structure that describes the geometric relationships of key equipment in a distribution station area.

[0109] In this embodiment of the invention, it is first determined whether the initial point cloud density meets the modeling requirements. If not, structured light scanning technology or deep neural network reconstruction technology is introduced to supplement and optimize the initial point cloud to generate a denser point cloud with higher density. Then, the initial point cloud or the enhanced dense point cloud is organized into a unified spatial model. At the same time, a globally unified world coordinate system is constructed, and the point cloud data of all devices are aligned to this coordinate system and the scale and orientation are unified to construct a spatial feature model.

[0110] S16. Optimize the spatial feature model to obtain a three-dimensional image model.

[0111] Mesh reconstruction refers to the process of converting discrete point cloud data into a continuous surface mesh model, which can represent the surface morphology of equipment. Noise reduction refers to the process of removing isolated points and abnormal surfaces from the mesh model, which can improve the surface smoothness of the model and the recognizability of the equipment's shape. A 3D stereoscopic image model refers to an image model that can accurately and intuitively display the 3D shape and spatial relationships of key equipment in a distribution substation area.

[0112] In this embodiment of the invention, a boundary extraction algorithm is used to reconstruct the point cloud data in the spatial feature model to generate a continuous surface mesh model of the device; then the mesh model is simplified to reduce complexity, and noise reduction is performed at the same time. Through these operations, the spatial feature model is optimized to finally obtain a three-dimensional image model.

[0113] S17. Extract the contour information of key equipment based on the three-dimensional image model, and perform segmentation and semantic recognition on the contour information to obtain equipment instances with semantic labels.

[0114] Simplification and noise reduction refer to removing isolated points and outliers from the mesh model to improve its smoothness and recognizability. Pre-trained device classification models are models that have been trained beforehand and possess the ability to classify device types. Semantic labels are tags that identify semantic information such as device type.

[0115] Segmentation and semantic recognition refer to the process of first dividing the contour information of key equipment in a 3D image model into several independent regions (each region corresponds to one or some equipment) using image segmentation algorithms, and then using semantic recognition technology to perform joint processing of category judgment and attribute labeling on each segmented region. The purpose is to separate different equipment objects from image data and assign them clear semantic information.

[0116] A device instance refers to a specific device individual with clear boundaries and semantic labels obtained through segmentation and semantic recognition in a distribution substation scenario. It is a digital mapping of key power equipment such as transformers, switchgear, and cables that actually exist in the physical world. Each device instance contains information such as its spatial location, geometric shape, and semantic labels, and is the basic unit for constructing a spatial device feature set.

[0117] In this embodiment of the invention, based on a three-dimensional image model, the shape contour is first extracted and the boundary is reconstructed to identify the spatial structural features of various power equipment. Then, the boundary extraction algorithm is used to reconstruct the initial point cloud data into a grid to obtain a continuous surface. Next, the grid model is simplified and noise is reduced. Then, key contour lines are extracted based on the grid topology to identify the contour shape of the equipment. Finally, the contour information is segmented and semantically recognized. Combined with a pre-trained equipment classification model, the type classification of various power equipment in the distribution substation is completed, thereby obtaining equipment instances with semantic labels.

[0118] S18. Perform vector encoding on the device instances and summarize the encoding results to generate spatial device feature sets for each key device.

[0119] Device feature vector encoding refers to converting a device instance into a vector containing its geometric parameters, such as size, volume, and spatial coordinate range, as well as semantic labels, such as device type and functional attributes. Initial feature set integration refers to summing up all device feature vectors to form a complete initial spatial device feature set.

[0120] In this embodiment of the invention, each identified device instance is converted into a vector form for encoding, and then the feature vectors of all device instances are summarized to generate a spatial device feature set for each key device.

[0121] Step 102: Perform multimodal fusion encoding on the feature set of space equipment to obtain coupled feature vectors.

[0122] In this embodiment of the invention, the feature set of space equipment is subjected to structured encoding processing, and the coupling relationship between the electrical parameter change features and image features of each device is extracted based on the multimodal fusion algorithm to obtain the coupling feature vector.

[0123] Furthermore, step 102 may include the following sub-steps:

[0124] S21. Perform structured encoding on the spatial equipment feature set to obtain a structured dataset of equipment features.

[0125] A space equipment feature set refers to a collection of feature vectors summarizing key equipment. Structured coding refers to the process of transforming unstructured or semi-structured equipment feature data into structured data with a fixed format, facilitating computer processing and analysis. A structured equipment feature dataset refers to a set of equipment features presented in a standardized data structure after structured coding, enabling more efficient support for subsequent data processing and analysis operations.

[0126] In this embodiment of the invention, the feature vectors of each key device in the spatial device feature set are formatted, organized, and stored according to a preset data structure specification, thereby obtaining a structured dataset of device features.

[0127] S22. Based on the structured dataset of equipment features, electrical parameters are collected to obtain a time series dataset of electrical parameters.

[0128] Electrical parameters refer to parameters such as voltage, current, and frequency that reflect the electrical operating status of power equipment. An electrical parameter time series dataset refers to a collection of electrical parameter data organized in chronological order, demonstrating the changing patterns of electrical parameters over time.

[0129] In this embodiment of the invention, an edge electrical parameter acquisition module is configured for each type of key equipment within the distribution transformer area. A multi-functional power parameter acquisition instrument is used to record electrical parameters such as voltage, current, power factor, frequency, and harmonics of each key equipment in real time, according to a typical acquisition cycle of 1s to 60s that can reflect changes in operating status. The acquired electrical parameter data is accompanied by a high-precision timestamp and a unique device identifier ID is recorded. Then, this electrical parameter data is uploaded to an edge computing node or a central server, stored in a time series format, and combined with a structured dataset of device features. The electrical parameter data is aligned with the image features of the corresponding device through timestamps and normalized to obtain an electrical parameter time series dataset.

[0130] S23. Establish a device image feature dataset based on a three-dimensional image model.

[0131] Static image features refer to the inherent visual characteristics of an image presented by a device in a stable state, such as shape and texture. Dynamic change features refer to the visual changes in a device image over time or during operation.

[0132] The device image feature dataset refers to the collection formed by summarizing the static and dynamic image features of each device node.

[0133] In this embodiment of the invention, an image feature database is established for each device node in the three-dimensional image model, and static image features and dynamic change features are extracted. Based on this, a device image feature dataset containing these features is established based on the three-dimensional image model.

[0134] S24. Map the device image feature dataset to the electrical parameter time series dataset to generate an initial multimodal dataset.

[0135] Device ID refers to the unique identification code used to identify each key device within a distribution area. Initial multimodal dataset refers to a collection that integrates different modalities of data, such as images and electrical parameters, reflecting the device status from multiple dimensions.

[0136] In this embodiment of the invention, the image features of each device in the device image feature dataset are mapped and associated with the electrical parameter time series dataset by using spatial coordinates and device ID, thereby generating an initial multimodal dataset.

[0137] S25. Align and standardize the initial multimodal dataset to obtain a standard multimodal dataset.

[0138] Standardization refers to the process of converting data with different dimensions and value ranges into a unified scale.

[0139] In this embodiment of the invention, timestamps are used to synchronize and align image frames and electrical parameter time series in the initial multimodal dataset. For example, linear interpolation or window averaging is used to correct incompletely aligned time points. Then, the aligned data is standardized to obtain a standard multimodal dataset.

[0140] S26. A multimodal neural network with a pre-set fusion attention mechanism is used to jointly model the standard multimodal dataset to obtain the coupling feature vectors of each key device.

[0141] Joint modeling refers to the process of integrating and modeling two different types (modalities) of data: image features and electrical parameter change trends, using a multimodal neural network with a fusion attention mechanism.

[0142] The coupled feature vector refers to the image-electrical parameter joint state of the device at the current moment, which is used as a multimodal input feature to construct the load state feature model of the distribution area load state.

[0143] In this embodiment of the invention, a pre-built fusion attention mechanism multimodal neural network with a dual-branch neural network structure is used. One branch processes the image feature input from the standard multimodal dataset, and the other branch processes the electrical parameter time series. The outputs of the two networks are then fed into the multimodal attention fusion module. By dynamically weighting the influence of different modalities through channel attention and temporal attention, a joint representation vector is constructed to capture the potential coupling pattern between image changes and electrical parameter anomalies. During the training phase, supervised learning is used (the labels can be equipment operating status, fault level, or load level). Finally, a coupling feature vector with a fixed dimension is output. This is used to jointly model the standard multimodal dataset to obtain the coupling feature vectors of each key device.

[0144] Step 103: Construct a load state feature model using coupled feature vectors, and introduce a dynamic partitioning mechanism to mark load fluctuation areas in real time.

[0145] The dynamic zoning mechanism refers to the mechanism that automatically divides the distribution area into multiple dynamically changing regional units based on the real-time load status reflected by the coupling characteristic vectors of each key device within the distribution area, combined with preset load fluctuation thresholds and spatial topology relationships, and marks areas where the load exceeds the normal fluctuation range in real time. This mechanism can adaptively adjust the zoning boundaries according to the dynamic changes in load status, accurately locating the area of ​​abnormal load fluctuations.

[0146] In this embodiment of the invention, a load status feature model integrating multimodal information is constructed using the coupled feature vectors of each key device. This model can output the load status parameters of the entire distribution area and each device in real time. At the same time, a dynamic partitioning mechanism is introduced. Based on the output of the load status feature model and the spatial distribution of the devices, the distribution area is divided into different regions according to the degree of load fluctuation. When the load fluctuation of the devices in a certain region exceeds the set threshold, the region is immediately marked in real time, so as to realize the rapid identification and location of abnormal load regions.

[0147] Furthermore, step 103 may include the following sub-steps:

[0148] S31. Using coupled feature vectors as input to the preset graph neural network architecture model, construct the initial device topology graph.

[0149] The initial equipment topology diagram refers to a graph that can truly reflect the structural dependencies between power equipment, serving as the basic structure for subsequent feature propagation and aggregation.

[0150] In this embodiment of the invention, the coupling feature vectors of each power device extracted and fused from image information and electrical parameters are used as node inputs of a pre-set graph neural network architecture model, with each device corresponding to a node in the model network. Then, based on the actual layout information of the devices in the transformer area, all nodes are numbered and a correspondence with the device entities is established. Then, the edges between nodes are constructed according to the physical connection logic of the power devices in the transformer area, thereby generating an initial device topology graph.

[0151] S32. Based on the initial device topology diagram, embed the actual physical connection relationship between each key device, and define weights to represent the coupling strength to obtain the target device topology diagram.

[0152] The target device topology diagram refers to the graph obtained by embedding the actual physical connection relationships between key devices on the basis of the initial device topology diagram, and defining weights to represent the coupling strength.

[0153] In this embodiment of the invention, the actual physical connection relationships between key devices are embedded in the initial device topology diagram. For any pair of device nodes with a connection relationship in the diagram, their actual physical connection attributes are further embedded, such as defining the structural basic connection strength based on whether it is a direct electrical connection, connection voltage level, connection distance, or functional role. Then, the coupling degree at the feature level is calculated by combining the numerical similarity between the coupling feature vectors of the two devices. After fusing these two types of indicators, the final edge weight is defined. The edge weight is used to represent the importance of device connection to information propagation. The higher the weight, the closer the connection between the two devices in terms of structure and function, and the greater the possibility of mutual influence of features. In this way, the target device topology diagram is obtained.

[0154] S33. Iteratively aggregate the local neighborhood features of the target device topology graph through graph convolution operations to generate a load state feature model.

[0155] Graph convolution operations refer to convolution operations performed on graph-structured data. By weighting and aggregating the features of the target node and its neighboring nodes according to the connection weights, the features of the nodes in the graph are updated and information is transmitted. It can effectively capture the topological relationships and feature dependencies between nodes in the graph and is one of the core operations for processing graph-structured data.

[0156] Iterative aggregation refers to the process in graph neural networks where feature aggregation operations such as graph convolution are repeatedly performed in multiple rounds. This allows node features to continuously integrate information from their neighboring nodes, gradually expanding from a local neighborhood to a larger range, ultimately achieving the integration of global information and improving the completeness and representativeness of feature representations.

[0157] Device local neighborhood features refer to the feature information contained in the neighboring device nodes that are directly or indirectly connected to a certain device node in the topology diagram of the target device. These features include the coupling characteristics, connection attributes, and operating status of the neighboring devices, reflecting the status and mutual influence of related devices around the target device.

[0158] The load status feature model refers to the model generated by iteratively aggregating the local neighborhood features of the target equipment topology graph through graph convolution operations. It integrates individual equipment attributes, spatial topology relationships, coupling strength and multimodal features, and can comprehensively characterize the load status of the entire distribution substation area and each equipment. It can be used for tasks such as load level assessment, trend prediction and anomaly detection.

[0159] In this embodiment of the invention, local neighborhood features of the target device topology graph are first aggregated using graph convolution operations to form an initial load state feature model with global perception capabilities. After constructing the weighted topology graph, graph convolution operations are used to iteratively update the features of the device nodes. In each round of graph convolution, the features of a device node and the features of its neighboring nodes are weighted and aggregated according to the connection weights, thereby realizing the extraction and fusion of neighborhood information. By stacking multiple layers of graph convolution structures, each node can not only perceive the information of directly connected devices, but also gradually perceive the transformer area structure and operating status in a larger range, thereby obtaining a more representative load feature representation. Finally, after processing by the graph neural network, the high-dimensional feature vector of each device node constitutes the initial load state feature model. This model comprehensively considers the individual attributes, spatial location, functional role of the device, and its coupling relationship in the overall transformer area structure, and has global perception capabilities, which can provide high-quality input for subsequent power load analysis and anomaly detection tasks. Then, the initial load state feature model is further processed. The model is optimized as follows: An adaptive graph attention mechanism is embedded in the graph neural network model to dynamically adjust the edge weights in the adjacency matrix based on the similarity between node features, thereby improving the accuracy of identifying the coupling relationship of key equipment; a multi-level graph convolutional network structure is designed to extract load features from different topological ranges, and a unified representation is achieved through a feature fusion module, enhancing the model's ability to jointly model local and global information; to avoid the oversmoothing problem caused by deep graph convolution, a residual connection structure is used to maintain the original node features, while layer normalization technology is combined to improve the model's training stability and generalization ability; different data perturbations are applied to the nodes in the graph to construct positive and negative sample pairs, and the consistency of node representation is optimized by comparing loss functions to improve the model's recognition accuracy under abnormal load conditions; multiple task branches are set in the model output layer to predict the current load level, future change trend, and potential anomaly probability, and a weighted loss function is introduced to achieve joint optimization, thereby improving the model's multi-dimensional adaptability in actual load monitoring scenarios and ultimately generating a load status feature model.

[0160] S34. The load state characteristic model is divided into grids to obtain multiple spatial sub-regions of the distribution substation.

[0161] A spatial sub-region refers to a region that, after being divided, contains a certain number of power equipment nodes and serves as the basic spatial unit for dynamic analysis.

[0162] In this embodiment of the invention, the load state characteristic model is divided into grids. Based on the spatial coordinate system of the constructed load state characteristic model, the entire distribution substation area is processed into regular grids. According to the equipment distribution density, power topology or actual geographical boundary, an appropriate division scale is selected to divide the entire power supply substation area into several non-overlapping spatial sub-regions. Each sub-region contains a certain number of power equipment nodes. The division result of the sub-regions will be used as the basic spatial unit for dynamic analysis, thereby obtaining multiple spatial sub-regions of the distribution substation area.

[0163] S35. Based on each preset time step, calculate the rate of change of coupling characteristics of multiple spatial sub-regions in real time.

[0164] The coupling characteristic change rate refers to a value used to quantify load state fluctuations within a region and serves as a key indicator for fluctuation identification. It reflects the degree of change in the coupling characteristics of equipment within a sub-region over time.

[0165] In this embodiment of the invention, after entering the dynamic monitoring stage, the latest round of device coupling feature vectors is obtained in each time step according to the set sampling period. For each spatial sub-region, the feature vectors of all device nodes in the region are counted and compared with the corresponding features of the previous time step. The coupling feature change rate of the sub-region is comprehensively calculated by using a multi-dimensional change rate measurement method. Based on each preset time step, the coupling feature change rate of multiple spatial sub-regions is calculated in real time.

[0166] For example, for a certain spatial sub-region, first take the average value of the coupling feature vectors of all devices in the region within the current time step as the current region feature value, and then take the average feature value of the corresponding region in the previous time step as the historical region feature value; then calculate the Euclidean distance (or Manhattan distance) between the two, divide the distance by the time step interval (e.g., 2 minutes), and the result is the rate of change of coupling features of the spatial sub-region in the current time step. The larger the value, the more significant the load state fluctuation in the region.

[0167] S36. When the rate of change of coupling characteristics is greater than the preset change threshold, the spatial sub-region is marked as a load fluctuation region.

[0168] In this embodiment of the invention, a set of dynamic or static thresholds are set as the basis for fluctuation judgment. When the rate of change of coupling characteristics of a certain spatial sub-region exceeds the corresponding threshold, the load fluctuation judgment mechanism is triggered, and the sub-region is immediately marked as a load fluctuation region. The fluctuation intensity and duration are recorded, including the rate of change value, start time and duration. At the same time, the marking status is continuously updated through a sliding window mechanism. Based on a sliding time window of a set length, the historical rate of change data of each sub-region is dynamically updated and smoothed. The mean, maximum value or trend index within the window is statistically analyzed to comprehensively determine whether the current marking status should be maintained or canceled. If the rate of change of fluctuation of a certain region is lower than the threshold in multiple consecutive time steps, the fluctuation marking is automatically canceled, thereby realizing real-time tracking of dynamic load fluctuation regions.

[0169] It is worth mentioning that the sliding window mechanism enables continuous, stable, and real-time tracking of load fluctuation areas within the distribution area, enhancing the system's ability to respond to local load anomalies.

[0170] Step 104: Based on the labeling results, dynamically monitor the load trend index sequence and determine the coupling result between the load trend index sequence and the image change trajectory feature sequence of the load fluctuation area.

[0171] Furthermore, step 104 may include the following sub-steps:

[0172] S41. Parse the labeling results to obtain the load status time-series output sequence of the load fluctuation area.

[0173] Load status time-series output sequence refers to a data sequence arranged in chronological order that reflects changes in load status within a load fluctuation area.

[0174] In this embodiment of the invention, for each load fluctuation region, after it is marked, the system continuously records the load state feature model output of the region. At each time step, the system extracts the aggregation result of the initial load state feature model of the sub-region and stores it in the time series cache. Each dynamic region maintains a corresponding time series cache, and then the marking result is parsed to obtain the load state time series output sequence of the load fluctuation region.

[0175] S42. A load change time series model is constructed by inputting the load state time series output sequence into a pre-set spatial-temporal long short-term memory network model.

[0176] Spatial-temporal Long Short-Term Memory (LSTM) networks are network models that integrate spatial structure inputs such as the current region's spatial coupling features or graph convolution embedding results, as well as time-series inputs such as the state output values ​​from the previous time step and several historical time steps. Through a gating mechanism, they memorize the long-term trends and short-term disturbances of the region's state, outputting a hidden state vector that reflects the current state trend. Load variation time-series models, based on spatial-temporal LSM networks, integrate spatial structure and time-series inputs, using a gating mechanism to memorize the long-term trends and short-term disturbances of the region's load, and are used to model the time-series evolution of load fluctuation sub-regions.

[0177] In this embodiment of the invention, for each marked load fluctuation region, the load state time series output sequence is input into a preset spatial-temporal long short-term memory network to construct a load change time series model that can track the evolution trajectory of model parameters and model the time series evolution law of load state in that region.

[0178] S43. Monitor the first trend index of the load change time series model output to obtain the load trend index sequence.

[0179] The first trend indicator refers to the set of indicators used to quantify the characteristics of load state changes in the output of the load change time series model. It mainly includes trend slope (reflecting the rate of increase or decrease of load state), fluctuation frequency (reflecting the number of high and low frequency switching of state per unit time), and change amplitude (measuring the intensity of fluctuation). These indicators can intuitively capture the evolution trend of load state over time.

[0180] Load trend index sequence refers to a data sequence recorded sequentially over time steps, including trend slope, fluctuation frequency, and change amplitude, used to measure the changing trend of regional load status.

[0181] In this embodiment of the invention, based on the predicted output of the load change time series model, the system extracts a first trend index at each time step, including the trend slope (by linearly fitting the model output values ​​of several time steps to calculate the rate of increase or decrease of load state change), fluctuation frequency (based on the Fourier transform or zero-crossing analysis of the model output time series, to count the number of high-frequency and low-frequency state switching per unit time), and change amplitude (extracting the range or standard deviation of the model output within the current sliding window to measure the intensity of fluctuation), and arranges these three indices in chronological order to obtain a load trend index sequence.

[0182] S44. Extract the image change trajectory feature sequence of each key device within the load fluctuation area.

[0183] Image change trajectory feature sequence refers to a data sequence of image-related features recorded in chronological order, such as changes in the color histogram of key equipment, displacement of structural contours, and texture difference maps.

[0184] In this embodiment of the invention, color histogram analysis is performed on the image regions of each key device within the load fluctuation area to extract the color distribution change trajectory over time in RGB or HSV space, reflecting possible physical state changes such as heating, aging, and stain accumulation. An edge detection algorithm is used to extract the edge contour of the device, and an image registration algorithm is used to track the contour position changes between consecutive time points, recording the deformation or displacement trajectory. Based on texture feature methods such as local binary mode or gray-level co-occurrence matrix, the texture differences of the same device in images at different times are analyzed to generate a time-series texture difference map, identifying subtle changes such as surface wear and condensation, thereby extracting the image change trajectory feature sequence of each key device within the load fluctuation area.

[0185] S45. Normalize the load trend index sequence and the image change trajectory feature sequence.

[0186] In this embodiment of the invention, the load trend index sequence and the image change trajectory feature sequence are normalized to map their values ​​to a unified numerical range for subsequent fusion analysis and other operations.

[0187] It is worth mentioning that scale normalization was performed on the load trend index sequence and the image change trajectory feature sequence respectively, so that data with different dimensions and distributions can be effectively compared.

[0188] S46. The similarity index between the normalized load trend index sequence and the normalized image change trajectory feature sequence is calculated using the dynamic time warping matching algorithm.

[0189] The similarity index is an indicator calculated using a dynamic time warping algorithm to compare the feature sequence of image change trajectories with the load state output sequence. First, the dynamic time warping algorithm is used to obtain the minimum cumulative distance between the two sequences. Then, based on this distance, a similarity score is converted using a formula, which can be normalized to the 0-1 range. A higher value indicates that the image change is more likely to be associated with abnormal load fluctuations, and it is used to measure the degree of correlation between image changes and load fluctuations.

[0190] In this embodiment of the invention, the minimum cumulative distance between the normalized load trend index sequence and the normalized image change trajectory feature sequence is calculated using a dynamic time warping algorithm, and the formula is as follows:

[0191]

[0192] In the formula, Represents the feature sequence of image change trajectory. This represents the load status output sequence. Represents the set of all possible alignment paths. Indicates the first The image features and the first The distance between each load state characteristic.

[0193] Then, calculate the similarity index between image changes and load fluctuations using the following formula:

[0194]

[0195] In the formula, Represents the similarity index. This represents the smoothing coefficient.

[0196] By using the exponential decay transformation of the dynamic time warping distance, the original non-negative distance is converted into a similarity score, which has good alignment flexibility and can adapt to the time lag or rate inconsistency between image changes and load fluctuations. It is suitable for correlation analysis scenarios between multimodal time series. At the same time, through the dynamic time warping algorithm, the image change trajectory sequence and the load state sequence are flexibly aligned on the time axis to find the minimum distance path to match the nonlinear change process. Based on the shortest distance value obtained by dynamic time warping matching, a similarity index between image changes and load fluctuations is defined. The index can be normalized to the range of 0-1, where the higher the value, the more likely the image change is to be associated with abnormal load fluctuations.

[0197] S47. When the similarity index is greater than or equal to the preset index threshold, it is determined that the image changes of the key equipment are highly coupled with the abnormal load fluctuations in the load fluctuation area.

[0198] In this embodiment of the invention, when the similarity index calculated by the dynamic time warping matching algorithm is greater than or equal to a preset index threshold, it is determined that the image changes of the key equipment are highly coupled with the abnormal load fluctuations in the load fluctuation area.

[0199] S48. When the similarity index is less than the preset index threshold, it is determined that the image changes of the key equipment and the abnormal load fluctuations in the load fluctuation area are not highly coupled.

[0200] In this embodiment of the invention, when the similarity index is less than a preset index threshold, it is determined that the image changes of the key equipment and the abnormal load fluctuations in the load fluctuation area are not highly coupled.

[0201] Step 105: Identify anomalies in load fluctuation areas where the coupling determination result indicates high coupling, and perform multi-dimensional feature inversion on potential anomaly areas to generate load anomaly response maps.

[0202] Furthermore, step 105 may include the following sub-steps:

[0203] S51. Calculate the second trend index for load fluctuation regions with high coupling.

[0204] The second trend indicator refers to a set of indicators used to measure the direction, rate, and volatility of load fluctuations in highly coupled regions, including the average slope of the trend, the standard deviation of the rate of change, and the root mean square deviation.

[0205] In this embodiment of the invention, for a load fluctuation region with high coupling, within a set time window, the regional load state characteristic sequence and the image change trajectory sequence are linearly fitted to obtain the average slope of the change trend (used to measure the direction and rate of change). At the same time, statistical indicators such as the standard deviation and root mean square deviation of the rate of change per unit time in the time series are calculated to quantify the volatility of the rate of change. These indicators together constitute the second change trend index, thereby calculating the second change trend index of the load fluctuation region.

[0206] S52. The second trend indicator is used as input to the preset multi-parameter stability assessment model to conduct a stability assessment and obtain a stability score.

[0207] Stability score refers to the score obtained by a multi-parameter stability assessment model based on the similarity between current regional characteristics and historical stable samples to quantitatively assess the stability of a region, and is used to measure the degree of stability of the region.

[0208] In this embodiment of the invention, samples of changing parameters of different power equipment under normal and stable conditions are first preset to form a typical model set of stable conditions. Then, the second trend index (which will not be repeated if explained in the previous steps) is constructed into a multi-dimensional feature vector. The multi-dimensional feature vector is then input into a preset multi-parameter stability assessment model. By comparing the similarity between the features of the current region and historical stable samples, a stability score is output. Thus, the second trend index is used to input into the preset multi-parameter stability assessment model to perform stability assessment and obtain a stability score.

[0209] For example, in a region with highly coupled load fluctuations, the second trend indicator is processed to form a multi-dimensional feature vector. The average slope of the trend is 0.2 (the average slope of historical stable samples is 0.1, with a standard deviation of 0.05), and the standard deviation of the rate of change is 0.1 (the standard deviation of historical stable samples is 0.08, with a standard deviation of 0.02). This vector is input into a multi-parameter stability assessment model. The model calculates a similarity of 0.8 with historical stable samples, and outputs a stability score of 80 (out of 100; a higher score indicates greater stability). This indicates that the current state of the region is relatively close to its historical stable state, but there are still some fluctuations.

[0210] S53. When the stability score is greater than or equal to the preset score threshold, and the second trend indicator meets the preset stability characteristic conditions, the load fluctuation area is marked as a stable change area.

[0211] Preset stable characteristic conditions refer to the characteristic standards used to determine whether a load fluctuation area is in a stable state of change. They usually refer to the slope and fluctuation frequency of the second trend indicator tending to be gentle within a specified time window, without obvious abrupt changes, reflecting the stability of the coordinated change of regional load and image features.

[0212] In this embodiment of the invention, when the stability score output by the multi-parameter stability assessment model is greater than or equal to the preset score threshold, and the second trend index meets the preset stability characteristic conditions (i.e., the slope and fluctuation frequency tend to level off within a specified time window and no longer exhibit abrupt changes), the load fluctuation area is marked as a stable change area.

[0213] S54. When the stability score is less than the preset score threshold and the second trend indicator meets the preset abnormal characteristic conditions, the load fluctuation area is marked as a potential abnormal area.

[0214] Preset abnormal feature conditions refer to the feature standards used to determine whether there are potential anomalies in the load fluctuation area. Generally, they refer to the continuous strengthening of the change trend in the second change trend indicator, or the appearance of periodic abnormal fluctuations and other features that do not conform to the stable state, reflecting the abnormality of the coordinated change of regional load and image features.

[0215] In this embodiment of the invention, when the stability score is less than a preset score threshold and the second trend indicator meets the preset abnormal characteristic conditions (i.e., the trend of change continues to increase or periodic anomalies occur), the load fluctuation area is marked as a potential abnormal area.

[0216] S55. Extract multimodal features from potential abnormal regions and generate joint feature vectors.

[0217] Multimodal features refer to a set of features extracted from potential anomaly areas that cover different types and can reflect the state of the area from multiple dimensions. Specifically, they include image features (such as histograms of texture changes in images of key equipment, differences in structural edge contours, image illumination and color shifts, etc.), load features (such as load change trends, load model prediction errors, etc.), and electrical parameter features (such as changes in voltage, current, power factor, phase angle shift, etc. within a set time window). These features provide a basis for anomaly analysis from different dimensions such as visual, load change, and electrical parameters.

[0218] The joint feature vector refers to the vector obtained by fusing multimodal features such as images, loads, and electrical parameters of a potential anomaly region, and is used to comprehensively characterize the anomaly-related information of that region.

[0219] In this embodiment of the invention, multimodal features are extracted for potential abnormal regions. Image features include texture change histograms, structural edge contour differences, image illumination and color shifts extracted from images of key equipment within the potential abnormal regions. Load features include load change trends and load model prediction errors. Electrical parameter features include changes in voltage, current, power factor, phase angle shift, etc., within a set time window. These multimodal features are then fused using a weighted fusion method to generate a joint feature vector.

[0220] S56. Calculate the regional risk index score of the joint feature vector.

[0221] The regional risk index score, a value between 0 and 1, is the output of the joint feature vector after processing by a multilayer perceptron. It represents the potential risk level of abnormal load in the current region; a higher value indicates a greater risk. Dropout, an overfitting prevention mechanism, is a technique used during neural network training to randomly and temporarily discard some neurons and their connections with a certain probability. This reduces excessive dependence between neurons, forcing the network to learn more robust and generalizable features, thus effectively preventing the model from overfitting to the training data and improving its adaptability to new data.

[0222] In this embodiment of the invention, a multilayer perceptron architecture (typically consisting of an input layer, hidden layer 1, hidden layer 2, and output layer, with each layer containing an activation function and a Dropout anti-overfitting mechanism) is used to input the joint feature vector into the multilayer perceptron network, thereby calculating the regional risk index score.

[0223] For example, the joint feature vector of a potential anomaly region contains multimodal feature data such as image texture changes, load change trends, and voltage changes. This vector is input into a multilayer perceptron (structure: input layer - hidden layer 1 - hidden layer 2 - output layer, each layer having an activation function and a Dropout anti-overfitting mechanism). After network processing, the output region risk index score is 0.8, indicating a high degree of potential load anomaly risk in this region.

[0224] S57. Perform layer mapping on the regional risk index score, overlay the visualization results onto the three-dimensional image model, and generate a load anomaly response map.

[0225] The load anomaly response map refers to a map that uses layer mapping to overlay regional risk index scores onto a 3D image model in a pseudo-color or other visual format. It can intuitively display the distribution and degree of load anomaly risk in different regions, and each marked region can display dynamic risk index labels, supporting interactive clicking to view detailed load and image change records for that region.

[0226] In this embodiment of the invention, the calculated regional risk index score is first mapped to the spatial coordinates of the abnormal area. Then, a pseudo-color mapping function is used (e.g., red represents high risk, yellow represents medium risk, and green represents low risk). Next, the risk layer is superimposed on the system's three-dimensional image model, and transparency control is used to achieve layer fusion, ultimately generating a load anomaly response map.

[0227] Step 106: Based on the load anomaly response map, conduct a standardized assessment of potential anomaly areas and generate a distribution area load status monitoring and analysis report.

[0228] Furthermore, step 106 may include the following sub-steps:

[0229] S61. Analyze the image features, electrical parameter change features, and time-series evolution data of each potential abnormal region in the load anomaly response spectrum.

[0230] In this embodiment of the invention, the image features, electrical parameter change features, and time-series evolution data of each potential abnormal region in the load anomaly response spectrum are analyzed.

[0231] S62. Risk assessment is conducted by inputting image features, electrical parameter change features, and time-series evolution data of potential anomaly regions into a pre-set expert rule base.

[0232] The expert rule base refers to a set of rules that includes professional rules such as the operation specifications and equipment standards of distribution substations, used to assess and classify the risk level of abnormal areas.

[0233] In this embodiment of the invention, these multi-dimensional feature data (image features, electrical parameter change features, and time-series evolution data) are input into a pre-set expert rule base for risk assessment.

[0234] S63. Standardize and format the risk assessment results to generate a distribution area load status monitoring and analysis report.

[0235] A distribution area load status monitoring and analysis report refers to a report that includes the area number, risk level, recommended handling measures, and subsequent monitoring strategies, and is used to present the results of distribution area load status monitoring and anomaly analysis.

[0236] In this embodiment of the invention, the risk assessment results are standardized and formatted to generate a distribution area load status monitoring and analysis report.

[0237] Please see Figure 2 , Figure 2 This is a structural block diagram of a distribution substation load status monitoring system based on stereoscopic image analysis, provided as an embodiment of the present invention.

[0238] This invention provides a distribution substation load status monitoring system based on stereoscopic image analysis, comprising: a construction module 201, used to construct a three-dimensional stereoscopic image model using multi-view image data of key equipment within the distribution substation, and generate a spatial equipment feature set based on the three-dimensional stereoscopic image model; an encoding module 202, used to perform multimodal fusion encoding on the spatial equipment feature set to obtain coupled feature vectors; a marking module 203, used to construct a load status feature model using the coupled feature vectors, and introduce a dynamic partitioning mechanism to mark load fluctuation areas in real time; a monitoring module 204, used to dynamically monitor the load trend index sequence based on the marking results, and determine the coupling judgment result between the load trend index sequence and the image change trajectory feature sequence of the load fluctuation area; an identification module 205, used to identify anomalies in load fluctuation areas where the coupling judgment result indicates high coupling, and perform multi-dimensional feature inversion on potential anomaly areas to generate a load anomaly response map; and an evaluation module 206, used to perform standardized evaluation of potential anomaly areas based on the load anomaly response map, and generate a distribution substation load status monitoring and analysis report.

[0239] Furthermore, the construction module 201 includes: an image preprocessing submodule, used to acquire multi-view image data of each key device in the distribution area and perform image preprocessing on the multi-view image data; a feature matching point pair submodule, used to perform feature point matching on the preprocessing results to obtain feature matching point pairs; a spatial three-dimensional coordinate submodule, used to perform image registration and transformation on the feature matching point pairs, and perform coordinate calculation in combination with a preset spatial projection model to determine the spatial three-dimensional coordinates of the feature matching points; an initial point cloud submodule, used to integrate all spatial three-dimensional coordinates using a sparse reconstruction method to generate an initial point cloud; a spatial feature model submodule, used to enhance the density of the initial point cloud and construct a spatial feature model using the enhanced initial point cloud; a three-dimensional stereo image model submodule, used to optimize the spatial feature model to obtain a three-dimensional stereo image model; a device instance submodule, used to extract the contour information of key devices based on the three-dimensional stereo image model, and perform segmentation and semantic recognition on the contour information to obtain device instances with semantic labels; and a spatial device feature set submodule, used to perform vector encoding on the device instances and summarize the encoding results to generate a spatial device feature set for each key device.

[0240] Furthermore, the encoding module 202 includes: a device feature structured dataset submodule, used to perform structured encoding on the spatial device feature set to obtain a device feature structured dataset; an electrical parameter time series dataset submodule, used to collect electrical parameters based on the device feature structured dataset to obtain an electrical parameter time series dataset; a device image feature dataset submodule, used to establish a device image feature dataset based on a three-dimensional image model; an initial multimodal dataset submodule, used to map the device image feature dataset and the electrical parameter time series dataset to generate an initial multimodal dataset; a standard multimodal dataset submodule, used to align and standardize the initial multimodal dataset to obtain a standard multimodal dataset; and a coupled feature vector submodule, used to jointly model the standard multimodal dataset using a multimodal neural network with a pre-set fusion attention mechanism to obtain coupled feature vectors for each key device.

[0241] Further, the labeling module 203 includes: an initial equipment topology graph submodule, used to construct an initial equipment topology graph by inputting coupling feature vectors into a preset graph neural network architecture model; a target equipment topology graph submodule, used to embed the actual physical connection relationships between key equipment based on the initial equipment topology graph and define weights to represent coupling strength to obtain a target equipment topology graph; a load state feature model submodule, used to iteratively aggregate the local neighborhood features of the target equipment topology graph through graph convolution operations to generate a load state feature model; a spatial sub-region submodule, used to perform grid division on the load state feature model to obtain multiple spatial sub-regions of the distribution substation; a coupling feature change rate submodule, used to calculate the coupling feature change rate of multiple spatial sub-regions in real time based on each preset time step; and a load fluctuation region submodule, used to label the spatial sub-region as a load fluctuation region when the coupling feature change rate is greater than a preset change threshold.

[0242] Further, the monitoring module 204 includes: a load status time-series output sequence submodule, used to parse the labeling results to obtain the load status time-series output sequence of the load fluctuation area; a load change time-series model submodule, used to input the load status time-series output sequence into a preset spatial-temporal long short-term memory network model to construct a load change time-series model; a load trend index sequence submodule, used to monitor the first change trend index of the load change time-series model output results to obtain a load trend index sequence; an image change trajectory feature sequence submodule, used to extract the image change trajectory feature sequence of each key device in the load fluctuation area; a normalization submodule, used to normalize the load trend index sequence and the image change trajectory feature sequence; a similarity index submodule, used to calculate the similarity index between the normalized load trend index sequence and the normalized image change trajectory feature sequence using a dynamic time warping matching algorithm; a first processing submodule, used to determine that the image change of the key device and the abnormal load fluctuation in the load fluctuation area are highly coupled when the similarity index is greater than or equal to a preset index threshold; and a second processing submodule, used to determine that the image change of the key device and the abnormal load fluctuation in the load fluctuation area are not highly coupled when the similarity index is less than a preset index threshold.

[0243] Further, the identification module 205 includes: a second trend index submodule, used to calculate a second trend index for load fluctuation areas with high coupling; a stability scoring submodule, used to input the second trend index into a preset multi-parameter stability assessment model to perform stability assessment and obtain a stability score; a third processing submodule, used to mark the load fluctuation area as a stable change area when the stability score is greater than or equal to a preset scoring threshold and the second trend index meets preset stability feature conditions; a fourth processing submodule, used to mark the load fluctuation area as a potential abnormal area when the stability score is less than a preset scoring threshold and the second trend index meets preset abnormal feature conditions; a joint feature vector submodule, used to extract multimodal features of potential abnormal areas and generate a joint feature vector; a regional risk index numerator module, used to calculate the regional risk index score of the joint feature vector; and a load anomaly response map submodule, used to perform layer mapping on the regional risk index score, overlay the visualization results onto a three-dimensional image model, and generate a load anomaly response map.

[0244] Furthermore, the evaluation module 206 includes: a parsing submodule, used to parse the image features, electrical parameter change features, and time-series evolution data of each potential anomaly region in the load anomaly response map; a risk assessment submodule, used to input the image features, electrical parameter change features, and time-series evolution data of the potential anomaly regions into a pre-set expert rule base for risk assessment; and a standardization and formatting submodule, used to standardize and format the risk assessment results to generate a distribution area load status monitoring and analysis report.

[0245] Please see Figure 3 , Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of the present invention.

[0246] An electronic device according to an embodiment of the present invention includes: a memory 301 and a processor 302. The memory 301 stores a computer program. When the computer program is executed by the processor 302, the processor 302 executes the distribution area load status monitoring method based on stereoscopic image analysis as described in any of the above embodiments.

[0247] Memory 301 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 301 has storage space 303 for program code 313 for performing any of the method steps described above. For example, storage space 303 for program code may include various program codes 313 for implementing the various steps in the methods described above. These program codes may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. These program codes may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When these codes are run by a computing processing device, the device performs the various steps in the distribution area load status monitoring method based on stereoscopic image analysis described above.

[0248] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the distribution area load status monitoring method based on stereoscopic image analysis as described in any of the above embodiments.

[0249] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0250] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0251] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0252] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0253] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0254] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the load status of a distribution area based on stereoscopic image analysis, characterized by, The method comprises the following steps: adopting the multi-view image data of each key equipment in the power distribution area to construct a three-dimensional image model, and generating a spatial equipment feature set based on the three-dimensional image model; performing multi-modal fusion coding on the spatial equipment feature set to obtain a coupling feature vector; constructing a load state feature model using the coupling feature vector, and introducing a dynamic partition mechanism to mark the load fluctuation area in real time; based on the marking result, dynamically monitoring the load trend index sequence, and determining the coupling determination result between the load trend index sequence and the image change trajectory feature sequence of the load fluctuation area; performing abnormal identification on the load fluctuation area with high coupling in the coupling determination result, and performing multi-dimensional feature inversion on the potential abnormal area to generate a load abnormal response atlas; according to the load abnormal response atlas, performing standardized evaluation on the potential abnormal area to generate a power distribution area load state monitoring analysis report.

2. The method of claim 1, wherein, The method comprises the following steps: adopting the multi-view image data of each key equipment in the power distribution area to construct a three-dimensional image model, and generating a spatial equipment feature set based on the three-dimensional image model; obtaining multi-view image data of each key equipment in the power distribution area, and performing image preprocessing on the multi-view image data; performing feature point matching on the preprocessing result to obtain a feature matching point pair; performing image registration and transformation on the feature matching point pair, and combining a preset space projection model to calculate coordinates to determine the spatial three-dimensional coordinates of the feature matching points; integrating all the spatial three-dimensional coordinates using a sparse reconstruction method to generate an initial point cloud; enhancing the density of the initial point cloud, and constructing a spatial feature model using the enhanced initial point cloud; optimizing the spatial feature model to obtain a three-dimensional image model; extracting contour information of the key equipment according to the three-dimensional image model, and performing segmentation and semantic recognition on the contour information to obtain a device instance with a semantic label; 3. The method for monitoring the load status of a distribution feeder based on stereoscopic image analysis according to claim 1, wherein, vector coding the device instance, and summarizing the coding result to generate a spatial equipment feature set of each key equipment. The method comprises the following steps: performing structured coding on the spatial equipment feature set to obtain a device feature structured data set; performing electric parameter acquisition according to the device feature structured data set to obtain an electric parameter time series data set; establishing a device image feature data set based on the three-dimensional image model; mapping the device image feature data set and the electric parameter time series data set to generate an initial multi-modal data set; aligning and standardizing the initial multi-modal data set to obtain a standard multi-modal data set; 4. The method for monitoring the load status of a distribution feeder based on stereoscopic image analysis according to claim 1, wherein, using a multi-modal neural network with a pre-set fusion attention mechanism to jointly model the standard multi-modal data set to obtain a coupling feature vector of each key equipment. The method comprises the following steps: The coupling feature vector is input into a preset graph neural network architecture model to construct an initial device topology relationship graph; Actual physical connection relationships between the key devices are embedded based on the initial device topology relationship graph, and a weight representing coupling strength is defined to obtain a target device topology relationship graph; A load state feature model is generated by iteratively aggregating device local neighborhood features of the target device topology relationship graph through graph convolution operations; The load state feature model is divided into a grid to obtain a plurality of spatial sub-regions of the power distribution area; Coupling feature change rates of the spatial sub-regions are calculated in real time based on each preset time step; When the coupling feature change rate is greater than a preset change threshold, the spatial sub-region is marked as a load fluctuation region.

5. The stereoscopic image analysis based distribution feeder load status monitoring method as claimed in claim 1, wherein, The load trend index sequence is dynamically monitored based on the marking result, and a coupling determination result between the load trend index sequence and an image change trajectory feature sequence of the load fluctuation region is determined, including: The marking result is analyzed to obtain a load state time sequence output sequence of the load fluctuation region; The load state time sequence output sequence is input into a preset spatial-temporal long short-term memory network model to construct a load change time sequence model; A first change trend index of an output result of the load change time sequence model is monitored to obtain a load trend index sequence; An image change trajectory feature sequence of each key device in the load fluctuation region is extracted; The load trend index sequence and the image change trajectory feature sequence are normalized; A dynamic time warping matching algorithm is used to calculate a similarity index between the normalized load trend index sequence and the normalized image change trajectory feature sequence; When the similarity index is greater than or equal to a preset index threshold, it is determined that there is a high coupling between the image change of the key device and the abnormal load fluctuation of the load fluctuation region; When the similarity index is less than the preset index threshold, it is determined that there is no high coupling between the image change of the key device and the abnormal load fluctuation of the load fluctuation region.

6. The stereoscopic image analysis based distribution feeder load status monitoring method as claimed in claim 1, wherein, The load fluctuation region with high coupling is subjected to abnormal identification, and a multi-dimensional feature inversion is performed on a potential abnormal region to generate a load abnormal response atlas, including: A second change trend index of the load fluctuation region with high coupling is calculated; The second change trend index is input into a preset multi-parameter stability evaluation model for stability evaluation to obtain a stability score; When the stability score is greater than or equal to a preset score threshold, and the second change trend index meets a preset stable feature condition, the load fluctuation region is marked as a stable change region; When the stability score is less than the preset score threshold, and the second change trend index meets a preset abnormal feature condition, the load fluctuation region is marked as a potential abnormal region; A joint feature vector is generated by extracting multi-modal features of the potential abnormal region; A regional risk index score of the joint feature vector is calculated; The regional risk index score is subjected to layer mapping, and a visual result is superimposed on the three-dimensional image model to generate a load abnormal response atlas.

7. The stereoscopic image analysis based distribution feeder load status monitoring method as claimed in claim 1, wherein, The potential abnormal area is standardized and evaluated according to the load abnormal response atlas, and a power distribution area load state monitoring analysis report is generated, including: Analyzing the image features, electrical parameter change features, and time sequence evolution data of each potential abnormal area in the load abnormal response atlas; Using the image features, electrical parameter change features, and time sequence evolution data of the potential abnormal area to input a preset expert rule library for risk assessment; Standardizing and formatting the risk assessment results to generate a power distribution area load state monitoring analysis report.

8. A system for monitoring the status of loads in a distribution area based on stereoscopic image analysis, characterized by Including: A construction module for constructing a three-dimensional image model using the multi-view image data of each key device in the power distribution area, and generating a spatial device feature set based on the three-dimensional image model; An encoding module for multi-modal fusion encoding of the spatial device feature set to obtain a coupled feature vector; A labeling module for constructing a load state feature model using the coupled feature vector and introducing a dynamic partition mechanism to real-time label the load fluctuation area; A monitoring module for dynamically monitoring the load trend index sequence based on the labeling results and determining the coupling determination results between the load trend index sequence and the image change trajectory feature sequence of the load fluctuation area; An identification module for abnormal identification of the load fluctuation area with high coupling in the coupling determination results, and multi-dimensional feature inversion of the potential abnormal area to generate a load abnormal response atlas; An evaluation module for standardized evaluation of the potential abnormal area according to the load abnormal response atlas to generate a power distribution area load state monitoring analysis report.

9. An electronic device, comprising: A memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the power distribution area load state monitoring method based on the stereoscopic image analysis according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed to realize the power distribution area load state monitoring method based on the stereoscopic image analysis according to any one of claims 1-7.

Images