Substation operation risk real-time management and control system based on 3D Gaussian point cloud

Abstract

The invention provides a transformer substation operation risk real-time management and control system based on 3D Gaussian point cloud, and the system comprises a space protection region dynamic delineation module which is used for enabling a user to dynamically delineate a space protection region based on a high-precision 3D Gaussian point cloud model of a transformer substation; and the real-time registration and distance measurement module is used for performing millisecond-level space registration on real-time point cloud data of the space protection area acquired by the multi-view panoramic intelligent measurement device and a high-precision 3D Gaussian point cloud model through a point cloud-Gaussian model real-time registration algorithm, and tracking the three-dimensional position of a dangerous source in real time with sub-meter-level precision. Dynamically calculating the safety distance between the live equipment and the live equipment in the space protection area; and the multi-modal risk linkage early warning module is used for generating multi-modal risk data through millisecond-level linkage when the safety distance triggers an early warning threshold value, and pushing the multi-modal risk data to the unified video management and control platform through a wireless network. The flexibility of space protection area definition, the accuracy of danger source tracking and the real-time performance of risk early warning are obviously improved.

Description

Real-time risk management system for substation operations based on 3D Gaussian point clouds Technical Field

[0001] This invention relates to the field of Gaussian point cloud technology, and in particular to a real-time risk management system for substation operations based on 3D Gaussian point clouds. Background Technology

[0002] Currently, with the rapid development of power systems and the continuous improvement of substation automation levels, substation operation safety management faces increasingly complex challenges. Traditional substation risk control mainly relies on manual inspections and two-dimensional video monitoring, which makes it difficult to accurately locate and provide real-time early warnings of hazards in complex three-dimensional spaces. For example, critical areas such as high-voltage equipment and cables within a substation require precise spatial protection zone delineation, but traditional methods lack efficient interactive definition tools, resulting in insufficient flexibility in protection zone settings. Furthermore, existing technologies for real-time tracking of hazards (such as personnel and vehicles) are typically based on two-dimensional image processing, which struggles to achieve sub-meter accuracy and has slow registration speeds, failing to meet millisecond-level real-time requirements. Simultaneously, the rapid movement of dynamic targets and complex environmental interference in substation operation scenarios further increase the difficulty of risk identification. While existing multimodal data fusion technologies can provide video streams and location information, their data linkage efficiency is low, making it difficult to achieve millisecond-level early warning pushes, thus limiting the real-time performance and reliability of risk control.

[0003] To address these issues, there is an urgent need for a solution based on a high-precision 3D model. This solution should utilize a robust registration algorithm and dynamic target recognition technology to enable flexible definition of spatial protection zones, accurate tracking of hazard sources, and rapid linkage of multimodal risk data, thereby improving the safety and efficiency of substation operations. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time risk management system for substation operations based on 3D Gaussian point clouds, in order to solve the problems mentioned in the background art.

[0005] The substation operation risk real-time management and control system based on 3D Gaussian point cloud provided in this invention includes: a dynamic spatial protection area delineation module, used by users to dynamically delineate the spatial protection area based on a high-precision 3D Gaussian point cloud model of the substation; a real-time registration and ranging module, used to perform millisecond-level spatial registration of the real-time point cloud data of the spatial protection area collected by the multi-view panoramic intelligent measurement device with the high-precision 3D Gaussian point cloud model through a point cloud-Gaussian model real-time registration algorithm, and to track the three-dimensional position of the hazard source in real time with sub-meter accuracy, and dynamically calculate its safe distance from the live equipment in the spatial protection area; and a multi-modal risk linkage early warning module, used to generate multi-modal risk data in millisecond-level linkage when the safe distance triggers the early warning threshold, and push it to a unified video management and control platform via a wireless network.

[0006] Optionally, the dynamic delineation module for the spatial protection area allows users to dynamically delineate the spatial protection area based on a high-precision 3D Gaussian point cloud model of the substation. This includes: loading and displaying the high-precision 3D Gaussian point cloud model of the substation to the user; receiving zero-code, arbitrary-shape visual delineation operations performed by the user on the high-precision 3D Gaussian point cloud model; and determining the spatial protection area based on the delineated area on the high-precision 3D Gaussian point cloud model based on the visual delineation operation.

[0007] Optionally, the real-time registration algorithm using point cloud-Gaussian model to perform millisecond-level spatial registration of real-time point cloud data of the spatial protection area collected by the multi-view panoramic intelligent measurement device with a high-precision 3D Gaussian point cloud model includes: utilizing the expressive characteristics of the 3D Gaussian point cloud model, extracting shared features between the real-time point cloud data and the 3D Gaussian point cloud model through a highly robust registration algorithm; calculating spatial transformation parameters from the real-time point cloud coordinate system to the 3D Gaussian point cloud model coordinate system based on the shared features; and mapping and registering the real-time point cloud data to the 3D Gaussian point cloud model with millisecond-level, high-precision mapping according to the spatial transformation parameters, achieving centimeter-level registration accuracy.

[0008] Optionally, the real-time tracking of the three-dimensional position of the hazard source with sub-meter accuracy includes: performing dynamic target segmentation and identification on the real-time point cloud data in the registered 3D Gaussian point cloud model coordinate system to extract point cloud clusters representing the hazard source; based on the point cloud clusters of the hazard source, calculating and outputting the sub-meter accuracy three-dimensional coordinates of the hazard source in the real three-dimensional space of the substation in real time through centroid calculation, bounding box fitting or target tracking algorithms.

[0009] Optionally, the multimodal risk data includes: high-definition video streams or image snapshots of the scene at the moment the early warning is triggered, collected by a multi-view panoramic intelligent measurement device, as well as the three-dimensional location of the hazard source and its safe distance from the electrical equipment within the space protection area, which are tracked in real time.

[0010] Optionally, the real-time risk management and control system for substation operations based on 3D Gaussian point clouds further includes: a multi-expert collaborative intelligent scheduling and optimization module, used to construct a spatiotemporal coupled prediction model based on historical hazard trajectory data in the high-precision 3D Gaussian point cloud model of the substation and sub-meter-level three-dimensional position information output by the real-time registration and ranging module; the spatiotemporal coupled prediction model adopts a hybrid architecture of graph convolutional neural network and long short-term memory network to dynamically associate and model the motion characteristics of hazard sources with the topology of substation equipment, thereby predicting the future movement status of each target hazard source in the next control cycle; based on the predicted future movement status, combined with the safe distance threshold of energized equipment and the dynamic intrusion rate of hazard sources, a control time limit prediction model is used to predict the latest control time limit of each target hazard source; all target hazard sources are arranged in ascending order from nearest to farthest according to their latest control time limit. A dynamic hazard source sequence is generated. Simultaneously, for each target hazard source in the sequence, its necessary value for expert autonomous control is analyzed. Each target hazard source is traversed sequentially according to the order of the dynamic hazard source sequence. When traversing to any target hazard source, the control rhythm is determined based on the collaborative processing history of multiple experts before the traversal time. The control rhythm is defined as the expected time window for experts to reach the hazard source location and the operational intensity parameter. Furthermore, based on this control rhythm, it is predicted whether each expert can complete control before the latest control deadline for the target hazard source. If the prediction result indicates that control cannot be completed, or the necessary value for expert autonomous control of the target hazard source is lower than a preset dynamic threshold, a multi-level auxiliary mechanism is immediately triggered. The collaborative processing history includes the expert's current task load, movement trajectory, and historical response time. The dynamic threshold adaptively fluctuates based on the substation's real-time load rate and historical false alarm rate.

[0011] Optionally, the motion characteristics of the hazard source include velocity vector, rate of change of acceleration, and environmental constraint factor; the environmental constraint factor quantifies the feasibility space of the hazard source movement by analyzing the electromagnetic field distribution heat map and physical protection boundary of the energized equipment in the Gaussian point cloud model, ensuring that the prediction results conform to the physical laws of substation operation.

[0012] Optionally, the latest control time limit is defined as the critical time point of the safety distance threshold for the intrusion of hazardous sources, and a risk attenuation coefficient is introduced for dynamic correction. The risk attenuation coefficient is adaptively adjusted according to the type of hazardous source and the voltage level of the equipment to eliminate prediction deviations caused by environmental noise.

[0013] Optionally, the necessary value for expert autonomous control is calculated using a multi-dimensional synergistic gain evaluation model. This model comprehensively considers the spatial topological relationship between the expert's location and the hazard source, the matching degree between the expert's skill tags and the hazard source type, and the chain-like gain effect of expert autonomous control on subsequent hazard source control tasks. The chain-like gain effect is quantified as a synergistic value coefficient, calculated as follows: Synergistic Value Coefficient = w1 × Spatial Overlap + w2 × Skill Transfer Rate + w3 × Temporal Dependence, where w1, w2, and w3 are weight coefficients, representing the contribution of the spatial, skill, and temporal dimensions, respectively. Spatial overlap measures the degree of optimization of the control path for adjacent hazard sources after the expert handles the current hazard source, skill transfer rate assesses the efficiency of expert skill reuse, and temporal dependency analyzes the compression effect of the current control on the time window of subsequent tasks, thereby ensuring that the necessary value accurately reflects the global value of expert intervention.

[0014] Optionally, the multi-level assistance mechanism executes the following chain-like optimization process: Based on the control rhythm and the latest control deadline, a strategy framework is planned to complete the assistance within a conservative time; the strategy framework divides the assistance process into N consecutive time windows, each window is assigned a differentiated assistance plan and cognitive objective; the conservative time is a buffer period of a preset duration before the latest control deadline, dynamically expanded according to the risk level of the hazard source; within each time window, the assigned assistance plan is executed and expert behavior data is monitored in real time; if the expert fails to achieve the corresponding cognitive objective, a chain-like feedback optimization engine is activated: this engine, based on the reasons for failure in the current window, combines the assistance plan and cognitive objective of the remaining time window to construct a multi-objective optimization problem; an improved particle swarm optimization algorithm is used for fast solution, dynamically redistributing the assistance resource intensity and cognitive objective threshold of subsequent windows to ensure that the optimization process meets real-time constraints; based on the chain-like optimization results, subsequent assistance processes are updated and continuously iterated until all target hazard sources are effectively controlled within the latest control deadline, or the sequence traversal is completed.

[0015] This invention achieves the following beneficial effects: Through the above steps, the system realizes real-time risk management and control of substation operations based on 3D Gaussian point clouds, significantly improving the flexibility of spatial protection zone definition, the accuracy of hazard source tracking, and the real-time nature of risk warning. Steps A1-A3 generate precise spatial protection zones by using a high-precision 3D Gaussian point cloud model and zero-code delineation operations, covering key equipment such as transformers and reducing the risk of misoperation. Steps B1-B3 utilize highly robust registration algorithms and spatial transformation parameters to achieve centimeter-level registration between real-time point clouds and high-precision models, accurately tracking the location of hazard sources and ensuring the reliability of safe distance calculations. Steps C1-C2 use technologies such as dynamic target segmentation and recognition, and centroid calculation to calculate the sub-meter-level three-dimensional coordinates of hazard sources in real time, and support the generation and push of multimodal risk data, significantly reducing the risk of collisions and electric shocks during operations and ensuring the safe operation of substations.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 is a schematic diagram of a real-time risk management system for substation operations based on 3D Gaussian point clouds in an embodiment of the invention; Figure 2 is a flowchart of the steps performed by each module of the system in an embodiment of the invention; Figure 3 is another flowchart of the steps performed by each module of the system in an embodiment of the invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] The research and development approach of this application is based on 3D Gaussian point cloud technology. Addressing the real-time, accurate, and flexible requirements of risk management in substation operations, it designs a system solution integrating dynamic delineation of spatial protection areas, real-time registration and ranging, and multimodal risk linkage early warning. By loading a high-precision 3D Gaussian point cloud model and supporting zero-code visual delineation operations, the system achieves user-friendly spatial protection area definition, meeting the flexibility requirements of complex substation scenarios. Utilizing a highly robust registration algorithm and spatial transformation parameter calculation, it achieves centimeter-level registration between the real-time point cloud and the high-precision model, ensuring sub-meter-level accuracy in hazard source tracking. Through dynamic target segmentation and recognition technology, combined with multimodal data fusion, the system can generate and push risk data containing high-definition video streams, 3D coordinates, and safety distances within milliseconds, thereby significantly improving the real-time performance and reliability of risk warnings and ensuring substation operational safety.

[0021] Figure 1 is a schematic diagram of a real-time substation operation risk management system based on 3D Gaussian point cloud provided in an embodiment of this application. As shown in Figures 1 and 2, the system includes: a spatial protection area dynamic delineation module 100, used to execute step A, allowing the user to dynamically delineate the spatial protection area based on a high-precision 3D Gaussian point cloud model of the substation. Step A specifically includes the following sub-steps: A1, loading and displaying the high-precision 3D Gaussian point cloud model of the substation to the user.

[0022] A high-precision 3D Gaussian point cloud model refers to a digital model generated by scanning a substation scene using LiDAR or multi-view vision equipment. It represents the position, color, and scale of each point in a Gaussian distribution, including point coordinates (x, y, z), a Gaussian covariance matrix (describing the uncertainty of the points), and color attributes (RGB values). The loading process involves reading the model data from the storage medium and decoding it into a renderable format. Rendering is accelerated using a graphics processing unit (GPU), presenting the point cloud model to the user in an interactive 3D view. The user interface must support real-time zooming, rotation, and translation operations, ensuring users can view details of substation equipment, cables, and the environment from any perspective. Key technical parameters for model loading include the number of point clouds (typically in the millions, reflecting model resolution), the Gaussian covariance matrix (defining the spatial distribution characteristics of the point cloud, calculated based on the laser scanning accuracy during point cloud acquisition), and the rendering frame rate (must reach at least 30 frames per second to ensure smooth interaction). The point cloud points are obtained by counting the total amount of point cloud data recorded by the scanning device. The covariance matrix is ​​calculated using the precision parameters of the point cloud acquisition device (such as the angular resolution of 0.01 degrees for LiDAR). The rendering frame rate is determined by the GPU performance and the display resolution (usually 1080p).

[0023] A2. Receive zero-code, arbitrary-shape visualization delineation operations performed by users on high-precision 3D Gaussian point cloud models.

[0024] Zero-code visual delimitation refers to users directly delimiting spatial protection areas on a 3D Gaussian point cloud model through a graphical user interface (GUI) by dragging, clicking, or drawing, without writing any code. The operation involves capturing the user's input of 2D screen coordinates (based on mouse or touch events), converting them into a 3D coordinate sequence in the 3D Gaussian point cloud model using an inverse projection algorithm, forming closed or open geometric shapes (such as polygons or free curves). Key technical parameters include screen coordinate accuracy (determined by monitor resolution, such as 1920x1080 pixels), inverse projection error (determined by the camera projection matrix and the accuracy of the point cloud model, typically ±1 mm), and the geometric complexity of the delimited area (expressed in terms of the number of vertices, such as 50 vertices). Screen coordinates are obtained through GUI event listeners, inverse projection error is calculated using camera intrinsic parameters (such as focal length and lens distortion coefficients) and the covariance matrix of the point cloud model, and geometric complexity is determined by the number of vertices drawn by the user. The system needs to provide real-time feedback on the boundary visualization effect of the delimited area (such as highlighting) and support user adjustment or undo operations.

[0025] A3. Based on the visualization delineation operation, the delineation area on the high-precision 3D Gaussian point cloud model is used to determine the spatial protection area.

[0026] The spatial protection zone refers to a geometric area defined in the three-dimensional space of a substation for risk management, based on a user-defined subset of 3D point clouds. It is represented as a 3D polygon or volumetric region in the point cloud model. The determination process involves converting the user-defined 2D screen coordinate sequence into a closed geometry in the 3D point cloud model, generating spatial protection zone data containing boundary point sets and volumetric extent. Key technical parameters include the number of points in the point cloud subset of the defined area (determined by the defined extent and point cloud density, e.g., 100,000 points), the topological connectivity of the boundary point set (calculated using a neighborhood search algorithm to ensure boundary closure), and the volume of the volumetric extent (calculated using a voxelization algorithm, in cubic meters). The number of points in the point cloud subset is obtained by statistically analyzing the number of points within the defined area. Topological connectivity is analyzed using the k-nearest neighbor algorithm (k=8) to determine the neighborhood relationships of the point cloud. The volume of the volumetric extent is calculated by summing the points after pixelating the point cloud subset into a 0.01 cubic meter resolution grid. The system must ensure the accuracy and real-time performance of the area definition to support subsequent registration and risk assessment.

[0027] The real-time registration and ranging module 200 is used to execute step B, which involves using a point cloud-Gaussian model real-time registration algorithm to perform millisecond-level spatial registration between the real-time point cloud data of the space protection area collected by the multi-view panoramic intelligent measurement device and a high-precision 3D Gaussian point cloud model. It also tracks the three-dimensional position of the hazard source in real time with sub-meter accuracy and dynamically calculates the safe distance between the hazard source and the live equipment within the space protection area. Step B specifically includes the following sub-steps: B1. Utilizing the expressive characteristics of the 3D Gaussian point cloud model, a highly robust registration algorithm is used to extract the shared features between the real-time point cloud data and the 3D Gaussian point cloud model.

[0028] The expressive characteristics of a 3D Gaussian point cloud model refer to the Gaussian distribution features of each point cloud point, including position coordinates (x, y, z), covariance matrix (describing spatial uncertainty), and color attributes (RGB values). These characteristics can serve as geometric and textural features for registration. Shared features refer to the overlap between real-time point cloud data (acquired by a multi-view panoramic intelligent measurement device) and the 3D Gaussian point cloud model in spatial geometry (e.g., curvature, edges) and texture (e.g., color gradient). Highly robust registration algorithms (such as the Iterative Closest Point Algorithm combined with normal vector constraints, ICP-Normal) extract a set of shared feature points by comparing the geometric and textural features of the two sets of point clouds. Key technical parameters include the number of feature points (typically in the thousands, reflecting registration accuracy), feature extraction time (must be less than 50 milliseconds to support real-time performance), and feature matching error (calculated using the Euclidean distance between points, with a target of ±1 cm). The number of feature points is determined by thresholding the curvature and color gradient of the point cloud. The extraction time is determined by the number of algorithm iterations (usually 10) and hardware performance (such as GPU parallel computing power). The matching error is calculated by the convergence residual of the ICP algorithm.

[0029] B2. Based on shared features, calculate the spatial transformation parameters from the real-time point cloud coordinate system to the 3D Gaussian point cloud model coordinate system.

[0030] Spatial transformation parameters refer to the rigid transformation matrices required to align the real-time point cloud coordinate system to the 3D Gaussian point cloud model coordinate system. These include rotation matrices (3x3 matrices describing pose changes) and translation vectors (3D vectors describing position offsets). The calculation process is based on a shared feature point set, using the least squares method to optimize the rotation and translation parameters, minimizing the Euclidean distance between feature point pairs in the two sets of point clouds. Key technical parameters include the accuracy of the transformation matrix (measured by residual error, with a target of ±1 mm), computation time (needs to be less than 20 milliseconds to support real-time performance), and the number of feature point pairs (provided by the B1 step, affecting optimization stability). The accuracy of the transformation matrix is ​​calculated using the convergence residuals from least squares optimization. The computation time is determined by the number of feature point pairs and hardware performance (such as CPU floating-point arithmetic capabilities), with the number of feature point pairs directly inherited from the B1 step. The system must ensure the robustness of the transformation parameters to cope with noise, occlusion, and other interference.

[0031] B3. Based on the spatial transformation parameters, the real-time point cloud data is mapped and registered to the 3D Gaussian point cloud model with millisecond-level high precision, achieving centimeter-level registration accuracy.

[0032] Mapping and registration refers to using the spatial transformation parameters calculated in the B2 step to map the coordinates of each point in the real-time point cloud data to the coordinate system of the 3D Gaussian point cloud model through rotation and translation transformations, achieving spatial alignment of the two sets of point clouds. Registration accuracy is measured by the Euclidean distance error between the transformed points, with a target of centimeter-level (±1 cm). Key technical parameters include registration time (must be less than 10 milliseconds to support real-time performance), registration error (calculated by the average distance of the overlapping areas of the point clouds), and the number of point clouds (real-time point clouds are typically in the millions, affecting computational complexity). Registration time is determined by the multiplication operations of the transformation matrix and the parallel capabilities of the GPU. The registration error is verified by the Random Sample Consensus (RANSAC) algorithm to check the distance of the overlapping areas of the point clouds. The number of point clouds is counted by the acquisition frequency of the multi-view measurement device (e.g., 10 Hz). The system must ensure the real-time performance and high accuracy of the mapping process to support subsequent hazard source tracking.

[0033] The multimodal risk linkage early warning module 300 is used to execute step C: when the safety distance triggers the early warning threshold, it generates multimodal risk data in milliseconds and pushes it to the unified video management and control platform via wireless network. The multimodal risk data includes: high-definition video streams or image snapshots of the scene at the moment the early warning is triggered, collected by the multi-view panoramic intelligent measurement device, as well as the three-dimensional position of the real-time tracked hazard source and its safe distance from the live equipment in the space protection area. Step C specifically includes the following sub-steps: C1: In the registered 3D Gaussian point cloud model coordinate system, dynamic target segmentation and identification are performed on the real-time point cloud data to extract point cloud clusters representing the hazard source.

[0034] Dynamic target segmentation and recognition refers to the process of segmenting real-time point cloud data into independent target point cloud clusters using clustering algorithms (such as DBSCAN) within a registered 3D Gaussian point cloud model coordinate system, and then using machine learning models (such as PointNet) to identify point cloud clusters representing hazard sources (such as people and vehicles). A point cloud cluster is a subset of point clouds that shares spatial connectivity and similar characteristics (such as color and density). Key technical parameters include the number of point cloud clusters (determined by the clustering algorithm, reflecting scene complexity), segmentation time (needs to be less than 50 milliseconds to support real-time performance), and recognition accuracy (measured by the F1 score of the classification model, with a target of 95%). The number of point cloud clusters is determined by the neighborhood radius (typically 0.1 meters) and minimum number of points (e.g., 50) parameters of the DBSCAN algorithm. Segmentation time is determined by the number of point cloud points and GPU performance. Recognition accuracy is obtained by testing a pre-trained model on a labeled dataset. The system must handle the rapid movement of dynamic targets and complex background interference. C2. Based on the point cloud clusters of hazard sources, the substation's actual three-dimensional coordinates of the hazard sources in the real three-dimensional space are calculated and output in real time using centroid calculation, bounding box fitting, or target tracking algorithms.

[0035] The 3D coordinate calculation of a hazard source refers to determining its location in the actual 3D space of the substation based on the point cloud clusters extracted in step C1, through centroid calculation (taking the average coordinates of the point cloud clusters), bounding box fitting (generating the minimum bounding rectangle), or target tracking algorithms (such as Kalman filtering). Sub-meter accuracy means the coordinate error is less than 1 meter, typically ±10 centimeters. Key technical parameters include coordinate accuracy (calculated from the number of points and spatial distribution of the point cloud clusters), calculation time (must be less than 20 milliseconds to support real-time performance), and tracking continuity (measured by the state update frequency of the Kalman filter, such as 10Hz). Coordinate accuracy is determined by the number of points in the point cloud cluster (typically in the thousands) and the Gaussian covariance matrix; calculation time is determined by algorithm complexity and hardware performance (such as GPU parallel computing); and tracking continuity is measured by the prediction-update cycle statistics of the Kalman filter. The system must ensure the real-time performance and stability of the coordinate output to support risk warning.

[0036] Through the above steps, the system achieves real-time risk management and control of substation operations based on 3D Gaussian point clouds, significantly improving the flexibility of spatial protection zone definition, the accuracy of hazard source tracking, and the real-time nature of risk warnings. Steps A1-A3 generate precise spatial protection zones by using a high-precision 3D Gaussian point cloud model and zero-code delineation operations, covering critical equipment such as transformers and reducing the risk of misoperation. Steps B1-B3 utilize highly robust registration algorithms and spatial transformation parameters to achieve centimeter-level registration between the real-time point cloud and the high-precision model, accurately tracking the location of hazard sources and ensuring the reliability of safe distance calculations. Steps C1-C2 use technologies such as dynamic target segmentation and recognition, and centroid calculation to calculate the sub-meter-level three-dimensional coordinates of hazard sources in real time, and support the generation and push of multimodal risk data, significantly reducing the risk of collisions and electric shocks during operations and ensuring the safe operation of the substation.

[0037] As shown in Figure 3, in some embodiments, the real-time risk management and control system for substation operations based on 3D Gaussian point clouds further includes: a multi-expert collaborative intelligent scheduling and optimization module, used to execute steps including: D. Constructing a spatiotemporal coupled prediction model based on historical hazard trajectory data in the high-precision 3D Gaussian point cloud model of the substation and sub-meter-level three-dimensional position information output by the real-time registration and ranging module; the spatiotemporal coupled prediction model adopts a hybrid architecture of graph convolutional neural network and long short-term memory network to dynamically associate the motion characteristics of the hazard source with the topology of the substation equipment, thereby predicting the future movement status of each target hazard source in the next control cycle; the motion characteristics of the hazard source include velocity vector, rate of change of acceleration and environmental constraint factor; the environmental constraint factor quantifies the feasibility space of hazard source movement by analyzing the electromagnetic field distribution heat map and physical protection boundary of the energized equipment in the Gaussian point cloud model, ensuring that the prediction results conform to the physical laws of substation operations.

[0038] The spatiotemporal coupled prediction model refers to a model that uses a hybrid architecture of graph convolutional neural network (GCN) and long short-term memory network (LSTM) to dynamically correlate the motion characteristics of hazard sources with the topology of substation equipment, thereby predicting the future movement of each target hazard source in the next control cycle. The motion characteristics of the hazard source include velocity vector (defined as the displacement rate of the hazard source in three-dimensional space, in meters per second), rate of change of acceleration (defined as the second derivative of the velocity vector, in meters per second²), and environmental constraint factor (defined as a weighting coefficient that quantifies the feasibility space of hazard source movement, ranging from 0 to 1). The environmental constraint factor is calculated by analyzing the electromagnetic field distribution heatmap (represented by electromagnetic field intensity gradient, in volts / meter) of the energized equipment in a high-precision 3D Gaussian point cloud model and the physical protection boundary (generated from the set of physical obstacle boundary points marked in the point cloud model). The specific steps are as follows: First, a subset of the point cloud containing the energized equipment (e.g., transformers) is extracted from the point cloud model, and the spatial distribution of the electromagnetic field intensity on the equipment surface is calculated using the Gaussian covariance matrix to generate a heatmap. Then, based on topological connectivity analysis of the point cloud (e.g., k-nearest neighbor algorithm, k=8), the three-dimensional coordinates of the physical protection boundary (e.g., fence, wall) are identified. Finally, the environmental constraint factor is calculated by weighted superposition of electromagnetic field intensity and boundary distance (e.g., a constraint factor of 0.2 at 1 meter from the energized equipment and 0.8 at 2 meters). The velocity vector and rate of change of acceleration are estimated using a Kalman filter algorithm with a 1-second filtering window, based on the sub-meter level three-dimensional position sequence (10 samples per second) output by the real-time registration and ranging module, to smooth out noise effects. Graph Convolutional Neural Networks (GNNs) model the substation equipment topology as a graph structure (nodes represent equipment, edges represent spatial adjacency relationships between equipment). LSTM captures the temporal dependencies of hazard source trajectories. The hybrid architecture fuses graph features and temporal features through fully connected layers, outputting a hazard source location prediction (3D coordinates, accuracy ±10 cm) for the next control cycle (typically 5 seconds). Model training data comes from historical hazard source trajectories (long-term recordings by the point cloud registration module, containing at least 1000 trajectories) and real-time location input. Training uses a mean squared error loss function, with Adam as the optimizer, a learning rate of 0.001, and 1000 training iterations. The prediction results are verified against the physical constraints of the point cloud model to ensure compliance with the physical laws of substation operations (e.g., hazard sources cannot cross physical obstacles).

[0039] In one implementation example, in a practical application at a 110kV substation, the system loads a high-precision 3D Gaussian point cloud model (5 million points, covariance matrix accuracy ±1 mm), extracts 1000 personnel and vehicle trajectories (each containing at least 100 location points, sampling frequency 10 Hz) from a historical hazard source trajectory database, and combines this with the 3D location of the hazard source (accuracy ±10 cm) output by the real-time registration and ranging module. The point cloud model is analyzed using the k-nearest neighbor algorithm (k=8) to identify the spatial distribution of transformers (electromagnetic field strength 1000 volts / meter, constraint factor 0.2 at 1 meter distance) and fences (boundary point set containing 100,000 points), generating an environmental constraint factor heatmap (resolution 0.1 meters). The Kalman filter algorithm calculates velocity vectors (e.g., personnel movement speed 1.2 m / s) and acceleration change rate (0.5 m / s²) based on the real-time location sequence (10 points per second). The spatiotemporal coupled prediction model employs a hybrid architecture of GCN (50 nodes, 200 edges, reflecting device topology) and LSTM (64-dimensional hidden layer, 10-step time). The fused fully connected layer outputs the predicted location of the hazard source within the next 5 seconds (e.g., a person moving from coordinates (10,5,2) to (12,6,2), with an error of ±8 cm). The model was trained 1000 times on a GPU (NVIDIA RTX 3090) in 2 hours, with a prediction time of 20 milliseconds. Verification of the prediction results with the fence boundary of the point cloud model revealed that a person's predicted path was 1.5 meters from a transformer. The environmental constraint factor was adjusted to 0.3, and the path was corrected to avoid high-risk areas, ensuring the prediction conforms to physical laws.

[0040] E. Based on the predicted future movement patterns, combined with the safe distance threshold of energized equipment and the dynamic intrusion rate of hazardous sources, a control time limit prediction model is used to predict the latest control time limit for each target hazardous source. The latest control time limit is defined as the critical time point of the hazardous source intrusion safe distance threshold, and a risk attenuation coefficient is introduced for dynamic correction. The risk attenuation coefficient is adaptively adjusted according to the type of hazardous source and the voltage level of the equipment to eliminate prediction deviations caused by environmental noise.

[0041] The control time limit prediction model refers to predicting the critical time point of the intrusion safety distance threshold for each target hazard source by analyzing the future movement trend of the hazard source (output from step D, including predicted position and velocity vectors), the safe distance threshold of the live equipment (defined as the minimum allowable distance between the hazard source and the equipment, e.g., 2 meters for a transformer), and the dynamic intrusion rate (defined as the rate at which the hazard source approaches the live equipment, in meters per second). The dynamic intrusion rate is calculated using the time difference of the predicted position sequence. Specifically, it extracts the distance change between adjacent time points from the predicted position sequence for the next 5 seconds (sampling frequency 10Hz), divides it by the time interval (0.1 seconds), and obtains the intrusion rate (e.g., 0.5 meters per second). The safe distance threshold is determined by the substation equipment specifications (e.g., 2 meters for a 110kV transformer, 3 meters for a 220kV transformer) and stored in the equipment attribute table of the point cloud model. The risk attenuation coefficient (range 0 to 1) is used to correct prediction bias. It is adaptively adjusted based on hazard type (0.8 for personnel, 0.6 for vehicles) and equipment voltage level (0.9 for 110kV, 0.7 for 220kV). Specifically, it is obtained through a lookup table: the voltage level is extracted from the equipment attribute table, the type is extracted from the hazard identification results (step C1), and mapped to a preset coefficient table. The control time limit prediction model uses a linear regression model. The inputs are the intrusion rate, current distance (provided by the real-time registration module, accuracy ±10 cm), and risk attenuation coefficient. The output is the latest control time limit, calculated as: Time Limit = Current Distance / (Intrusion Rate × Risk Attenuation Coefficient). The model training data comes from historical hazard trajectory records (1000) and actual control time limit records. Training uses mean squared error loss and iterates 500 times. The prediction process must be completed within 10 milliseconds, with the error controlled within ±0.5 seconds. Environmental noise (such as wind speed and light) is considered and dynamically corrected using the risk attenuation coefficient to ensure the robustness of the prediction results.

[0042] Continuing with the above implementation example, in the 110kV substation scenario, the system obtains the future movement status of a personnel hazard source from step D (predicting movement from 3 meters to 1.8 meters from the transformer within 5 seconds, with an intrusion rate of 0.24 m / s), combined with the equipment attribute table of the point cloud model (110kV transformer safety distance threshold of 2 meters). The real-time registration module provides the current distance as 3 meters (accuracy ±10 cm), the hazard source type as personnel (risk attenuation coefficient 0.8), and the voltage level as 110kV (coefficient 0.9). The comprehensive risk attenuation coefficient is calculated as 0.8 × 0.9 = 0.72 using the table lookup method. The control time limit prediction model (linear regression, training data of 1000 trajectories, 500 iterations) inputs the current distance of 3 meters, the intrusion rate of 0.24 m / s, and the risk attenuation coefficient of 0.72, and calculates the latest control time limit as (3-2) / (0.24 × 0.72) ≈ 5.8 seconds, with an error of ±0.4 seconds. The prediction process took 8 milliseconds on the CPU (Intel Xeon 2.4GHz). Considering environmental noise (such as wind speed of 1 m / s potentially causing positional deviation), the system corrected for this with a risk attenuation coefficient and recalculated the time limit to 5.6 seconds to ensure the prediction results were not overly optimistic. The results showed that personnel hazards needed to be controlled within 5.6 seconds to prevent intrusion into the 2-meter safety distance. The system then passed this data to step F for sequence sorting.

[0043] F. Arrange all target hazards in ascending order from nearest to farthest according to the latest control deadline to generate a dynamic hazard sequence. Simultaneously, analyze the necessary value for expert autonomous control for each target hazard in the sequence. This necessary value is calculated using a multi-dimensional collaborative gain evaluation model. This model comprehensively considers the spatial topological relationship between the expert's location and the hazard, the matching degree between the expert's skill tags and the hazard type, and the chain-like gain effect of expert autonomous control on subsequent hazard control tasks. The chain-like gain effect is quantified as a collaborative value coefficient, calculated as: Collaborative Value Coefficient = w1 × Spatial Overlap + w2 × Skill Transfer Rate + w3 × Temporal Dependence, where w1, w2, and w3 are weighting coefficients, representing the contribution of the spatial, skill, and temporal dimensions, respectively. Spatial overlap measures the degree of optimization of the control path for adjacent hazards after the expert handles the current hazard; skill transfer rate assesses the efficiency of expert skill reuse; and temporal dependency analyzes the compression effect of the current control on the time window of subsequent tasks, thus ensuring that the necessary value accurately reflects the global value of expert intervention.

[0044] The dynamic hazard source sequence refers to an ordered list formed by arranging all target hazard sources in ascending order from nearest to farthest according to their latest control time limit (output from step E, in seconds). This list guides subsequent traversal and control priority allocation. The necessary value for expert autonomous control (range 0 to 1) is calculated using a multi-dimensional synergistic gain evaluation model, comprehensively considering the spatial topological relationship between the expert's location and the hazard source (defined as the Euclidean distance from the expert to the hazard source, in meters), the matching degree between the expert's skill tags and the hazard source type (defined as the skill matching score, range 0 to 1), and the chain gain effect (quantified as a synergistic value coefficient, range 0 to 1). The spatial topological relationship is obtained by calculating the Euclidean distance between the expert's location (accuracy ±10 cm) and the hazard source location using the real-time registration module; for example, a distance less than 5 meters indicates high priority. The skill matching degree is obtained by looking up a table based on a preset matching matrix between the expert's skill tags (e.g., electrical maintenance, mechanical operation) and the hazard source type (personnel, vehicle). (e.g., electrical maintenance has a matching degree of 0.9 for personnel and 0.6 for vehicles.) The chain gain effect is calculated using the synergy value coefficient, defined as: Synergy Value Coefficient = w1 × Spatial Overlap + w2 × Skill Transfer Rate + w3 × Temporal Dependence, where w1, w2, and w3 are weighting coefficients (e.g., 0.4, 0.3, 0.3, determined by historical data regression analysis). Spatial overlap (range 0 to 1) is calculated by analyzing the path optimization of adjacent hazards after expert processing of the current hazard, based on the topological adjacency relationship (k-nearest neighbor, k=8) of the point cloud model. Skill transfer rate (range 0 to 1) is calculated by comparing the reuse efficiency of expert skills across different hazard sources; for example, the reuse rate of electrical maintenance skills between personnel and vehicles is 0.8. Temporal dependency (range 0 to 1) is calculated by analyzing the compression effect of current control on the time window of subsequent tasks, based on historical control timeliness data (1000 records). The multi-dimensional synergy gain evaluation model uses a weighted linear model, with training data being historical control records, trained 1000 times, outputting necessary values ​​with an error of ±0.05. The calculation process must be completed within 20 milliseconds to ensure real-time performance.

[0045] Continuing the above implementation example, in a 110kV substation, the system detects three hazard sources: Personnel A (latest control time limit 5.6 seconds), Vehicle B (8.2 seconds), and Personnel C (10.5 seconds). A dynamic hazard source sequence is generated by ascending order: [Personnel A, Vehicle B, Personnel C]. For Personnel A, the system calculates the necessary value for expert autonomous control: the real-time registration module provides the location of Expert 1 (4 meters from Personnel A, calculated using Euclidean distance, accuracy ±10 cm), with the skill tag "Electrical Maintenance" (matching degree 0.9, obtained by checking the matching matrix). Spatial overlap is analyzed using a point cloud model; Personnel A and Vehicle B are 3 meters apart (calculated using the k-nearest neighbor algorithm). After Expert 1 processes Personnel A, the path is optimized to Vehicle B, resulting in a spatial overlap of 0.7. Skill transfer rate is based on the reuse efficiency of the electrical maintenance skill between personnel and vehicles (0.8, obtained by checking the reuse matrix). Temporal dependency is analyzed using historical control records (1000 records); processing Personnel A can reduce the control time of Vehicle B by 0.5 seconds, resulting in a dependency of 0.6. The collaborative value coefficient is calculated as 0.4×0.7+0.3×0.8+0.3×0.6=0.7, with a necessary value of 0.7 (weighted linear model, error ±0.04, time 15 milliseconds). Similarly, the necessary values ​​for vehicle B and person C are 0.6 and 0.5, respectively. The sequence and necessary values ​​are stored as inputs for subsequent steps G, ensuring that high-risk, high-needs-value hazards are prioritized.

[0046] G. Traverse each target hazard source sequentially according to the dynamic hazard source sequence; when traversing to any target hazard source, determine the control rhythm based on the collaborative processing status history of multiple experts before the traversal time; the control rhythm is defined as the expected time window and operation intensity parameter of the expert reaching the hazard source location; then, predict whether each expert can complete the control before the latest control deadline of the target hazard source based on the control rhythm; if the prediction result is that the control cannot be completed, or the necessary value of expert autonomous control of the target hazard source is lower than the preset dynamic threshold, then immediately trigger the multi-level auxiliary mechanism; wherein, the collaborative processing status history includes the expert's current task load, movement trajectory and historical response time; the dynamic threshold is adaptively fluctuated according to the substation's real-time load rate and historical false alarm rate. The multi-level auxiliary mechanism executes the following chain optimization process: G1. Based on the control rhythm and the latest control deadline, plan a strategy framework to complete the assistance within a conservative time; the strategy framework divides the assistance process into N consecutive time windows, each window is assigned a differentiated assistance scheme and cognitive target; the conservative time is a buffer period of a preset length before the latest control deadline, which is dynamically expanded according to the hazard source risk level.

[0047] The strategy framework divides the assistance process into N consecutive time windows (N is determined by the latest control deadline and real-time requirements, e.g., 5). Each window is assigned a differentiated assistance plan (defined as a set of resource allocation and guidance instructions, such as pushing navigation paths and adjusting task priorities) and a cognitive goal (defined as the behavioral state that the expert needs to achieve within the window, such as reaching a designated location or completing a specific operation). The conservative time refers to a pre-set buffer period (e.g., 1 second) before the latest control deadline, dynamically expanded according to the hazard risk level (high, medium, low, determined based on the intrusion rate and distance in step E). It is calculated as: Conservative Time = Latest Control Deadline - Buffer Duration, where the buffer duration is obtained by looking up the risk level (1 second for high risk, 0.5 seconds for medium risk, and 0.2 seconds for low risk). The control rhythm is defined as the expected time window (in seconds, e.g., arriving within 2 seconds) for the expert to reach the hazard location and the operational intensity parameter (defined as the complexity of the operation per unit time, e.g., processing 1 instruction per second). The time window is divided based on the latest control deadline (e.g., 5.6 seconds divided into 5 windows of 1.12 seconds each). The auxiliary plan for each window is generated by analyzing the collaborative processing history (including expert task load, movement trajectory, and historical response time). Task load is calculated using the current number of assigned tasks (e.g., 3 tasks), movement trajectory is provided by the real-time registration module (accuracy ±10 cm), and response time is calculated as the average of 1000 historical control records (e.g., 2 seconds). Cognitive objectives are generated by matching expert skill tags with hazard types to produce specific behavioral requirements (e.g., an electrical maintenance expert needs to check the safety equipment of personnel A). The planning process uses a rule-based scheduling algorithm. The inputs are the control rhythm, the latest control deadline, and the collaborative status history; the outputs are the auxiliary plans for N windows and the cognitive objectives. The calculation takes 20 milliseconds with an error of ±0.1 seconds.

[0048] Continuing the above implementation example, in a 110kV substation, the dynamic hazard source sequence is traversed. Currently, personnel A is being handled (latest control time limit 5.6 seconds, high risk level). The control rhythm is based on the collaborative processing status history (expert 1 has 2 tasks, movement trajectory shows a distance of 4 meters from personnel A, historical response time 2 seconds), determining the expected time window as 2 seconds, and the operation intensity as 1 instruction per second. The high risk level corresponds to a buffer time of 1 second, with a conservative time of 5.6-1=4.6 seconds, divided into 5 time windows (each 0.92 seconds). The system analyzes expert 1's electrical maintenance skills (match degree 0.9) and personnel A's type, generating auxiliary plans: window 1 pushes the navigation path (moving from 4 meters to 2 meters), window 2 guides equipment inspection, and windows 3-5 monitor the approach distance. The cognitive objective is: window 1 reaches the 2-meter position, and window 2 completes the equipment inspection. The planning process takes 18 milliseconds on the CPU. The rule-based scheduling algorithm references historical records (1000 entries) to ensure that the plan matches the expert's capabilities. The result generates a strategy framework containing 5 windows, which is then passed to G2 for execution.

[0049] G2. Within each time window, execute the assigned auxiliary scheme and monitor expert behavior data in real time. If the expert fails to achieve the corresponding cognitive goal, the chain feedback optimization engine is activated. This engine constructs a multi-objective optimization problem based on the reasons for failure in the current window, combined with the auxiliary schemes and cognitive goals of the remaining time windows. An improved particle swarm optimization algorithm is used for fast solution, and the auxiliary resource intensity and cognitive goal threshold of subsequent windows are dynamically reallocated to ensure that the optimization process meets real-time constraints.

[0050] Within each time window, the system executes the assigned assistance plan (e.g., pushing navigation paths, sending voice commands) and monitors expert behavior data (including position and action status, accuracy ±10 cm, sampling frequency 10 Hz) in real time through a multi-view panoramic intelligent measurement device. The behavior data is compared with the cognitive objective (e.g., reaching a designated location, completing an operation) to determine if it has been achieved. The achievement criterion is that the deviation between the behavior data and the objective is less than a preset threshold (e.g., position deviation ±0.5 meters). If not achieved, a chain feedback optimization engine is activated. This engine, based on the reasons for not achieving the objective in the current window (e.g., path deviation, operation delay), combines the assistance plan for the remaining time window with the cognitive objective to construct a multi-objective optimization problem (optimization objectives include minimizing time deviation and maximizing resource utilization). The optimization problem is solved using an improved particle swarm optimization (PSO) algorithm with 50 particles, 20 iterations, and adaptively adjusted inertia weight (initially 0.9, decreasing to 0.4), with a solution time of 10 milliseconds. The reasons for non-compliance are obtained through behavioral data analysis; for example, positional deviation is calculated using real-time point cloud registration, and operational latency is estimated using an action recognition model (PointNet, F1 score 95%). The optimization results dynamically reallocate the intensity of auxiliary resources for subsequent windows (e.g., increasing navigation frequency from 1 to 2 times / second) and the cognitive target threshold (e.g., relaxing positional deviation from 0.5 meters to 0.7 meters) to ensure real-time constraints (total optimization time less than 20 milliseconds). The optimization process considers expert task load (provided by step G) and the substation's real-time load factor (obtained through equipment sensors, e.g., 80%) to balance resource allocation.

[0051] Continuing with the above implementation example, in the management of personnel A, the system executes an auxiliary plan in the first 0.92-second window: pushing a navigation path (moving from 4 meters to 2 meters), monitoring the position of expert 1 through a multi-view measurement device (real-time point cloud registration, accuracy ±10 cm), and finding that expert 1 only moved to 3 meters, failing to achieve the cognitive objective (position deviation 1 meter, threshold 0.5 meters). The chain feedback optimization engine is activated, analyzing the reason for the failure to meet the target as path deviation (calculated through point cloud registration), and combining the plans (guiding equipment inspection, monitoring distance) and objectives (reaching 1 meter, completing the inspection) of the remaining 4 windows (total duration 3.68 seconds). The multi-objective optimization problem is defined as minimizing the time deviation (target 0 seconds) and maximizing the utilization of navigation resources. The PSO algorithm (50 particles, 20 iterations) takes 8 milliseconds on the CPU, outputting the optimization result: increasing the navigation frequency to 2 times / second in window 2, and relaxing the position deviation threshold to 0.6 meters. After optimization, the system updates the plans for subsequent windows, continues to push navigation instructions and monitor expert 1's behavior, ensuring that management is completed within 4.6 seconds. The substation load rate of 80% (sensor acquisition) restricts the allocation of navigation resources. The optimization process balances resource usage and takes 15 milliseconds.

[0052] G3. Based on the chain optimization results, update the subsequent auxiliary processes and continue to iterate until all target hazards are effectively controlled within the latest control time limit, or the sequence traversal is completed.

[0053] Based on the chained optimization results of step G2, the system updates the auxiliary processes for subsequent time windows, including adjusting auxiliary schemes (e.g., navigation paths, command frequencies) and cognitive objectives (e.g., location thresholds, operation completion rates), and continues iteratively until all hazards are controlled or the sequence is traversed within the latest control time limit. The update process reallocates resources by mapping the optimization results (resource intensity, target thresholds) to the remaining windows; for example, adjusting the navigation frequency from 2 times / second to 1.5 times / second to reduce load. The iterative process monitors expert behavior data (position, actions, accuracy ±10 cm, sampling 10 Hz) and compares it with the updated cognitive objectives to determine whether control is complete (the standard is that the distance between the hazard and energized equipment is greater than a safety threshold, e.g., 2 meters). If a hazard is controlled (e.g., personnel A moves to a safe area), the system traverses the next hazard in the sequence (e.g., vehicle B) and repeats steps G1-G2. If not completed, the system continues to iterate and optimize, adjusting the dynamic threshold (e.g., from 0.6 to 0.65) based on the substation's real-time load rate (obtained from sensors, e.g., 80%) and historical false alarm rate (statistically calculated from 1000 records, e.g., 5%) to ensure the robustness of the auxiliary processes. Each iteration takes 15 milliseconds, and the total process is controlled within the latest control time limit (e.g., 5.6 seconds). After all hazard sources are controlled, the system generates a control report, recording the time limit, auxiliary solutions, and effects for each hazard source, for subsequent optimization.

[0054] Continuing the above implementation example, in the management of personnel A, the navigation frequency of window 2, adjusted by optimization result G2, is 1.5 times / second, and the position threshold is 0.6 meters. The system updates the auxiliary processes for the subsequent three windows (total duration 2.76 seconds): window 2 pushes the optimized path, and windows 3-4 monitor the completion of equipment checks. Real-time monitoring shows that expert 1 reaches 1.8 meters in window 2 (deviation 0.2 meters, meeting the standard), window 3 completes the equipment check, and personnel A is guided to a safe area of ​​2.5 meters, completing the management (time 3.5 seconds, less than 5.6 seconds). The system then iterates through the next hazard source, vehicle B (latest time limit 8.2 seconds), repeating G1-G2, planning four windows (each 18 seconds), and pushing navigation and operation instructions. Vehicle B moves to a safe distance of 3 meters within 7 seconds, completing the management. All hazard sources (personnel A, vehicle B, personnel C) are managed within the time limit, and the system generates a report recording information such as personnel A's navigation frequency of 1.5 times / second and management duration of 3.5 seconds. A substation load rate of 80% and a false alarm rate of 5% raise the dynamic threshold to 0.65, ensuring process robustness. The total iteration time is 4.8 seconds (15 milliseconds per window), meeting real-time requirements.

[0055] Through the above steps, the substation operation risk real-time management and control system based on 3D Gaussian point cloud of this invention achieves efficient scheduling and optimization through multi-expert collaboration. Step D uses a spatiotemporal coupling prediction model, combined with historical trajectories and real-time locations, to accurately predict the movement of hazardous sources, providing a reliable basis for subsequent management and control. Step E utilizes a management and control time limit prediction model, comprehensively considering intrusion rate and risk attenuation coefficient, to accurately calculate the latest management and control time limit, ensuring priority handling of high-risk hazardous sources. Step F optimizes resource allocation through dynamic sequence sorting and expert autonomous management and control necessity value evaluation, improving scheduling efficiency by 20% through collaborative value coefficient. Step G and sub-steps G1-G3, through management and control rhythm planning, chain optimization, and iterative assistance, ensure that all hazardous sources are effectively managed within the time limit, improving overall management and control timeliness and reducing false alarm rate. Overall, the system significantly reduces the risk of collisions and electric shocks in substation operations, ensuring operational safety and stable equipment operation, and is suitable for complex multi-hazard scenarios.

[0056] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time risk management system for substation operations based on 3D Gaussian point clouds, characterized in that, include: The dynamic delineation module for spatial protection areas is used by users to dynamically delineate spatial protection areas based on a high-precision 3D Gaussian point cloud model of the substation. The real-time registration and ranging module is used to perform millisecond-level spatial registration between the real-time point cloud data of the spatial protection area collected by the multi-view panoramic intelligent measurement device and the high-precision 3D Gaussian point cloud model through a point cloud-Gaussian model real-time registration algorithm. It also tracks the three-dimensional position of the hazard source in real time with sub-meter accuracy and dynamically calculates the safe distance between the hazard source and the live equipment in the spatial protection area. The multimodal risk linkage early warning module is used to generate multimodal risk data in milliseconds when the safe distance triggers the early warning threshold, and push it to the unified video management and control platform via wireless network.

2. The real-time substation operation risk management system based on 3D Gaussian point cloud as described in claim 1, characterized in that, The dynamic spatial protection area delineation module allows users to dynamically delineate spatial protection areas based on a high-precision 3D Gaussian point cloud model of the substation. This includes: loading and displaying the high-precision 3D Gaussian point cloud model of the substation to the user; receiving zero-code, arbitrary-shape visual delineation operations performed by the user on the high-precision 3D Gaussian point cloud model; and determining the spatial protection area based on the delineated area on the high-precision 3D Gaussian point cloud model.

3. The real-time substation operation risk management system based on 3D Gaussian point cloud as described in claim 1, characterized in that, The real-time registration algorithm using point cloud-Gaussian model performs millisecond-level spatial registration between real-time point cloud data of the spatial protection area collected by the multi-view panoramic intelligent measurement device and a high-precision 3D Gaussian point cloud model. This includes: utilizing the expressive characteristics of the 3D Gaussian point cloud model, extracting shared features between the real-time point cloud data and the 3D Gaussian point cloud model using a highly robust registration algorithm; calculating spatial transformation parameters from the real-time point cloud coordinate system to the 3D Gaussian point cloud model coordinate system based on the shared features; and mapping and registering the real-time point cloud data to the 3D Gaussian point cloud model with millisecond-level, high-precision mapping according to the spatial transformation parameters, achieving centimeter-level registration accuracy.

4. The real-time substation operation risk control system based on 3D Gaussian point cloud as described in claim 3, characterized in that, The method of tracking the three-dimensional position of the hazard source in real time with sub-meter accuracy includes: performing dynamic target segmentation and identification on the real-time point cloud data in the registered 3D Gaussian point cloud model coordinate system to extract point cloud clusters that represent the hazard source; and based on the point cloud clusters of the hazard source, calculating and outputting the three-dimensional coordinates of the hazard source in the real three-dimensional space of the substation with sub-meter accuracy in real time through centroid calculation, bounding box fitting or target tracking algorithms.

5. The real-time substation operation risk control system based on 3D Gaussian point cloud as described in claim 3, characterized in that, The multimodal risk data includes: high-definition video streams or image snapshots of the scene at the moment the early warning is triggered, collected by a multi-view panoramic intelligent measurement device, as well as the three-dimensional location of the hazard source and its safe distance from the electrical equipment in the space protection area, which are tracked in real time.

6. The real-time substation operation risk control system based on 3D Gaussian point cloud as described in claim 1, characterized in that, Also includes: A multi-expert collaborative intelligent scheduling and optimization module is used to: construct a spatiotemporal coupled prediction model based on historical hazard trajectory data from a high-precision 3D Gaussian point cloud model of the substation and sub-meter-level 3D position information output by a real-time registration and ranging module; the spatiotemporal coupled prediction model adopts a hybrid architecture of graph convolutional neural network and long short-term memory network to dynamically associate the motion characteristics of the hazard source with the substation equipment topology, thereby predicting the future movement status of each target hazard source in the next control cycle; based on the predicted future movement status, combined with the safe distance threshold of energized equipment and the dynamic intrusion rate of the hazard source, a control time limit prediction model is used to predict the latest control time limit of each target hazard source; all target hazard sources are arranged in ascending order from nearest to farthest according to their latest control time limit to generate a dynamic hazard source sequence; simultaneously, for... For each target hazard in the sequence, the necessary value for expert autonomous control is analyzed; each target hazard is traversed sequentially according to the dynamic hazard sequence; when traversing to any target hazard, the control rhythm is determined based on the collaborative processing history of multiple experts before the traversal time; the control rhythm is defined as the expected time window for experts to reach the hazard location and the operation intensity parameter; then, based on the control rhythm, it is predicted whether each expert can complete the control before the latest control deadline of the target hazard; if the prediction result is that the control cannot be completed, or the necessary value for expert autonomous control of the target hazard is lower than the preset dynamic threshold, a multi-level auxiliary mechanism is immediately triggered; wherein, the collaborative processing history includes the expert's current task load, movement trajectory and historical response time; the dynamic threshold is adaptively fluctuated according to the substation's real-time load rate and historical false alarm rate.

7. The real-time substation operation risk control system based on 3D Gaussian point cloud as described in claim 6, characterized in that, The motion characteristics of a hazard source include velocity vector, rate of change of acceleration, and environmental constraint factors. The environmental constraint factors quantify the feasibility space for the movement of the hazard source by analyzing the electromagnetic field distribution heat map and physical protection boundary of the energized equipment in the Gaussian point cloud model, ensuring that the prediction results conform to the physical laws of substation operation.

8. The real-time substation operation risk control system based on 3D Gaussian point cloud as described in claim 6, characterized in that, The latest control time limit is defined as the critical time point of the safety distance threshold for the intrusion of hazardous sources, and a risk attenuation coefficient is introduced for dynamic correction. The risk attenuation coefficient is adaptively adjusted according to the type of hazardous source and the voltage level of the equipment to eliminate prediction deviations caused by environmental noise.

9. The real-time substation operation risk control system based on 3D Gaussian point cloud as described in claim 6, characterized in that, The necessary value for expert autonomous control is calculated using a multi-dimensional synergistic gain evaluation model. This model comprehensively considers the spatial topological relationship between the expert's location and the hazard source, the matching degree between the expert's skill tags and the hazard source type, and the chain-like gain effect of expert autonomous control on subsequent hazard source control tasks. The chain-like gain effect is quantified as a synergistic value coefficient, calculated as follows: Synergistic Value Coefficient = w1 × Spatial Overlap + w2 × Skill Transfer Rate + w3 × Temporal Dependence, where w1, w2, and w3 are weight coefficients representing the contributions of the spatial, skill, and temporal dimensions, respectively. Spatial overlap measures the degree of optimization of the control path for adjacent hazard sources after the expert handles the current hazard source, skill transfer rate assesses the efficiency of expert skill reuse, and temporal dependency analyzes the compression effect of the current control on the time window of subsequent tasks, thereby ensuring that the necessary value accurately reflects the global value of expert intervention.

10. The real-time substation operation risk management system based on 3D Gaussian point cloud as described in claim 6, characterized in that, The multi-level assistance mechanism executes the following chain-like optimization process: Based on the control rhythm and the latest control deadline, a strategy framework is planned to complete the assistance within a conservative timeframe; the strategy framework divides the assistance process into N consecutive time windows, each window is assigned a differentiated assistance plan and cognitive objective; the conservative time is a buffer period preset before the latest control deadline, dynamically expanded according to the risk level of the hazard source; within each time window, the assigned assistance plan is executed and expert behavior data is monitored in real time; if the expert fails to achieve the corresponding cognitive objective, a chain-like feedback optimization engine is activated: this engine, based on the reasons for failure in the current window, combines the assistance plan and cognitive objective of the remaining time windows to construct a multi-objective optimization problem; an improved particle swarm optimization algorithm is used for rapid solution, dynamically redistributing the assistance resource intensity and cognitive objective threshold of subsequent windows to ensure that the optimization process meets real-time constraints; based on the chain-like optimization results, subsequent assistance processes are updated and continuously iterated until all target hazard sources are effectively controlled within the latest control deadline, or the sequence traversal is completed.

Images