Power semantic constraint and space-time graph neural network-based power distribution room early warning method

Abstract

This invention belongs to the field of intelligent operation and maintenance technology for power systems. It discloses a method for early warning of substations based on power semantic constraints and spatiotemporal graph neural networks. The method involves acquiring video streams and electrical status signals from the substation, constructing a virtual synchronization layer to map and generate synchronous multimodal events, extracting entity objects as graph nodes, injecting electrical potential attenuation attributes into equipment nodes to calculate non-static safety boundaries, and constructing a dynamic semantic graph. The dynamic semantic graph sequence is input into the spatiotemporal graph neural network, embedding a physical constraint layer to perform physical consistency correction on personnel trajectory and arc propagation predictions. Based on the microscopic time difference between operations and responses, cross-modal micro-causal edges are constructed as attention constraints for time encoding. Finally, the method outputs the risk evolution probability and generates hierarchical early warning commands. This invention achieves deep fusion of multimodal data, significantly improving the accuracy and rationality of risk prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and more specifically, to a method for early warning of substations based on power semantic constraints and spatiotemporal graph neural networks. Background Technology

[0002] Patent application CN109470912A discloses a method and system for early warning of faults in the power distribution room of a power-protected user. It includes a current measuring device, a voltage measuring device, a switch status detection device, a wireless communication device, a display device, an alarm device, and a microprocessor. The current measuring device, voltage measuring device, and switch status detection device are all assigned an identification ID. The current measuring device and voltage measuring device are installed at measurement points in the power distribution room of the power-protected user, located at the starting point of the power distribution branch and the connection point of the power distribution equipment. The switch status detection device is installed on the switch of the power distribution line of the power-protected user. The wireless communication device is installed inside the power distribution room of the power-protected user. The beneficial effects are: it can promptly detect and warn of faults in the power distribution lines and equipment inside the power-protected user's premises, guide fault handling and repair, and accelerate the speed of restoring power to the user.

[0003] However, in typical power distribution room operation scenarios, when maintenance personnel enter the room for routine inspections or operations, existing video surveillance systems and electrical monitoring systems often operate independently, making it impossible to establish effective spatiotemporal correlation analysis. While video systems can record personnel movement trajectories, the inconsistency in acquisition frequency and time reference with the electrical system means that when personnel operate equipment such as circuit breakers, the system cannot accurately correlate the operational actions with subsequent changes in electrical state, resulting in missed opportunities to establish a causal relationship between personnel behavior and electrical risks. Existing technologies generally adopt static safety distance standards, keeping the safety boundary fixed regardless of changes in equipment load, voltage level, or operating conditions. This one-size-fits-all approach may underestimate risks during high-load operation or excessively restrict normal personnel operations during light-load operation. In predictive analysis, traditional methods lack necessary physical constraints, often producing predictions that violate basic physical principles, such as predicting personnel moving at speeds exceeding human limits or electric arcs spreading in ways that do not conform to heat diffusion laws. These unrealistic predictions severely impact the scientific rigor of risk assessments. When common situations arise in the power distribution room, such as equipment obstruction, changes in lighting, or personnel briefly leaving the area of ​​sight, existing video tracking systems often cannot maintain continuous monitoring, leading to interruptions in risk assessment at critical moments. Furthermore, existing early warning systems typically rely on a single risk threshold for simple judgments, failing to assess the feasibility and effectiveness of intervention measures such as power outages under the current risk conditions. This often results in either overly sensitive warnings with frequent false alarms, or slow responses that miss the optimal intervention window.

[0004] In view of this, the present invention proposes a power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural networks to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural networks, comprising: Asynchronous multimodal data from the power distribution room is acquired. The asynchronous multimodal data includes video stream frame sequences and electrical topology state change signals. A virtual synchronization layer is constructed, and the electrical topology state change signals are mapped onto the time axis of the video stream frame sequences to generate synchronous multimodal events. Based on synchronous multimodal events, entity objects in the power distribution room scenario are extracted as graph nodes, and electrical potential energy decay attributes are injected into equipment nodes. Non-static safety boundary distances are calculated based on electrical potential energy decay attributes. Dynamic semantic graphs are constructed by combining the spatial positional relationships between graph nodes and the electrical topology connectivity relationships. The dynamic semantic graphs at multiple consecutive time points are input into the spatiotemporal graph neural network. A physical law constraint layer is embedded in the spatial feature encoding process of the spatiotemporal graph neural network. The physical law constraint layer is used to perform physical consistency verification and correction on the prediction of personnel movement trajectory and the prediction of arc discharge propagation. In the dynamic semantic graph, based on the micro time difference between the time of operation and the time of electrical state response determined by the virtual synchronization layer, cross-modal micro-causal edges are constructed, and these cross-modal micro-causal edges are used as attention weight constraints for the temporal feature encoding of the spatiotemporal graph neural network. The spatiotemporal graph neural network outputs the risk evolution probability based on the corrected spatial features and constrained temporal features, and generates early warning instructions based on the risk evolution probability and the preset hierarchical strategy.

[0006] The technical effects and advantages of the power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural networks in this invention are as follows: This invention improves the overall performance and reliability of power distribution room safety monitoring, achieving deep fusion of multimodal sensing information. It innovatively realizes dynamic adaptive adjustment of safety boundaries, intelligently adjusting the protection range based on the real-time operating status of equipment, significantly improving the targeting and rationality of safety protection. Regarding prediction accuracy, this invention ensures that all risk prediction results conform to real physical laws and establishes a precise causal relationship identification mechanism, accurately capturing the micro-temporal correlation between personnel operation behavior and electrical equipment response, greatly improving the scientific rigor and credibility of risk identification. This invention implements an intelligent hierarchical early warning strategy, enabling scientific judgment based on risk evolution trends and intervention feasibility, effectively avoiding false alarms and missed alarms. Through intelligent utilization of historical safety events and successful handling experiences, this invention achieves continuous optimization of the early warning strategy, significantly enhancing the robustness of the monitoring system. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a logical architecture diagram of the intelligent monitoring and security analysis of the present invention; Figure 3 This is a schematic diagram of the spatiotemporal graph neural network architecture of the present invention; Figure 4 This is a flowchart illustrating the technical process architecture for inferring the intent of this invention. Figure 5 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] Example 1: like Figure 1 As shown, embodiments of the present invention propose a power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural networks. The method includes the following steps: Step 1: Acquire asynchronous multimodal data in the power distribution room. The asynchronous multimodal data includes video stream frame sequences and electrical topology state change signals. Construct a virtual synchronization layer and map the electrical topology state change signals onto the time axis of the video stream frame sequences to generate synchronous multimodal events. Step 2: Based on synchronous multimodal events, extract entity objects in the power distribution room scenario as graph nodes, inject electrical potential energy decay attributes into equipment nodes, calculate non-static safety boundary distances based on electrical potential energy decay attributes, and construct a dynamic semantic graph by combining the spatial positional relationships between graph nodes and the electrical topology connectivity relationships. The "dynamic semantic graph" constructed in this step and the reasoning logic behind it can be found in the appendix. Figure 2 (Intelligent Monitoring and Security Analysis Logical Architecture Diagram) provides an intuitive understanding. As shown in the attached diagram, this architecture fully demonstrates the entire process from raw data to advanced semantic reasoning: The top-level "Video Stream Acquisition and Preprocessing" module acquires video frames; the middle-level "Semantic Entity Extraction and Spatiotemporal Association" module extracts semantic nodes such as "personnel," "equipment," and "area" through target detection and segmentation (such as YOLO / Mask R-CNN algorithms), and calculates dynamic attributes such as "speed = 0.6m / s²" and "distance = 0.65m," ultimately constructing a dynamic graph connecting "spatial edges" and "security rule edges"; the bottom-level "Advanced Semantic Reasoning and Security Situation Understanding" module performs risk analysis based on graph neural networks. The specific parameters marked in the diagram are precisely the instantiation of the calculation of "non-static security boundary distance" and subsequent physical law constraints.

[0010] Step 3: Input the dynamic semantic graphs from multiple consecutive time points into the spatiotemporal graph neural network. Embed a physical law constraint layer during the spatial feature encoding process of the spatiotemporal graph neural network. Use the physical law constraint layer to perform physical consistency verification and correction on the prediction of personnel movement trajectory and the prediction of arc discharge propagation. The core architecture of the "Spatiotemporal Graph Neural Network (ST-GNN)" can be found in the appendix. Figure 3 (Schematic diagram of spatiotemporal graph neural network architecture). As shown in the figure, this network takes a "continuous multi-time dynamic semantic graph sequence" as input, and its core consists of a "spatial feature encoding module" and a "temporal feature encoding module". The spatial encoding module includes a "Graph Attention Network (GAT) layer" and a key "spatial encoding rule block". This rule block is combined with the "physical law constraint layer" below (integrating the limits of human biomechanics and the laws of electromagnetic thermal diffusion) to ensure the physical credibility of the feature encoding. The temporal encoding module processes temporal dependencies through a "Temporal Convolutional Network (TCN)" and introduces a "dynamic dilation factor" to achieve adaptive temporal perception. The "cross-modal intent correction interface" and "micro-causal attention weight constraint" clearly marked at the bottom of the figure reflect the design of using external information (such as intent vectors) and causal relationships to correct predictions.

[0011] Step 4: In the dynamic semantic graph, based on the micro time difference between the time of operation and the time of electrical state response determined by the virtual synchronization layer, a cross-modal micro-causal edge is constructed, and the cross-modal micro-causal edge is used as the attention weight constraint for the temporal feature encoding of the spatiotemporal graph neural network. Step 5: The spatiotemporal graph neural network outputs the risk evolution probability based on the corrected spatial features and constrained temporal features, and generates early warning instructions based on the risk evolution probability and the preset hierarchical strategy.

[0012] In this embodiment of the invention, the invention employs a technique of acquiring asynchronous multimodal data within a power distribution room and constructing a virtual synchronization layer to achieve time-axis mapping between electrical signals and video streams. It extracts entity objects as graph nodes based on synchronous multimodal events and injects electrical potential attenuation attributes into device-type nodes to calculate non-static safety boundary distances. A physical constraint layer is embedded in the spatiotemporal graph neural network to verify the consistency between motion trajectory and arc propagation predictions. Cross-modal micro-causal edges are constructed using micro-time differences as attention weight constraints. Finally, the invention outputs risk evolution probabilities and generates tiered early warning commands. This effectively overcomes the technical problems in existing power distribution room safety monitoring schemes where video monitoring and electrical monitoring are independent and cannot establish effective correlation analysis, and where early warning schemes based on static safety distances cannot adapt to dynamic electrical state changes and lack physical constraints leading to unreliable prediction results. This achieves deep fusion and spatiotemporal correlation analysis of multimodal data, improves prediction accuracy through dynamic safety boundaries and physical constraints, and establishes a precise spatiotemporal attention mechanism based on micro-causal relationships, ultimately realizing intelligent perception and tiered early warning of power distribution room safety risks.

[0013] In a preferred embodiment of the present invention, the above-described construction of a virtual synchronization layer, which maps electrical topology state change signals onto the time axis of a video stream frame sequence to generate synchronous multimodal events, includes: Step 1.1 involves dividing the video stream frame sequence into equally spaced time slots based on the acquisition timestamps. Specifically, this includes deploying a high-definition video acquisition device in the power distribution room security monitoring system. This device continuously acquires video stream frame sequences of the internal scene of the power distribution room at a fixed frame rate. Each video frame carries a precise acquisition timestamp, with millisecond-level accuracy to meet the needs of refined time series analysis. After receiving the video stream frame sequence, the acquisition timestamp of each video frame is extracted as a time reference to establish a continuous time axis. Based on the extracted timestamp information, equally spaced time slots are divided according to a preset time interval. The spacing of the time slots is determined based on the operation response characteristics of the power distribution room, for example, set to 100 milliseconds, to ensure that the micro-temporal correlation between personnel operation actions and electrical equipment responses can be captured. Each time slot has a clear start timestamp, end timestamp, and center time point, providing a precise time reference framework for the time mapping of subsequent electrical topology state change signals.

[0014] Step 1.2: Extract the signal occurrence timestamp carried by the electrical topology state change signal, and calculate the deviation value between the signal occurrence timestamp and the center point of the adjacent time slot. Specifically, the electrical monitoring system in the power distribution room collects state change signals of various electrical equipment in real time, including circuit breaker opening and closing status changes, current and voltage fluctuations, and protection device operation signals. Each electrical topology state change signal carries an accurate signal occurrence timestamp, recording the specific time of signal generation. After receiving these electrical topology state change signals, the signal occurrence timestamp carried by each signal is extracted one by one, and the position of the signal on the time axis is determined according to the timestamp. The extracted signal occurrence timestamp is matched with the established equally spaced time slots to find the adjacent time slot closest to the signal occurrence timestamp. The time deviation value between the signal occurrence timestamp and the center point of the adjacent time slot is calculated. The deviation value reflects the degree of synchronization between the electrical signal and the video frame time reference.

[0015] Step 1.3: If the deviation value is less than the preset jitter threshold, the electrical topology state change signal is embedded into the corresponding adjacent time slot. If the deviation value is greater than or equal to the preset jitter threshold, linear interpolation is performed within the adjacent time slot based on the previous and next states of the electrical topology state change signal to generate a virtual electrical state signal to complete the time slot, thereby outputting a synchronized multimodal event. Specifically, this includes setting a preset jitter threshold, for example, 50 milliseconds, based on the response characteristics of the electrical equipment in the power distribution room, to determine the acceptable level of synchronization between the electrical signal and the video time slot. When the calculated deviation value is less than 50 milliseconds, it is determined that the electrical topology state change signal has a good time correspondence with the adjacent time slot, and the electrical signal is directly embedded into the corresponding adjacent time slot to establish a time mapping relationship between the electrical signal and the video frame. When the deviation value is greater than or equal to 50 milliseconds, a significant time asynchrony is determined, and a virtual signal generation mechanism is initiated. First, the preceding and following state values ​​of the electrical topology state change signal are acquired, such as the state transition of a circuit breaker from closed to open. Then, within the time range of adjacent time slots, linear interpolation calculations are performed based on the preceding and following states to generate the virtual electrical state signal corresponding to that time slot. Finally, the generated virtual electrical state signal is filled into the time slot to achieve signal completion. Through the above processing, each time slot contains corresponding video frame information and electrical state information, forming a time-synchronized multimodal event sequence, and outputting a synchronized multimodal event.

[0016] In this embodiment of the invention, by employing a technique of dividing time slots into equal intervals based on the acquisition timestamps of the video stream frame sequence, extracting the timestamps of electrical topology state change signals and calculating the deviation values ​​from the center points of adjacent time slots, and embedding electrical signals into time slots or generating virtual signals to complete time slots through linear interpolation based on the comparison results of the deviation values ​​and preset jitter thresholds, the technical problems of inconsistent acquisition frequencies of video data and electrical data, asynchronous timestamps leading to the inability to establish accurate time correlations, and the lack of effective time alignment mechanisms causing difficulties in multimodal data fusion in existing multimodal monitoring systems are overcome. This achieves the technical effect of realizing accurate time synchronization of asynchronous multimodal data, establishing a unified time reference for video streams and electrical signals through a virtual synchronization layer, providing a reliable time basis for subsequent cross-modal correlation analysis, and ensuring the temporal continuity and integrity of multimodal events.

[0017] In a preferred embodiment of the present invention, the above-mentioned injection of electrical potential energy attenuation attributes into device nodes, and the calculation of non-static safety boundary distance based on the electrical potential energy attenuation attributes, includes: Step 2.1: Read the real-time voltage level, current load current, and circuit breaker opening / closing status corresponding to the equipment nodes to construct an electrical potential energy field feature vector. Specifically, during the construction of the dynamic semantic graph of the power distribution room, first identify the equipment nodes in the synchronous multimodal events, including key electrical equipment such as transformers, switchgear, circuit breakers, and cable joints. For each equipment node, the electrical data acquisition interface reads its corresponding electrical operating parameters in real time: voltage level parameters are collected through voltage transformers to obtain the actual operating voltage value of the equipment, such as 10kV, 35kV, etc.; current load current parameters are collected through current transformers to obtain the real-time load current carried by the equipment, reflecting the load status of the equipment; circuit breaker opening / closing status parameters are collected through status sensors to obtain the current open or closed status information of the circuit breaker, determining the continuity of the electrical circuit. Based on the collected multidimensional electrical operating parameters, an electrical potential energy field feature vector is constructed. This vector contains normalized voltage level values, load current values, and circuit breaker status codes, forming a multidimensional vector representation describing the distribution characteristics of the equipment's electrical potential energy field.

[0018] Step 2.2 involves inputting the electrical potential energy field feature vector into a pre-trained potential energy decay estimation model and outputting the potential energy decay coefficient at the current moment. Specifically, this includes pre-training the potential energy decay estimation model based on a large amount of historical operating data from power distribution rooms. This model employs a deep neural network architecture to learn the nonlinear mapping relationship between the electrical potential energy field feature vector and the potential energy decay coefficient. The training data includes the potential energy decay patterns of electrical equipment under different voltage levels, load states, and switching state combinations, as well as corresponding safety accident case analysis data, ensuring that the model can accurately estimate the potential energy decay characteristics under different electrical states. During real-time early warning, the constructed electrical potential energy field feature vector is input into the pre-trained potential energy decay estimation model. The model performs forward calculations based on the input feature vector and outputs the potential energy decay coefficient at the current moment. This potential energy decay coefficient reflects the changing trend of the equipment's danger level under the current electrical state, with a value range from 0 to 1. A value close to 1 indicates slow potential energy decay and high danger, while a value close to 0 indicates fast potential energy decay and relative safety.

[0019] Step 2.3: Obtain the basic static safety distance corresponding to the equipment type node. Use the potential energy attenuation coefficient as a dynamic adjustment factor. Calculate the scaling of the basic static safety distance based on this dynamic adjustment factor to obtain the non-static safety boundary distance. The non-static safety boundary distance evolves non-linearly with the change in the potential energy attenuation coefficient. Specifically, this includes: pre-configuring the basic static safety distance corresponding to each type of equipment node according to electrical safety specifications and equipment technical parameters. For example, the basic static safety distance for a 10kV switchgear is 1.5 meters, and the basic static safety distance for a 35kV transformer is 3.0 meters. During the dynamic safety boundary calculation, the basic static safety distance corresponding to the current equipment type node is obtained as the calculation basis. Using the potential energy attenuation coefficient obtained in the previous step as a dynamic adjustment factor, a non-linear scaling function is designed to adjust the basic static safety distance: when the potential energy attenuation coefficient is large (close to 1), it indicates a high level of equipment risk, and the scaling function increases the safety boundary distance; when the potential energy attenuation coefficient is small (close to 0), it indicates relatively safe equipment, and the scaling function decreases the safety boundary distance. The scaling calculation uses an exponential function to ensure that the change in the safety boundary distance reflects the non-linear characteristics of electrical risk. The resulting non-static safety boundary distance is dynamically adjusted in real time as the electrical state of the equipment changes, providing an accurate dynamic reference for subsequent personnel safety distance judgment.

[0020] In this embodiment of the invention, the technical means of constructing an electrical potential energy field feature vector by reading the real-time voltage level, current load current, and circuit breaker opening and closing status of equipment nodes, inputting this feature vector into a pre-trained potential energy decay estimation model to output a potential energy decay coefficient, and then using the potential energy decay coefficient as a dynamic adjustment factor to perform nonlinear scaling calculation on the basic static safety distance to obtain a non-static safety boundary distance, overcomes the technical problems of existing power distribution room safety monitoring systems that use fixed static safety distances, cannot dynamically adjust safety boundaries according to the real-time electrical status of equipment, and lack electrical potential energy decay law modeling leading to inaccurate safety boundary settings. This achieves the technical effect of dynamically adaptively adjusting the safety boundary based on the actual electrical operating status of equipment, accurately reflecting the time-varying characteristics of electrical risks through potential energy decay modeling, providing precise dynamic spatial boundary constraints for personnel safety protection, and effectively improving the accuracy and real-time performance of power distribution room safety early warning.

[0021] In a preferred embodiment of the present invention, the above-mentioned physical consistency verification and correction of the prediction of personnel movement trajectory and the prediction of arc discharge propagation using the physical law constraint layer includes: Step 3.1 involves performing a graph convolution operation on spatial feature encoding in the spatiotemporal graph neural network. This is followed by obtaining the displacement vectors of personnel nodes at adjacent time points for predicting personnel movement trajectories, and obtaining the arc influence range expansion vector of equipment nodes for predicting arc discharge propagation. Specifically, this includes using a graph convolutional neural network architecture to spatially aggregate and extract features from the node features in the dynamic semantic graph. After completing the graph convolution operation, motion-related information is extracted from the output node features: for personnel nodes, the coordinate difference between adjacent time points is calculated to obtain the displacement vector of the personnel node at adjacent time points. This vector contains displacement components in the x and y axes, reflecting the predicted movement trajectory of personnel within the power distribution room. For equipment nodes, the expansion trend of the arc influence range is calculated based on the electrical potential energy decay attribute, resulting in the arc influence range expansion vector of the equipment node. This vector describes the propagation direction and speed prediction of arc discharge in space. This vector information provides crucial motion prediction data for subsequent physical consistency verification.

[0022] Step 3.2 involves calculating the relative approximation velocity between the displacement vector and the arc influence range expansion vector, and extracting the motion acceleration features of the personnel node. Specifically, this includes receiving the personnel node displacement vector and the equipment node arc influence range expansion vector, and then calculating their relative motion relationship. First, the angle between the displacement vector and the arc expansion vector is calculated to determine the spatial relationship between the personnel's movement direction and the arc propagation direction. Then, based on the vector magnitude, the relative approximation velocity is calculated, representing the speed at which the personnel approach the arc danger zone. This velocity reflects the urgency of the potential safety risk. Simultaneously, by performing differential calculations on the personnel displacement vectors at multiple consecutive moments, the motion acceleration features of the personnel node are extracted, including the magnitude and direction of the acceleration. These motion acceleration features reflect the changing trends of the personnel's motion state, providing an important basis for judging the physical rationality of the motion prediction.

[0023] Step 3.3: Input the relative approximation velocity and motion acceleration characteristics into the physical constraint judgment logic. If the motion acceleration characteristics exceed the human biomechanical limit threshold, or the relative approximation velocity violates the physical laws of arc heat diffusion, a physical inconsistency is determined. Specifically, this includes: pre-setting human biomechanical limit thresholds and arc heat diffusion physical law parameters. The human biomechanical limit thresholds are determined based on human kinematics research, including limiting parameters such as the maximum acceleration and maximum movement speed of a person in a normal working environment; the arc heat diffusion physical law parameters are determined based on the physical mechanism of arc discharge, including constraints such as the maximum speed of arc diffusion and the range of heat propagation. The input motion acceleration characteristics are compared with the human biomechanical limit thresholds. If the magnitude of the acceleration exceeds the range of normal human movement ability, or the change in the direction of acceleration is too drastic and violates the continuity of human movement, a physical inconsistency is determined in the prediction of human motion. Simultaneously, the relative approximation velocity is compared with the physical laws of arc heat diffusion. If the approximation velocity exceeds the physically possible range of arc diffusion, an arc propagation prediction is determined to violate physical laws. When any physical inconsistency is detected, a judgment result indicating a physical inconsistency is output.

[0024] Step 3.4: When physical inconsistencies exist, a regularization penalty is applied to the node feature vector output by the graph convolution operation based on the human biomechanical limit threshold or the physical law of arc heat diffusion. This forces the predicted personnel movement trajectory or arc discharge propagation to be pulled back to the physically permissible range. Specifically, when the physical constraint judgment logic detects physical inconsistencies, the physical law constraint layer initiates a feature correction mechanism. For physical inconsistencies in personnel movement trajectory prediction, the constraint layer designs a regularization penalty term based on the human biomechanical limit threshold. This penalty term constrains the personnel node feature vector output by the graph convolution operation, forcing the predicted motion acceleration back to the range of the human biomechanical limit threshold, ensuring that the movement trajectory prediction conforms to the physical possibility of human movement. For physical inconsistencies in arc discharge propagation prediction, the constraint layer designs a corresponding regularization penalty term based on the physical law of arc heat diffusion, constraining and correcting the device node feature vector, adjusting the arc influence range expansion vector to a range that conforms to the physical law of arc heat diffusion. Regularization penalty is applied to the training process of spatiotemporal graph neural network through gradient backpropagation mechanism, continuously optimizing network parameters to output physically reliable prediction results, and finally realizing the physical consistency verification and correction of personnel movement trajectory prediction and arc discharge propagation prediction.

[0025] In this embodiment of the invention, by employing the method of obtaining the personnel displacement vector and the equipment arc spread vector as motion prediction after performing graph convolution operation on the spatiotemporal graph neural network, calculating the relative approximation velocity between the two and extracting the personnel motion acceleration features, and using physical constraint judgment logic to check whether it violates the limits of human biomechanics or the physical laws of arc heat diffusion, and applying regularization penalties to force the prediction results to return to the physically permissible range when physical inconsistencies exist, the technical problems of existing graph neural network prediction methods lacking physical constraints, resulting in prediction results that do not conform to the real physical laws, and failing to guarantee the credibility of motion trajectory and arc spread predictions, as well as the technical problems of traditional early warning systems failing to effectively integrate human biomechanics and arc physical characteristics, are overcome. Thus, the technical effect of ensuring the physical credibility of prediction results through a physical law constraint layer, effectively avoiding erroneous predictions that violate physical laws, improving the prediction accuracy and reliability of spatiotemporal graph neural networks in power distribution room safety early warning, and providing a scientific and credible prediction basis for safety risk assessment is achieved.

[0026] In a preferred embodiment of the present invention, the construction of a cross-modal micro-causal edge based on the micro-time difference between the occurrence time of the operation action and the electrical state response time determined by the virtual synchronization layer includes: Step 4.1 involves extracting video frame moments of personnel interaction with device nodes from the virtual synchronization layer as the operation action occurrence time, and extracting timestamps of electrical topology state change signals corresponding to device nodes as electrical state response times. Specifically, this includes extracting the timing information of personnel-device interaction from the synchronous multimodal events constructed by the virtual synchronization layer. For extracting the operation action occurrence time, video frame analysis identifies the interaction behaviors between personnel and device nodes, including actions such as personnel approaching the device, hand operation, and tool use. The specific video frame moments of these interactions are recorded as the operation action occurrence time. Interaction behavior recognition is based on human posture detection and action recognition algorithms, accurately locating the precise time point of personnel operating the device. For extracting the electrical state response time, electrical topology state change signals corresponding to device nodes related to personnel operation are obtained from the synchronous multimodal events. These signals include circuit breaker action, current changes, and voltage fluctuations. The timestamps carried by each electrical topology state change signal are extracted as the electrical state response time. By accurately extracting the operation action occurrence time and the electrical state response time, an accurate temporal basis is provided for constructing cross-modal micro-causal edges.

[0027] Step 4.2 calculates the microscopic time difference between the time of the operation and the time of the electrical state response. This includes receiving the extracted timestamps of the operation and the electrical state response, and then performing a precise time difference calculation. First, ensure that both timestamps use a unified time base and precision standard to eliminate potential clock synchronization errors. Then, calculate the time difference between the electrical state response and the operation, reflecting the response delay between the personnel's operation and the change in the state of the electrical equipment. The calculation accuracy of the microscopic time difference reaches the millisecond level, enabling the capture of microscopic causal timing relationships in power distribution room operations. A positive time difference indicates that the change in electrical state lags behind the operation, consistent with normal causal logic; a negative time difference may indicate anticipated operations or system clock deviations, requiring further analysis and verification.

[0028] Step 4.3: If the micro-time difference is within a preset causal relationship time window, a cross-modal micro-causal edge is constructed between the personnel node and the equipment node. Specifically, this involves: pre-setting a causal relationship time window based on the operating characteristics of the power distribution room. This window is determined based on the typical response time of electrical equipment, for example, set to 0 to 5 seconds, covering the normal response range from the operation action to the change in electrical state. The calculated micro-time difference is compared with the preset causal relationship time window to determine whether there is a reasonable causal relationship between the operation action and the change in electrical state. If the micro-time difference is positive and within the time window, it indicates that the change in electrical state is indeed caused by the personnel operation action, and there is a clear causal relationship. If the micro-time difference exceeds the time window range, it is determined that there is no direct causal relationship between the operation action and the change in electrical state. When a causal relationship is confirmed, a cross-modal micro-causal edge is constructed between the corresponding personnel node and equipment node. This edge connects the personnel node in the video modality and the equipment node in the electrical modality, reflecting the cross-modal causal relationship.

[0029] Step 4.4 determines the initial edge weights of cross-modal micro-causal edges based on the reciprocal of the micro-time difference. The smaller the micro-time difference, the larger the initial edge weight, indicating a higher confidence level in the causal relationship between the operational behavior and the electrical response. Specifically, this includes designing a weight allocation mechanism for cross-modal micro-causal edges based on the micro-time difference. The reciprocal of the micro-time difference is used as the basic weight calculation method to ensure that causal relationships with smaller time differences receive higher weights. The specific calculation formula is: Initial edge weight = 1 / (micro-time difference + ε), where ε is a very small positive number to prevent division by zero. This weight allocation method reflects the principle of temporal closeness of causal relationships: the shorter the time interval between the operational action and the electrical response, the more direct and reliable the causal relationship between the two, and the larger the corresponding causal edge weight. Normalization is also introduced to ensure that the weights of all cross-modal micro-causal edges are within a reasonable range, avoiding extreme weight values ​​from affecting subsequent attention mechanism calculations. The calculated initial edge weights serve as an important attribute of cross-modal micro-causal edges, guiding the attention weight allocation in the temporal feature encoding process of the spatiotemporal graph neural network.

[0030] In this embodiment of the invention, the technical means of using the timestamps of video frames and electrical topology state change signals extracted from the virtual synchronization layer as the time of operation and the time of electrical state response, respectively, to calculate the micro time difference between the two, construct cross-modal micro-causal edges when the time difference is within a preset causal association time window, and determine the weight of the causal edge based on the reciprocal of the micro time difference to characterize the confidence of causal association, overcomes the technical problems of existing multimodal fusion methods that cannot establish accurate causal relationships between video and electrical signals, lack of micro-temporal correlation analysis leading to inaccurate cross-modal association, and traditional attention mechanisms that cannot reflect the causal closeness between operation behavior and electrical response. Thus, it achieves the technical effect of accurately identifying and quantifying the micro-causal relationship between personnel operation and electrical state change, establishing accurate video-electrical association through cross-modal micro-causal edges, providing physically meaningful attention weight constraints for spatiotemporal graph neural networks, and effectively improving the accuracy and interpretability of multimodal feature fusion.

[0031] In a preferred embodiment of the present invention, after inputting the dynamic semantic graph at multiple consecutive time points into the spatiotemporal graph neural network and embedding the physical law constraint layer during the spatial feature encoding process of the spatiotemporal graph neural network, the method further includes: Step 3.5: During the temporal feature encoding process in the spatiotemporal graph neural network, the non-static safety boundary distance of the device nodes at the current moment is obtained, and a difference calculation is performed between this distance and the safety boundary distance at the previous moment to obtain the boundary evolution rate. Specifically, this includes: real-time monitoring of the dynamic changes in the safety boundaries of the device nodes during the temporal feature aggregation process. The non-static safety boundary distance of each device node is extracted from the dynamic semantic graph at the current moment; this distance value reflects the safety protection range of the device under the current electrical state. Simultaneously, the safety boundary distance values ​​of the same device nodes at the previous moment are obtained from the time series cache to establish a time comparison benchmark. By performing a difference calculation between the safety boundary distances at the current moment and the previous moment, the boundary evolution rate is obtained; this rate reflects the changing trend and speed of the safety boundary. A positive boundary evolution rate indicates that the safety boundary is expanding and the device's danger level is increasing; a negative boundary evolution rate indicates that the safety boundary is contracting and the device's danger level is decreasing. The boundary evolution rate provides important dynamic adjustment information for temporal feature encoding.

[0032] Step 3.6 uses the boundary evolution rate as the dynamic expansion factor of the temporal convolution kernel when aggregating temporal features of personnel nodes. Specifically, when the boundary evolution rate indicates an expanding safety boundary, the dynamic expansion factor is increased to broaden the temporal receptive field; when the boundary evolution rate indicates a contracting safety boundary, the dynamic expansion factor is decreased to focus on short-term dependencies. This includes receiving boundary evolution rate information and converting it into the dynamic expansion factor of the temporal convolution kernel. When the boundary evolution rate is positive, indicating an expanding safety boundary, the dynamic expansion factor is increased, expanding the receptive field of the temporal convolution kernel and enabling the aggregation of personnel node feature information over longer time series. This is because a expanding safety boundary implies an increasing risk impact range, requiring consideration of personnel behavior patterns over a longer time period. When the boundary evolution rate is negative, indicating a contracting safety boundary, the dynamic expansion factor is decreased, causing the temporal convolution kernel to focus on short-term temporal dependencies, emphasizing the analysis of recent changes in personnel node features. This is because a contracting safety boundary indicates a decreasing risk, requiring greater attention to the current immediate state. The dynamic expansion factor is adjusted using a nonlinear function to ensure that the expansion degree and the change in the boundary evolution rate maintain a reasonable correspondence, thereby achieving adaptive adjustment of the temporal receptive field.

[0033] In this embodiment of the invention, because the non-static security boundary distance of the device node at the current moment is obtained during the temporal graph neural network's temporal feature encoding, and the boundary evolution rate is calculated by difference between the distance at the current moment and the distance at the previous moment, and this evolution rate is used as the dynamic expansion factor of the temporal convolution kernel, the size of the temporal receptive field is adaptively adjusted according to the expansion or contraction of the boundary, thus overcoming the technical problems of existing temporal convolutional neural networks that use a fixed expansion factor and cannot dynamically adjust the temporal focus range according to the changes in security status, and lack correlation with changes in physical security boundaries, resulting in insufficient accuracy in temporal feature extraction, the invention achieves the technical effect of adaptive adjustment of the temporal receptive field based on dynamic changes in security boundaries, expanding the temporal focus range to capture long-term risk trends when risks expand, and focusing on short-term changes to improve response sensitivity when risks shrink, effectively improving the targeting and accuracy of temporal feature encoding in the spatiotemporal graph neural network.

[0034] In a preferred embodiment of the present invention, the above-mentioned output of risk evolution probability by the spatiotemporal graph neural network based on the corrected spatial features and constrained temporal features includes: Step 5.1 involves fusing the corrected spatial features with the constrained temporal features to obtain fused spatiotemporal features. Specifically, this includes receiving the corrected spatial features and constrained temporal features output from the spatiotemporal graph neural network and performing a multi-level feature fusion operation. The corrected spatial features include node spatial location information, spatial relationship features, and motion prediction features with physical consistency, verified and corrected by the physical constraint layer. The constrained temporal features include temporal evolution features constrained by cross-modal micro-causal edge attention weights, time-dependent features adjusted by dynamic expansion factors, and causal relationship encoding features. An attention mechanism is used to perform weighted fusion of spatial and temporal features, where the attention weights are dynamically calculated based on the confidence of cross-modal micro-causal edges. The fusion process consists of two stages: the first stage aligns feature dimensions, projecting spatial and temporal features of different dimensions onto a unified feature space; the second stage performs weighted fusion, calculating fusion weights based on feature importance and causal correlation strength to generate fused spatiotemporal features containing spatiotemporal correlation information.

[0035] Step 5.2: Based on the fused spatiotemporal features, extract the trajectory prediction intersection points from personnel nodes to equipment nodes, and calculate the first probability component of the trajectory prediction intersection points falling within the non-static safety boundary distance. Specifically, this includes: based on the personnel node motion trajectory prediction information in the fused spatiotemporal features, combined with the spatial location of the equipment nodes, calculating the spatial geometric relationship between the personnel motion trajectory and the equipment safety boundary. First, extract the motion trajectory prediction vector of the personnel node, and predict the possible motion path of the personnel in the future time period based on the current position and motion trend; then, extract the spatial location of the equipment node and the distance to the non-static safety boundary to construct a dynamic safety boundary region; next, calculate the geometric intersection points of the personnel motion trajectory and the equipment safety boundary region to determine the spatial coordinates of the trajectory prediction intersection points. Based on the spatial relationship between the trajectory prediction intersection points and the safety boundary, a probabilistic reasoning method is used to calculate the probability that the trajectory prediction intersection points fall within the non-static safety boundary distance. This probability considers the uncertainty of motion prediction, the dynamic changes of the safety boundary, and the complexity of the spatial geometric relationship, and outputs the first probability component.

[0036] Step 5.3 involves extracting the second probability component indicating whether a cross-modal micro-causal edge exists between personnel nodes and equipment nodes, and using the initial edge weight of the cross-modal micro-causal edge as the adjustment coefficient for the second probability component. Specifically, this includes analyzing the cross-modal micro-causal edge connections between personnel nodes and equipment nodes in the dynamic semantic graph. For each pair of personnel and equipment nodes, it checks whether a cross-modal micro-causal edge connection exists: if a cross-modal micro-causal edge exists, the second probability component is set to 1, indicating a causal relationship; if no cross-modal micro-causal edge exists, the second probability component is set to 0, indicating a lack of a clear causal relationship. When a cross-modal micro-causal edge exists, its initial edge weight is extracted as the adjustment coefficient for the second probability component. This weight reflects the confidence level of the causal relationship. The adjustment coefficient adjusts the contribution of the second probability component to the final risk probability: a high-weight causal edge indicates a strong causal relationship, enhancing the influence of the second probability component; a low-weight causal edge indicates a weak causal relationship, weakening the influence of the second probability component.

[0037] Step 5.4 involves nonlinearly fusing the first probability component with the second probability component weighted by an adjustment coefficient to output the risk evolution probability. Specifically, this includes receiving the first probability component and the second probability component weighted by an adjustment coefficient, and performing nonlinear probability fusion calculation. First, the second probability component is weighted by an adjustment coefficient: Weighted Second Probability Component = Second Probability Component × Adjustment Coefficient, yielding a weighted probability value reflecting the strength of the causal relationship. Then, a nonlinear fusion function is used to fuse the first probability component and the weighted second probability component. The fusion function design considers the different characteristics of spatial risk and causal risk: when both probability components are high, the risk evolution probability exhibits nonlinear growth, reflecting the superposition effect of risks; when there are differences in the probability components, the fusion function balances the contributions of different risk dimensions. The final output risk evolution probability integrates the spatial relationship risk between personnel trajectory and safety boundary, as well as the causal relationship risk between operational behavior and electrical response, providing a comprehensive risk assessment result for power distribution room safety early warning.

[0038] In this embodiment of the invention, the technical means of fusing modified spatial features with constrained temporal features to obtain fused spatiotemporal features, extracting the intersection of personnel trajectory predictions based on the fused features and calculating the first probability component of its falling into the nonstatic safety boundary, extracting the existence of cross-modal micro-causal edges as the second probability component and adjusting it with causal edge weights, and finally nonlinearly fusing the two probability components to output the risk evolution probability, overcomes the technical problems of existing risk assessment methods that only consider single spatial or temporal relationships and cannot comprehensively assess multi-dimensional risk factors, as well as the technical problems of traditional probability fusion methods that use linear combinations and cannot reflect the nonlinear superposition characteristics of risks. Thus, it achieves comprehensive risk assessment by integrating spatial trajectory risk and causal relationship risk, accurately reflects the interaction of different risk dimensions through nonlinear fusion, provides scientific and accurate risk evolution probability for power distribution room safety early warning, and effectively improves the comprehensiveness and accuracy of risk prediction.

[0039] In a preferred embodiment of the present invention, the above-mentioned generation of early warning instructions based on risk evolution probability and preset hierarchical strategy includes: Step 5.5 compares the risk evolution probability with a preset risk threshold. When the risk evolution probability exceeds the preset risk threshold, an intervention feasibility assessment is triggered. This specifically includes receiving the risk evolution probability output by the spatiotemporal graph neural network and comparing it numerically with the pre-configured preset risk threshold. The preset risk threshold is determined based on the power distribution room safety management standards and historical accident analysis. For example, it is set to 0.7, meaning that an early warning response needs to be initiated when the risk evolution probability exceeds 70%. When the risk evolution probability value is greater than the preset risk threshold, it is determined that the current safety situation has a high risk, and it is necessary to assess whether effective intervention can be carried out through technical means, thus triggering the intervention feasibility assessment process. The intervention feasibility assessment aims to analyze the effectiveness and feasibility of taking intervention measures such as power outages under the current risk situation, providing a decision-making basis for subsequent early warning level determination.

[0040] Step 5.6: Obtain the estimated arrival time of the trajectory prediction intersection point to the non-static safety boundary distance, and generate a simulated power outage command to be sent to the virtual electrical topology model. Obtain the completion time of the safety boundary contraction after the simulated power outage. Specifically, this includes: calculating the estimated arrival time of the trajectory prediction intersection point to the non-static safety boundary distance based on the spatial relationship between the personnel movement trajectory prediction and the safety boundary. Considering the current movement speed, movement direction, and possible movement changes of the personnel, a physical motion model is used to predict the time for the personnel to reach the safety boundary. Simultaneously, a simulated power outage command is generated for the relevant electrical equipment, and simulation calculations are performed. The virtual electrical topology model, based on the technical parameters of the electrical equipment and the power outage operation process, simulates the change process of the electrical equipment state after the execution of the power outage command, including the circuit breaker action time, arc extinction time, and electrical potential energy decay time, and calculates the completion time for the safety boundary to contract from the current state to the minimum safe state. The simulation process considers the response delay of the electrical equipment and the dynamic change law of the safety boundary, outputting an accurate contraction completion time.

[0041] Step 5.7: If the estimated arrival time is longer than the contraction completion time, the risk is determined to be preventable, and a Level 1 warning instruction is generated. The Level 1 warning instruction is linked to regular audio-visual alerts and platform push notifications, specifically including: a comparative analysis of the estimated arrival time and the contraction completion time. When the estimated arrival time is longer than the contraction completion time, it indicates that the safety boundary can be contracted to a safe state by performing a power outage operation before personnel reach the danger zone, and the risk can be effectively prevented by technical means. Based on the preventable determination, a Level 1 warning instruction is generated. This level of warning instruction has a relatively mild response, aiming to remind relevant personnel to pay attention to safety and prepare to take preventative measures. The Level 1 warning instruction includes the following response actions: activating the audio-visual alarm equipment in the power distribution room to emit sound and light signals to remind on-site personnel; pushing warning information to the safety monitoring platform to notify monitoring personnel to pay attention to the current safety status; and recording the warning event log to provide data support for subsequent safety analysis.

[0042] Step 5.8: If the estimated arrival time is less than or equal to the contraction completion time, the risk is determined to be unstoppable, and a second-level warning instruction is generated. The second-level warning instruction is linked to physical interlocking and emergency remote power outage. Specifically, it includes: when the estimated arrival time is less than or equal to the contraction completion time, it is determined that even if a power outage is immediately executed, the safety boundary cannot be contracted before personnel reach the danger zone, and the risk cannot be blocked by conventional technical means. Based on the unstoppable determination, a second-level warning instruction is generated. This level of warning instruction employs more stringent and urgent response measures, aiming to prevent safety accidents through mandatory means. The response actions included in the second-level warning instruction are: activating the physical interlocking function of the power distribution room access control system to prevent personnel from entering or to force personnel to evacuate the danger zone; executing an emergency remote power outage to immediately cut off the power supply to relevant electrical equipment and eliminate electrical safety hazards; simultaneously initiating the emergency response process and notifying emergency management personnel to arrive at the scene; recording the emergency warning event and initiating subsequent safety investigation procedures.

[0043] In this embodiment of the invention, by employing a technical approach that compares the probability of risk evolution with a preset risk threshold to trigger an intervention feasibility assessment, obtains the estimated arrival time of the trajectory prediction intersection point to the safety boundary, and simulates the completion time of the safety boundary contraction after a power outage using a virtual electrical topology model, and determines the risk blockability based on the comparison results of the two times and generates a corresponding level of warning instruction, the invention overcomes the technical problems of existing warning systems that use a single threshold for judgment, cannot assess the feasibility of risk intervention leading to inaccurate warning responses, and lack of time-based risk blockability analysis leading to weak targeting of warning measures. This achieves the technical effect of accurately determining the warning level based on the probability of risk evolution and the intervention time window, providing a scientific assessment of the intervention effect through virtual electrical topology model simulation, ensuring the matching degree between the warning response measures and the actual risk situation, and effectively improving the accuracy and practicality of power distribution room safety warnings.

[0044] In a preferred embodiment of the present invention, before inputting the dynamic semantic graph at multiple consecutive time points into the spatiotemporal graph neural network, the method further includes: Step 3.7: When a person node is detected to be missing due to occlusion in the video stream frame sequence, the historical motion trajectory vector of the occluded person node before its loss and the physical boundary constraints of its current location are obtained. Specifically, this includes real-time monitoring of the visibility status of the person node during the video stream frame sequence analysis. When a person node is detected to be missing in a video frame due to occlusion or other reasons, the information reconstruction process for the occluded person node is initiated. First, the historical motion trajectory vector of the occluded person node within a certain period before its loss is obtained from the historical trajectory database. This vector contains information such as the person's movement path, speed, and direction within the power distribution room, providing a motion trend reference for predicting the person's current possible location. Simultaneously, the last visible location of the occluded person node before its loss is analyzed. Combined with the physical layout information of the power distribution room, the physical boundary constraints of the current location are obtained, including spatial limitations imposed by physical obstacles such as walls, equipment, and passageways. These constraints limit the possible range of activity for the person.

[0045] Step 3.8: Based on the maximum movement speed of personnel constrained by the physical constraint layer, determine the possible location distribution circle of the occluded personnel node at the current moment. Specifically, this includes: calculating the maximum spatial range that the personnel could reach within the lost time period based on the maximum movement speed limit set in the physical constraint layer and the historical movement trajectory vector of the occluded personnel node. Using kinematic calculation methods, with the last known position before the personnel's loss as the center and the maximum movement speed multiplied by the loss time as the radius, determine the theoretical maximum possible activity range of the personnel at the current moment. Considering the uncertainty and complexity of personnel movement, and combining the directional and continuous characteristics of historical movement trajectories, the theoretical circle is modified to generate a possible location distribution circle that better conforms to the actual movement laws. This circle integrates the physical constraints of personnel movement, historical behavior patterns, and spatial geometric constraints, providing a probabilistic spatial range for estimating the position of the occluded personnel node.

[0046] Step 3.9: The center point coordinates of the possible location distribution circle are used as the virtual location coordinates of the occluded person node. The occluded person node is then completed in the dynamic semantic graph. The radius variance of the possible location distribution circle is added to the node features as an uncertainty attribute of the virtual location coordinates. Specifically, this includes: setting the calculated center point coordinates of the possible location distribution circle as the virtual location coordinates of the occluded person node, representing the most likely current location of the person; creating virtual person nodes in the dynamic semantic graph and using the virtual location coordinates as the spatial location attribute of the nodes to ensure the integrity of the graph structure; calculating the radius variance of the possible location distribution circle to accurately reflect the uncertainty of the virtual location, which reflects the reliability of the location prediction: a smaller radius indicates a more accurate location prediction and lower uncertainty; a larger radius indicates higher uncertainty. The radius variance is added as an uncertainty attribute to the feature vector of the virtual person node to provide location reliability information for subsequent risk assessment. During the calculation process of the spatiotemporal graph neural network, this uncertainty attribute is used to adjust the contribution weight of the virtual node to the risk evolution probability, ensuring that the risk assessment under occlusion conditions still has reasonable reliability.

[0047] In this embodiment of the invention, by employing the technical means of obtaining the historical motion trajectory vector and physical boundary constraints of a person node when occlusion is detected in a video stream frame sequence, determining the possible location distribution circular domain based on the maximum motion speed limit of the person in the physical constraint layer, using the center point of the circular domain as virtual location coordinates to complete the person node in the dynamic semantic graph, and attaching the radius variance as an uncertainty attribute to the node features, the technical problems of existing video surveillance systems such as loss of node tracking under person occlusion, inability to maintain graph structure integrity leading to decreased accuracy of risk assessment, and lack of a location prediction mechanism based on physical constraints that cannot reasonably estimate the possible location of occluded persons are overcome. Thus, the technical effects of maintaining the structural integrity of the dynamic semantic graph under person node occlusion, achieving reasonable location prediction through physical constraints and historical trajectories, and quantifying prediction credibility through uncertainty attributes are achieved, ensuring the continuity and reliability of risk assessment under occlusion are achieved.

[0048] In a preferred embodiment of the present invention, after the spatiotemporal graph neural network outputs the risk evolution probability based on the corrected spatial features and constrained temporal features, the method further includes: Step 5.9 involves graph-structure encoding the current dynamic semantic graph and its corresponding risk evolution probability to generate a current risk situation vector. This includes receiving the complete dynamic semantic graph structure information at the current moment, including feature vectors of all nodes, edge connections, and edge weights, while simultaneously acquiring the corresponding risk evolution probability values. A graph neural network encoder is used to structurally encode the dynamic semantic graph, compressing node features, edge relationships, and global graph topology information into a fixed-dimensional graph representation vector. The encoding process considers key information such as the different characteristics of personnel and equipment nodes, the weight distribution of cross-modal micro-causal edges, and safety boundary distances to ensure the encoding result comprehensively reflects the current risk state. Furthermore, the risk evolution probability is used as a global risk feature and concatenated with the graph structure encoding result to generate a comprehensive risk situation vector containing both graph structure and probability information. This vector serves as a compact representation of the current risk state of the power distribution room.

[0049] Step 5.10: Retrieve the historical risk situation vector with the highest cosine similarity to the current risk situation vector from the historical risk situation database. Obtain the historical handling strategies and their handling result scores associated with the historical risk situation vector. Specifically, this includes using the generated current risk situation vector as the query vector and performing a similarity search in the pre-built historical risk situation database. The historical database stores a large number of risk situation vectors, corresponding handling strategies, and handling effect scores for historical safety events in power distribution rooms, providing historical experience references for the current risk situation. Using the cosine similarity calculation method, the similarity between the current risk situation vector and each historical risk situation vector in the database is calculated one by one. The closer the cosine similarity value is to 1, the more similar the two situations are. Select the historical risk situation vector with the highest cosine similarity as the best matching result, and extract the handling strategy information corresponding to this historical situation, including the warning level adopted, the safety measures implemented, and the personnel evacuation plan; at the same time, obtain the handling result score, which is comprehensively evaluated based on factors such as the final safety effect, response time, and handling cost of the historical event.

[0050] Step 5.11: Package historical handling strategies and early warning instructions with scores higher than a preset threshold and simultaneously send them to the power distribution room monitoring terminal. This includes: quality screening of the retrieved historical handling strategies, setting a preset threshold for the handling result score (e.g., 8.0 out of 10), and selecting high-quality historical handling strategies with scores higher than the preset threshold. These high-scoring strategies represent effective safety handling solutions proven in practice under similar risk situations, possessing high reference value and a high success rate. Integrate and package the selected high-quality historical handling strategies with the currently generated early warning instructions to form a comprehensive handling plan that includes the current early warning instructions and historical successful experiences. The packaged information includes: current risk assessment results, automatically generated early warning levels and response measures, successful handling strategies for similar historical situations, and suggested optimization measures. Finally, simultaneously send the packaged comprehensive handling plan to the power distribution room on-site monitoring terminal, providing on-site safety management personnel with current early warning information and historical experience references to support scientific decision-making and efficient response.

[0051] Furthermore, to further improve the semantic understanding and accuracy of risk assessment, embodiments of this invention introduce the following... Figure 4 The intent inference mechanism is shown in the (Intent Inference Technology Process Architecture Diagram).

[0052] like Figure 4 As shown, this intent inference mechanism receives "trajectory encoding" and "image features" from video analysis as input. First, the multimodal features are transformed into sequence information through a "serialization and position encoding" process (including word embedding mapping and positional information injection). Then, this sequence is input into a "Transformer encoder" for deep feature extraction and fusion. Its core "multi-head attention mechanism" effectively captures long-term dependencies and key patterns in the sequence of human behavior. After feature transformation and stabilization through a "feedforward network and layer normalization," the final output is an "intent vector" representing the behavioral intent. This "intent vector" can be connected to... Figure 3 The "intent correction interface" in the ST-GNN architecture shown performs feature fusion or confidence weighting correction with the initial risk probability output by the spatiotemporal graph neural network, so that the semantics of the character's behavior (such as "inspection", "repair", "accidental entry") can more accurately affect the final risk assessment result.

[0053] In this embodiment of the invention, by employing a technical approach of encoding the current dynamic semantic graph and risk evolution probability into a risk situation vector using a graph structure, retrieving the historical risk situation vector with the highest similarity from the historical database and obtaining the corresponding historical handling strategy and score, and packaging the high-scoring historical strategy with the current warning instruction and simultaneously sending it to the monitoring terminal, the technical problems of existing warning systems—which only provide current risk information, lack historical experience reference leading to a lack of scientific basis for handling decisions, and are unable to fully utilize historical successful cases to optimize the current warning response strategy—are overcome. This achieves the technical effect of optimizing warning strategies based on experience driven by a historical risk situation database, providing scientific historical handling references through similar situation matching, and providing comprehensive decision support integrating current warnings and historical experience for on-site safety management personnel, effectively improving the practicality and handling effect of power distribution room safety warnings.

[0054] In a preferred embodiment of the present invention, the above-mentioned power distribution room early warning method is extended to the safe operation scenario of intelligent robots, specifically including: When the intelligent robot performs autonomous inspection, maintenance, or assists personnel in a power distribution room, the asynchronous multimodal data acquired in step 1 includes not only video stream frame sequences and electrical topology state change signals, but also robot sensor data, such as the robot's real-time position coordinates, joint angle states, end effector load, and task execution progress. In step 2, when constructing the dynamic semantic graph, the robot is treated as a special type of intelligent node, and mechanical dynamics constraint attributes are injected into the robot node, including the maximum extension radius of the robotic arm, joint movement speed limits, and load safety thresholds. The safe operating boundary distance of the robot is dynamically calculated based on its current task state and equipment interaction requirements. In the physical constraint layer of step 3, in addition to constraints on personnel movement... In addition to applying biomechanical constraints to the trajectory, mechanical dynamic constraints are also applied to the robot nodes to ensure that the predicted robot motion trajectory conforms to the kinematic constraints of the robotic arm, the joint angular velocity constraints, and the collision avoidance principle. In step 4, when constructing cross-modal micro-causal edges, control delay modeling between the robot command issuance time and the execution action response time is added to construct command-execution causal edges. In step 5, risk assessment comprehensively considers personnel safety risks, robot operation safety risks, and human-robot collaborative interaction risks. When it is detected that the robot and personnel are simultaneously approaching the same equipment and the distance is less than the preset collaborative safety distance, a coordinated early warning strategy of robot pause command and personnel evacuation warning is generated first to realize intelligent safety early warning and risk management in the human-robot collaborative environment.

[0055] Example 2:

[0056] Please see Figure 5As shown, this embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural network provided above.

[0057] Since the electronic device described in this embodiment is the one used to implement the power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural networks in this application embodiment, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural networks described in this application embodiment. Therefore, how the electronic device implements the method in this application embodiment will not be described in detail here. Any electronic device used by those skilled in the art to implement the power distribution room early warning method based on power semantic constraints and spatiotemporal graph neural networks in this application embodiment falls within the scope of protection of this application.

[0058] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0059] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0060] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for early warning of substations based on power semantic constraints and spatiotemporal graph neural networks, characterized in that, include: Asynchronous multimodal data in the power distribution room is acquired. The asynchronous multimodal data includes video stream frame sequences and electrical topology state change signals. The time slots are divided at equal intervals based on the acquisition timestamp of the video stream frame sequences. The electrical topology state change signal is embedded into the adjacent time slot closest to the timestamp of the signal it carries, forming a time-synchronized multimodal event sequence; Based on the time-synchronized multimodal event sequence, entity objects in the power distribution room scenario are extracted as graph nodes, and a dynamic semantic graph is constructed by combining the spatial positional relationships and electrical topology connectivity relationships between the graph nodes; the graph nodes include personnel nodes and equipment nodes; Read the real-time voltage level, current load current and circuit breaker opening / closing status of the device node to construct an electrical potential energy field feature vector. The electrical potential energy field feature vector is input into the pre-trained potential energy decay estimation model, and the potential energy decay coefficient at the current moment is output. Obtain the basic static safety distance corresponding to the device type node, use the potential energy attenuation coefficient as a dynamic adjustment factor, and perform scaling calculation based on the dynamic adjustment factor to obtain the non-static safety boundary distance; The dynamic semantic graph at multiple consecutive time points is input into the spatiotemporal graph neural network. After the graph convolution operation of the spatial feature encoding of the spatiotemporal graph neural network, a physical law constraint layer is embedded. The physical law constraint layer performs physical consistency verification and correction on the prediction of personnel movement trajectory and the prediction of arc discharge spread obtained from the node feature vectors output by the graph convolution operation based on the human biomechanical limit threshold or the physical law of arc heat diffusion. In the dynamic semantic graph, based on the micro time difference between the time of operation and the time of electrical state response determined by the time-synchronized multimodal event sequence, a cross-modal micro-causal edge is constructed between personnel nodes and equipment nodes. The initial edge weight of the cross-modal micro-causal edge is determined based on the reciprocal of the micro time difference. The cross-modal micro-causal edge is used as the attention weight constraint for the temporal feature encoding of the spatiotemporal graph neural network. The spatiotemporal graph neural network outputs the risk evolution probability based on the corrected spatial features and constrained temporal features, and generates early warning instructions based on the risk evolution probability and a preset hierarchical strategy. The process of outputting the risk evolution probability includes: The modified spatial features are fused with the constrained temporal features to obtain fused spatiotemporal features. Based on the fused spatiotemporal features, the trajectory prediction intersection point of personnel nodes pointing to equipment nodes is extracted, and the first probability component of the trajectory prediction intersection point falling within the non-static safety boundary distance is calculated. Extract the second probability component of whether the cross-modal micro-causal edge exists between the personnel node and the device node, and use the initial edge weight of the cross-modal micro-causal edge as the adjustment coefficient of the second probability component; The first probability component is nonlinearly fused with the second probability component weighted by the adjustment coefficient to output the risk evolution probability.

2. The method according to claim 1, characterized in that, The step of embedding the electrical topology state change signal into the adjacent time slot closest to the timestamp of the signal it carries, forming a time-synchronized multimodal event sequence, includes: Calculate the deviation between the signal occurrence timestamp and the center point of the adjacent time slot; If the deviation value is less than the preset jitter threshold, the signal occurrence timestamp carried by the electrical topology state change signal is embedded into the corresponding adjacent time slot; If the deviation value is greater than or equal to the preset jitter threshold, then linear interpolation is performed on the previous and next states of the electrical topology state change signal in the adjacent time slot to generate a virtual electrical state signal to complete the time slot. Each time slot contains the corresponding video frame information and electrical state information, forming a time-synchronized multimodal event sequence.

3. The method according to claim 1, characterized in that, The physical consistency verification and correction includes: After performing a graph convolution operation to encode spatial features in the spatiotemporal graph neural network, the displacement vectors of personnel nodes at adjacent time points are obtained as predictions of personnel movement trajectories, and the arc influence range expansion vectors of equipment nodes are obtained as predictions of arc discharge propagation. Calculate the relative approximation velocity between the displacement vector and the arc influence range expansion vector, and extract the motion acceleration features of the personnel node; The relative approximation velocity and the motion acceleration feature are input into the physical constraint judgment logic. If the motion acceleration feature exceeds the human biomechanical limit threshold, or the relative approximation velocity violates the physical law of arc heat diffusion, then it is determined that there is a physical inconsistency. When physical inconsistencies exist, a regularization penalty is applied to the node feature vector output by the graph convolution operation based on the human biomechanical limit threshold or the physical law of arc heat diffusion, so as to force the prediction of the human motion trajectory or the prediction of arc discharge spread back to the physically permissible range.

4. The method according to claim 1, characterized in that, The micro-time difference between the occurrence time of the operational action and the electrical state response time determined based on the time-synchronized multimodal event sequence is used to construct a cross-modal micro-causal edge between personnel nodes and equipment nodes, including: The video frame moments when personnel nodes interact with device nodes are extracted as the operation action occurrence moments, and the timestamps of the electrical topology state change signals corresponding to the device nodes are extracted as the electrical state response moments. Calculate the microscopic time difference between the time when the operation occurs and the time when the electrical state response occurs; If the micro-time difference is within a preset causal correlation time window, then a cross-modal micro-causal edge is constructed between the personnel node and the device node.

5. The method according to claim 1, characterized in that, After embedding a physical constraint layer following the graph convolution operation of the spatial feature encoding of the spatiotemporal graph neural network, the following further includes: When the spatiotemporal graph neural network performs time feature encoding, it obtains the non-static security boundary distance of the device node at the current time and performs differential calculation with the security boundary distance at the previous time to obtain the boundary evolution rate. The boundary evolution rate is used as the dynamic expansion factor of the temporal convolution kernel when aggregating temporal features of personnel nodes. When the boundary evolution rate indicates the expansion of the safe boundary, the dynamic expansion factor is increased to expand the temporal receptive field. When the boundary evolution rate indicates the contraction of the safe boundary, the dynamic expansion factor is decreased to focus on short-term dependencies.

6. The method according to claim 1, characterized in that, The generation of early warning instructions based on the risk evolution probability and the preset grading strategy includes: The probability of risk evolution is compared with a preset risk threshold. When the probability of risk evolution exceeds the preset risk threshold, an intervention feasibility assessment is triggered. Obtain the estimated arrival time of the trajectory prediction intersection point to the distance to the non-static safety boundary, generate a simulated power outage command and send it to the virtual electrical topology model, and obtain the completion time of the shrinkage of the safety boundary after the simulated power outage; If the estimated arrival time is longer than the contraction completion time, the risk is determined to be preventable, and a first-level warning instruction is generated. The first-level warning instruction is associated with regular sound and light reminders and platform push notifications. If the estimated arrival time is less than or equal to the contraction completion time, the risk is determined to be unstoppable, and a second-level warning instruction is generated. The second-level warning instruction is associated with physical locking and emergency remote power outage.

7. The method according to claim 1, characterized in that, Before inputting the dynamic semantic graph from multiple consecutive time points into the spatiotemporal graph neural network, the following steps are also included: When it is detected that a person node is occluded and lost in the video stream frame sequence, the historical motion trajectory vector of the occluded person node before the loss and the physical boundary constraints of the current area are obtained. Based on the maximum movement speed of personnel limited by the physical law constraint layer, the possible location distribution circle of the occluded personnel node at the current moment is determined. The coordinates of the center point of the possible location distribution circle are used as the virtual location coordinates of the occluded person node. The occluded person node is completed in the dynamic semantic graph, and the radius variance of the possible location distribution circle is added to the node feature as the uncertainty attribute of the virtual location coordinates.

8. The method according to claim 1, characterized in that, After the spatiotemporal graph neural network outputs the risk evolution probability based on the corrected spatial features and constrained temporal features, the method further includes: The dynamic semantic graph at the current moment and the corresponding risk evolution probability are encoded using a graph structure to generate the current risk situation vector; Retrieve the historical risk situation vector with the highest cosine similarity to the current risk situation vector from the historical risk situation database, and obtain the historical handling strategies and their handling result scores associated with the historical risk situation vector; The historical handling strategies with scores higher than a preset threshold are packaged together with the early warning instructions and simultaneously sent to the power distribution room monitoring terminal.

Images