A method and system for intelligent sensing of environmental safety monitoring in power distribution rooms

Abstract

This invention relates to the field of power distribution room environmental safety monitoring technology, and discloses an intelligent sensing method and system for power distribution room environmental safety monitoring. The method includes: acquiring multi-source sensing data and generating a layered thermal field distribution map and a spatial propagation delay matrix; analyzing the spatial occlusion relationship between sensing channels and generating an occlusion coefficient matrix; calculating the expected measurement sequence of each sensor channel based on a graph regularized constrained environmental field interpolation algorithm; distinguishing between spatial inhomogeneity and sensor drift and outputting micro-environmental correction sensing data; extracting the thermal anomaly evolution dynamics characteristics of each device; calculating the environmental sensing dynamic thermal coupling propagation coefficient matrix; recursively predicting the thermodynamic state evolution trajectory of each device based on a spatiotemporal graph convolutional cascade prediction network; and generating a comprehensive monitoring report of power distribution room environmental safety.

Description

Technical Field

[0001] This invention relates to the field of power distribution room environmental safety monitoring technology, specifically to an intelligent sensing method and system for power distribution room environmental safety monitoring. Background Technology

[0002] The power distribution room is densely packed with equipment such as transformers, busbars, and switchgear. There are ventilation dead zones in areas such as the top of the cabinets, narrow passages behind the cabinets, and cable trench openings. The temperature and humidity in these dead zones exhibit obvious stratification and delayed diffusion characteristics, meaning that point sensors and area infrared thermal imaging equipment do not observe the same spatial layer.

[0003] Existing methods for monitoring the environment of power distribution rooms typically employ a combination of point-type temperature and humidity sensors and infrared thermal imagers to monitor the indoor temperature and humidity field and the thermal state of equipment. The static thermal coupling coefficient is then calculated based on the equipment spacing to predict the cascading propagation of heat.

[0004] However, existing methods assume that all sensors in the same area perceive the same environmental state, which can easily misinterpret real spatial inhomogeneities as sensor drift or mistake location-related drift for local environmental anomalies, leading to inaccurate sensing data. Furthermore, the actual propagation behavior of heat—rapid lateral diffusion along the heat accumulation layer in dead zones or slow infiltration along obstructed channels—deviates significantly from the static thermal coupling coefficient calculated based on device spacing. This causes path misjudgments and time-of-arrival estimation errors in cascade propagation prediction, failing to provide effective environmental safety warnings in the early stages of thermal runaway cascading. Summary of the Invention

[0005] This invention provides an intelligent sensing method and system for monitoring the environmental safety of a power distribution room, solving the technical problems in related technologies such as inaccurate thermal anomaly identification and the inability to predict the cascading propagation path of thermal runaway caused by the spatial non-uniformity and sensor drift of multi-sensor sensing data in power distribution rooms.

[0006] This invention discloses an intelligent sensing method for monitoring the environmental safety of a power distribution room, comprising: acquiring multi-source sensing data of the power distribution room, performing rasterization processing on the power distribution room space based on sensor coordinates and cabinet layout, and generating a layered thermal field distribution map and a spatial propagation delay matrix; Calculate the occlusion coefficient matrix between each sensing channel based on sensor coordinates and cabinet geometric parameters; The layered thermal field distribution map, spatial propagation delay matrix, and occlusion coefficient matrix are input into the graph regularized constrained environment field interpolation algorithm, which outputs the expected measurement sequence of each sensor channel. Based on the positional correlation residual between the expected measurement sequence and the original sensor sequence, spatial non-uniformity and sensor drift are distinguished, and micro-environment correction sensing data is output. Based on the microenvironment correction and sensing data, the thermal anomaly evolution dynamic characteristics of each device are extracted, and the environmental sensing dynamic thermal coupling propagation coefficient matrix is ​​calculated based on the spatial propagation delay matrix and the layered thermal field distribution map. The dynamic characteristics of thermal anomaly evolution and the dynamic thermal coupling propagation coefficient matrix of environmental perception are input into the spatiotemporal graph convolutional cascade prediction network, which outputs the probability time series of each device entering the critical state of thermal runaway and the cascade propagation path.

[0007] Furthermore, the graph regularized constrained environment field interpolation algorithm includes: Based on the spatial coordinates of each grid cell in the hierarchical thermal field distribution map, the power distribution room space is represented as an undirected graph structure, with each grid cell serving as a graph node and graph edges established between adjacent grid cells. The weights of each graph edge are calculated based on the propagation delay values ​​between each node pair in the spatial propagation delay matrix and the corresponding element values ​​in the occlusion coefficient matrix. The graph edge weights are inversely proportional to the propagation delay between node pairs and inversely proportional to the occlusion coefficient. The graph Laplacian matrix is ​​generated based on the graph edge weights. A graph regularization constraint term is applied to the column vector formed by the temperature and humidity field values ​​of each node in the interpolation region. The graph regularization constraint term requires that nodes that are closely connected in space have similar temperature and humidity field values.

[0008] Furthermore, the graph regularized constrained environment field interpolation algorithm also includes: Using known sensor measurements as observation constraints and convective flux fields as boundary flux constraints, the objective function is minimized by combining the graph regularization constraint term. The minimized objective function is composed of a weighted combination of three terms: observation error term, graph regularization constraint term, and boundary flux constraint term. The interpolated temperature and humidity field values ​​of each grid node are obtained by taking the derivative of the minimization objective function with respect to the column vector of temperature and humidity field values ​​and setting the derivative to zero. Based on the interpolated temperature and humidity field values ​​of the grid nodes where each sensor is located, and combined with the sensor's installation height and orientation parameters, the expected measurement sequence of each sensor channel is calculated.

[0009] Furthermore, the distinction between spatial non-uniformity and sensor drift includes: Calculate the time-by-time difference between the desired measurement sequence and the original sensor sequence to generate a position-correlated residual sequence; The location-related residual sequence is matched with the historical aging curves of each sensor. If the location-related residual shows a monotonically slow change trend consistent with the historical aging curve, the residual component is identified as sensor aging drift. If the location-related residual shows fluctuation characteristics related to the environmental load cycle or there is a consistent residual pattern between spatially adjacent sensors, the residual component is identified as environmental differences caused by spatial inhomogeneity. The aging drift component of the sensor is extracted as a location correlation compensation parameter. The location correlation compensation parameter is subtracted from the original sensor sequence to output microenvironment correction sensing data.

[0010] Furthermore, the matching analysis also includes spatial consistency verification: For multiple sensors within the same grid area, if the correlation coefficient between their location-related residual sequences exceeds a preset threshold, it is determined that the residual in that area is mainly caused by spatial inhomogeneity; if the correlation between the residual sequence of a certain sensor and the residual sequences of other sensors in the same area is lower than the preset threshold, it is determined that the sensor has individual drift.

[0011] Furthermore, the extraction of the thermal anomaly evolution dynamics characteristics of each device includes: The temperature and humidity sequences of the grid where each device is located in the microenvironment correction sensing data and the real-time load current data of each device are input into the thermodynamic state space model of the corresponding device. The thermodynamic state space model is a linear state space equation pre-calibrated based on the device's heat capacity, heat dissipation coefficient and load heating power parameters. Calculate the residual vector between the current actual thermal state vector of each device and the model-predicted thermal state vector; Sliding window trajectory analysis is performed on the residual vector sequence. Within each sliding window, the motion direction, acceleration, and curvature features of the residual offset vector in the state space are extracted and combined to generate the thermal anomaly evolution dynamics features of each device.

[0012] Furthermore, the computational environment-aware dynamic thermal coupling propagation coefficient matrix includes: Based on the convection channel distribution in the layered thermal field distribution map, the actual heat propagation path between each device is identified. The actual heat propagation path includes the lateral propagation path along the heat accumulation layer and the penetration path along the shielding channel. Based on the spatial propagation delay matrix and the occlusion coefficient matrix, the equivalent heat conduction time constant of each device along the actual heat propagation path is calculated. After taking the reciprocal of the equivalent heat conduction time constant between each pair of devices, mean normalization based on the range is performed, and the reciprocal values ​​of each pair of devices are mapped to the interval between 0 and 1 to generate the dynamic thermal coupling propagation coefficient matrix of environmental perception. The dynamic thermal coupling propagation coefficient matrix of the environment perception is dynamically adjusted as the micro-environment correction perception data is updated.

[0013] Furthermore, the spatiotemporal graph convolutional cascaded prediction network uses devices as nodes and environmental perception dynamic thermal coupling propagation coefficients as edge weights, and includes spatial graph convolutional units, temporal gated recurrent units, and cascaded inference units. The spatial graph convolution unit uses the current thermal anomaly evolution dynamics of each device as node features and the environmental perception dynamic thermal coupling propagation coefficient matrix as adjacency weights to perform graph convolution operations, and outputs a device state representation vector that integrates spatial coupling relationships. The time-gated loop unit models the device state representation vector sequence along the time dimension and outputs the predicted state vector sequence of each device at multiple future time steps. The cascaded inference unit includes a thermal runaway probability output layer and a cascaded path inference layer. The thermal runaway probability output layer maps the predicted state vector to the probability value of each device entering the thermal runaway critical state at the corresponding time step. The cascaded path inference layer determines the propagation order of thermal anomalies among devices based on edge weights and the order in which each device reaches the critical state.

[0014] Furthermore, it also includes generating a comprehensive environmental safety monitoring report for the power distribution room: Based on the cascaded propagation path and the thermal runaway probability time series of each device, the interventionable time window of each node device on the propagation link is calculated. The interventionable time window is the time difference between the current moment and the time step corresponding to the first time when the probability value in the probability time series exceeds the preset critical probability threshold. Local environmental security alarm levels are divided based on the length of the interventionable time window and the coverage of the cascading propagation path. The highest alarm level is generated when the interventionable time window is shorter than the first time threshold and the number of devices covered by the cascading propagation path exceeds the first quantity threshold. The medium alarm level is generated when the interventionable time window is between the first time threshold and the second time threshold. The general alarm level is generated when the interventionable time window is longer than the second time threshold.

[0015] This invention provides an intelligent sensing-based power distribution room environmental safety monitoring system, comprising: The layered thermal field generation module is used to acquire multi-source sensing data of the power distribution room, and to perform rasterization processing on the power distribution room space based on sensor coordinates and cabinet layout to generate a layered thermal field distribution map and a spatial propagation delay matrix. The occlusion analysis module is used to calculate the occlusion coefficient matrix between each sensing channel based on sensor coordinates and cabinet geometric parameters. The environmental field interpolation module is used to input the layered thermal field distribution map, spatial propagation delay matrix and occlusion coefficient matrix into the graph regularized constrained environmental field interpolation algorithm, and output the expected measurement sequence of each sensor channel; The drift differentiation and correction module is used to differentiate spatial non-uniformity and sensor drift based on the position correlation residual between the expected measurement sequence and the original sensor sequence, and outputs micro-environment correction sensing data. The feature extraction module is used to extract the thermal anomaly evolution dynamics features of each device based on microenvironment correction sensing data; The coupling coefficient calculation module is used to calculate the dynamic thermal coupling propagation coefficient matrix of environmental perception based on the spatial propagation delay matrix and the hierarchical thermal field distribution map. The cascaded prediction module is used to input the dynamic characteristics of thermal anomaly evolution and the dynamic thermal coupling propagation coefficient matrix of environmental perception into the spatiotemporal graph convolutional cascaded prediction network, and output the probability time series of each device entering the critical state of thermal runaway and the cascaded propagation path.

[0016] This invention reconstructs the layered thermal and humidity field of dead-angle regions using a graph regularized constrained environmental field interpolation algorithm. Spatial propagation delay and occlusion coefficient are used as graph edge weights to constrain the interpolation process, ensuring that the interpolation results follow the physical characteristics of heat propagation along actual channels. This allows for the acquisition of the expected measurement values ​​of each sensor in a real micro-environment. Based on the positional correlation residual analysis between the expected measurement sequence and the original sensor sequence, environmental differences caused by spatial inhomogeneity are distinguished from individual sensor aging drift. This avoids misjudging real environmental differences as sensor malfunctions or sensor drift as environmental anomalies, ensuring that subsequent monitoring is based on accurate sensing data.

[0017] This invention calculates the dynamic thermal coupling propagation coefficient matrix of environmental perception based on a layered thermal field distribution map and a spatial propagation delay matrix, replacing the static coefficient based on device spacing. This ensures that the information aggregation path of the spatiotemporal graph convolutional cascade prediction network is consistent with the actual propagation path of heat spreading laterally along the heat accumulation layer or penetrating along the shielding channel. As a result, the cascade prediction can capture the unique thermal propagation patterns of dead zone areas and provide early warning information including propagation path and remaining intervention time in the early stages of thermal runaway cascade development. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent sensing method for monitoring the environmental safety of a power distribution room provided in an embodiment of the present invention; Figure 2 This is a schematic diagram comparing the original temperature and standardized temperature of a typical node provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatial propagation delay distribution of a typical grid pair provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a typical sensor occlusion coefficient matrix provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the expected measurement value of a typical node with the original measurement value provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the correlation residuals and distinction results of each node position provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the thermal coupling propagation coefficient and equivalent time constant of the device provided in the embodiment of the present invention; Figure 8 This is a schematic diagram of the time series prediction results of the thermal runaway probability of each device provided in the embodiments of the present invention; Figure 9 This is a schematic diagram comparing the dynamic characteristics of thermal anomaly evolution of various devices provided in the embodiments of the present invention. Detailed Implementation

[0019] Example 1 In the scenario of environmental safety monitoring in power distribution rooms, indoor equipment such as transformers, busbars, and switchgear are densely arranged, and ventilation dead zones exist in areas such as the top of cabinets, narrow passages behind cabinets, and cable trench openings. Temperature and humidity in these dead zones exhibit obvious stratification and delayed diffusion characteristics, meaning that point sensors and area infrared thermal imaging devices do not observe the same spatial layer. Existing monitoring methods assume that all sensors in the same area perceive the same environmental state, easily misinterpreting actual spatial inhomogeneities as sensor drift, or mistaking location-related drift for local environmental anomalies, leading to inaccurate sensing data. Simultaneously, the actual propagation behavior of heat—rapid lateral diffusion along the heat accumulation layer or slow infiltration along obstructed channels in dead zones—deviates significantly from the static thermal coupling coefficient calculated based on equipment spacing. This causes path misjudgment and arrival time estimation errors in cascade propagation prediction, failing to provide effective environmental safety early warnings in the early stages of thermal runaway cascading development.

[0020] The hardware environment of this embodiment includes: a multi-point temperature and humidity sensor array deployed in various areas of the power distribution room, an SF6 gas concentration sensor, an infrared thermal imager, an equipment load monitoring module and a fan operation status acquisition terminal, as well as an edge computing server for performing data processing and predictive calculations.

[0021] According to an embodiment of the present invention, a method for intelligent sensing of the safety monitoring environment of a power distribution room is provided, such as... Figure 1 As shown, it includes the following steps: Step 1: Acquire multi-source sensing data from the power distribution room and generate a layered thermal field distribution map and a spatial propagation delay matrix.

[0022] Acquire multi-point temperature sequences, humidity sequences, SF6 gas concentration sequences, infrared thermal imaging temperature matrix sequences, equipment load sequences, fan operation sequences, and sensor coordinate data for the power distribution room. Perform timestamp alignment processing on the above multi-source data to unify them to the same sampling time base. Based on the sensor coordinate data and cabinet layout information, the power distribution room space is rasterized, and each sensor is assigned to a corresponding raster unit.

[0023] Before being processed, the multi-source sensing data, including temperature sequences, humidity sequences, SF6 gas concentration sequences, infrared thermal image temperature matrix sequences, and equipment load sequences, undergo Z-score standardization to eliminate the impact of differences in the dimensions of different physical quantities on subsequent weighted calculations.

[0024] Based on the gridded temperature sequence, the vertical temperature difference distribution between adjacent vertical grids is calculated. Based on the temperature gradient of adjacent transverse grids and the thermal conductivity of the corresponding material, the transverse heat flux density distribution is calculated. Combining the fan operation sequence and the cross-sectional parameters of the channels between grids, the local convection rate distribution of each grid cell is calculated.

[0025] The vertical temperature difference distribution, lateral heat flux density distribution, and local convection rate distribution are combined according to spatial grid coordinates to generate a layered thermal field distribution map. A temperature step perturbation is applied between each grid cell, and the time delay of the perturbation signal reaching the adjacent grid is recorded to generate a spatial propagation delay matrix.

[0026] It should be noted that the above-mentioned vertical temperature difference distribution refers to the sequence of temperature differences between adjacent grid levels along the height of the distribution room. In the top area of ​​the cabinet, due to the natural upward movement of hot air and its obstruction by the top, the vertical temperature difference usually exhibits a gradient distribution where the upper layers are higher than the lower layers. In the cable trench area, due to the sinking effect of cold air, the vertical temperature difference exhibits a reverse gradient. The layered thermal field distribution map represents the thermodynamic state parameters of each spatial layer in the distribution room in the form of a three-dimensional grid. Each grid cell contains temperature value, humidity value, heat flux density vector, and convection rate scalar.

[0027] Step 2: Analyze the spatial occlusion relationship between the sensing channels and generate an occlusion coefficient matrix.

[0028] Based on sensor coordinate data and the visible area of ​​the infrared thermal imager, the direct observability of each temperature sensor's grid is determined. Based on the geometric parameters and installation position of each cabinet in the cabinet layout grid, it is calculated whether the line-of-sight path between any two sensors is obstructed by a cabinet. For sensor pairs with obstruction, the proportion of the obstructed area to the total line-of-sight cross-section is calculated, generating an obstruction coefficient matrix.

[0029] It should be noted that the occlusion coefficient matrix is ​​a symmetric matrix, with element values ​​ranging from 0 to 1. A value of 0 indicates no occlusion between the two sensing channels, while a value of 1 indicates complete occlusion. For the occlusion coefficient between the infrared thermal imager and the point sensor, the reflection path of infrared radiation must also be considered, i.e., the proportion of thermal radiation indirectly received by the infrared thermal imager after being reflected from the cabinet surface. This reflection contribution should be subtracted from the occlusion coefficient.

[0030] Step 3: Calculate the expected measurement sequence for each sensor channel based on the graph regularized constrained environmental field interpolation algorithm.

[0031] The layered thermal field distribution map, spatial propagation delay matrix, and occlusion coefficient matrix are input into the graph-regularized constrained environmental field interpolation algorithm, which outputs the expected measurement sequence for each sensor channel. The graph-regularized constrained environmental field interpolation algorithm is a non-uniform interpolation algorithm for reconstructing the thermal and humidity field of a power distribution room, and its steps are as follows.

[0032] Step 301: Based on the spatial coordinates of each grid cell in the layered thermal field distribution map, the power distribution room space is represented as an undirected graph structure, with each grid cell serving as a graph node and graph edges established between adjacent grid cells.

[0033] Step 302: Based on the propagation delay values ​​between each node pair in the spatial propagation delay matrix and the corresponding element values ​​in the occlusion coefficient matrix, calculate the weight of each graph edge. Let the nodes... With nodes The propagation delay between them after Z-score normalization is The occlusion coefficient is , To prevent small positive numbers with a denominator of zero, the graph edge weights are... Calculate as follows: in, Let be the index of the first node in the graph. This is the index of the second node in the graph. For nodes With nodes The propagation delay between them after Z-score normalization For nodes With nodes The occlusion coefficient between them To prevent small positive numbers with a denominator of zero, For nodes With nodes The edge weights between graph edges. With propagation delay Inversely proportional to the occlusion coefficient Inversely proportional, the shorter the propagation delay and the less occlusion, the greater the connection weight between adjacent nodes.

[0034] Step 303: Generate the graph Laplacian matrix based on graph edge weights. The column vector formed by the temperature and humidity field values ​​of each node in the region to be interpolated Apply graph regularization constraints .in, This is a column vector composed of the temperature and humidity field values ​​of each node. The graph Laplacian matrix is ​​generated based on propagation delay and occlusion coefficient. This represents the transpose, and the graph regularization constraint term. Nodes that are closely connected in space are required to have similar temperature and humidity values.

[0035] Step 304: Using known sensor measurements as observation constraints and the convective flux field as boundary flux constraints, combine the graph regularization constraint term. Solve to minimize the objective function to obtain the interpolated temperature and humidity field values ​​and convective flux field values ​​for each grid node. Minimize the objective function. The form is: in, As column vectors Let be the objective function that minimizes the independent variable; The observation matrix, the observation matrix Each row corresponds to the indicator vector of the grid node where the known sensor measurement value is located; This is a column vector composed of the temperature and humidity field values ​​of each node; A column vector consisting of known sensor measurements; The weighting coefficients for graph regularization constraints; Indicates transpose; The graph is a Laplace matrix; Here is the boundary flux constraint matrix. The flux direction vector of each row corresponding to the boundary grid node; This is a column vector consisting of the known flux values ​​of the flux field at the boundary nodes; These are the weighting coefficients for the boundary flux constraints. For minimizing the objective function... Regarding column vectors By taking the derivative and setting it to zero, we obtain a system of linear equations. Solving this system of linear equations yields the interpolated temperature and humidity field values ​​for each grid node.

[0036] Step 305: Based on the interpolated temperature and humidity field values ​​of each sensor's grid node, combined with the sensor's installation height and orientation parameters, calculate the theoretical values ​​that the sensor should observe under the current environmental field, and output the expected measurement sequence for each sensor channel. The installation height parameter is used to extract the temperature and humidity values ​​of the corresponding height layer from the interpolated temperature and humidity field values, and the orientation parameter is used to determine the heat flux density component corresponding to the sensor's sensing surface. Together, they determine the sensor's theoretical observation values.

[0037] It should be noted that the graph regularization constraint term This ensures that the interpolation results follow the physical characteristics of heat diffusion along the actual propagation channel, rather than simple Euclidean distance interpolation. In dead zones, due to the larger occlusion coefficient and longer propagation delay, the corresponding graph edge weights are smaller, which weakens the interpolation coupling between the dead zone and the main channel region, thus reflecting the true physical characteristics of the difference between the environmental conditions in the dead zone and the main channel region.

[0038] Step 4: Distinguish between spatial non-uniformity and sensor drift, and output microenvironment correction sensing data.

[0039] Calculate the time-by-time difference between the desired measurement sequence and the original sensor sequence to generate a position-related residual sequence.

[0040] Historical aging curves for each sensor are obtained, representing the monotonic variation trend of sensor measurement deviation over service time. The location-related residual sequence is matched with the historical aging curves: if the location-related residuals exhibit a monotonic, gradual trend consistent with the historical aging curves, the residual component is identified as sensor aging drift; if the location-related residuals exhibit fluctuation characteristics related to environmental load cycles or if spatially adjacent sensors have consistent residual patterns, the residual component is identified as actual environmental differences caused by spatial inhomogeneity.

[0041] Based on the above distinction results, the sensor aging drift component is extracted as a location-related compensation parameter. The location-related compensation parameter is subtracted from the original sensor sequence to output microenvironment-corrected sensing data.

[0042] In this embodiment of the application, in order to improve the accuracy of distinguishing between sensor drift and environmental differences, the matching analysis of position-related residuals also includes spatial consistency verification: for multiple sensors in the same grid area, if the correlation coefficient between their position-related residual sequences exceeds a preset threshold, it is determined that the residuals in the area are mainly caused by spatial inhomogeneity; conversely, if the residual sequence of a certain sensor has a low correlation with the residual sequences of other sensors in the same area, it is determined that the sensor has individual drift.

[0043] Step 5: Extract the dynamic characteristics of thermal anomaly evolution of each device.

[0044] The temperature and humidity sequences of each device's grid in the microenvironment correction sensing data, along with the real-time load current data of each device, are input into the corresponding device's thermodynamic state-space model. The thermodynamic state-space model is a linear state-space equation pre-calibrated based on the device's heat capacity, heat dissipation coefficient, and load heating power parameters. The inputs to the thermodynamic state-space model are the load current and ambient temperature, and the output is the device's predicted thermal state vector.

[0045] Calculate the residual vector between the actual thermal state vector of each device and the model-predicted thermal state vector. Perform sliding window trajectory analysis on the residual vector sequence, with the window length covering multiple sampling periods. Within each sliding window, extract the motion direction, acceleration, and curvature features of the residual offset vector in the state space. The motion direction represents the development trend of the thermal anomaly, the acceleration represents the deterioration rate of the thermal anomaly, and the curvature represents the nonlinearity of the thermal anomaly evolution path. Combine these features to generate the thermal anomaly evolution dynamics characteristics of each device.

[0046] Before inputting the dynamic features of thermal anomaly evolution into the subsequent network, Z-score normalization is performed on the motion direction, acceleration, and curvature features to eliminate the influence of the difference in the dimensions of each feature component on the subsequent graph convolution operation.

[0047] It should be noted that the thermal state vector in the thermodynamic state-space model includes three components: the equipment surface temperature, the estimated internal winding temperature, and the radiator temperature. When the equipment is operating normally, the residual vector between the actual thermal state vector and the model-predicted thermal state vector fluctuates randomly around the origin. When the equipment exhibits early thermal anomalies, the residual vector continuously deviates from the origin and moves in a specific direction, with the acceleration of the residual vector turning from zero to positive, indicating that the anomaly is accelerating.

[0048] Step 6: Calculate the dynamic thermal coupling propagation coefficient matrix of environmental perception.

[0049] Based on the propagation delay values ​​between each pair of device nodes in the spatial propagation delay matrix, the fundamental time constant for heat propagation from one device to another is determined. Based on the convection channel distribution in the layered thermal field distribution map, the actual heat propagation paths between devices are identified, including lateral propagation paths along the heat accumulation layer and infiltration paths along obstruction channels. The fundamental time constant is corrected based on the degree of obstruction on the corresponding paths in the obstruction coefficient matrix.

[0050] For any two devices, the equivalent heat conduction time constant along the actual heat propagation path is calculated: on the lateral heat accumulation layer path, the equivalent heat conduction time constant is shorter, reflecting rapid lateral heat diffusion; on the shielded channel penetration path, the equivalent heat conduction time constant is longer, reflecting slow heat penetration. The reciprocal of the equivalent heat conduction time constant between each device pair is taken, and mean normalization based on the range is performed. This maps the reciprocal values ​​of each device pair to the interval between 0 and 1, generating an environment-aware dynamic thermal coupling propagation coefficient matrix, ensuring the comparability of the elements in the environment-aware dynamic thermal coupling propagation coefficient matrix.

[0051] In this embodiment, to reflect the impact of environmental state changes on the thermal coupling coefficient, the environmental sensing dynamic thermal coupling propagation coefficient matrix is ​​dynamically adjusted as the microenvironment correction sensing data is updated. When the fan operating state changes, the convection channel distribution changes accordingly, and the propagation paths and equivalent heat conduction time constants between devices are updated accordingly. When local heat accumulation causes changes in the layered structure, the effectiveness of the lateral propagation path is adjusted accordingly.

[0052] Step 7: Based on the spatiotemporal graph convolutional cascade prediction network, recursively predict the thermodynamic state evolution trajectory of each device.

[0053] The dynamic characteristics of thermal anomaly evolution of each device and the dynamic thermal coupling propagation coefficient matrix of environmental perception are input into the spatiotemporal graph convolutional cascade prediction network, which outputs the probability time series of each device entering the critical state of thermal runaway and the cascade propagation path.

[0054] The spatiotemporal graph convolutional cascaded prediction network uses devices as nodes and environmental perception dynamic thermal coupling propagation coefficients as edge weights. It consists of three components: spatial graph convolutional units, temporal gated recurrent units, and cascaded inference units. The data transfer relationships between these units are as follows.

[0055] The input to the spatial graph convolution unit is the thermal anomaly evolution dynamics characteristics of each device at the current moment (as node features) and the environmental perception dynamic thermal coupling propagation coefficient matrix (as adjacency weights). The spatial graph convolution unit performs graph convolution operations, aggregates the thermal state information of the neighboring nodes of each device, and outputs a device state representation vector that fuses spatial coupling relationships, which is then passed to the temporal gating loop unit.

[0056] The input to the temporally gated loop unit is the sequence of device state representation vectors output by the spatial graph convolution unit. The temporally gated loop unit models the temporal evolution of the thermal state of each device along the time dimension and outputs the predicted state vector sequence of each device for multiple future time steps, which is then passed to the cascaded inference unit.

[0057] The input to the cascaded inference unit is the sequence of predicted state vectors of each device output by the time-gated loop unit. The cascaded inference unit comprises two sub-layers: a thermal runaway probability output layer and a cascaded path inference layer. The thermal runaway probability output layer is a fully connected layer that maps the predicted state vector of each device to the probability value of that device entering the thermal runaway critical state at the corresponding time step. The activation function is the Sigmoid function, and the output value ranges from 0 to 1, forming the probability time series of each device. The cascaded path inference layer, based on the edge weights in the environmentally perceived dynamic thermal coupling propagation coefficient matrix and the temporal sequence of each device reaching the critical state, determines the propagation order of thermal anomalies among devices in descending order of edge weights, outputting the cascaded propagation path. The probability value of each time step in the probability time series directly corresponds to the probability that the device will enter the thermal runaway critical state at that moment, and is used for the calculation of the subsequent interveneable time window; the cascaded propagation path is represented by an ordered sequence of device nodes, directly serving as the data source for the spatially corrected cascaded path graph in step 8.

[0058] The spatiotemporal graph convolutional cascade prediction network uses the historical thermal anomaly evolution dynamic feature sequence and the corresponding environmental perception dynamic thermal coupling propagation coefficient matrix sequence as training inputs, and the actual time label of each device entering the thermal runaway critical state and the actual cascade propagation order as supervision labels. The prediction error of the probability time series is calculated using the cross-entropy loss function, and the parameters are updated using the Adam optimization algorithm.

[0059] It should be noted that the adjacency weights used by the spatial graph convolutional unit are environmentally perceptual dynamic thermal coupling propagation coefficients, rather than fixed weights based on the Euclidean distance between devices. This characteristic ensures that the information aggregation path of the spatial graph convolutional unit is consistent with the actual heat propagation path in the dead zone area. The device state representation vector output by the spatial graph convolutional unit can reflect the coupling effects of two different propagation modes: rapid lateral conduction along the heat accumulation layer and slow penetration along the obstruction channel.

[0060] Step 8: Generate a comprehensive environmental safety monitoring report for the power distribution room.

[0061] Based on the cascading propagation path and the thermal runaway probability time series of each device, the interventionable time window for each node device in the propagation link is calculated. The interventionable time window is the time difference between the current moment and the moment when the device is expected to reach the thermal runaway critical state. The moment when the device is expected to reach the thermal runaway critical state is the time step corresponding to the first time when the probability value in the probability time series exceeds the preset critical probability threshold.

[0062] By combining the temperature values ​​and dew point temperatures of each grid cell in the microenvironment correction sensing data, a condensation risk index is calculated. Based on the local accumulation concentration values ​​in the dead zone areas of the SF6 gas concentration sequence, a local gas accumulation concentration index is calculated. Based on the temperature-time change rate of each device's grid in the layered thermal field distribution map, a hotspot temperature rise index is calculated.

[0063] Based on the above indicators, a comprehensive environmental safety monitoring report for the power distribution room is generated. This report includes the identification of thermal anomaly source equipment, a spatially corrected cascade path diagram, the estimated critical state time for each device, the remaining intervention time, and the local environmental safety alarm level.

[0064] In this embodiment, the classification of local environmental security alarm levels is based on the length of the interventionable time window and the coverage of the cascading propagation path: when the interventionable time window is shorter than a first time threshold and the number of devices covered by the cascading propagation path exceeds a first quantity threshold, the highest alarm level is generated; when the interventionable time window is between the first time threshold and the second time threshold, a medium alarm level is generated; when the interventionable time window is longer than the second time threshold, a general alarm level is generated.

[0065] This implementation reconstructs the layered thermal and humidity field of the dead-angle area in the power distribution room using a graph regularized constrained environmental field interpolation algorithm. Spatial propagation delay and occlusion coefficient are used as graph edge weights to constrain the interpolation process, ensuring that the interpolation results follow the physical characteristics of heat propagation along actual channels rather than the assumption of uniform space. This allows for the acquisition of the expected measurement values ​​of each sensor in the real micro-environment. Based on the positional correlation residual analysis between the expected measurement sequence and the original sensor sequence, environmental differences caused by spatial inhomogeneity are distinguished from individual sensor aging drift. This ensures that subsequent monitoring is based on accurate sensing data, avoiding misjudging real environmental differences as sensor malfunctions or sensor drift as environmental anomalies.

[0066] Meanwhile, this implementation calculates an environmentally perceptible dynamic thermal coupling propagation coefficient matrix reflecting the actual heat propagation behavior in dead-end areas based on a layered thermal field distribution map and a spatial propagation delay matrix. This replaces the static coefficients based on device spacing, ensuring that the information aggregation path of the spatiotemporal graph convolutional cascade prediction network aligns with the actual propagation path of heat either rapidly diffusing laterally along the heat accumulation layer or slowly penetrating along obstruction channels. Therefore, cascade prediction can capture the unique heat propagation patterns in dead-end areas, providing early warning information including propagation paths and remaining intervention times in the early stages of thermal runaway cascade development. This overcomes the limitations of existing methods that only perform static assessments of the overall temperature field gradient and cannot accurately predict cascade propagation paths and arrival times.

[0067] The following is an example of an application of the present invention, such as... Figure 2-9 As shown, the implementation process is as follows: The power distribution room of an industrial park (hereinafter referred to as the "target power distribution room") is responsible for the main power supply of the park. It houses three dry-type transformers (Transformer A, Transformer B, and Transformer C), two busbars, and four switchgear cabinets. The narrow passage behind the cabinets is approximately 0.6 meters wide, and the net height from the top of the cabinets to the ceiling is approximately 0.4 meters. The cable trench is located in the southeast corner of the power distribution room. The room contains 12 temperature and humidity sensor nodes, two infrared thermal imagers, three SF6 gas concentration sensors, and one fan operation status acquisition terminal. The edge computing server has a sampling period of 30 seconds.

[0068] On a summer day in 20XX, the park's electricity load remained high, and a suspected thermal anomaly signal appeared in the area where transformer A was located, triggering a full-process monitoring operation.

[0069] Step 1: Acquire multi-source sensing data and generate a hierarchical thermal field distribution map and a spatial propagation delay matrix; The system collects raw readings from each sensor at the current moment, including 12 temperature and humidity nodes, 3 SF6 concentration nodes, temperature matrices output from 2 infrared thermal imagers, load current of each device, and fan speed. After aligning the timestamps to the same sampling time base, the power distribution room space is rasterized, dividing the indoor space into 72 grid units of 6×4×3 (length×width×height), with the height divided into three levels: low (0 to 0.8 meters), middle (0.8 to 1.8 meters), and high (above 1.8 meters).

[0070] After Z-score normalization of temperature, humidity, SF6 concentration, infrared temperature matrix, and load current, the temperature difference between adjacent vertical grids, lateral heat flux density, and local convection rate are calculated and combined to generate a layered thermal field distribution map. After applying a temperature step perturbation to each grid, the time delay of the perturbation reaching adjacent grids is recorded to generate a spatial propagation delay matrix.

[0071] Table 1. Raw data from multi-source sensing and Z-score normalization results (typical nodes) Table 2. Typical raster pair spatial propagation delay matrix (partial) The lateral propagation delay of the heat accumulation layer on the top of the cabinet (nodes 1 to 4) is only 18 seconds, while the delay of the penetration path through the shielding channel (nodes 4 to 5) is 137 seconds, a difference of about 7.6 times, which reflects the lag characteristic of heat propagation in dead corner areas.

[0072] Step 2: Analyze the spatial occlusion relationship between sensing channels and generate an occlusion coefficient matrix; Based on sensor coordinates and cabinet geometry parameters, the line-of-sight obstruction ratio between each sensor pair is calculated. Infrared thermal imager 1 covers areas A and B of transformers, while infrared thermal imager 2 covers areas C of transformers and the switchgear. There is cabinet obstruction between the rear aisle node 4 and infrared thermal imager 1, with the obstructed area accounting for 83% of the line-of-sight cross-section. After deducting the contribution of cabinet surface reflection (approximately 9%), the final obstruction coefficient is calculated. The value is 0.74. The path from node 4 to node 5 is in the same obstruction channel behind the cabinet. The obstruction is the side wall of the switch cabinet, accounting for 79% of the obstruction area. After deducting the 5% reflection compensation, the obstruction coefficient is... It is 0.74.

[0073] Table 3. Typical sensor occlusion coefficient matrix (partial) Step 3: Calculate the expected measurement sequence for each sensor channel based on the graph regularized constrained environmental field interpolation algorithm; Using 72 grid cells as graph nodes, graph edges are established between adjacent grid cells. Taking the calculation of the graph edge weight between node 1 and node 4 as an example, the normalized propagation delay is substituted. Occlusion coefficient , : The graph edge weights from node 4 to node 5 (the occlusion channel penetration path) are substituted into... , , : The edge weights of nodes 1 to 4 (4.17) are much greater than those of nodes 4 to 5 (0.35), resulting in strong lateral coupling of the heat accumulation layer and weak permeation coupling of the shading channel during the interpolation process, which is consistent with the actual heat propagation characteristics.

[0074] Based on the Tulaplace matrix Given the sensor measurements, solve for the objective function that minimizes the given values. The interpolated temperature and humidity field values ​​of each grid node are obtained, and then combined with the installation height and orientation parameters of each sensor, the desired measurement sequence is output.

[0075] Table 4 Comparison of expected measurement sequences and original measurement values ​​for typical nodes The expected temperature at node 4 (54.8℃) is higher than the original reading (51.9℃), indicating that the actual heat accumulation in this dead zone is higher than the level reflected by the direct reading of the point sensor. The difference comes from the spatial non-uniformity caused by occlusion, rather than sensor drift.

[0076] Step 4: Differentiate between spatial non-uniformity and sensor drift, and output micro-environment correction sensing data; Calculate the residual sequence associated with each node location and perform matching analysis with the historical aging curve.

[0077] Table 5. Location-related residual analysis and differentiation results The residuals of nodes 2 and 3 show a monotonically increasing trend and have a correlation coefficient of less than 0.3 with the residuals of other nodes in the same region. This is determined to be sensor aging drift, and the corresponding compensation parameters are subtracted from the original sequence to output microenvironment correction sensing data. The residuals of nodes 4 and 5 are highly correlated with the residuals of nodes in the same region (correlation coefficient exceeds 0.75), which is determined to be differences in the real environment caused by spatial inhomogeneity, and no compensation is performed.

[0078] Step 5: Extract the dynamic characteristics of thermal anomaly evolution for each device; The temperature and humidity sequences and real-time load current of the grids containing transformers A, B, and C in the microenvironment correction sensing data are input into their respective thermodynamic state space models. The residual vector between the actual thermal state vector and the predicted thermal state vector is calculated. The motion direction, acceleration, and curvature features are extracted from the residual vector sequence within the sliding window (window length is 10 sampling periods, i.e., 5 minutes), and Z-score standardization is performed.

[0079] Table 6. Dynamic characteristics of thermal anomaly evolution for each device (current moment) The residual acceleration of transformer A (1.83) is significantly higher than that of transformers B and C, indicating that its thermal state is deviating from the model prediction trajectory at an accelerated rate, and it has early thermal anomaly characteristics.

[0080] Step 6: Calculate the dynamic thermal coupling propagation coefficient matrix of environmental perception; Based on the spatial propagation time delay matrix and the hierarchical thermal field distribution map, the actual heat propagation path between each pair of devices is identified, the equivalent heat conduction time constant is calculated, and after taking the reciprocal, mean normalization based on the range is performed to generate the environmental perception dynamic thermal coupling propagation coefficient matrix.

[0081] Table 7. Equivalent heat conduction time constant and dynamic thermal coupling propagation coefficient of the equipment The thermal coupling propagation coefficient from transformer A to transformer B (0.87) is much higher than that from transformer A to the cable trench opening (0.12), reflecting that the thermal propagation efficiency of the heat accumulation layer lateral path is significantly better than that of the shielding channel penetration path.

[0082] Step 7: Recursively predict the thermodynamic state evolution trajectory of each device based on the spatiotemporal graph convolutional cascade prediction network; Using transformers A, B, and C as graph nodes and the thermal coupling propagation coefficients in Table 7 as edge weights, the dynamic characteristics of thermal anomaly evolution of each device are input into a spatiotemporal graph convolutional cascaded prediction network. After the spatial graph convolutional unit aggregates the thermal state information of neighboring nodes, the temporally gated recurrent unit predicts the state evolution over the next 12 time steps (6 minutes) along the time dimension, and the cascaded inference unit outputs the time series of thermal runaway probability of each device and the cascaded propagation path.

[0083] Table 8. Time series prediction results of thermal runaway probability for each device (next 6 minutes) When transformer A's probability first exceeds the preset critical probability threshold of 0.5 at +2 minutes, and transformer B exceeds the threshold between +4 and +5 minutes, the cascaded path inference layer determines the propagation order according to the edge weight from largest to smallest as follows: transformer A → transformer B → transformer C.

[0084] Step 8: Generate a comprehensive environmental safety monitoring report for the power distribution room; Based on probabilistic time series, the estimated time for transformer A to reach the critical state of thermal runaway is the current time + 2 minutes, with an intervention window of 2 minutes; for transformer B, it is + 4 minutes, with an intervention window of 4 minutes. Condensation risk indicators (dew point difference in area 5 is 3.2℃, indicating condensation risk), SF6 local gas concentration indicators (concentration at the cable trench opening is 1180ppm, exceeding the alarm threshold), and hotspot temperature rise indicators (temperature rise rate at the top of transformer A cabinet is 0.38℃ / min) are calculated using microenvironmental correction and sensing data.

[0085] Transformer A’s intervention window (2 minutes) is shorter than the first time threshold (5 minutes), and the number of devices covered by the cascade propagation path (3 units) exceeds the first quantity threshold (2 units), triggering the highest alarm level.

[0086] Table 9 Core Contents of the Comprehensive Environmental Safety Monitoring Report for the Power Distribution Room The entire data stream originates from raw readings of 12 sensor nodes. Z-score normalization and rasterization generate a layered thermal field distribution map and a spatial propagation delay matrix. Then, occlusion coefficient matrix calculation and graph regularization constraint interpolation yield the desired measurement sequence. Location-related residual analysis removes the aging drift (+1.3℃ and +0.8℃) of nodes 2 and 3 from spatial inhomogeneity, outputting microenvironmentally corrected sensing data. The corrected data drives a thermodynamic state-space model to extract the early thermal anomaly dynamics of transformer A. Combined with a dynamic thermal coupling propagation coefficient matrix calculated based on the actual propagation path, a spatiotemporal graph convolutional cascade prediction network outputs a top-level warning report containing the propagation path and remaining intervention time two minutes before thermal runaway occurs, achieving a complete transformation from raw sensing data to interventionable warning information.

[0087] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent sensing-based environmental safety monitoring in a power distribution room, characterized in that, include: Acquire multi-source sensing data from the power distribution room, and perform rasterization processing on the power distribution room space based on sensor coordinates and cabinet layout to generate a layered thermal field distribution map and a spatial propagation delay matrix; Calculate the occlusion coefficient matrix between each sensing channel based on sensor coordinates and cabinet geometric parameters; The layered thermal field distribution map, spatial propagation delay matrix, and occlusion coefficient matrix are input into the graph regularized constrained environment field interpolation algorithm, which outputs the expected measurement sequence of each sensor channel. Based on the positional correlation residual between the expected measurement sequence and the original sensor sequence, spatial non-uniformity and sensor drift are distinguished, and micro-environment correction sensing data is output. Based on the microenvironment correction and sensing data, the thermal anomaly evolution dynamic characteristics of each device are extracted, and the environmental sensing dynamic thermal coupling propagation coefficient matrix is ​​calculated based on the spatial propagation delay matrix and the layered thermal field distribution map. The dynamic characteristics of thermal anomaly evolution and the dynamic thermal coupling propagation coefficient matrix of environmental perception are input into the spatiotemporal graph convolutional cascade prediction network, which outputs the probability time series of each device entering the critical state of thermal runaway and the cascade propagation path.

2. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 1, characterized in that, The graph regularized constrained environment field interpolation algorithm includes: Based on the spatial coordinates of each grid cell in the hierarchical thermal field distribution map, the power distribution room space is represented as an undirected graph structure, with each grid cell serving as a graph node and graph edges established between adjacent grid cells. The weights of each graph edge are calculated based on the propagation delay values ​​between each node pair in the spatial propagation delay matrix and the corresponding element values ​​in the occlusion coefficient matrix. The graph edge weights are inversely proportional to the propagation delay between node pairs and inversely proportional to the occlusion coefficient. The graph Laplacian matrix is ​​generated based on the graph edge weights. A graph regularization constraint term is applied to the column vector formed by the temperature and humidity field values ​​of each node in the interpolation region. The graph regularization constraint term requires that nodes that are closely connected in space have similar temperature and humidity field values.

3. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 2, characterized in that, The graph regularized constrained environment field interpolation algorithm also includes: Using known sensor measurements as observation constraints and the convective flux field as boundary flux constraints, the objective function is minimized by combining the graph regularization constraint term. The minimized objective function is composed of a weighted combination of the observation error term, the graph regularization constraint term, and the boundary flux constraint term. The interpolated temperature and humidity field values ​​of each grid node are obtained by taking the derivative of the minimization objective function with respect to the column vector of temperature and humidity field values ​​and setting the derivative to zero. Based on the interpolated temperature and humidity field values ​​of the grid nodes where each sensor is located, and combined with the sensor's installation height and orientation parameters, the expected measurement sequence of each sensor channel is calculated.

4. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 1, characterized in that, The distinction between spatial non-uniformity and sensor drift includes: Calculate the time-by-time difference between the desired measurement sequence and the original sensor sequence to generate a position-correlated residual sequence; The location-related residual sequence is matched with the historical aging curves of each sensor. If the location-related residual shows a monotonically slow change trend consistent with the historical aging curve, the residual component is identified as sensor aging drift. If the location-related residual shows fluctuation characteristics related to the environmental load cycle or there is a consistent residual pattern between spatially adjacent sensors, the residual component is identified as environmental differences caused by spatial inhomogeneity. The aging drift component of the sensor is extracted as a location correlation compensation parameter. The location correlation compensation parameter is subtracted from the original sensor sequence to output microenvironment correction sensing data.

5. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 4, characterized in that, The matching analysis also includes spatial consistency verification: For multiple sensors within the same grid area, if the correlation coefficient between their location-related residual sequences exceeds a preset threshold, it is determined that the residual in that area is mainly caused by spatial inhomogeneity; if the correlation between the residual sequence of a certain sensor and the residual sequences of other sensors in the same area is lower than the preset threshold, it is determined that the sensor has individual drift.

6. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 1, characterized in that, The extracted thermal anomaly evolution dynamics characteristics of each device include: The temperature and humidity sequences of the grid where each device is located in the microenvironment correction sensing data and the real-time load current data of each device are input into the thermodynamic state space model of the corresponding device. The thermodynamic state space model is a linear state space equation pre-calibrated based on the device's heat capacity, heat dissipation coefficient and load heating power parameters. Calculate the residual vector between the current actual thermal state vector of each device and the model-predicted thermal state vector; Sliding window trajectory analysis is performed on the residual vector sequence. Within each sliding window, the motion direction, acceleration, and curvature features of the residual offset vector in the state space are extracted and combined to generate the thermal anomaly evolution dynamics features of each device.

7. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 1, characterized in that, The computational environment-aware dynamic thermal coupling propagation coefficient matrix includes: Based on the convection channel distribution in the layered thermal field distribution map, the actual heat propagation path between each device is identified. The actual heat propagation path includes the lateral propagation path along the heat accumulation layer and the penetration path along the shielding channel. Based on the spatial propagation delay matrix and the occlusion coefficient matrix, the equivalent heat conduction time constant of each device along the actual heat propagation path is calculated. After taking the reciprocal of the equivalent heat conduction time constant between each pair of devices, mean normalization based on the range is performed, and the reciprocal values ​​of each pair of devices are mapped to the interval between 0 and 1 to generate the dynamic thermal coupling propagation coefficient matrix of environmental perception. The dynamic thermal coupling propagation coefficient matrix of the environment perception is dynamically adjusted as the micro-environment correction perception data is updated.

8. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 1, characterized in that, The spatiotemporal graph convolutional cascaded prediction network uses devices as nodes and environmental perception dynamic thermal coupling propagation coefficients as edge weights, and includes spatial graph convolutional units, temporal gated recurrent units, and cascaded inference units. The spatial graph convolution unit uses the current thermal anomaly evolution dynamics of each device as node features and the environmental perception dynamic thermal coupling propagation coefficient matrix as adjacency weights to perform graph convolution operations, and outputs a device state representation vector that integrates spatial coupling relationships. The time-gated loop unit models the device state representation vector sequence along the time dimension and outputs the predicted state vector sequence of each device at multiple future time steps. The cascaded inference unit includes a thermal runaway probability output layer and a cascaded path inference layer. The thermal runaway probability output layer maps the predicted state vector to the probability value of each device entering the thermal runaway critical state at the corresponding time step. The cascaded path inference layer determines the propagation order of thermal anomalies among devices based on edge weights and the order in which each device reaches the critical state.

9. The intelligent sensing method for monitoring the environmental safety of a power distribution room according to claim 1, characterized in that, This also includes generating a comprehensive environmental safety monitoring report for the power distribution room: Based on the cascaded propagation path and the thermal runaway probability time series of each device, the interventionable time window of each node device on the propagation link is calculated. The interventionable time window is the time difference between the current moment and the time step corresponding to the first time when the probability value in the probability time series exceeds the preset critical probability threshold. Local environmental security alarm levels are divided based on the length of the interventionable time window and the coverage of the cascading propagation path. The highest alarm level is generated when the interventionable time window is shorter than the first time threshold and the number of devices covered by the cascading propagation path exceeds the first quantity threshold. The medium alarm level is generated when the interventionable time window is between the first time threshold and the second time threshold. The general alarm level is generated when the interventionable time window is longer than the second time threshold.

10. A smart sensing power distribution room environmental safety monitoring system, used to execute the smart sensing power distribution room environmental safety monitoring method according to any one of claims 1 to 9, characterized in that, include: The layered thermal field generation module is used to acquire multi-source sensing data of the power distribution room, and to perform rasterization processing on the power distribution room space based on sensor coordinates and cabinet layout to generate a layered thermal field distribution map and a spatial propagation delay matrix. The occlusion analysis module is used to calculate the occlusion coefficient matrix between each sensing channel based on sensor coordinates and cabinet geometric parameters. The environmental field interpolation module is used to input the layered thermal field distribution map, spatial propagation delay matrix and occlusion coefficient matrix into the graph regularized constrained environmental field interpolation algorithm, and output the expected measurement sequence of each sensor channel; The drift differentiation and correction module is used to differentiate spatial non-uniformity and sensor drift based on the position correlation residual between the expected measurement sequence and the original sensor sequence, and outputs micro-environment correction sensing data. The feature extraction module is used to extract the thermal anomaly evolution dynamics features of each device based on microenvironment correction sensing data; The coupling coefficient calculation module is used to calculate the dynamic thermal coupling propagation coefficient matrix of environmental perception based on the spatial propagation delay matrix and the hierarchical thermal field distribution map. The cascaded prediction module is used to input the dynamic characteristics of thermal anomaly evolution and the dynamic thermal coupling propagation coefficient matrix of environmental perception into the spatiotemporal graph convolutional cascaded prediction network, and output the probability time series of each device entering the critical state of thermal runaway and the cascaded propagation path.

Images