Low-orbit satellite multi-source sensing disaster monitoring method for power system

Abstract

The invention is suitable for the technical field of disaster monitoring, and provides a power system-oriented low-orbit satellite multi-source sensing disaster monitoring method, which comprises the following steps of: acquiring heterogeneous observation data and time sequence monitoring data, mapping the acquired data into a space-time diagram structure, and generating a multi-dimensional sensing field of a surrounding environment of power equipment; the method comprises the following steps: analyzing disaster characteristics at a low-orbit satellite end through a lightweight federated learning model based on a multi-dimensional sensing field, generating a semantic instruction which can be executed by a machine, distributing the instruction to an unmanned aerial vehicle cluster and an edge computing node through an inter-satellite link, and triggering a self-adaptive observation strategy; predicting a potential disaster chain reaction based on a pre-trained causal inference model; when a disaster chain triggers a threshold value, hierarchical response is activated autonomously; and a self-adaptive monitoring strategy configuration file is generated according to a disaster response result and is used for task planning of a next monitoring period, so that the early warning precision and response timeliness of disasters such as mountain fire and flood are improved, and meanwhile, the stability of communication and power grid operation in an extreme environment is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of disaster monitoring, and in particular to a low-orbit satellite multi-source sensing disaster monitoring method for power systems. Background Art

[0002] Compared to traditional power systems, new power systems operate in a more complex environment and rely significantly more heavily on monitoring and communication networks. This poses significant challenges to grid stability, particularly during extreme natural disasters or emergencies. Natural disasters can directly lead to the collapse of transmission lines, damage to substation equipment, and disruption to distribution facilities, triggering widespread power outages and even secondary disasters. In extreme cases, large-scale grid failures often have consequences beyond a brief interruption in power supply. They can also lead to the loss of power to communication base stations, the inability to operate automated dispatch systems, and the paralysis of critical infrastructure, exacerbating socioeconomic losses.

[0003] Current power system disaster monitoring relies primarily on ground-based sensor networks and drone inspections. Ground-based sensors are primarily deployed on transmission lines, substations, and distribution network facilities to monitor equipment operating status and environmental changes. However, their spatial coverage is limited, and the sensor equipment itself may be damaged or lose communication capabilities in extreme disaster situations, resulting in data transmission interruptions. Although drone inspections offer flexibility and high-resolution monitoring capabilities, their inspection range is limited by their endurance, flight environment, and communication conditions. After a large-scale disaster, it is difficult to cover all affected areas in a short period of time. Therefore, the current power grid disaster monitoring system still has significant deficiencies in terms of coverage, data acquisition continuity, and emergency response efficiency, making it difficult to meet the global disaster monitoring needs under extreme conditions. Summary of the Invention

[0004] The present application provides a low-orbit satellite multi-source sensing disaster monitoring method for power systems, which is used to solve the problem that existing monitoring methods are difficult to meet the global disaster monitoring needs under extreme conditions.

[0005] This application provides a low-orbit satellite multi-source sensing disaster monitoring method for power systems, including:

[0006] Obtain heterogeneous observation data from polar-orbit and inclined-orbit satellites, and collect time-series monitoring data based on drones and ground sensors;

[0007] Mapping the heterogeneous observation data and time series monitoring data into a spatiotemporal graph structure to generate a multi-dimensional perception field of the surrounding environment of the power equipment;

[0008] Based on the multi-dimensional perception field, a lightweight federated learning model is used on the low-orbit satellite to analyze disaster characteristics, generate machine-executable semantic instructions, and distribute the instructions to the drone cluster and edge computing nodes through inter-satellite links to trigger an adaptive observation strategy.

[0009] Based on a pre-trained causal inference model, the causal relationship chain between disaster triggering factors and power equipment status in the multi-dimensional perception field is analyzed to predict potential disaster chain reactions; when the disaster chain triggers a threshold, a hierarchical response is automatically activated;

[0010] Generate an adaptive monitoring strategy profile based on the disaster response results for mission planning in the next monitoring cycle.

[0011] Furthermore, the mapping of the heterogeneous observation data and time series monitoring data into a spatiotemporal graph structure to generate a multi-dimensional perception field of the surrounding environment of the power equipment includes:

[0012] Dynamically correct the timestamps of multi-source data through the Kalman filter model, and align and optimize the spatial coordinate systems of multi-source data based on the nonlinear least squares method;

[0013] High-priority data is screened through data quality assessment formulas, spatial correlation features of multi-source data are extracted using feature matching algorithms, and disaster evolution paths are predicted based on hybrid time series models to generate a multi-dimensional perception field that includes equipment status, environmental risk heat maps, and disaster chain probabilities.

[0014] Furthermore, based on the multi-dimensional perception field, a lightweight federated learning model is used on the low-orbit satellite to analyze disaster characteristics, generate machine-executable semantic instructions, and distribute the instructions to the drone cluster and edge computing nodes through the inter-satellite link to trigger an adaptive observation strategy, including:

[0015] The knowledge distillation technique is used to compress the pre-trained disaster identification benchmark model into a lightweight model suitable for spaceborne edge computing.

[0016] A distributed training network is built based on inter-satellite communication. Each satellite node uploads local model gradients through a differential privacy protection mechanism, and the central aggregation server updates the global model parameters.

[0017] The spatiotemporal graph data of the multi-dimensional perception field is input, the spatiotemporal graph convolutional network extracts the device status features, the visual transformer extracts the disaster abnormal area features from the satellite image, and the fully connected layer fuses the device status features and the disaster abnormal area features;

[0018] The fused feature vector is mapped into a semantic instruction encoding space, and a machine-parseable instruction label is output.

[0019] Furthermore, the multi-dimensional perception field is used to analyze disaster characteristics on the low-orbit satellite side through a lightweight federated learning model, generate machine-executable semantic instructions, and distribute the instructions to the drone cluster and edge computing nodes through inter-satellite links to trigger an adaptive observation strategy, which also includes:

[0020] The ST-GCN network is used to extract the spatiotemporal correlation characteristics of temperature and current fluctuations in power equipment, and the visual transformer is used to identify the fire spread contours and flooded areas in satellite images.

[0021] A disaster risk score is calculated based on the fused features, and a semantic encoder generates an instruction when the disaster risk score is greater than a preset threshold.

[0022] Furthermore, the multi-dimensional perception field is used to analyze disaster characteristics on the low-orbit satellite side through a lightweight federated learning model, generate machine-executable semantic instructions, and distribute the instructions to the drone cluster and edge computing nodes through inter-satellite links to trigger an adaptive observation strategy, which also includes:

[0023] A low-latency intersatellite communication protocol is used to encapsulate semantic instructions into priority data packets;

[0024] Send command copies to several neighboring satellites of the target UAV cluster through a multi-path transmission strategy to ensure that at least one link is successfully received;

[0025] After the instruction reaches the edge server, it triggers local data caching and preprocessing tasks.

[0026] Furthermore, the adaptive observation strategy includes:

[0027] Dynamically adjust satellite observation frequency and resolution based on disaster risk levels, prioritizing intensive monitoring in high-risk areas and reducing monitoring intensity in low-risk areas;

[0028] When planning drone inspection routes, they proactively avoid high-risk disaster areas and generate safe detour trajectories based on real-time environmental data.

[0029] Dynamically allocate data processing tasks based on the real-time load status of satellite-borne and edge nodes.

[0030] Furthermore, the pre-trained causal inference model includes:

[0031] Based on historical disaster data and power equipment status parameters, a dynamic causal graph is constructed with disaster triggering factors as nodes and causal relationship strength as directed edges.

[0032] Dynamic Bayesian networks are used to model the temporal dependency between disaster factors and equipment status, and variational inference is used to calculate conditional probability distributions and identify causal chains across time steps.

[0033] The spatiotemporal graph features of the multidimensional perception field are cross-modally aligned with the node embedding vectors of the causal graph to output the probabilistic propagation path of the disaster chain.

[0034] Furthermore, the pre-trained causal inference model analyzes the causal relationship chain between the disaster triggering factors and the power equipment status in the multi-dimensional perception field and predicts potential disaster chain reactions, including:

[0035] Starting from the currently detected disaster triggering factor node, traverse the directed edges in the causal graph and extract all potential propagation paths;

[0036] Calculate the propagation probability of each path based on dynamic Bayesian network;

[0037] According to the propagation probability and the equipment status deviation threshold, the disaster chain is divided into three risk levels: low risk, medium risk and high risk.

[0038] Furthermore, when the disaster chain triggers a threshold, the hierarchical response is autonomously activated, including:

[0039] The first-level response reconstructs the communication link through the satellite-borne edge node, prioritizes the allocation of a preset proportion of bandwidth to transmit substation damage and line break data, and initiates satellite revisit monitoring of the source area of ​​the disaster chain in seconds;

[0040] The second-level response sends encrypted coordinate instructions to the drone cluster, executes the equipment's infrared temperature measurement and image acquisition tasks, and the edge server starts data caching and abnormal current fast Fourier transform analysis;

[0041] The third-level response generates a grid topology isolation plan, which is sent to the substation control terminal via low-orbit satellite to disconnect the transmission lines associated with the disaster chain.

[0042] Furthermore, the generation of an adaptive monitoring strategy configuration file based on the disaster response results for task planning for the next monitoring cycle includes:

[0043] Analyze the actual blocking effect of the multi-level response strategy based on the disaster response execution log, including the reduction ratio of the disaster chain spread range, the timeliness of key equipment repair and the efficiency of resource consumption;

[0044] Adjust monitoring strategy parameters based on the assessment results, including satellite image revisit cycles, drone inspection path priorities, and grid isolation rule triggering thresholds;

[0045] The optimized strategy parameters are encapsulated into a structured configuration file and distributed to onboard edge nodes and ground control centers via low-orbit satellites to guide mission planning for the next monitoring cycle.

[0046] It can be seen from the above technical solutions that this application has the following advantages:

[0047] This application constructs a multi-dimensional perception field of the spatiotemporal map of the environment around power equipment by integrating multi-orbit satellite heterogeneous observation data and air-ground time-series monitoring data, realizing a holographic analysis of disaster characteristics; generates executable semantic instructions in real time based on the satellite-borne lightweight federated learning model, driving drones and edge nodes to dynamically adjust observation strategies; combines causal inference models to predict disaster chain reactions and trigger graded responses, significantly improving the early warning accuracy and blocking efficiency of disasters such as wildfires and floods; and finally optimizes the continuous iterative monitoring scheme through adaptive strategy optimization, reducing the disaster risk and operation and maintenance costs of the power system in complex environments, and realizing an intelligent, efficient, and autonomous disaster monitoring and emergency response system. This system can meet the global disaster monitoring needs under extreme conditions while improving response efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The figure is a flow chart of an embodiment of a method for multi-source sensing disaster monitoring of low-orbit satellites for power systems in the present invention. DETAILED DESCRIPTION

[0049] The terms "first," "second," "third," "fourth," etc. (if any) in the specification and claims of the present application and in the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential sequence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "corresponding to," and any variations thereof, are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or apparatus.

[0050] Example 1

[0051] The implementation method in this embodiment can be implemented in the system, can be implemented in the server, and can also be implemented in the terminal, and the specific implementation is not clearly limited. The following will introduce the low-orbit satellite multi-source perception disaster monitoring method for the power system in this application from the perspective of system implementation. Figure 1 , the method provided in the embodiment of the present application includes the following steps:

[0052] S11. Acquire heterogeneous observation data from polar-orbit and inclined-orbit satellites, and collect time-series monitoring data from drones and ground sensors.

[0053] Polar-orbiting satellites are deployed to acquire wide-area SAR imagery and multispectral data. Inclined-orbit satellites provide high-resolution infrared thermal imaging and intensive observation data for specific areas, allowing for real-time reception of heterogeneous observation data streams. Synchronously controlled drone swarms equipped with high-precision LiDAR and optical cameras perform periodic inspections along transmission corridors, collecting high-frequency video streams of equipment surface temperature, structural deformation, and the state of surrounding vegetation. Ground sensor networks are deployed at key nodes such as substations and transmission towers, continuously collecting time-series data on environmental parameters (temperature, humidity, wind speed) and equipment operating status (current, voltage, vibration). GPS timing modules are used to unify the time base of multi-source data, combined with Kalman filtering to dynamically correct clock drift. Spatial registration algorithms are used to eliminate coordinate system differences, ultimately constructing a spatiotemporally aligned heterogeneous dataset, providing highly consistent input data for the generation of multidimensional perception fields.

[0054] S12. Mapping heterogeneous observation data and time-series monitoring data into a spatiotemporal graph structure to generate a multidimensional perception field of the environment surrounding the power equipment;

[0055] In this embodiment, generating a multi-dimensional perception field includes:

[0056] 1. Dynamically correct the timestamps of multi-source data using the Kalman filter model, and simultaneously optimize the alignment of the spatial coordinate systems of multi-source data using the nonlinear least squares method;

[0057] To address the time inconsistency issue of data collected by sensors in multi-source disaster monitoring of power systems, the timestamps of different data sources are dynamically corrected to align them to a unified time base. This can be done using the following Kalman filter formula:

[0058] X k =AX k―1 +Bu k +W k ,Z k =HX k +V k

[0059] Where: X k and X k―1 are the estimated values ​​of the time offset at the current moment and the next moment, respectively, which characterize the deviation between the sensor clock and the unified time reference. A is the state transfer matrix, which describes the dynamic change law of the time offset. Bu k As the control input, the calibration value is provided by GPS timing signal or satellite-borne atomic clock, W k is the process noise, random error of analog clock crystal drift or network transmission delay, Z k is the observed value, that is, the timestamp error actually reported by each sensor, H is the observation matrix, mapping the state variables to the observation space, V kis the observation noise, which represents the timestamp measurement uncertainty. By iteratively updating the state estimation, the time synchronization of multi-source data is achieved.

[0060] The nonlinear least squares method is used to optimize spatial registration, and the objective function is:

[0061]

[0062] Where: R is the rotation matrix, which corrects the viewing angle deviation between the UAV LiDAR point cloud and the satellite image; t is the translation vector, which eliminates the offset between the ground sensor data and the power grid GIS coordinate system; p i is the coordinate of the matching point of satellite image or UAV LiDAR, q i are the coordinates of the corresponding points on the ground sensor or GIS basemap. The Levenberg-Marquardt algorithm is used to iteratively optimize R and t until the error converges, achieving spatial consistency among multi-source data.

[0063] 2. Filter high-priority data through data quality assessment formulas, use feature matching algorithms to extract spatial correlation features of multi-source data, and predict disaster evolution paths based on hybrid time series models to generate a multi-dimensional perception field that includes equipment status, environmental risk heat maps, and disaster chain probabilities.

[0064] Quantify the value of data through the priority formula:

[0065] B i =α i U(D i )+β i L(D i )+γ i S(D i )

[0066] Where: U(D i ) is the data urgency, such as the sudden temperature rise of the equipment is marked as 0.9, and the historical data is marked as 0.2. i ) is the spatial resolution, such as 1.0 for satellite images with a resolution of 0.5 meters and 0.3 for ground sensors with a resolution of 10 meters. i ) is the signal-to-noise ratio, normalized to the range of 0 to 1, α i , β i , γ i is the weight coefficient, which is set according to the historical disaster impact analysis, such as α i =0.6,β i =0.3,γ i =0.1), filter priority B i High-value data ≥0.7 enters subsequent processing.

[0067] The SIFT algorithm is used to match key points between satellite imagery and drone LiDAR data. Spatial topological relationships are then constructed based on ground sensor locations to extract spatial correlation features. Based on the LSTM-Transformer hybrid time series model architecture, time series data (device temperature and current fluctuations) and spatial features are input to predict disaster evolution paths. The LSTM layer of the LSTM-Transformer hybrid time series model captures short-term temporal dependencies in device states, such as temperature trends; the Transformer layer models long-term cross-regional correlations, such as the causal relationship between wildfire spread and power line failure. The final output is structured data containing the following layers: a device status layer (device health score, ranging from 0 to 1); an environmental risk heat map (risk level of surface temperature and flood water level, low / medium / high); and a disaster chain probability map (probability of wildfire-induced power grid failure propagation, ranging from 0% to 100%).

[0068] The above steps solve the consistency problem of multi-source heterogeneous data through spatiotemporal alignment and data screening; combine feature matching with hybrid time series models to achieve accurate prediction of disaster evolution and provide a holographic perception basis for graded response.

[0069] S13. Based on the multi-dimensional perception field, a lightweight federated learning model is used on the low-orbit satellite to analyze disaster characteristics, generate machine-executable semantic instructions, and distribute the instructions to the drone cluster and edge computing nodes via inter-satellite links to trigger adaptive observation strategies.

[0070] In this embodiment, a lightweight federated learning model is used to efficiently analyze disaster characteristics at the onboard edge. The knowledge distillation technology is used to compress the pre-trained disaster identification benchmark model (based on the ResNet-Transformer hybrid architecture) into a lightweight model, including the following:

[0071] 1. Using knowledge distillation technology to compress the pre-trained disaster identification benchmark model into a lightweight model suitable for spaceborne edge computing;

[0072] 2. Build a distributed training network based on inter-satellite communication. Each satellite node uploads local model gradients through a differential privacy protection mechanism, and a central aggregation server updates the global model parameters.

[0073] 3. Input the spatiotemporal graph data of the multi-dimensional perception field, the spatiotemporal graph convolutional network extracts the device status features, the visual transformer extracts the disaster anomaly area features from the satellite image, and the fully connected layer fuses the device status features and the disaster anomaly area features;

[0074] 4. Map the fused feature vector into a semantic instruction encoding space and output a machine-parseable instruction label.

[0075] Specifically, knowledge distillation: Through a teacher-student model framework, knowledge from the teacher model is transferred to the student model, preserving key feature extraction capabilities and reducing model size. A federated learning framework: A distributed training network based on intersatellite communication is constructed. When each satellite node trains the model locally, gradient data is protected through differential privacy (adding Laplace noise with a noise scale of 0.1). A central server aggregates global parameters, with an update cycle of every 24 hours.

[0076] Spatiotemporal Graph Convolutional Network (ST-GCN): Inputs spatiotemporal graph data of a multi-dimensional perception field and extracts spatiotemporal correlation features of temperature and current fluctuations of power equipment (the time window is set to 10 minutes, and the spatial adjacency radius is 500 meters). Vision Transformer: Processes satellite images in blocks (block size 16×16 pixels) and uses a self-attention mechanism to identify fire contours and flooded areas. Fully Connected Layer Fusion: Concatenates the ST-GCN output (dimension 256) with the Vision Transformer output (dimension 512), and reduces the dimension to a 128-dimensional feature vector through a fully connected layer. Semantic Encoder: Uses a multi-layer perceptron (MLP) to map the fused features to a semantic instruction encoding space, outputting instruction labels such as "fire source coordinates (35.2N, 118.8E), risk level III."

[0077] In this embodiment, disaster feature analysis and instruction generation are based on quantifying disaster risk based on fused features. The triggering conditions for instruction generation include the following:

[0078] 1. The ST-GCN network is used to extract the spatiotemporal correlation characteristics of temperature and current fluctuations in power equipment, and the visual transformer is used to identify fire spread contours and flooded areas in satellite images.

[0079] 2. Calculate the disaster risk score based on the fused features, and generate instructions when the disaster risk score is greater than the preset threshold.

[0080] Specifically, the ST-GCN network here extracts the spatiotemporal dependency of the device status by inputting the device temperature and current time series data, and outputs a feature dimension of 256. The visual transformer identifies the fire spread outline and the flooded area by inputting satellite images. When the risk score is greater than 0.7, that is, the high risk threshold, the trigger instruction is generated. Semantic instructions include observation instructions and emergency instructions, such as "starting the second-level revisit of the SAR image of the coordinate (35.2N, 118.8E) area" is an observation instruction, and "the drone cluster goes to the coordinate (35.2N, 118.8E) and performs infrared temperature measurement of the equipment (frequency 10Hz)" is an emergency instruction.

[0081] In this embodiment, ISL command distribution ensures that commands are efficiently and reliably transmitted to the target node, including the following:

[0082] 1. Using a low-latency intersatellite communication protocol to encapsulate semantic instructions into priority data packets;

[0083] 2. Send command copies to several neighboring satellites of the target drone cluster through a multi-path transmission strategy to ensure that at least one link is successfully received;

[0084] 3. After the instruction reaches the edge server, it triggers local data caching and preprocessing tasks.

[0085] The low-latency communication protocol uses SpaceWire-C, with data packets encapsulated in a header and payload format. Multipath redundant transmission involves sending command copies to three neighboring satellites in the target drone cluster. The link selection strategy prioritizes satellites with an orbital altitude difference of ≤100 km. If no confirmation signal is received within 10 seconds, the transmission is automatically retransmitted. Edge node tasks are triggered by the edge server upon command arrival. High-priority tasks involve real-time FFT analysis of current data (with a 1024-point window and 50% overlap); low-priority tasks involve downsampling environmental data to 1Hz for storage in HDF5 format.

[0086] In this embodiment, the adaptive observation strategy execution dynamically optimizes resource allocation and task execution, including the following:

[0087] 1. Dynamically adjust satellite observation frequency and resolution based on disaster risk levels, prioritizing intensive monitoring in high-risk areas and reducing monitoring intensity in low-risk areas;

[0088] 2. When planning drone inspection routes, the system proactively avoids high-risk areas and generates safe detour trajectories based on real-time environmental data.

[0089] 3. Dynamically allocate data processing tasks based on the real-time load status of the satellite and edge nodes.

[0090] Specifically, satellite observations in high-risk areas will be adjusted to maintain a SAR image revisit period of ≤30 seconds and infrared resolution increased to 0.5 meters. Satellite observations in low-risk areas will be reduced to hourly monitoring with a resolution of 5 meters. Drone path planning includes Bezier curve obstacle avoidance, whereby path control points avoid areas listed in the risk heat map. Furthermore, priority is set for paths within a 500-meter radius of the fire source, with a weight of 0.9 and for paths within the general area of ​​the fire source, a weight of 0.3. The load balancing strategy involves proportionally offloading tasks to edge servers when the satellite CPU utilization rate is ≥80%. Finally, the onboard and edge node load status is collected every 5 minutes to update the task allocation table.

[0091] The above steps achieve on-board real-time analysis through lightweight models, combined with redundant communications and dynamic strategies to ensure the accurate generation and efficient execution of disaster commands.

[0092] S14. Based on a pre-trained causal inference model, analyze the causal relationship chain between disaster triggering factors and power equipment status in a multi-dimensional perception field, and predict potential disaster chain reactions. When the disaster chain triggers the threshold, autonomously activate a hierarchical response.

[0093] In this embodiment, the pre-trained causal inference model includes:

[0094] 1. Based on historical disaster data and power equipment status parameters, a dynamic causal graph is constructed with disaster triggering factors as nodes and causal relationship strength as directed edges;

[0095] 2. Use dynamic Bayesian networks to model the temporal dependency between disaster factors and equipment status, calculate conditional probability distributions through variational inference, and identify causal chains across time steps;

[0096] 3. Cross-modally align the spatiotemporal graph features of the multidimensional perception field with the causal graph node embedding vectors to output the probabilistic propagation path of the disaster chain.

[0097] Specifically, the nodes in the dynamic causal graph are defined as disaster triggers, including surface temperature anomalies (>50°C), vegetation moisture content (<30%), and current fluctuation amplitude (>±10% of rated value). Power equipment status parameters include transformer temperature and transmission line vibration amplitude. Edge weights are based on historical disaster data, using Granger causality tests to quantify the strength of causal relationships between factors. For example, the causal strength of surface temperature anomalies on wildfire risk is [missing value]. Furthermore, the edge weights in the causal graph are updated every 24 hours based on the latest monitoring data.

[0098] The hidden variables in the dynamic Bayesian network structure represent the state of the disaster chain (e.g., "wildfire triggered"), while the observed variables represent the device states and environmental parameters of the multidimensional sensory field. By maximizing the evidence lower bound (ELBO), the posterior probability distribution is approximated for variational inference. Cross-modal alignment uses a graph attention network (GAT) to embed spatiotemporal graph features (e.g., the spatiotemporal distribution of device temperature) into a 128-dimensional vector. This embedding is then aligned with the node vectors in the causal graph using cosine similarity to output the disaster chain propagation path, such as "surface temperature anomaly → wildfire → power line failure."

[0099] In this embodiment, predicting potential disaster chain reactions includes the following:

[0100] 1. Starting from the currently detected disaster triggering factor node, traverse the directed edges in the causal graph and extract all potential propagation paths;

[0101] 2. Calculate the propagation probability of each path based on the dynamic Bayesian network;

[0102] 3. Based on the propagation probability and the equipment status deviation threshold, the disaster chain is divided into three risk levels: low risk, medium risk and high risk.

[0103] Specifically, starting from the currently detected trigger factor node, a depth-first search (DFS) is performed along the directed edges of the causal graph to extract all possible propagation paths. Based on the conditional probability chain rule of the dynamic Bayesian network, the path propagation probability is calculated. For example, the probability of the path "temperature anomaly → wildfire → line fuse" is P = P (wildfire | temperature anomaly) × P (line fuse | wildfire). Risk levels are divided into low risk, medium risk and high risk. If the low risk probability is less than 0.3, only logs are recorded and no active response is triggered; if the medium risk probability is greater than 0.3 and less than 0.7, the second-level response is initiated and equipment diagnosis is performed; if the high risk probability is greater than 0.7, the first-level response is activated immediately and resources are invested first.

[0104] In this embodiment, when a disaster chain triggers a threshold, a hierarchical response is automatically activated, including the following:

[0105] 1. Level 1 response: Reconstruct the communication link through the satellite-borne edge node, prioritize the allocation of a preset proportion of bandwidth to transmit substation damage and line break data, and initiate satellite revisit monitoring of the source area of ​​the disaster chain in seconds;

[0106] Level 1 response is a high-risk response. Communication reconstruction and priority scheduling classify data through the priority function K(O) = ε1T1(O) + ε2T2(O) + ε3T3(O), where: T1(O) is the severity of the disaster, such as substation damage is marked as 1.0 and equipment temperature abnormality is marked as 0.5. T2(O) is the data type, such as video stream weight 0.8 and sensor data weight 0.3. T3(O) is the bandwidth status, such as sufficient bandwidth weight 0.2 and insufficient bandwidth weight 0.7. Generate a priority queue Q = {O1, O2, ..., O n}, ensuring that high-priority data (K(O)≥0.7) is allocated 90% of the bandwidth first.

[0107] Intelligent link switching is based on the objective function L opt =arg max(δ1Q link,i +δ2G i ―δ3Z i ) Select the optimal link, where: Q link,i Score the link quality, such as 0.9 for satellite link, 0.7 for terrestrial link, and 0.9 for G i For the available bandwidth, such as 100Mbps for satellite link, 50Mbps for terrestrial link, i For transmission delay, such as 200ms for satellite link and 50ms for ground link, with weights δ1 = 0.6, δ2 = 0.3, and δ3 = 0.1, dynamic switching to low-orbit satellite link to transmit key data ensures that the delay is ≤ 300ms.

[0108] 2. The second-level response sends encrypted coordinate instructions to the drone cluster, executes the equipment's infrared temperature measurement and image acquisition tasks, and the edge server initiates data caching and abnormal current fast Fourier transform analysis;

[0109] The second-level response is a medium-risk response, which uses reinforcement learning-driven data scheduling, that is, using a reinforcement learning framework:

[0110]

[0111] Where: state C1 = (bandwidth utilization, computing load, disaster risk level), action C2 = (data compression rate, transmission path, task priority), reward R i Based on the data transmission success rate and response timeliness, the data compression rate is dynamically adjusted through the Q-learning algorithm.

[0112] 3. The third-level response generates a grid topology isolation plan, which is sent to the substation control terminal via a low-orbit satellite to disconnect the transmission lines associated with the disaster chain.

[0113] The third-level response is a low-risk response. It combines the priority queue and link switching log to optimize the monitoring parameters for the next cycle. The satellite revisit period is shortened from 30 seconds to 20 seconds, and the drone path weight increases the avoidance coefficient of high-risk areas to 0.9.

[0114] The above steps embed data classification, reinforcement learning scheduling and intelligent link switching technologies into a hierarchical response strategy, which not only achieves accurate blocking of disaster chains but also optimizes communication resource allocation, which is significantly different from the static response mechanism of existing patents.

[0115] S15. Generate an adaptive monitoring strategy configuration file based on the disaster response results for task planning in the next monitoring cycle.

[0116] The effectiveness of the multi-level response strategy is assessed through quantitative analysis of key indicators in disaster response execution logs. These include the reduction in the disaster chain's spread, the average timeliness of repairing critical equipment, and the efficiency of satellite and drone resource consumption. Monitoring parameters are dynamically adjusted based on these results: the satellite image revisit period for high-risk areas is shortened to a preset lower limit (e.g., ≤ 20 seconds for SAR images), the drone inspection path priority algorithm is optimized, and the trigger threshold for power grid isolation rules is lowered, for example, from a risk probability of 0.7 to 0.6. The optimized parameters are then packaged into a structured configuration file (XML / JSON format) and broadcast to onboard edge nodes and ground control centers via the L-band communication link of the low-orbit satellite, driving the automated deployment of task planning for the next monitoring cycle.

[0117] The above-mentioned embodiments build a multidimensional perception field by fusing data from multi-orbit satellites and air-ground sensors. Combined with federated learning and causal inference models, they achieve precise prediction and intelligent blocking of disaster chains, significantly improving the accuracy of early warnings and the timeliness of responses to disasters such as wildfires and floods. A dynamic closed-loop optimization mechanism adaptively adjusts monitoring strategies and resource allocation, reducing power system operation and maintenance costs while ensuring the stability of communications and grid operations in extreme environments.

[0118] It is understandable that those skilled in the art can, under the guidance of the above embodiments, combine various implementation methods in the above embodiments to obtain technical solutions of multiple implementation methods.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A low-orbit satellite multi-source sensing disaster monitoring method for power systems, characterized in that: include: Obtain heterogeneous observation data from polar-orbit and inclined-orbit satellites, and collect time-series monitoring data based on drones and ground sensors; Mapping the heterogeneous observation data and time series monitoring data into a spatiotemporal graph structure to generate a multi-dimensional perception field of the surrounding environment of the power equipment; Based on the multi-dimensional perception field, a lightweight federated learning model is used on the low-orbit satellite to analyze disaster characteristics, generate machine-executable semantic instructions, and distribute the instructions to the drone cluster and edge computing nodes through inter-satellite links to trigger an adaptive observation strategy. Based on the pre-trained causal inference model, the causal relationship chain between the disaster triggering factors and the power equipment status in the multi-dimensional perception field is analyzed to predict the potential disaster chain reaction; When a disaster chain triggers a threshold, a hierarchical response is automatically activated; Generate an adaptive monitoring strategy profile based on the disaster response results for mission planning in the next monitoring cycle.

2. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 1 is characterized in that: Mapping the heterogeneous observation data and time series monitoring data into a spatiotemporal graph structure to generate a multi-dimensional perception field of the surrounding environment of the power equipment includes: Dynamically correct the timestamps of multi-source data through the Kalman filter model, and align and optimize the spatial coordinate systems of multi-source data based on the nonlinear least squares method; High-priority data is screened through data quality assessment formulas, spatial correlation features of multi-source data are extracted using feature matching algorithms, and disaster evolution paths are predicted based on hybrid time series models to generate a multi-dimensional perception field that includes equipment status, environmental risk heat maps, and disaster chain probabilities.

3. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 1 is characterized in that: The method analyzes disaster characteristics on the low-orbit satellite side through a lightweight federated learning model based on the multi-dimensional perception field, generates machine-executable semantic instructions, and distributes the instructions to the drone cluster and edge computing nodes through inter-satellite links to trigger an adaptive observation strategy, including: The knowledge distillation technique is used to compress the pre-trained disaster identification benchmark model into a lightweight model suitable for spaceborne edge computing. A distributed training network is built based on inter-satellite communication. Each satellite node uploads local model gradients through a differential privacy protection mechanism, and the central aggregation server updates the global model parameters. The spatiotemporal graph data of the multi-dimensional perception field is input, the spatiotemporal graph convolutional network extracts the device status features, the visual transformer extracts the disaster abnormal area features from the satellite image, and the fully connected layer fuses the device status features and the disaster abnormal area features; The fused feature vector is mapped into a semantic instruction encoding space, and a machine-parseable instruction label is output.

4. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 3 is characterized in that: The method further includes: analyzing disaster characteristics on the low-orbit satellite side through a lightweight federated learning model based on the multi-dimensional perception field, generating machine-executable semantic instructions, and distributing the instructions to the drone cluster and edge computing nodes through inter-satellite links to trigger an adaptive observation strategy. The ST-GCN network is used to extract the spatiotemporal correlation characteristics of temperature and current fluctuations in power equipment, and the visual transformer is used to identify the fire spread contours and flooded areas in satellite images. A disaster risk score is calculated based on the fused features, and a semantic encoder generates an instruction when the disaster risk score is greater than a preset threshold.

5. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 4 is characterized in that: The method further includes: analyzing disaster characteristics on the low-orbit satellite side through a lightweight federated learning model based on the multi-dimensional perception field, generating machine-executable semantic instructions, and distributing the instructions to the drone cluster and edge computing nodes through inter-satellite links to trigger an adaptive observation strategy. A low-latency intersatellite communication protocol is used to encapsulate semantic instructions into priority data packets; Send command copies to several neighboring satellites of the target UAV cluster through a multi-path transmission strategy to ensure that at least one link is successfully received; After the instruction reaches the edge server, it triggers local data caching and preprocessing tasks.

6. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 5 is characterized in that: The adaptive observation strategy includes: Dynamically adjust satellite observation frequency and resolution based on disaster risk levels, prioritizing intensive monitoring in high-risk areas and reducing monitoring intensity in low-risk areas; When planning drone inspection routes, they proactively avoid high-risk disaster areas and generate safe detour trajectories based on real-time environmental data. Dynamically allocate data processing tasks based on the real-time load status of satellite-borne and edge nodes.

7. The method for monitoring disasters using multi-source sensing of low-orbit satellites for power systems according to claim 1, characterized in that: The pre-trained causal inference model includes: Based on historical disaster data and power equipment status parameters, a dynamic causal graph is constructed with disaster triggering factors as nodes and causal relationship strength as directed edges. Dynamic Bayesian networks are used to model the temporal dependency between disaster factors and equipment status, and variational inference is used to calculate conditional probability distributions and identify causal chains across time steps. The spatiotemporal graph features of the multidimensional perception field are cross-modally aligned with the node embedding vectors of the causal graph to output the probabilistic propagation path of the disaster chain.

8. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 7 is characterized in that: The pre-trained causal inference model analyzes the causal relationship chain between the disaster triggering factors and the power equipment status in the multi-dimensional perception field and predicts potential disaster chain reactions, including: Starting from the currently detected disaster triggering factor node, traverse the directed edges in the causal graph and extract all potential propagation paths; Calculate the propagation probability of each path based on dynamic Bayesian network; According to the propagation probability and the equipment status deviation threshold, the disaster chain is divided into three risk levels: low risk, medium risk and high risk.

9. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 8, characterized in that: When the disaster chain triggers a threshold, the hierarchical response is automatically activated, including: The first-level response reconstructs the communication link through the satellite-borne edge node, prioritizes the allocation of a preset proportion of bandwidth to transmit substation damage and line break data, and initiates satellite revisit monitoring of the source area of ​​the disaster chain in seconds; The second-level response sends encrypted coordinate instructions to the drone cluster, executes the equipment's infrared temperature measurement and image acquisition tasks, and the edge server starts data caching and abnormal current fast Fourier transform analysis; The third-level response generates a grid topology isolation plan, which is sent to the substation control terminal via low-orbit satellite to disconnect the transmission lines associated with the disaster chain.

10. The method for multi-source sensing disaster monitoring of low-orbit satellites for power systems according to claim 1, characterized in that: The method of generating an adaptive monitoring strategy configuration file based on the disaster response results for use in task planning for the next monitoring cycle includes: Analyze the actual blocking effect of the multi-level response strategy based on the disaster response execution log, including the reduction ratio of the disaster chain spread range, the timeliness of key equipment repair and the efficiency of resource consumption; Adjust monitoring strategy parameters based on the assessment results, including satellite image revisit cycles, drone inspection path priorities, and grid isolation rule triggering thresholds; The optimized strategy parameters are encapsulated into a structured configuration file and distributed to onboard edge nodes and ground control centers via low-orbit satellites to guide mission planning for the next monitoring cycle.

Images