Power distribution system disaster risk early warning method based on low earth orbit satellite internet

Abstract

The invention is suitable for the technical field of disaster risk early warning, and provides a power distribution system disaster risk early warning method based on a low earth orbit satellite internet, and the method comprises the steps: obtaining a multi-dimensional observation data set, and carrying out the normalization processing; inputting the normalized data set into a preset hierarchical time sequence feature extraction model, and outputting a fused multi-dimensional feature vector; performing dynamic weight fusion on the multi-dimensional feature vectors, mapping the multi-dimensional feature vectors to a risk interval, and determining a disaster early warning signal according to a risk level result divided by a preset threshold value; determining a power distribution network node weight coefficient based on the disaster early warning signal, and constructing a multi-modal resource collaborative optimization decision model according to the power distribution network node weight coefficient; and solving the multi-modal resource collaborative optimization decision model based on a deep reinforcement learning algorithm to obtain an optimal resource deployment strategy, and issuing the optimal resource deployment strategy to a terminal for execution. According to the method, the real-time adaptive decision of resource deployment and power grid recovery can be realized while the early warning accuracy is improved, and the post-disaster recovery efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of disaster risk warning technology, and in particular to a disaster risk warning method for a power distribution system based on low-orbit satellite Internet. Background Art

[0002] With the increasing frequency of extreme disasters, power distribution systems are facing increasingly severe impacts. Traditional disaster warning methods often rely on a single data source, such as a weather station or ground-based sensor, making it difficult to fully determine the multi-dimensional impact of disasters on the power grid, such as the complex coupling relationship between meteorological conditions, geographical environment, and equipment status. In addition, existing data fusion technologies typically use shallow models or single-layer time series analysis, which cannot effectively extract the deep spatiotemporal characteristics of high-dimensional heterogeneous data, resulting in insufficient warning accuracy. In the post-disaster recovery phase, resource scheduling often uses static planning or heuristic algorithms, which makes it difficult to dynamically coordinate the deployment of multimodal resources such as satellites, drones, and emergency power vehicles with the real-time needs of distribution network restoration, which can easily lead to resource conflicts, excessive costs, or recovery delays.

[0003] In view of this, a distribution system disaster risk warning method based on low-orbit satellite Internet is needed. Summary of the Invention

[0004] This application provides a distribution system disaster risk warning method based on low-orbit satellite Internet, which is used to solve problems such as resource conflicts, excessive costs or recovery delays caused by insufficient warning accuracy.

[0005] This application provides a power distribution system disaster risk early warning method based on low-orbit satellite internet, including:

[0006] Acquire a multidimensional observation data set through a space-based remote sensing system, an air-based observation system, and a ground-based sensing system, and perform normalization processing on the multidimensional observation data set;

[0007] The normalized dataset is input into a preset hierarchical time series feature extraction model. A layer-by-layer LSTM network is used to extract cross-dimensional time series features of meteorological, geographical, and power grid equipment, and outputs a fused multidimensional feature vector.

[0008] Dynamically weighting the multidimensional feature vectors and mapping them to risk intervals, and determining disaster warning signals based on risk level results divided by preset thresholds;

[0009] Determine the weight coefficients of the distribution network nodes based on the disaster warning signal, and construct a multimodal resource collaborative optimization decision model based on the weight coefficients of the distribution network nodes; wherein the objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, and the constraints include resource deployment constraints, distribution network load recovery constraints, and information network recovery constraints;

[0010] The multimodal resource collaborative optimization decision model is solved based on the deep reinforcement learning algorithm to obtain the optimal resource deployment strategy, which is then sent to the terminal for execution.

[0011] Furthermore, the normalized data set is input into a preset hierarchical time series feature extraction model, and the cross-dimensional time series features of meteorological, geographical and power grid equipment are extracted through a layer-by-layer progressive LSTM network, and the fused multi-dimensional feature vector is output, including:

[0012] The first time series layer of the hierarchical time series feature extraction model uses a long short-term memory network to capture the short-term dependency of meteorological parameters time-step by time step and outputs a meteorological feature vector;

[0013] The second time series layer takes the meteorological feature vector as input, extracts the long-term evolution law of geographical parameters through the long short-term memory network, and outputs the geographical feature vector after dimensionality reduction;

[0014] The third time series layer takes the geographic feature vector as input, fuses the cross-dimensional correlation features of meteorological, geographical and power grid equipment parameters through the long short-term memory network, and outputs the power grid equipment status feature vector;

[0015] Among them, the output features of each time series layer are reduced in dimension by the fully connected layer to generate a fused multi-dimensional feature vector.

[0016] Furthermore, the multi-dimensional feature vectors are dynamically weighted and mapped to risk intervals, and a disaster warning signal is determined according to the risk level results divided by a preset threshold, including:

[0017] Construct parallel time series network branches to independently extract time series patterns of meteorological, geographical and power grid equipment characteristics, and dynamically integrate data-driven weights with prior weights generated by expert scores to generate comprehensive risk weights;

[0018] The comprehensive risk weight is mapped to a preset continuous risk interval, and a disaster warning signal is output according to the low risk, medium risk and high risk level thresholds divided in the risk interval.

[0019] Furthermore, the parallel time series network branches are constructed to independently extract the time series patterns of meteorological, geographical and power grid equipment characteristics, and dynamically integrate the data-driven weights with the prior weights generated by expert scores to generate comprehensive risk weights, including:

[0020] y=α·y1+(1―α)·y2

[0021]

[0022] Where: y is the comprehensive risk weight, y1 is the weight obtained by taking the weighted average method to fuse data features, y2 is the weight obtained after normalization, α∈[0,1] is the relative proportion between the LSTM network prediction result and the expert score, x ζ is the value of the ζth feature, ω ζ is the weight of the ζth feature, and λ is the number of eigenvalues.

[0023] Furthermore, the objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, including:

[0024] Minimize post-disaster load loss L loss The expression:

[0025]

[0026] Where: T is the total time, N P is the set of distribution network nodes, ω i,t is the weight coefficient of node i at time t, Δt is the time interval, is the demand load of node i at time t, is the actual recovery load of node i at time t.

[0027] Furthermore, the objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, and also includes:

[0028] Minimize resource deployment cost c deploy The expression:

[0029]

[0030] Where: c S 、c U 、c G 、c CV and c PV are the deployment costs of satellites, drones, generators, communication vehicles, and emergency power supply vehicles, respectively. and They are respectively node sets that can choose to deploy satellites, drones, generators, communication vehicles, and emergency power supply vehicles. and are 0-1 variables indicating whether node i has deployed satellites, drones, generators, communication vehicles, and emergency power supply vehicles, respectively. 1 indicates deployed, and 0 indicates not deployed.

[0031] Furthermore, the objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, and also includes:

[0032] The expression of the objective function F is:

[0033] F=minLloss +βc deploy

[0034] Where: β is the deployment cost and load conversion coefficient.

[0035] Furthermore, the constraints include resource deployment constraints, distribution network load recovery constraints, and information network recovery constraints, including:

[0036] The resource deployment constraints include mutual exclusion constraints of node deployment, an upper limit on the number of global resources, satellite deployment dependence on ground communication station conditions, and timing constraints that communication coverage takes precedence over power restoration;

[0037] The load recovery constraints of the distribution network include active output range restrictions of generators and emergency power supply vehicles, AC power flow balance of node power conservation, voltage safe operation range constraints, and maximum phase angle difference restrictions between adjacent nodes;

[0038] The information network recovery constraints include the node traffic conservation requirement and the dynamic limitation of communication link bandwidth due to the deployment of satellites, drones and communication vehicles.

[0039] Furthermore, the multimodal resource collaborative optimization decision model is solved based on a deep reinforcement learning algorithm to obtain an optimal resource deployment strategy, which is then sent to the terminal for execution, including:

[0040] The multimodal resource collaborative optimization decision model is mapped into a Markov decision process; the state space is a combination of the node's power recovery state, communication coverage state, and resource deployment state; the action space is a discrete action set of the node deploying generators, emergency power supply vehicles, satellites, drones, communication vehicles, or waiting; and the load loss, resource deployment cost, and constraint violation penalty are fed back through the reward function.

[0041] Furthermore, the method of solving the multimodal resource collaborative optimization decision model based on a deep reinforcement learning algorithm to obtain an optimal resource deployment strategy and sending it to the terminal for execution further includes:

[0042] A dual-depth Q-network is used for strategy optimization. The interaction data is stored and randomly sampled through the experience replay mechanism. Invalid actions that violate resource mutual exclusivity and communication coverage priority are dynamically filtered out by combining action masks. The Q-network parameters are iteratively updated to generate the optimal deployment strategy that meets the constraints.

[0043] It can be seen from the above technical solutions that this application has the following advantages:

[0044] This application collects multivariate data through low-orbit satellite internet, normalizes it, and then uses a multi-layer parallel LSTM network structure to extract time series features. It generates a comprehensive risk assessment based on historical experience and expert scores, categorizes risk levels according to thresholds, establishes a decision-making model for collaborative optimization of multimodal resources in distribution networks after disasters, and uses deep reinforcement learning to solve complex optimization models. This application addresses the limitations of traditional single data sources, while enhancing the reliability of risk assessments, enabling real-time adaptive decision-making for resource deployment and grid restoration, and significantly improving post-disaster recovery efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of an embodiment of a method for early warning of power distribution system disaster risks based on low-orbit satellite internet in the present invention;

[0046] Figure 2 Schematic diagram of the multivariate data collection scenario in the present invention;

[0047] Figure 3 This is the framework diagram of the DDQN algorithm in the present invention. DETAILED DESCRIPTION

[0048] 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.

[0049] Example 1

[0050] 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, without specific limitation. The following will introduce the power distribution system disaster risk warning method based on low-orbit satellite Internet 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:

[0051] S11. Acquire a multidimensional observation dataset through space-based remote sensing systems, air-based observation systems, and ground-based sensing systems, and perform normalization on the multidimensional observation dataset;

[0052] In this embodiment, the impact mechanisms of different types of extreme disasters on power grid components are different, as shown in Table 1 below, which includes the impact of different types of extreme disasters on power grid components and key parameter analysis. In order to better cope with the impact of extreme disasters on the power distribution system, this step uses low-orbit satellite Internet to extract features of key parameters covering three dimensions of meteorology, geography, and power grid equipment, including temperature, precipitation, air pressure, wind speed, vegetation cover, soil type, terrain, equipment operation years, voltage level, tower material, and other data information closely related to various extreme disasters and power grid components, to form multi-layer feature data, and use long short-term memory networks (LSTM) to perform risk warnings based on multivariate data fusion. Reference Figure 2 , Figure 2 This is a schematic diagram of a multi-data collection scenario based on low-orbit satellite internet. A detailed description is given below:

[0053] Table 1 Impact of different types of extreme disasters on power grid components and analysis of key parameters

[0054]

[0055]

[0056] In terms of meteorology, infrared sensors and microwave radars carried by remote sensing satellites monitor the spatial distribution of air temperature, surface temperature, and precipitation. UAVs provide supplementary low-altitude observations to obtain high-resolution meteorological parameters. Furthermore, a ground-based sensor network collects real-time data on temperature, humidity, and air pressure, which is calibrated with satellite observations to improve monitoring accuracy. In terms of geography, multispectral imaging technology carried by remote sensing satellites dynamically monitors changes in vegetation cover and identifies soil types and terrain features. Combined with precise location information provided by navigation satellites, drones are used to collect geographic information in disaster-prone areas. Furthermore, the ground-based sensor network uses precise sampling to supplement and verify satellite observation data, generating comprehensive geographic information. Regarding power grid equipment, high-resolution imagery from remote sensing satellites is used to identify the distribution and operating status of power grid equipment. Combined with drone inspection technology, this technology can identify aging, corrosion, or potential failure risks in electrical equipment. Furthermore, smart sensors are deployed to monitor power grid operating parameters in real time. Low-orbit satellite internet is used to transmit this diverse data to the cloud and ultimately back to the data processing center.

[0057] The acquired data on temperature, precipitation, air pressure, wind speed, vegetation cover, soil type, terrain, equipment operation age, voltage level, and tower material are processed using the Min-Max normalization method to map the resulting values ​​to the [0,1] interval. The conversion function is as follows:

[0058]

[0059] Where: x is the original data, min(x) is the minimum value in the data set, and max(x) is the maximum value in the data set.

[0060] S12. Input the normalized dataset into a preset hierarchical time series feature extraction model. A layer-by-layer LSTM network is used to extract cross-dimensional time series features of meteorological, geographical, and power grid equipment, and output a fused multidimensional feature vector.

[0061] In this embodiment, the hierarchical temporal feature extraction model includes the following:

[0062] 1. The first temporal layer uses a long short-term memory network to capture the short-term dependencies of meteorological parameters at each time step and outputs a meteorological feature vector;

[0063] 2. The second time series layer takes the meteorological feature vector as input, extracts the long-term evolution of geographic parameters through the long short-term memory network, and outputs the reduced-dimensional geographic feature vector;

[0064] 3. The third time series layer takes the geographic feature vector as input, fuses the cross-dimensional correlation features of meteorological, geographical, and power grid equipment parameters through a long short-term memory network, and outputs the power grid equipment status feature vector;

[0065] 4. The output features of each time series layer are reduced in dimension by the fully connected layer to generate a fused multi-dimensional feature vector.

[0066] Specifically, the normalized data dimensions are [data size, time step, parameters], where parameters include meteorological, geographical, and power grid equipment parameters. The number of hidden units in the first LSTM layer is set to 128, and the weight matrix and bias term are initialized. The activation functions for the input, forget, and output gates are defined as Sigmoid, and the activation function for the candidate memory units is defined as Tanh. Data is input sequentially by time step, and the output and hidden state are calculated for each time step. The first layer outputs the hidden state for each time step, which is used to capture short-term temporal dependencies (such as hourly meteorological changes). The final feature extracted is the hidden state of the last time step, with a dimension of [data size, 128]. The second LSTM layer inputs the hidden state sequence output by the first layer, with a dimension of [data size, time step, 128]. The number of hidden units is increased to 256 to capture more complex long-term dependencies (such as daily geographical changes). The input of the third LSTM layer is the output of the second layer, and the number of hidden units is set to 512, which is used to abstract cross-dimensional features (such as the relationship between weather and electrical equipment). The output of each LSTM layer is reduced in dimensionality through a fully connected layer. The output of the second layer is reduced to 64 dimensions, and the output of the third layer is reduced to 32 dimensions. The final output is a three-dimensional feature vector: [meteorological features, geographical features, electrical equipment features].

[0067] At the same time, considering the high-dimensional diversity of meteorological, geographical, and power grid equipment data, a parallel LSTM structure is adopted to assign an independent LSTM network to each meteorological, geographical, and power grid equipment data source, so that each network can focus on learning the time dependency of its corresponding data source to maintain the independence between different data streams, reduce interference between different source data, and improve processing efficiency.

[0068] S13. Dynamically weight the multi-dimensional feature vectors and map them to risk intervals, and determine disaster warning signals based on the risk level results divided by preset thresholds;

[0069] In this embodiment, determining the disaster warning signal includes the following:

[0070] 1. Construct parallel time series network branches to independently extract time series patterns of meteorological, geographical, and power grid equipment characteristics, and dynamically integrate data-driven weights with prior weights generated by expert scoring to generate comprehensive risk weights;

[0071] 2. Map the comprehensive risk weight to the preset continuous risk interval, and output the disaster warning signal according to the low-risk, medium-risk and high-risk level thresholds divided in the risk interval.

[0072] For the output features from different LSTM networks, we use the weighted average method to fuse the data features to obtain the weight y1, and use the expert scoring method to score the importance of the features (out of 10), and then normalize them to obtain the weight y2. The dynamic fusion weights are used to obtain the final result y. The specific formula is:

[0073] y=α·y1+(1―α)·y2

[0074]

[0075] Where: α∈[0,1] is the relative weight between the LSTM network prediction result and the expert score, x ζ is the value of the ζth feature, ω ζ is the weight of the ζth feature, λ is the number of eigenvalues, and the value of α is controlled by the amount of meteorological, geographical, and power grid equipment data.

[0076] Thresholds are set to distinguish different risk levels: low risk (0, 0.33), medium risk (0.33, 0.66), and high risk (0.66, 1). Each risk level corresponds to a different warning state. The output layer performs risk assessment based on the fused features and outputs a final risk warning.

[0077] S14. Determine the weight coefficients of distribution network nodes based on the disaster warning signals, and construct a multimodal resource collaborative optimization decision model based on the weight coefficients of the distribution network nodes; wherein the objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, and the constraints include resource deployment constraints, distribution network load recovery constraints, and information network recovery constraints;

[0078] Under the risk warning model of multivariate data fusion in the above satellite-ground fusion mode, this step is based on the load loss L during the post-disaster recovery period. loss and multimodal resource deployment cost c deploy Taking minimization as the objective function, a multi-modal resource collaborative optimization decision-making model for distribution network after disaster is established.

[0079] The expression of the objective function F is:

[0080] F=minL loss +βc deploy

[0081] Where: β is the deployment cost and load conversion coefficient.

[0082] Among them, minimizing the post-disaster load loss L loss The expression:

[0083]

[0084] Where: T is the total time, N P is the set of distribution network nodes, ω i,t is the weight coefficient of node i at time t, Δt is the time interval, is the demand load of node i at time t, is the actual recovery load of node i at time t.

[0085] Minimize resource deployment cost c deploy The expression:

[0086]

[0087] Where: c S 、c U 、c G 、c CV and c PV are the deployment costs of satellites, drones, generators, communication vehicles, and emergency power supply vehicles, respectively. and They are respectively node sets that can choose to deploy satellites, drones, generators, communication vehicles, and emergency power supply vehicles. and are 0-1 variables indicating whether node i has deployed satellites, drones, generators, communication vehicles, and emergency power supply vehicles, respectively. 1 indicates deployed, and 0 indicates not deployed.

[0088] In this embodiment, the constraints include resource deployment constraints, distribution network load recovery constraints, and information network recovery constraints;

[0089] 1. Resource deployment constraints include the mutual exclusion constraint of node deployment, the upper limit of the global resource quantity, the satellite deployment's dependence on ground communication station conditions, and the timing constraint that communication coverage takes precedence over power restoration:

[0090] The mutual exclusivity constraint of node deployment considers the mutual exclusivity of physical resources and information resources in node deployment: it means that at most one physical resource (generator, emergency power supply vehicle) can be deployed on the same node: Indicates that at most one information resource (satellite, drone, communication vehicle) can be deployed on the same node:

[0091] The upper limit of global resource quantity takes into account the global quantity limit of various resources: in, and They represent the maximum number of deployed satellites, drones, generators, communication vehicles, and emergency power supply vehicles respectively.

[0092] Satellite deployment depends on the conditions of ground communication stations: Among them, A i is a 0-1 variable, A i =0 means that node i has no satellite ground communication station deployed, otherwise it means that node i has a satellite ground communication station.

[0093] To ensure the observability and controllability of the distribution network, communication coverage of nodes must precede power restoration: Among them, M is a maximum constant, D i,t is a 0-1 variable, D i,t = 0 means that node i is not covered by communication resources at time t, D i,t =1 means that node i has been covered by communication resources at time t.

[0094] 2. Distribution network load recovery constraints include active output range limits for generators and emergency power supply vehicles, AC power flow balance for node power conservation, voltage safety operating range constraints, and maximum phase angle difference limits between adjacent nodes:

[0095] Active power output range limits of generators and emergency power supply vehicles: in, are the upper and lower limits of the active output of the generator at node i, is the actual output of the generator at node i at time t, are the upper and lower limits of the active power output of the emergency power supply vehicle at node i, is the actual output of the emergency power supply vehicle at node i at time t.

[0096] The AC power flow balance with node power conservation uses the AC power flow model to constrain the distribution network: Among them, P ij,t is the active power flow of line (i, j) at time t.

[0097] The voltage safe operation range constraint and the maximum phase angle difference limit between adjacent nodes take into account that the voltage and phase angle of the distribution network nodes after fault recovery must be within a safe range: in, are the upper and lower limits of the voltage at node i, U i,t is the voltage of node i at time t, θ i,t and θ j,t is the voltage phase angle between nodes i and j at time t, Δθ max The maximum phase angle difference allowed between adjacent nodes.

[0098] 3. Information network recovery constraints include node traffic conservation requirements and the dynamic limitations of communication link bandwidth due to the deployment of satellites, drones, and communication vehicles:

[0099] Node traffic conservation requirements: in, is the communication flow of communication link (i, j) at time t, is the data demand of node i at time t.

[0100] The communication link bandwidth is dynamically limited by the deployment of satellites, drones, and communication vehicles. Considering that the total link traffic does not exceed the bandwidth limit: in, is the maximum bandwidth of the communication link (i, j).

[0101] S15. Solve the multimodal resource collaborative optimization decision model based on the deep reinforcement learning algorithm, obtain the optimal resource deployment strategy, and send it to the terminal for execution.

[0102] In this embodiment, solving the multimodal resource collaborative optimization decision model includes the following:

[0103] 1. The multimodal resource collaborative optimization decision model is mapped into a Markov decision process. The state space is a combination of the node's power restoration state, communication coverage state, and resource deployment state. The action space is a set of discrete actions for the node to deploy generators, emergency power supply vehicles, satellites, drones, communication vehicles, or wait. The reward function is used to feedback load loss, resource deployment costs, and constraint violation penalties.

[0104] 2. Use a dual-depth Q network for strategy optimization. The experience replay mechanism stores interaction data and randomly samples it. Combined with action masks, it dynamically filters invalid actions that violate resource mutual exclusivity and communication coverage priority. It iteratively updates the Q network parameters to generate the optimal deployment strategy that meets the constraints.

[0105] Specifically, the multimodal resource collaborative optimization decision-making model for distribution network post-disasters poses severe challenges to traditional optimization methods due to its high dimensionality, nonlinearity, dynamicity and multi-objective characteristics. However, after being converted into a Markov decision process (MDP), through the deep reinforcement learning framework, it can efficiently handle complex constraints, dynamic time series and uncertainties in a data-driven manner, while achieving adaptive balance of multiple objectives.

[0106] In the decision process, the MDP consists of a tuple of five elements<S,A,R,P,γ> These five elements explain the impact of each control variable in the decision-making process from different dimensions. The state space variable S represents the set of operating environments; the action space A represents the set of actions taken by each agent in the MDP; the reward function R represents the interaction between the agent and the environment; the state transition probability P represents the impact of the agent's actions on the interaction with the environment; and the discount factor γ represents the discount coefficient for the agent's attention to future rewards.

[0107] The state space involves the node's power recovery status, communication coverage status, resource deployment status, and environmental parameters. Therefore, the system's state space can be expressed as:

[0108]

[0109] The actions in the action space are to deploy resources at the node or wait, which must satisfy the mutual exclusion constraints of node deployment and the upper limit of the global resource quantity. Deploy the generator at node i or emergency power supply vehicle Deploy satellites at node i or drones or communication vehicle Waiting: No resources are deployed, only time steps are advanced. Therefore, the action space of the system can be expressed as:

[0110]

[0111] The reward function directly maps the objective function and the constraint penalty:

[0112] Basic Reward: R basic (s t ,a t )=―(L loss +βc deploy ); Constraint penalty: If the action violates resource mutual exclusivity, a penalty R is imposed penalty = -100; if the power flow or voltage exceeds the limit, a penalty R is imposed penalty =-50.

[0113] Therefore, the total reward function of the system can be expressed as:

[0114] R(s t ,a t )=R basic +∑R penalp

[0115] Specifically, the state transition of the system is driven by the following rules:

[0116] Resource deployment: If action a t To deploy physical resources, update the resource status of the corresponding node And trigger power recovery; if action a t To deploy communication resources, update the communication coverage status D of the corresponding node i,t =1.

[0117] Power system dynamics: Update node power balance based on power flow equations; correct voltage and phase angle based on voltage constraints.

[0118] Time advancement: Each time an action is performed (including waiting), the time step t increases until t=T.

[0119] Deep reinforcement learning combines the perception ability of deep learning and the decision-making ability of reinforcement learning, enabling the intelligent agent to autonomously learn the optimal strategy in a complex environment through interaction with the environment. Therefore, this embodiment uses a dual deep Q network (DDQN) to solve the above MDP problem. Figure 3 , Figure 3 This is the DDQN algorithm framework diagram.

[0120] Agent interacts with the environment: According to the current state s t (Node restores load, voltage, communication coverage status, etc.), select action a through DDQN t (Deploy resources or wait). After executing the action, return to the next state s t+1 Reward R t , and update the distribution network status. Experience storage and sampling: store the four-tuple (s t ,a t ,s t+1,R t ), breaking data correlation by random sampling. Record the valid action set during storage and sampling, and filter conflicting actions. Q network update: input state s t , output the Q(s,a;θ) value of all actions, filter invalid actions through the mask layer. Calculate the target Q value Synchronize the random parameters θ of the Q network every C steps ′ =θ, using mean square error L DQN (θ)=E[(y―Q(s,a;θ)) 2 ], update the current network parameters through back propagation. The action mask prohibits the deployment of conflicting resources on the same node. If the node does not cover the communication resources, that is, D i,t = 0, prohibiting power restoration work. Dynamically update the number of remaining deployable resources and limit over-deployment through masks.

[0121] The above-mentioned embodiment significantly improves the spatiotemporal accuracy of disaster risk warnings through multi-dimensional data fusion and hierarchical time series feature extraction, and realizes dynamic collaborative perception of meteorological, geographical and power grid equipment status; combined with the adaptive optimization of deep reinforcement learning and the dynamic masking mechanism of resource deployment constraints, it quickly generates resource scheduling strategies that take into account both economy and safety in complex post-disaster scenarios, reduces load losses and improves recovery efficiency.

[0122] 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.

[0123] 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 power distribution system disaster risk early warning method based on low-orbit satellite internet, characterized in that: include: Acquire a multidimensional observation data set through a space-based remote sensing system, an air-based observation system, and a ground-based sensing system, and perform normalization processing on the multidimensional observation data set; The normalized dataset is input into a preset hierarchical time series feature extraction model. A layer-by-layer LSTM network is used to extract cross-dimensional time series features of meteorological, geographical, and power grid equipment, and outputs a fused multidimensional feature vector. Dynamically weighting the multidimensional feature vectors and mapping them to risk intervals, and determining disaster warning signals based on risk level results divided by preset thresholds; Determine the weight coefficients of the distribution network nodes based on the disaster warning signal, and construct a multimodal resource collaborative optimization decision model based on the weight coefficients of the distribution network nodes; wherein the objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, and the constraints include resource deployment constraints, distribution network load recovery constraints, and information network recovery constraints; The multimodal resource collaborative optimization decision model is solved based on the deep reinforcement learning algorithm to obtain the optimal resource deployment strategy, which is then sent to the terminal for execution.

2. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 1 is characterized in that: The normalized data set is input into a preset hierarchical time series feature extraction model, and the cross-dimensional time series features of meteorological, geographical and power grid equipment are extracted through a layer-by-layer progressive LSTM network, and the fused multi-dimensional feature vector is output, including: The first time series layer of the hierarchical time series feature extraction model uses a long short-term memory network to capture the short-term dependency of meteorological parameters time-step by time step and outputs a meteorological feature vector; The second time series layer takes the meteorological feature vector as input, extracts the long-term evolution law of geographical parameters through the long short-term memory network, and outputs the geographical feature vector after dimensionality reduction; The third time series layer takes the geographic feature vector as input, fuses the cross-dimensional correlation features of meteorological, geographical and power grid equipment parameters through the long short-term memory network, and outputs the power grid equipment status feature vector; Among them, the output features of each time series layer are reduced in dimension by the fully connected layer to generate a fused multi-dimensional feature vector.

3. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 1 is characterized in that: The step of dynamically weighting and fusing the multi-dimensional feature vectors and mapping them to risk intervals, and determining a disaster warning signal based on a risk level result divided by a preset threshold, includes: Construct parallel time series network branches to independently extract time series patterns of meteorological, geographical and power grid equipment characteristics, and dynamically integrate data-driven weights with prior weights generated by expert scores to generate comprehensive risk weights; The comprehensive risk weight is mapped to a preset continuous risk interval, and a disaster warning signal is output according to the low risk, medium risk and high risk level thresholds divided in the risk interval.

4. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 3 is characterized in that: The parallel time series network branches are constructed to independently extract the time series patterns of meteorological, geographical and power grid equipment characteristics, dynamically integrate data-driven weights with prior weights generated by expert scores, and generate comprehensive risk weights, including: y=α·y1+(1―α)·y2 Where: y is the comprehensive risk weight, y1 is the weight obtained by taking the weighted average method to fuse data features, y2 is the weight obtained after normalization, α∈[0,1] is the relative proportion between the LSTM network prediction result and the expert score, x ζ is the value of the ζth feature, ω ζ is the weight of the ζth feature, and λ is the number of eigenvalues.

5. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 1 is characterized in that: The objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, including: Minimize post-disaster load loss L loss The expression: Where: T is the total time, N P is the set of distribution network nodes, ω i,t is the weight coefficient of node i at time t, Δt is the time interval, is the demand load of node i at time t, is the actual recovery load of node i at time t.

6. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 5 is characterized in that: The objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, and also includes: Minimize resource deployment cost c deploy The expression: Where: c S 、c U 、c G 、c CV and c PV are the deployment costs of satellites, drones, generators, communication vehicles, and emergency power supply vehicles, respectively. and They are respectively node sets that can choose to deploy satellites, drones, generators, communication vehicles, and emergency power supply vehicles. and are 0-1 variables indicating whether node i has deployed satellites, drones, generators, communication vehicles, and emergency power supply vehicles, respectively. 1 indicates deployed, and 0 indicates not deployed.

7. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to any one of claims 5 or 6, characterized in that: The objective function includes minimizing post-disaster load loss and minimizing resource deployment costs, and also includes: The expression of the objective function F is: F=minL loss +βc deploy Where: β is the deployment cost and load conversion coefficient.

8. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 1 is characterized in that: The constraints include resource deployment constraints, distribution network load recovery constraints, and information network recovery constraints, including: The resource deployment constraints include mutual exclusion constraints of node deployment, an upper limit on the number of global resources, satellite deployment dependence on ground communication station conditions, and timing constraints that communication coverage takes precedence over power restoration; The load recovery constraints of the distribution network include active output range restrictions of generators and emergency power supply vehicles, AC power flow balance of node power conservation, voltage safe operation range constraints, and maximum phase angle difference restrictions between adjacent nodes; The information network recovery constraints include the node traffic conservation requirement and the dynamic limitation of communication link bandwidth due to the deployment of satellites, drones and communication vehicles.

9. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 1, characterized in that: Solving the multimodal resource collaborative optimization decision model based on a deep reinforcement learning algorithm to obtain the optimal resource deployment strategy and sending it to the terminal for execution includes: The multimodal resource collaborative optimization decision model is mapped into a Markov decision process; the state space is a combination of the node's power recovery state, communication coverage state, and resource deployment state; the action space is a discrete action set of the node deploying generators, emergency power supply vehicles, satellites, drones, communication vehicles, or waiting; and the load loss, resource deployment cost, and constraint violation penalty are fed back through the reward function.

10. The power distribution system disaster risk early warning method based on low-orbit satellite internet according to claim 9, characterized in that: Solving the multimodal resource collaborative optimization decision model based on a deep reinforcement learning algorithm to obtain an optimal resource deployment strategy and sending it to the terminal for execution also includes: A dual-depth Q-network is used for strategy optimization. The interaction data is stored and randomly sampled through the experience replay mechanism. Invalid actions that violate resource mutual exclusivity and communication coverage priority are dynamically filtered out by combining action masks. The Q-network parameters are iteratively updated to generate the optimal deployment strategy that meets the constraints.

Images