A power-traffic coupled network resilience improvement method and medium

Abstract

The application discloses a power-traffic coupling network resilience improvement method and medium, and belongs to the technical field of power system automation. The method comprises the following steps: based on power-traffic coupling network parameters, a conventional Markov decision process and a fault Markov decision process are respectively constructed; a deep reinforcement learning method is used to solve the conventional Markov decision process, so that an optimal power-traffic coupling network control strategy under a conventional scene and a soft label data set are obtained; the soft label data set and the deep reinforcement learning method are used to solve the fault scene Markov decision process, so that an optimal power-traffic coupling network control strategy under a fault scene is obtained, and power-traffic coupling network control parameters are optimized, and the resilience of the power-traffic coupling network is strengthened. The application significantly shortens the power distribution network fault recovery time and improves the traffic efficiency, and exhibits good stability and optimization capacity under multi-scene coupling.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a method and medium for improving the resilience of power-transportation coupled networks. Background Technology

[0002] With the accelerated development of new power systems and intelligent transportation, the structures of distribution networks and transportation networks are becoming increasingly complex. New elements such as distributed power sources, electric vehicle charging loads, and flexible loads are deeply integrated into the system, significantly increasing operational uncertainty. On the power side, the high proportion of renewable energy integration leads to increased power flow fluctuations in the distribution network, significantly increasing the difficulty of voltage regulation. On the transportation side, the rapid growth in the number of electric vehicles has made the areas around charging stations new traffic bottlenecks, with some road sections experiencing peak-hour congestion indices approximately 200% higher than ordinary roads. Simultaneously, frequent natural disasters such as typhoons, earthquakes, and floods, along with factors like equipment aging and human error, continue to drive the diversification and complexity of distribution network faults. Of particular concern is that traffic congestion severely restricts repair efficiency; ignoring real-time traffic information can extend fault isolation time by up to 72%. Excessive recovery time for distribution network and transportation network faults results in significant economic losses. These problems and challenges highlight the urgency and necessity of improving the resilience of the power-transportation coupled network.

[0003] The existing methods for improving the resilience of power distribution networks and transportation systems are mainly limited in the following two aspects: First, traditional fault recovery methods, such as heuristic algorithms and mathematical optimization methods, are mostly limited to single-objective optimization, have poor adaptability to dynamically changing power grid topologies, have limited comprehensive processing capabilities for distributed power output and load uncertainties, and are difficult to guarantee obtaining the global optimal solution, thus failing to fully fit the power distribution network and transportation system network. Second, although artificial intelligence (AI) methods, which have received much attention in recent years, have strong environmental adaptability, their performance is highly dependent on massive amounts of fault data for training. However, in actual power systems, fault scenarios are scarce and effective samples are insufficient, which can easily lead to limited model generalization ability. Furthermore, training with a large amount of data will significantly prolong the model convergence time, further restricting the rapid fault recovery capability of power distribution networks and transportation systems. Summary of the Invention

[0004] The purpose of this invention is to provide a method and medium for improving the resilience of power-transportation coupled networks, which solves the problems disclosed in the background art.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0006] On one hand, the present invention provides a method for improving the resilience of a power-transportation coupled network, comprising:

[0007] Based on the acquired power-transportation coupled network parameters, a heterogeneous coupling graph of the power-transportation coupled network is constructed;

[0008] A two-layer graph convolutional network is used to extract features from the heterogeneous coupling graph of the power-transportation coupling network to obtain the aggregated feature vector of the power-transportation coupling network.

[0009] Based on the aggregated feature vectors, a conventional Markov decision process and a faulty Markov decision process are constructed respectively.

[0010] We use deep reinforcement learning to solve the conventional Markov decision process and obtain the optimal power-transportation coupled network control strategy under conventional scenarios.

[0011] Based on the conventional action value function of the optimal power-transportation coupled network control strategy in conventional scenarios, a soft-label dataset is generated using the softmax function.

[0012] In response to the occurrence of a fault in the power-transportation coupled network, based on a soft-label dataset, a knowledge distillation deep Q-network algorithm based on transfer-reinforcement co-training is used to solve the Markov decision process of the fault scenario, thereby obtaining the optimal power-transportation coupled network control strategy under the fault scenario; wherein, the knowledge distillation deep Q-network algorithm based on transfer-reinforcement co-training solves the Markov decision process of the fault scenario through the collaborative solution of the knowledge distillation method and the deep Q-network method.

[0013] Based on the optimal power-traffic coupled network control strategy under normal scenarios and the optimal power-traffic coupled network control strategy under fault scenarios, the control parameters of the power-traffic coupled network are optimized to achieve enhanced resilience of the power-traffic coupled network. Among them, the power-traffic coupled network control parameters include: power output parameters of distribution network generators, voltage control parameters, and tidal lane direction control parameters of traffic network.

[0014] Optional, heterogeneous coupling diagram of power-transportation coupled network G =( V , E ) set of nodes V Includes distribution network nodes and transportation network nodes, edge set E It includes the distribution network edge set, which characterizes the power transmission relationship between distribution network nodes, and the transportation network edge set, which characterizes the road connection relationship between transportation network nodes.

[0015] Optionally, the aggregated feature vector of the power-transportation coupled network includes regular aggregated features and fault aggregated features; the structured data of the heterogeneous coupling graph of the power-transportation coupled network includes adjacency matrix and feature matrix;

[0016] A two-layer graph convolutional network is used to extract features from the heterogeneous coupling graph of the power-transportation coupled network, resulting in an aggregated feature vector for the power-transportation coupled network, including:

[0017] Structured data from the heterogeneous coupled graph of the power-transportation coupled network is input into a two-layer graph convolutional network, and conventional aggregated features are output through graph convolution operations.

[0018] Based on the fault probability matrix of the power-transportation coupled network, the fault adjacency matrix is ​​calculated using the adjacency matrix.

[0019] Based on the fault probability matrix, the fault feature matrix is ​​calculated using the feature matrix.

[0020] The fault adjacency matrix and fault feature matrix are input into a two-layer graph convolutional network, and the fault aggregation features are output through graph convolution operations.

[0021] Optionally, the construction of the failure probability matrix includes:

[0022] The failure probability of traffic segments and the fault probability of lines are calculated based on the obtained power-transportation coupled network parameters.

[0023] The failure probability of the charging station is calculated based on the failure probability of the lines in the area where the charging station is located.

[0024] A failure probability matrix is ​​constructed based on the failure probability of traffic segments, the failure probability of lines, and the failure probability of charging stations.

[0025] Among them, traffic sections ( i , j Failure probability The calculation expression is:

[0026] ;

[0027] In the formula: q ij For traffic sections ( i , j Traffic flow, C ij For traffic sections ( i , j Basic traffic capacity For adjustment coefficients,

[0028] line( u , y Failure probability The calculation expression is:

[0029] ;

[0030] In the formula:S uy For the line ( u , y The trend of ) For the line ( u , y Maximum transmission capacity α 1 is a parameter that controls the rate of change of probability. β 1 represents the overload threshold.

[0031] Optionally, the state space of a conventional Markov decision process and a faulty Markov decision process is defined as follows: t The operational status of the power distribution network and transportation network under normal conditions. Fault scenario state ;in, These are the regular aggregated features extracted by a two-layer graph convolutional network. Fault aggregation features extracted by a two-layer graph convolutional network;

[0032] The action space of a conventional Markov decision process and a faulty Markov decision process is defined as the action space of the agent in... t Actions in real time; actions in typical scenarios Including routine power distribution network operations Traffic network actions in conventional scenarios ; where Δ P Indicates generator output adjustment, Δ V This indicates that the transformer tap changer adjusts the voltage according to the direction of the average voltage deviation across the entire network. d r Indicates the direction of the tidal flow lanes in the transportation network. R tid For the collection of tidal flow lanes, r For tidal flow lanes; actions during fault scenarios Including distribution network actions in fault scenarios Traffic network actions in fault scenarios ;in, δ uy Indicates the line ( u , y The on / off state of ) ε PG This is the set of all controllable lines.

[0033] Optionally, the reward function for an agent in a conventional Markov decision process is defined as:

[0034] ;

[0035] ;

[0036] ;

[0037] In the formula: For a typical scenario reward function, For regular distribution network rewards, Rewards for regular transportation networks;

[0038] Objective function of conventional distribution network f PG,nor Minimize the rewards for a conventional distribution network, including minimizing voltage deviation. L vol and minimize active power loss P los The specific expression is:

[0039] ;

[0040] ;

[0041] ;

[0042] In the formula: V u For the first u Voltage of each distribution network node u =1,2,…, M nod , M nod This represents the total number of distribution network nodes. V nom Rated voltage, R h and I h The first h The resistance and current of the line, h =1,2,…, N lin , N lin This represents the total number of lines.

[0043] Objective function of conventional transportation network f TN,nor Minimize the reward for a regular transportation network, including minimizing travel time. T tra And maximize road capacity C A The specific expression is:

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] In the formula: x a It is a traffic section a Traffic flow on the road a ∈ A roa , A roa This includes all traffic sections. t a It is a travel time function. This refers to travel time in a free-flowing state. α and β For calibration parameters, C a It is a traffic section a Traffic capacity, and Traffic sections a The number of lanes and the saturation level of a single lane;

[0050] Solving conventional Markov decision processes using deep reinforcement learning methods includes:

[0051] Based on the current operating status of the distribution network and transportation network, S11, the agent outputs control actions according to the power-transportation coupled network control strategy of the conventional scenario, with the reward function as the objective, and updates the operating status of the distribution network and the transportation network.

[0052] S12 stores the control action, the updated power distribution network operation status, and the updated transportation network operation status in the conventional scenario experience pool.

[0053] S13 Iteratively optimizes the control strategy for the power-transport coupled network in the conventional scenario based on the experience pool of the conventional scenario;

[0054] Repeat steps S11, S12 and S13 until the preset solution cutoff condition is met, and output the optimal power-transportation coupled network control strategy under normal scenarios.

[0055] The conventional action value function is constructed using the state space, action space, and reward function of a conventional Markov decision process.

[0056] Optionally, based on the conventional action value function, the calculation formula for generating the soft-labeled dataset using the softmax function is as follows:

[0057] ;

[0058] In the formula: s This is the current state. It is a temperature coefficient used to smooth probability distributions. This is a soft-label dataset. This is the value function for regular actions.

[0059] Optionally, the reward function for the agent in a faulty Markov decision process is defined as:

[0060] ;

[0061] In the formula, For fault scenario reward function;

[0062] Distribution network rewards for fault scenarios Including positive incentives for distribution networks and distribution network constraint penalty items :

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] In the formula: The objective function is the function to be applied when a fault occurs in the distribution network. It is a logical discrimination function; for t A network connectivity constraint penalty is introduced if the reconstructed network does not satisfy the radial structure requirement. for t Timetable and route ( u , y The on / off state of ); for t Penalty for limiting the number of operations at any given time, if t The number of switching actions during time-based control exceeds the allowable number Δ max If so, then this penalty will be introduced; As a penalty for exceeding the line power limit, for any line after reconfiguration ( u , yIf the power exceeds the line ( u , y Maximum limit If so, then this penalty will be introduced; For distributed power source power constraint penalties, if t At any given time, the power of a distributed power source If the value is outside the allowed range, this penalty is imposed. for t The maximum power of a certain distributed power source at a given time. Let the minimum power of a certain distributed power source be defined at a given time. λ 1~ λ 4 is the penalty coefficient. N PG For the set of distribution network nodes, S uy,t for t Timetable and route ( u , y The power of )

[0071] Traffic network rewards for fault scenarios Including positive incentives for transportation networks and traffic network constraints and penalties :

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] In the formula: Let be the objective function for traffic network congestion. λ 5 and λ 6 is the penalty coefficient. They are respectively t Penalties for violation of flow balance constraints at any time t Penalties for violating road capacity constraints at any time x a,t for t Traffic sections a Traffic flow on the road N TN For transportation network integration, and To represent the nodes of the transportation network respectively u The set of entry and exit road sections for section 1;

[0078] Based on a soft-label dataset, a deep Q-network algorithm based on transfer-reinforcement co-training is used to solve the Markov decision process for fault scenarios, including:

[0079] Construct a fault action value function based on the state space, action space, and reward function of the fault Markov decision process;

[0080] The knowledge distillation loss function is calculated based on the fault action value function and the soft-label dataset.

[0081] Sample data for solving Markov decision processes are obtained based on the reinforcement learning DQN algorithm, and an experience pool for fault scenarios is constructed.

[0082] The reinforcement learning loss is calculated based on the sample data obtained from the experience pool of fault scenarios.

[0083] Construct a total policy loss function based on reinforcement learning loss and knowledge distillation loss;

[0084] Based on the overall policy loss function, the optimal power-transportation coupled network control strategy under fault scenarios is obtained by minimizing the loss through gradient descent.

[0085] Optionally, the expression for the total policy loss function in the failure scenario is:

[0086] ;

[0087] In the formula: α 2 is a hyperparameter used to balance the importance of the loss from knowledge distillation and reinforcement learning. L total Let be the total policy loss function under failure scenarios. θ fau For the network parameters of a faulty Markov decision process, L KD The knowledge distillation loss function is calculated using KL divergence:

[0088] ;

[0089] in, D KL To find the KL divergence function, This is a soft-label dataset. The fault-based policy distribution is obtained through the fault action value function, where || represents concatenation;

[0090] L RL The formula for calculating the reinforcement learning loss function is as follows:

[0091] ;

[0092] Among them, the target value y target The calculation method is as follows:

[0093] ;

[0094] In the formula: The current action is performed in the current state s. The instant reward obtained afterward It is a discount factor. The next state is... The next move is... It is the fault action value function. It is the target fault action value function. It indicates the desire for expectation.

[0095] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program / instructions are executed by a processor, the steps of the method for improving the resilience of power-transportation coupled networks as described in any of the first aspects above are implemented.

[0096] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0097] This invention constructs conventional Markov decision processes and fault Markov decision processes respectively, achieving coordinated optimization of power network reconfiguration and traffic flow control through coupled constraints. It can select the optimal power-traffic coupled network strategy under multiple optimization objectives. Furthermore, this invention improves the solution efficiency of the fault Markov decision process by constructing a soft-label dataset using the solved conventional Markov decision process, and significantly increases the solution stability and convergence speed of the optimal power-traffic coupled network control strategy under fault scenarios. Finally, this invention improves the recovery speed of the power-traffic coupled network during fault occurrences and enhances its resilience by controlling the generator output parameters, voltage control parameters, and tidal lane direction control parameters of the distribution network within the power-traffic coupled network.

[0098] This invention utilizes the soft-label dataset output by a conventional Markov decision process to provide a large number of soft-label samples for faulty Markov decision processes, thereby improving the model's generalization ability. The two-layer graph convolutional network can perform aggregated feature vector extraction in scenarios with frequently changing network topologies, improving the robustness and dynamic adjustment capability of the power-transportation coupled network resilience enhancement method in this invention for scenarios with uncertain distributed power output and load.

[0099] This invention includes optimization models for both conventional and fault scenarios. By coupling constraints, it achieves coordinated optimization of power network reconfiguration and traffic flow control, overcoming the decision-making conflict problem caused by the independent optimization of power network and traffic network systems in traditional methods.

[0100] This invention constructs a smart distillation optimization Markov decision process model for power-transportation networks. By accurately extracting the deep topological features of the coupled system under fault disturbances through GCN, the complex coupling relationship is transformed into a quantifiable state-space representation that reflects the vulnerability of the network, thus realizing dynamic collaborative optimization of distribution network load recovery and tidal lane adjustment. Attached Figure Description

[0101] Figure 1 This is an architecture diagram of the knowledge distillation-based strategy for enhancing the resilience of power-transportation coupled networks according to the present invention.

[0102] Figure 2 This is a flowchart of the knowledge distillation-based strategy for improving the resilience of power-transportation coupled networks according to the present invention.

[0103] Figure 3 This is a flowchart of the feature extraction process for the power-transportation coupled network based on Graph Convolutional Networks (GCN) of the present invention.

[0104] Figure 4 This is a flowchart of the algorithm based on Knowledge Distillation for Deep Q-Networks (KDQN) of the present invention. Detailed Implementation

[0105] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0106] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0107] Example 1: This example introduces a method for improving the resilience of a power-transportation coupled network, such as... Figure 2 As shown in this embodiment, teacher and student models are constructed based on the aggregated features of regular and extreme scenarios extracted by GCN convolution operations, respectively. Extreme scenarios represent fault scenarios in the power-transportation coupled network. The specific steps of the power-transportation coupled network resilience enhancement method are as follows:

[0108] Step 1: Initialize the power-transportation coupled network parameters based on the acquired power-transportation coupled network data. The power-transportation coupled network parameters include distribution network node parameters, distribution network load, distribution network voltage constraints, distribution network current constraints, distribution network power, transportation network node information, transportation network branches, transportation network routes, transportation network traffic flow, and transportation network tidal lane time periods.

[0109] Step 2: Based on the parameters of the power-transportation coupled network, construct a heterogeneous coupling graph of the power-transportation coupled network through graph structure definition; the heterogeneous coupling graph of the power-transportation coupled network is used to characterize the multi-layer topology and attribute features under normal and fault scenarios.

[0110] Step 3 involves using a two-layer GCN to extract features from the heterogeneous coupled graph data of the power-transportation coupled network, obtaining the aggregated feature vector of the power-transportation coupled network. The stacking of the two-layer GCN can capture deeper feature representations, giving the model better expressive power. The aggregated feature vector includes aggregated features for regular scenarios, i.e., aggregated feature vectors for regular scenarios and aggregated feature vectors for fault scenarios in the power-transportation coupled network. A fault Markov decision process is then constructed based on the aggregated feature vectors for fault scenarios.

[0111] Step 4: Construct a conventional Markov decision process based on aggregated feature vectors under conventional scenarios; solve the constructed Markov decision process using a deep reinforcement learning algorithm to obtain the optimal power-transportation coupled network control strategy under conventional scenarios.

[0112] Based on the conventional action value function of the optimal power-transportation coupled network control strategy in conventional scenarios, a soft-label dataset is obtained by using the softmax function to perform knowledge distillation (KD).

[0113] When a fault occurs in the power-transportation coupled network, the Knowledge Distillation for Deep Q-Networks (KDQN) algorithm based on transfer-reinforcement co-training is used to solve the Markov decision process of the fault scenario, and the optimal power-transportation coupled network control strategy under the fault scenario is obtained.

[0114] like Figure 2 and Figure 4The application scenario of the teacher model shown is a conventional scenario. In a conventional scenario, the agent within the teacher model continuously interacts with the environment to obtain the optimal power-transportation network optimization scheme under the conventional scenario. When the power-transportation coupled network fails and requires emergency recovery, the pre-trained teacher model softens the probability of each action in each state to form a soft-label dataset, which is then input into the student model. The student model is trained synchronously based on the soft-label dataset and DQN (Deep Q-Network). The knowledge distillation Deep Q-Network algorithm based on transfer-reinforcement co-training specifically obtains the probability distribution of all possible actions in different states of the power-transportation coupled network through an optimized conventional Markov decision process. This is used to help the faulty Markov decision process understand the relative advantages and disadvantages of different actions in a specific state. The conventional Markov decision process is regarded as the teacher model, and the faulty Markov decision process is regarded as the student model. The control strategy experience obtained by knowledge distillation based on the trained teacher model accelerates the training speed of the student model and ensures the stability of the student model's control strategy.

[0115] Step 5: Based on the optimal power-traffic coupled network control strategy under normal scenarios and the optimal power-traffic coupled network control strategy under fault scenarios, optimize the power-traffic coupled network control parameters; the power-traffic coupled network control parameters include: power output parameters of distribution network generators, voltage control parameters, and tidal lane direction control parameters of traffic network.

[0116] In summary, the conventional Markov decision process and the fault Markov decision process constructed in this embodiment can select the optimal power-transportation coupled network strategy under multiple optimization objectives, improve the solution efficiency of the fault Markov decision process, and significantly increase the solution stability and convergence speed of the optimal power-transportation coupled network control strategy under fault scenarios. This embodiment also improves the recovery speed of the power-transportation coupled network when faults occur, and enhances the resilience of the power-transportation coupled network.

[0117] Example 2, based on the same inventive concept as Example 1, provides a method for improving the resilience of power-transportation coupled networks based on knowledge distillation, such as... Figure 1The real-time network parameters of the power-transportation system include: distribution topology, distribution impedance parameters, node voltage and current, distributed power sources, road topology, lane traffic rules, road segment traffic density, and charging waiting queues. Based on the charging load interaction points, the real-time network parameters of the power-transportation system are dynamically coupled to obtain a power-transportation coupled network. This dynamic coupling operation includes physical constraint coupling and objective function coupling. Based on the real-time network parameters, an adjacency matrix and feature matrix are constructed through graph structure transformation to build a heterogeneous coupled graph of the power-transportation coupled network. Topological features of the circuit-transportation coupled network are extracted based on the GCN layer. These topological features are aggregated, and teacher and student models are constructed separately. Knowledge distillation is performed using the pre-trained teacher model to provide a large number of soft-label samples for the student model, accelerating the training speed and improving the generalization ability of the student model. Based on the trained student model, optimal distribution network interconnection switch action schemes and optimal tidal lane direction optimization schemes for fault scenarios are generated. The power-transportation coupled network resilience enhancement method based on knowledge distillation specifically includes the following steps:

[0118] like Figure 3 As shown, a heterogeneous coupling diagram of the power-transportation coupled network is constructed based on the obtained parameters of the power-transportation coupled network. G =( V,E ) set of nodes V Includes distribution network nodes and transportation network nodes, edge set E This includes the distribution network edge set, which characterizes the power transmission relationships between distribution network nodes, and the transportation network edge set, which characterizes the road connection relationships between transportation network nodes. The formula can be expressed as:

[0119] ;

[0120] ;

[0121] In the formula: For distribution network nodes, As a node in the transportation network, E PG This is the set of edges in the distribution network, representing the power transmission relationships between nodes in the distribution network; E TN Let be the set of edges of the transportation network, representing the road connections between nodes in the transportation network.

[0122] The heterogeneous coupling graph of the power-transportation coupled network characterizes the multi-layer topology and attribute characteristics of the system under normal and fault scenarios. Among these, the distribution network characteristics are as follows: for the... u Distribution network nodes Its eigenvectors Includes features such as voltage, load timing data, line impedance, and switch status:

[0123] ;

[0124] In the formula: V u ( t ), P u ( t ), Q u ( t ) represent the time t, respectively u Voltage, active power, and reactive power of each distribution network node; Z u and S u For the first u Each distribution network node is associated with the line impedance and the status of the associated switches.

[0125] The characteristics of the transportation network are: targeting the first j Transportation network nodes eigenvectors Features include lane flow, waiting time, road capacity, and current traffic flow.

[0126] ;

[0127] In the formula: F j,l ( t ), T j,l ( t ) represent the time t, respectively j Connecting lanes of a transportation network node l Traffic and waiting time; F j,k ( t ), C k They represent the time t, respectively. j The associated road segments of each transportation network node k Traffic flow and maximum permitted traffic flow.

[0128] The aggregated feature vector of the power-transportation coupled network includes regular aggregated features and fault aggregated features; the structured data of the heterogeneous coupled graph of the power-transportation coupled network includes the adjacency matrix and the feature matrix.

[0129] A two-layer graph convolutional network is used to extract features from the heterogeneous coupling graph of the power-transportation coupled network, resulting in an aggregated feature vector for the power-transportation coupled network, including:

[0130] Structured data from the heterogeneous coupled graph of the power-transportation coupled network is input into a two-layer graph convolutional network, and conventional aggregated features are output through graph convolution operations.

[0131] Based on the obtained fault probability matrix of the power-transportation coupled network, the fault adjacency matrix is ​​calculated using the adjacency matrix.

[0132] Based on the fault probability matrix, the fault feature matrix is ​​calculated using the feature matrix.

[0133] The fault adjacency matrix and fault feature matrix are input into a two-layer graph convolutional network, and the fault aggregated features are output through graph convolution operations.

[0134] Before feature extraction, the graph convolution model needs to construct an adjacency matrix. The aggregated features output by the graph convolution operation include aggregated features for regular scenarios and aggregated features for fault scenarios, corresponding to regular scenarios and fault scenarios respectively. During graph data extraction, the adjacency matrix of the power-transportation coupled network graph is constructed. Let m and n be the total number of nodes in the distribution network and the transportation network, respectively, and R be the dimension of a real number. Its elements... a num1,num2 It can be represented as:

[0135] ;

[0136] Constructed feature matrix ,in ψ The number of features for each node, X, can be defined as:

[0137] ;

[0138] A two-layer Graph Convolutional Network (GCN) model extracts graph data features from a power-transportation coupled network. The stacking of two GCN layers captures deeper feature representations, giving the model better expressive power. The input to the two-layer GCN is graph-structured data, and the inter-layer propagation rule for graph convolution operations is as follows:

[0139] ;

[0140] In the formula: H (ρ) For the first ρ The node feature matrix of the layer, initially H (0) =X. Add a self-loop to the adjacency matrix A, where I is the identity matrix, ensuring that each node includes its own features; D is... The degree matrix, where the diagonal elements are the sum of the elements in the corresponding row (or column); W (ρ) For the first ρ The trainable weight matrix of the layer; σ This is the activation function.

[0141] After graph convolution, an aggregated feature vector is output. This process can be divided into two cases depending on the scenario:

[0142] Typical scenario: The input to GCN is the original graph structure data G=(A,X). After graph convolution operation, the output is a conventional aggregated feature Z. nor .

[0143] Fault Scenario: To incorporate the impact of fault propagation into feature extraction, the original adjacency matrix A and feature matrix X are first multiplied by the fault probability matrix P obtained in the previous step to obtain the fault adjacency matrix A. fault =A*P and fault feature matrix X fault =X*P. Then, this new pair of matrices is used as input to the GCN. The GCN convolves the fault graph data to ultimately obtain a fault aggregation feature Z that reflects network vulnerability and fault states. fau .

[0144] Based on the acquired power-traffic coupled network parameters, the failure probability of traffic segments and the fault probability of power lines are calculated respectively; the fault probability of charging stations is calculated based on the fault probability of power lines in the area where charging stations are located; a fault probability matrix is ​​constructed based on the failure probabilities of traffic segments, power lines, and charging stations; where traffic segments ( i , j Failure probability The calculation expression is:

[0145] ;

[0146] In the formula: q ij For traffic sections ( i , j Traffic flow, C ij For traffic sections ( i , j Basic traffic capacity This is the adjustment coefficient;

[0147] line( u , y Failure probability The calculation expression is:

[0148] ;

[0149] In the formula: S uy For the line ( u , y The trend of ) For the line ( u ,y Maximum transmission capacity α 1 is a parameter that controls the rate of change of probability. β 1 represents the overload threshold.

[0150] like Figure 4 The diagram illustrates the construction of a conventional Markov decision process and a faulty Markov decision process based on aggregated feature vectors. The state spaces of the conventional and faulty Markov decision processes are defined as follows: t The operational status of the power distribution network and transportation network under normal conditions. Fault scenario state ;in, This refers to the regular feature information extracted by a two-layer graph convolutional network. The fault feature information is extracted by a two-layer graph convolutional network. The deep topological features of the coupled system under fault disturbance are accurately extracted by GCN, and the complex coupling relationship is transformed into a quantifiable state space representation that reflects the vulnerability of the network, realizing the dynamic collaborative optimization of distribution network load restoration and tidal lane adjustment.

[0151] The action space of a conventional Markov decision process and a faulty Markov decision process is defined as the action space of the agent in... t Actions in real time; actions in typical scenarios Including routine power distribution network operations Traffic network actions in conventional scenarios ; where Δ P Indicates generator output adjustment, Δ V This indicates that the transformer tap changer adjusts the voltage according to the direction of the average voltage deviation across the entire network. d r Indicates the direction of the tidal flow lanes in the transportation network. R tid For the collection of tidal flow lanes, r For tidal flow lanes; actions during fault scenarios Including distribution network actions in fault scenarios Traffic network actions in fault scenarios ,in, δ uy Indicates the line ( u , y The switch status (0 indicates closed, 1 indicates open). ε PG This is the set of all controllable lines.

[0152] The reward function of an agent in a conventional Markov decision process is defined as:

[0153] ;

[0154] ;

[0155] ;

[0156] In the formula: For a typical scenario reward function, For regular distribution network rewards, Rewards for regular transportation networks;

[0157] Objective function of conventional distribution network f PG,nor Minimize the rewards for a conventional distribution network, including minimizing voltage deviation. L vol and minimize active power loss P los The specific expression is:

[0158] ;

[0159] ;

[0160] ;

[0161] In the formula: V u For the first u Voltage of each distribution network node u =1,2,…, M nod , M nod This represents the total number of distribution network nodes. V nom Rated voltage, R h and I h The first h The resistance and current of the line, h =1,2,…, N lin , N lin This represents the total number of lines.

[0162] Objective function of conventional transportation network f TN,nor Minimize the reward for a regular transportation network, including minimizing travel time. T tra And maximize road capacity C A The specific expression is:

[0163] ;

[0164] ;

[0165] ;

[0166] ;

[0167] ;

[0168] In the formula: x a It is a traffic section a Traffic flow on the road a ∈ A roa , A roa This includes all traffic sections. t a It is a travel time function. This refers to travel time in a free-flowing state. α and β For calibration parameters, C a It is a traffic section a Traffic capacity, and Traffic sections a The number of lanes and the saturation level of a single lane;

[0169] When the coupled network does not require emergency recovery, deep reinforcement learning methods are used to solve conventional Markov decision processes, including:

[0170] Based on the current operating status of the distribution network and transportation network, S11, the agent outputs control actions according to the power-transportation coupled network control strategy of the conventional scenario, with the reward function as the objective, and updates the operating status of the distribution network and the transportation network.

[0171] S12 stores the control actions, the updated distribution network operating status, and the updated transportation network operating status in the regular scenario experience pool. D nor ;

[0172] S13 Iteratively optimizes the control strategy for the power-transport coupled network in a conventional scenario based on the experience pool of conventional scenarios;

[0173] Repeat steps S11, S12 and S13 until the preset solution cutoff condition is met, and output the optimal power-transportation coupled network control strategy under normal scenarios.

[0174] The conventional action value function is constructed using the state space, action space, and reward function of a conventional Markov decision process. The action value function represents the long-term reward an agent can obtain by taking an action in the current state and following the policy.

[0175] In state Actions of power-transportation coupled system scheduling The quality can be evaluated using the expected sum of future rewards over K time steps, expressed as:

[0176] ;

[0177] In the formula: Represents the value function of conventional actions; The current power-transportation coupled dispatch strategy; This is a discount factor, representing the importance of future rewards relative to current rewards. This represents the calculation of the expectation, s and These represent the current state and the current action, respectively. ts The time step is the time step. The goal of this scheduling problem is to find the optimal strategy. To obtain the optimal action-value function in a typical scenario, Find the expectation function for the current strategy , max represents the function for finding the maximum value;

[0178] The reward function in a typical scenario includes the following specific constraints and objective function:

[0179] Objective function in a typical scenario

[0180] Under normal operating conditions, the power network layer should ensure low losses and high stability in the distribution network. The objective function of a conventional distribution network is... f PG,nor This mainly includes minimizing voltage deviation. L vol and minimize active power loss P los :

[0181] ;

[0182] ;

[0183] ;

[0184] In the formula: V u For the first u Voltage of each distribution network node u =1,2,…, Mnod , M nod This represents the total number of distribution network nodes. V nom Rated voltage, R h and I h The first h The resistance and current of the line, h =1,2,…, N lin , N lin This represents the total number of lines.

[0185] In the transportation network layer, under normal circumstances, the transportation network should reduce users' travel time, improve traffic efficiency, and reduce vehicle congestion. An objective function for the conventional transportation network is established. f TN,nor Minimize the reward for a regular transportation network, including minimizing travel time. T tra And maximize road capacity C A :

[0186] ;

[0187] ;

[0188] ;

[0189] ;

[0190] ;

[0191] In the formula: x a It is a traffic section a Traffic flow on the road a ∈ A roa , A roa This includes all traffic sections. t a It is a travel time function, using the classic BPR function (Bureau of Public Roads function). This refers to travel time in a free-flowing state. α and β These are the calibration parameters, where α =0.15, β =0.4; C a It is a traffic sectiona Traffic capacity; and Traffic sections a The number of lanes and the saturation level of a single lane.

[0192] Common scenario constraints

[0193] Power flow balance constraints in the power network layer:

[0194] ;

[0195] ;

[0196] In the formula: and The first u Active and reactive power generation at each distribution network node; and The first u Active and reactive loads of each distribution network node; V u For the first u Voltage of each distribution network node; θ uy = θ u - θ y For the first u Voltage phase angle difference between distribution network nodes and distribution network node y; G uy and B uy The first u Distribution network nodes and distribution network nodes y The real and imaginary parts of the admittance of the intermediate circuit; N PG Let be the set of nodes in the distribution network, cos and sin be the sine and cosine functions respectively, and || be the amplitude. Here, is the summation function, and || is the concatenation symbol.

[0197] Node voltage constraints:

[0198] ;

[0199] In the formula: and They are the first u The lower and upper limits of the voltage at each distribution network node.

[0200] Branch current constraints:

[0201] ;

[0202] In the formula: It is the first h The maximum allowable current for this line.

[0203] Traffic network layer flow balance constraints:

[0204] ;

[0205] In the formula: and To represent the nodes of the transportation network respectively u The set of entry and exit road sections for section 1; N TN For a set of traffic nodes, C a It is a traffic section a Traffic capacity.

[0206] The road capacity constraint is that the traffic flow does not exceed its capacity.

[0207] Tidal flow lane time restrictions: Under normal operating conditions, tidal flow lanes are directionally controlled according to set time periods, meaning that vehicles are only allowed to travel in the prescribed direction during a certain time period.

[0208] ;

[0209] In the formula: d ij ( t ) represents the traffic segment at time t. i , j The status of the tidal flow lanes on the road is indicated by 1 for forward traffic, -1 for reverse traffic, and 0 for closure. and These are the sets of time periods during which the tidal flow lanes are "open in the forward direction" and "open in the reverse direction." A collection of traffic sections with tidal flow characteristics.

[0210] Based on the conventional action value function of the optimal power-transportation coupled network control strategy in conventional scenarios, a soft-label dataset is generated using the softmax function.

[0211] ;

[0212] In the formula: s This is the current state; It is a temperature coefficient used to smooth probability distributions. This is a soft-label dataset. This is the value function for regular actions.

[0213] like Figure 4As shown, in response to a fault in the power-transportation coupled network, a deep Q-network algorithm based on transfer-reinforcement co-training is used to solve the Markov decision process for the fault scenario, based on a soft-label dataset, to obtain the optimal power-transportation coupled network control strategy under the fault scenario. The fault Markov decision process is based on the soft labels of the teacher model. Q Ter ( a t |s t The probability distribution of the softened control action is used to select the control action, and the value of the action is evaluated based on the control action. Q Stu ( a t |s t Simultaneously, a faulty Markov decision process is trained using the DQN algorithm. This training includes obtaining the environmental feedback reward and the next state after the output control action based on the DQN algorithm, and storing the state transition information in the fault scenario experience pool. D fau In this process, a small batch of samples is sampled from the experience pool of fault scenarios, and the DQN loss is calculated. The total policy loss is then obtained jointly, and gradient descent is used to minimize the loss, leading to the optimal power grid optimization scheme for the fault scenario. The construction and iteration process of the fault Markov Decision Process is repeated until the maximum iteration time is reached, at which point the optimal resilience enhancement strategy for the power grid is output.

[0214] ;

[0215] In the formula, The reward function for fault scenarios is provided by the distribution network. Transportation network rewards constitute;

[0216] Distribution network rewards for fault scenarios Including positive incentives for distribution networks and distribution network constraint penalty items :

[0217] ;

[0218] ;

[0219] ;

[0220] ;

[0221] ;

[0222] ;

[0223] ;

[0224] In the formula: The objective function is the function to be applied when a fault occurs in the distribution network. This is a logical discrimination function, specifically, it is 1 if the logical condition is met, and 0 otherwise; for t A network connectivity constraint penalty is introduced if the reconstructed network does not satisfy the radial structure requirement. for t Timetable and route ( u , y The on / off state of ); for t Penalty for limiting the number of operations at any given time, if t The number of switching actions during time-based control exceeds the allowable number Δ max If so, then this penalty will be introduced; As a penalty for exceeding the line power limit, for any line after reconfiguration ( u , y If the power exceeds the line ( u , y Maximum limit If so, then this penalty will be introduced; For distributed power source power constraint penalties, if t At any given time, the power of a distributed power source If the value is outside the allowed range, this penalty is imposed. for t The maximum power of a certain distributed power source at a given time. for t Minimum power of a certain distributed power source at a given moment; λ 1~ λ 4 is the penalty coefficient. N PG For the set of distribution network nodes, S uy,t for t Timetable and route ( u , y The power of )

[0225] Traffic network rewards for fault scenarios Including positive incentives for transportation networks and traffic network constraints and penalties :

[0226] ;

[0227] ;

[0228] ;

[0229] ;

[0230] ;

[0231] In the formula: Let be the objective function for traffic network congestion. λ 5 and λ 6 is the penalty coefficient. They are respectively t Penalties for violation of flow balance constraints at any time t Penalties for violating road capacity constraints at any time x a,t for t Traffic sections a Traffic flow on the road N TN For transportation network integration, and To represent the nodes of the transportation network respectively u The set of entry and exit road sections for section 1;

[0232] The objective function for the fault scenario includes:

[0233] When a fault occurs at the power network layer, the distribution network needs to restore power quickly, maximizing load recovery while minimizing the number of switching operations to reduce impact. This necessitates constructing an objective function for the faulty distribution network. f PG,fau The main goal is to maximize load recovery. P res,loa and minimize the number of switching operations N sw :

[0234] ;

[0235] ;

[0236] ;

[0237] In the formula: It is the first kp The power of the restored load, kp =1,2,…, L res , L res Total number of nodes restored; Switch mor Indicates the first mor The operating status of each switch (0 indicates no operation, 1 indicates operation); S sw For a set of switches.

[0238] When traffic congestion occurs at the traffic network layer, it is necessary to optimize the tidal lane direction adjustment strategy to reduce road load and improve traffic efficiency. Construct the objective function for the faulty traffic network. f TN,con The main goal is to minimize traffic delays. D del and minimizing traffic flow on congested road sections Q con :

[0239] ;

[0240] ;

[0241] ;

[0242] In the formula: A con It is a collection of congested road sections.

[0243] The constraints of the failure scenario include:

[0244] Power network layer topology reconfiguration constraints ensure that the reconfigured network maintains a radial structure. In the formula: For the line ( u , y The switch status (0 for open, 1 for closed).

[0245] Operation count limit constraint In the formula: and The lines before and after reconstruction are respectively u , y The on / off state of ).

[0246] Line capacity constraints In the formula: For the line after topology reconstruction ( u , y The power of ) For the line ( u , y (The maximum power of)

[0247] DG power constraint In the formula: and For the first u The minimum and maximum active power of distributed generation at each distribution network node.

[0248] Transportation network layer

[0249] Flow balance constraints ;

[0250] Road capacity constraints mean that the traffic flow on a road segment cannot exceed its capacity. C a It is a traffic section a Traffic capacity:

[0251] Restrictions on the number of tidal lanes open In the formula: N tid,max This refers to the maximum number of tidal flow lanes that can be opened simultaneously on the transportation network. d ji ( t )for t Traffic sections at all times ( j , i The status of the tidal lane on the road.

[0252] Based on a soft-label dataset, a deep Q-network algorithm based on transfer-reinforcement co-training is used to solve the Markov decision process for fault scenarios, including:

[0253] Construct a fault action value function based on the state space, action space, and reward function of the fault Markov decision process;

[0254] In state The following are the actions of the power-transportation coupled system scheduling. The quality can be evaluated using the expected sum of future rewards over K time steps, expressed as:

[0255] ;

[0256] In the formula: The value function representing the fault action; The current power-transportation coupled dispatch strategy; This is a discount factor, representing the importance of future rewards relative to current rewards. To find the expected function, s and These represent the current state and the current action, respectively. ts The time step is the time step. The goal of this scheduling problem is to find the optimal strategy. To obtain the optimal action-value function under the fault scenario, Find the expectation function for the current strategy , max represents the function for finding the maximum value.

[0257] The knowledge distillation loss function is calculated based on the fault action value function and the soft-label dataset.

[0258] Sample data for solving Markov decision processes are obtained based on the reinforcement learning DQN algorithm, and an experience pool for fault scenarios is constructed.

[0259] The reinforcement learning loss is calculated based on the sample data obtained from the experience pool of fault scenarios.

[0260] Construct a total policy loss function based on reinforcement learning loss and knowledge distillation loss;

[0261] Based on the overall policy loss function, the optimal power-transportation coupled network control strategy under fault scenarios is obtained by minimizing the loss through gradient descent.

[0262] The expression for the total policy loss function in the failure scenario is:

[0263] ;

[0264] In the formula: α 2 is a hyperparameter used to balance the importance of the loss from knowledge distillation and reinforcement learning. L total Let be the total policy loss function under failure scenarios. θ fau For the network parameters of a faulty Markov decision process, L KD The knowledge distillation loss function is calculated using KL divergence:

[0265] ;

[0266] in, D KL To find the KL (Kullback-Leibler Divergence) divergence function, This is a soft-label dataset. The fault-based policy distribution is obtained through the fault action value function, where || represents concatenation;

[0267] L RL The formula for calculating the reinforcement learning loss function is as follows:

[0268] ;

[0269] Among them, the target value y target The calculation method is as follows:

[0270] ;

[0271] In the formula: The current action is performed in the current state s. The instant reward obtained afterward The next state is... The next move is... It is the fault action value function. It is the target fault action value function.

[0272] Based on the optimal power-traffic coupled network control strategy under normal scenarios and the optimal power-traffic coupled network control strategy under fault scenarios, the control parameters of the power-traffic coupled network are optimized.

[0273] The power-transportation coupled network control parameters include: generator output parameters and voltage control parameters of the distribution network, and tidal lane direction control parameters of the transportation network.

[0274] like Figure 4 The KDQN algorithm shown in this embodiment, with a teacher-student collaborative training structure at its core, achieves the dual goals of knowledge transfer and policy optimization by introducing policy distillation and DQN mechanisms. This method balances policy stability, model lightweightness, and cross-scenario generalization ability, providing an efficient and deployable intelligent decision-making solution for Markov decision process models, and obtaining optimal power network switching action strategies and traffic network tidal lane direction change strategies.

[0275] In summary, this embodiment includes a teacher model for conventional scenarios and a student model for fault scenarios. By coupling constraints, it achieves coordinated optimization of power network reconfiguration and traffic flow control, overcoming the decision-making conflict problem caused by independent optimization of power network and traffic network systems in traditional methods.

[0276] Example 3: This example also provides a computer storage medium applicable to the knowledge distillation-based method for improving the resilience of power-transportation coupled networks, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the knowledge distillation-based method for improving the resilience of power-transportation coupled networks as proposed in the above examples.

[0277] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0278] This embodiment proceeds by: constructing a conventional Markov decision process and a fault Markov decision process based on the acquired power-transportation coupled network parameters; solving the conventional Markov decision process using deep reinforcement learning to obtain the optimal power-transportation coupled network control strategy and soft-label dataset for the conventional scenario; solving the fault scenario Markov decision process using the soft-label dataset and deep reinforcement learning to obtain the optimal power-transportation coupled network control strategy for the fault scenario; and optimizing the power-transportation coupled network control parameters based on the optimal power-transportation coupled network control strategy for the conventional scenario and the optimal power-transportation coupled network control strategy for the fault scenario.

[0279] In summary, this invention achieves coordinated optimization of power network reconfiguration and traffic flow control through coupled constraints, enabling the selection of the optimal power-traffic coupled network strategy under multiple optimization objectives. By constructing a soft-label dataset using the solved conventional Markov decision process, the efficiency of solving the fault Markov decision process is improved. Furthermore, by controlling the generator output parameters, voltage control parameters, and tidal lane direction control parameters of the power-traffic coupled network, this invention enhances the recovery speed of the power-traffic coupled network during fault occurrences and improves its resilience.

[0280] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0281] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0282] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0283] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0284] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for improving the resilience of a power-transportation coupled network, characterized in that, include: Based on the acquired power-transportation coupled network parameters, a heterogeneous coupling graph of the power-transportation coupled network is constructed; A two-layer graph convolutional network is used to extract features from the heterogeneous coupling graph of the power-transportation coupling network to obtain the aggregated feature vector of the power-transportation coupling network. Based on the aggregated feature vectors, a conventional Markov decision process and a faulty Markov decision process are constructed respectively. We use deep reinforcement learning to solve the conventional Markov decision process and obtain the optimal power-transportation coupled network control strategy under conventional scenarios. Based on the conventional action value function of the optimal power-transportation coupled network control strategy in conventional scenarios, a soft-label dataset is generated using the softmax function. In response to a fault in the power-transportation coupled network, based on a soft-label dataset, a deep Q-network algorithm using knowledge distillation based on transfer-reinforcement co-training is employed to solve the Markov decision process for the fault scenario, yielding the optimal power-transportation coupled network control strategy under the fault scenario, including: Construct a fault action value function based on the state space, action space, and reward function of the fault Markov decision process; The knowledge distillation loss function is calculated based on the fault action value function and the soft-label dataset. Sample data for solving Markov decision processes are obtained based on the reinforcement learning DQN algorithm, and an experience pool for fault scenarios is constructed. The reinforcement learning loss is calculated based on the sample data obtained from the experience pool of fault scenarios. Construct a total policy loss function based on reinforcement learning loss and knowledge distillation loss; Based on the overall policy loss function, the optimal power-transportation coupled network control strategy under the fault scenario is obtained by minimizing the loss through gradient descent; wherein, the knowledge distillation deep Q network algorithm based on transfer-reinforcement co-training solves the Markov decision process of the fault scenario through the knowledge distillation method and the deep Q network method in a coordinated manner. The reward function of a faulty Markov decision process is defined as: ； In the formula, For fault scenario reward function; Distribution network rewards for fault scenarios Including positive incentives for distribution networks and distribution network constraint penalty items : ； ； ； ； ； ； ； In the formula: The objective function is the function to be applied when a fault occurs in the distribution network. It is a logical discrimination function; The penalty for network connectivity constraints at time t is introduced if the reconstructed network does not satisfy the radial structure. Let (u,y) be the switching state of line (u,y) at time t; The penalty for the number of operations at time t is the number of switching actions during control if it exceeds the allowed number Δ. max If so, then this penalty will be introduced; As a penalty for exceeding the line power limit, for any reconfigured line (u, y), if the power exceeds the maximum limit of line (u, y), a penalty will be imposed. If so, then this penalty will be introduced; As a distributed source power constraint penalty, if the power of a certain distributed source at time t... If the value is outside the allowed range, this penalty is imposed. Let t be the maximum power of a certain distributed power source. for The minimum power of a distributed power source at a given time; λ1~λ4 are penalty coefficients, N PG S is the set of distribution network nodes. uy,t Let be the power of line (u, y) at time t; Traffic network rewards for fault scenarios Including positive incentives for transportation networks and traffic network constraints and penalties : ； ； ； ； ； In the formula: Let be the objective function for traffic network congestion, and λ5 and λ6 be penalty coefficients. These represent the penalties for violating the flow balance constraint at time t and the penalties for violating the road capacity constraint at time t, respectively. a,t Let N be the traffic flow on road segment a at time t. TN For transportation network integration, and Let represent the sets of entry and exit road segments for node u1 in the transportation network, respectively; The expression for the total policy loss function is: ； In the formula: α2 is a hyperparameter used to balance the importance of the losses from knowledge distillation and reinforcement learning; L total Let θ be the total policy loss function under fault scenarios. fau For the network parameters of a faulty Markov decision process, L KD The knowledge distillation loss function is calculated using KL divergence: ； Among them, D KL To find the KL divergence function, This is a soft-labeled dataset. The fault-based policy distribution is obtained through the fault action value function, where || represents concatenation; L RL The formula for calculating the reinforcement learning loss function is as follows: ； Wherein, the target value y target The calculation method is as follows: ； In the formula: The current action is performed in the current state s. The instant reward obtained afterward It is a discount factor. The next state is... The next move is... It is the fault action value function. It is the target fault action value function. This indicates the intention to obtain the expected value; Based on the optimal power-traffic coupled network control strategy under normal scenarios and the optimal power-traffic coupled network control strategy under fault scenarios, the control parameters of the power-traffic coupled network are optimized to achieve enhanced resilience of the power-traffic coupled network. Among them, the power-traffic coupled network control parameters include: power output parameters of distribution network generators, voltage control parameters, and tidal lane direction control parameters of traffic network.

2. The method for improving the resilience of power-transportation coupled networks according to claim 1, characterized in that, In the heterogeneous coupled graph G=(V,E) of the power-transportation coupled network, the node set V includes distribution network nodes and transportation network nodes, and the edge set E includes the distribution network edge set used to characterize the power transmission relationship between distribution network nodes and the transportation network edge set used to characterize the road connection relationship between transportation network nodes.

3. The method for improving the resilience of power-transportation coupled networks according to claim 1, characterized in that, The aggregated feature vector of the power-transportation coupled network includes conventional aggregated features and fault aggregated features; The structured data of the heterogeneous coupling graph of the power-transportation coupled network includes the adjacency matrix and the feature matrix; A two-layer graph convolutional network is used to extract features from the heterogeneous coupling graph of the power-transportation coupled network, resulting in an aggregated feature vector for the power-transportation coupled network, including: Structured data from the heterogeneous coupled graph of the power-transportation coupled network is input into a two-layer graph convolutional network, and conventional aggregated features are output through graph convolution operations. Based on the fault probability matrix of the power-transportation coupled network, the fault adjacency matrix is ​​calculated using the adjacency matrix. Based on the fault probability matrix, the fault feature matrix is ​​calculated using the feature matrix. The fault adjacency matrix and fault feature matrix are input into a two-layer graph convolutional network, and the fault aggregation features are output through graph convolution operations.

4. The method for improving the resilience of power-transportation coupled networks according to claim 3, characterized in that, The construction of the failure probability matrix includes: The failure probability of traffic segments and the fault probability of lines are calculated based on the obtained power-transportation coupled network parameters. The failure probability of the charging station is calculated based on the failure probability of the lines in the area where the charging station is located. A failure probability matrix is ​​constructed based on the failure probability of traffic segments, the failure probability of lines, and the failure probability of charging stations. Among them, the failure probability of traffic segment (i,j) The calculation expression is: ； In the formula: q ij Let C be the traffic flow of traffic segment (i,j). ij Let (i,j) be the basic traffic capacity of traffic segment (i,j). For adjustment coefficients, Fault probability of line (u,y) The calculation expression is: ； In the formula: S uy Let (u,y) be the power flow of the line. Let α1 be the maximum transmission capacity of line (u,y), α1 be the parameter controlling the rate of change of probability, and β1 be the overload threshold.

5. The method for improving the resilience of power-transportation coupled networks according to claim 3, characterized in that, The state space of a conventional Markov decision process and a faulty Markov decision process is defined as the operating state of the distribution network and transportation network at time t. Under conventional scenarios, the state... Fault scenario state ;in, These are the regular aggregated features extracted by a two-layer graph convolutional network. Fault aggregation features extracted by a two-layer graph convolutional network; The action space of a conventional Markov decision process and a faulty Markov decision process is defined as the action of the agent at time t; the action in a conventional scenario... Including routine power distribution network operations Traffic network actions in conventional scenarios Where ΔP represents generator output adjustment, ΔV represents transformer tap adjustment of voltage according to the average deviation direction of the entire grid voltage, and d r Indicates the direction of the tidal flow lanes in the traffic network, R tid For the set of tidal lanes, r represents the tidal lane; actions in fault scenarios. Including distribution network actions in fault scenarios Traffic network actions in fault scenarios ; where δ uy ε represents the switching state of line (u,y). PG This is the set of all controllable lines.

6. The method for improving the resilience of power-transportation coupled networks according to claim 5, characterized in that, The reward function of an agent in a conventional Markov decision process is defined as: ； ； ； In the formula: For a typical scenario reward function, For regular distribution network rewards, Rewards for regular transportation networks; The objective function f of a conventional power distribution network PG,nor Minimize the reward for a conventional distribution network, including minimizing the voltage deviation L. vol and minimize active power loss P los The specific expression is: ； ； ； In the formula: V u Let u be the voltage of the u-th distribution network node, where u = 1, 2, ..., M nod M nod V represents the total number of nodes in the distribution network. nom R is the rated voltage. h and I h Let N be the resistance and current of the h-th line, where h = 1, 2, ..., N. lin N lin This represents the total number of lines; Objective function f of conventional transportation network TN,nor Minimize the reward for a regular transportation network, including minimizing the travel time T. tra And maximizing road capacity C A The specific expression is: ； ； ； ； ； In the formula: x a It is the traffic flow on traffic segment a, where a∈A roa A roa For the set of all traffic segments, t a It is a travel time function. The travel time is in free-flow condition, α and β are calibration parameters, and C is the travel time in free-flow condition. a It is the traffic capacity of traffic segment a. and These represent the number of lanes on traffic segment a and the saturation level of a single lane, respectively. Solving conventional Markov decision processes using deep reinforcement learning methods includes: Based on the current operating status of the distribution network and transportation network, S11, the agent outputs control actions according to the power-transportation coupled network control strategy of the conventional scenario, with the reward function as the objective, and updates the operating status of the distribution network and the transportation network. S12 stores the control action, the updated power distribution network operation status, and the updated transportation network operation status in the conventional scenario experience pool. S13 Iteratively optimizes the control strategy for the power-transport coupled network in the conventional scenario based on the experience pool of the conventional scenario; Repeat steps S11, S12 and S13 until the preset solution cutoff condition is met, and output the optimal power-transportation coupled network control strategy under normal scenarios. The conventional action value function is constructed using the state space, action space, and reward function of a conventional Markov decision process.

7. The method for improving the resilience of power-transportation coupled networks according to claim 5, characterized in that, Based on the conventional action value function, the calculation formula for generating a soft-labeled dataset using the softmax function is as follows: ； In the formula: s represents the current state. It is a temperature coefficient used to smooth probability distributions. This is a soft-labeled dataset. This is the value function for regular actions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method for improving the resilience of power-transportation coupled networks as described in any one of claims 1 to 7.

Images