Multi-type mobile emergency resource optimization scheduling and power distribution network toughness improvement strategy based on data-model hybrid driving method

Abstract

The invention provides a multi-type mobile emergency resource optimization scheduling and power distribution network toughness improvement strategy based on a data-model hybrid driving method, and the strategy starts from the dynamic interaction relationship between a mobile energy storage vehicle and an emergency repair team, and coordinates the mobile energy storage vehicle and the emergency repair team to carry out load recovery and road maintenance. Scheduling an emergency repair team to repair the damaged road of the traffic network under the condition of considering that the duration required for repairing the road is uncertain; scheduling the mobile energy storage vehicle to carry out load recovery on the power distribution network nodes under the condition of considering the dynamic change of the road state of the traffic network; the data driving part is supported by a graph neural network multi-agent reinforcement learning algorithm GATD3QN based on a graph attention mechanism so as to solve an optimal multi-type mobile emergency resource routing behavior strategy; and the model driving part constructs the scheduling model of the mobile energy storage vehicle and the emergency repair team, the power distribution network reconstruction model and the power distribution network optimal power flow model into an MISOCP model and solves an optimal multi-type mobile emergency resource scheduling behavior strategy.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network resilience enhancement, and specifically relates to a multi-type mobile emergency resource optimization scheduling and distribution network resilience enhancement strategy based on a data-model hybrid driving method. Background Art

[0002] As one of the key infrastructures, the distribution system is the basis for the normal operation of other infrastructures. Its safe and stable operation plays an extremely important role in the development of the national economy and social stability. However, in recent years, extreme events such as natural disasters and man-made attacks are increasingly threatening the safety of the distribution system. Extreme events have a wide range of impacts on the distribution network. In order to minimize the impact of extreme events, it is necessary to further improve the safe and stable operation capabilities of the distribution network. Research on strategies to improve the resilience of the distribution network has become an inevitable trend in the new era.

[0003] With the rapid development of urban electrified transportation, the number of charging stations in the transportation network has increased sharply, the energy exchange capacity between the transportation network and the distribution network has been continuously improved, and the mobile energy storage energy flowing in the transportation network has been continuously increased, forming a discrete energy network with huge scale, diverse forms and flexible topology, which provides a material and energy basis for enhancing the resilience of the distribution network. Mobile energy storage vehicles are vehicles that can store and flexibly transport electrical energy to different locations. In the event of natural disasters or power outages, mobile energy storage vehicles can quickly access charging station nodes to provide temporary power supply for important facilities and equipment. They are the most widely used emergency power supply resources for power grids. However, extreme events can also directly or indirectly affect the transportation network, causing road damage, traffic paralysis, etc. This greatly limits the mobility of mobile energy storage vehicles in the transportation network and affects the improvement of the resilience of the distribution network. Therefore, it is necessary to dispatch emergency repair teams to repair damaged roads. In the existing research on improving the resilience of distribution networks, most of them assume that the repair time of road faults in the transportation network is a fixed time, or even assume that the fault always exists to carry out resilience improvement research, which is inconsistent with the actual situation. In addition, existing research also ignores the impact of the dynamic interaction relationship of mobile emergency resources on their optimal scheduling decisions. Some studies choose pure model methods to solve the mathematical model of random optimization scheduling of mobile emergency resources. Although this method has the advantages of strong explanatory power and can obtain more accurate results, it has poor generalization ability and is affected by a large number of integer variables and nonlinear constraints, resulting in poor real-time performance and cannot meet the real-time decision-making needs of actual mobile emergency resources. Some technologies apply reinforcement learning methods (pure data methods) to the optimization and scheduling of mobile emergency resources, so that the intelligent agent has the ability to make real-time decisions, but there are shortcomings in the training process, such as long training time and the possibility of solving infeasible strategies. Summary of the invention

[0004] Therefore, in view of the defects and shortcomings of the above-mentioned existing technologies, an efficient and accurate method is needed to solve the multi-type mobile emergency resource scheduling strategy that takes into account the dynamic changes of traffic network road status and adjacency relationships, so as to improve the resilience of the distribution network after extreme events.

[0005] In view of this, the purpose of the present invention is to provide a multi-type mobile emergency resource scheduling and distribution network resilience improvement strategy based on a data-model hybrid drive method. The strategy starts from the dynamic interactive relationship between mobile energy storage vehicles and emergency repair teams, and coordinates mobile energy storage vehicles and emergency repair teams to perform load recovery and road maintenance. The strategy can consider the uncertainty of the time required to repair the road and dispatch the emergency repair team to repair the damaged roads of the traffic network. At the same time, it can consider the dynamic changes in the road status of the traffic network and dispatch mobile energy storage vehicles to perform load recovery on the distribution network nodes. In addition, this strategy innovatively combines data methods and model methods to solve the above-mentioned complex, high-dimensional, nonlinear, multi-type mobile emergency resource scheduling problem; among them, the data-driven part is supported by the graph neural network multi-agent reinforcement learning algorithm GATD3QN based on the graph attention mechanism. This algorithm combines the advantages of the graph neural network based on the attention mechanism, the Double DQN algorithm and the DuelingDQN algorithm to optimize the feature extraction ability and Q-value calculation of the traditional DQN algorithm, and adopts the priority experience replay strategy to improve the sampling efficiency and training effect of the algorithm. This algorithm can deeply mine the characteristic information in the power-transportation coupling network graph data and the adjacency relationship of mobile emergency resources, and solve the optimal multi-type mobile emergency resource routing behavior strategy; the model-driven part constructs the scheduling model of the mobile energy storage vehicle and the emergency repair team, the distribution network reconstruction model and the distribution network optimal power flow model into a MISOCP model and solves the optimal multi-type mobile emergency resource scheduling behavior strategy, including the power scheduling decision of the mobile energy storage vehicle, the road repair decision of the emergency repair team or the return to the warehouse decision, so as to achieve the improvement of the resilience of the distribution network.

[0006] The technical solution specifically adopted by the present invention to solve the technical problem is:

[0007] A multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on a data-model hybrid driven method. Starting from the dynamic interaction between mobile energy storage vehicles and emergency repair teams, the mobile energy storage vehicles and emergency repair teams are coordinated to restore load and repair roads at the same time. The emergency repair team is dispatched to repair damaged roads in the traffic network under the uncertainty of the time required to repair the roads. The mobile energy storage vehicles are dispatched to restore loads on distribution network nodes under the dynamic changes in the road status of the traffic network. The data method and the model method are combined to solve the above multi-type mobile emergency resource scheduling problems. The data-driven part is supported by the graph neural network multi-agent reinforcement learning algorithm GATD3QN based on the graph attention mechanism. The GATD3QN algorithm combines the graph neural network based on the attention mechanism, the Double DQN algorithm and the Dueling The DQN algorithm optimizes the feature extraction capability and Q-value calculation of the traditional DQN algorithm, and adopts the priority experience replay strategy to improve the sampling efficiency and training effect of the algorithm, so as to solve the optimal multi-type mobile emergency resource routing behavior strategy; the model-driven part constructs the scheduling model of the mobile energy storage vehicle and the emergency repair team, the distribution network reconstruction model and the distribution network optimal power flow model into a MISOCP model and solves the optimal multi-type mobile emergency resource scheduling behavior strategy, including the power scheduling decision of the mobile energy storage vehicle, the road repair decision of the emergency repair team or the return to the warehouse decision, so as to improve the resilience of the distribution network.

[0008] Furthermore, the strategy includes the following steps when executed:

[0009] Step S1: construct and initialize the power-traffic coupling network model, the mobile energy storage vehicle graph reinforcement learning model and the emergency repair team graph reinforcement learning model;

[0010] Step S2: Simulate the power-transportation coupling network failure scenario after the extreme event occurs, and set line disconnection in the distribution network and road damage in the transportation network;

[0011] Step S3: Initialize the graph reinforcement learning algorithm environment and the power-transportation coupling network fault scenario;

[0012] Step S4: Construct the mobile energy storage vehicle state of the graph neural network reinforcement learning algorithm based on the power-traffic coupling network information and the mobile emergency resource state information and the status of the emergency repair team Get the state matrix o t , construct the adjacency matrix A according to the adjacency relationship of mobile emergency resources t ;

[0013] Step S5: Calculate the mobile energy storage vehicle routing behavior strategy based on the ε-Greedy strategy and the graph neural network reinforcement learning algorithm as the data-driven part and emergency repair team routing behavior strategy

[0014] Step S6: Execute the mobile energy storage vehicle routing behavior strategy respectively and emergency repair team routing behavior strategy And judge and update the status of mobile energy storage vehicles and emergency repair teams;

[0015] Step S7: The model driving part calculates the power dispatching strategy and reward function r of the mobile energy storage vehicle according to the mobile energy storage vehicle dispatching model combined with the distribution network reconstruction model and the distribution network optimal power flow model. t M , calculate the maintenance scheduling strategy and reward function r of the emergency repair team according to the emergency repair team scheduling model t R ;

[0016] Step S8: Update the state of the mobile energy storage vehicle of the graph neural network reinforcement learning algorithm and the status of the emergency repair team Get the next state matrix o t+1 , update the adjacency matrix A t+1 ;

[0017] Step S9: The information of the current step and initial priority Stored in the mobile energy storage vehicle Figure reinforcement learning memory unit D M ,Will and initial priority Stored in the emergency repair team map reinforcement learning memory unit D R ;

[0018] Step S10: The priority experience replay strategy is used to sample the importance of the memory unit, and the weights of the graph neural network of the mobile energy storage vehicle and the emergency repair team are updated based on the stochastic gradient descent method;

[0019] Step S11: Determine whether the end time T has been reached End , if not, execute steps S3 to S10;

[0020] Step S12: Determine whether the training end episode number E has been reached End If not, execute steps S3 to S11; if yes, output the graph reinforcement learning algorithm parameters, routing and scheduling strategies of the mobile energy storage vehicle and the emergency repair team.

[0021] Furthermore, in step S1, the power-transportation coupling network model, the mobile energy storage vehicle graph reinforcement learning model and the emergency repair team graph reinforcement learning model are initialized, and the number of training episodes is initialized.

[0022] The nodes in the distribution network and the lines connecting the nodes are regarded as points and edges, forming a distribution network graph. Charging station nodes are added to form a distribution network graph model. Where N PDS represents the set of distribution network nodes, L PDS represents the set of distribution network lines, Represents a collection of charging station nodes;

[0023] The nodes in the traffic network and the roads connecting the nodes are regarded as points and edges to form a basic traffic network graph; charging station nodes and emergency repair team warehouses are added to form a traffic network graph model Where N TN represents the set of nodes in the transportation network; L TN Represents a collection of transportation network roads; Represents the set of charging station nodes in the transportation network; Represents the set of repair team warehouse nodes in the transportation network;

[0024] All mobile energy storage vehicles and emergency repair teams are regarded as agents, and each agent is regarded as a node, and their connections are regarded as edges to construct a dynamic network model of mobile emergency resources. Where M is the mobile energy storage vehicle set, W is the emergency repair team set, Mobile emergency resource adjacency The adjacency matrix constructed.

[0025] Furthermore, in step S2, some lines are randomly interrupted in the distribution network to construct a damaged line set L PDS,off , which is used to simulate the interruption of weak lines in the distribution network after an extreme event; and randomly set some road damage in the traffic network to construct a damaged road set This is to simulate the situation where some roads in the transportation network are blocked by obstacles caused by extreme events.

[0026] Furthermore, step S3 specifically includes: initializing various variables in the graph reinforcement learning algorithm environment and initializing the power-transportation coupling network fault scenario, and initializing the training time.

[0027] Further, in step S4:

[0028] According to the information of the power-traffic coupling network and the state information of the mobile emergency resources themselves, the state of the mobile energy storage vehicle graph neural network reinforcement learning algorithm is constructed Including the mobile energy storage vehicle's own status information Adjacent mobile emergency resource status information Traffic network information of your location Distribution network node load data P t ED、New Energy Output P t WT ;

[0029] According to the information of the power-traffic coupling network and the status information of the mobile emergency resources themselves, the state of the emergency repair team graph neural network reinforcement learning algorithm is constructed Including the mobile energy storage vehicle's own status information Adjacent mobile emergency resource status information Traffic network information of your location Traffic network road fault information and the length of time it takes to repair damaged roads

[0030] Then, the adjacency matrix A is constructed based on the adjacency relationship of mobile emergency resources. t , the adjacency matrix A t Used to input into the graph reinforcement learning algorithm to participate in the calculation of state-action values.

[0031] Further, in step S5, first, the random number p at time t is calculated t , whose value range is [0,1), and the ε-Greedy strategy is used to compare the current ε t value and random number p t , if p t <ε t , a random method is used to generate the mobile energy storage vehicle routing behavior strategy and emergency repair team routing behavior strategy

[0032] If there is p t ≥ε t , the routing behavior strategy of the mobile energy storage vehicle is generated through the graph neural network multi-agent reinforcement learning algorithm and emergency repair team routing behavior strategy

[0033] The neural network structure used in the mobile energy storage vehicle graph reinforcement learning algorithm and the emergency repair team graph reinforcement learning algorithm is the same: the same neural network is used to control multiple mobile energy storage vehicles, and another neural network is used to control multiple emergency repair teams; the neural network structure is: first, a fully connected input layer is used to input the mobile emergency resource state set o t Perform preliminary feature extraction and dimension conversion, output This is followed by two layers of graph attention network layers. These two layers of networks use the attention mechanism to and the mobile emergency resource adjacency matrix A t Calculate the attention coefficient and extract features based on the attention matrix. Output is based on adjacency relationship. Features The last layer adopts the strategy of Dueling DQN algorithm to split the feature information into two fully connected branches: the first branch outputs the state function scalar value V(o t ,A t ); The second branch output status The action advantage value function vector under

[0034] Furthermore, in step S7, a mobile energy storage vehicle dispatching model, a distribution network reconstruction model and a distribution network optimal power flow model are constructed.

[0035] A dispatch model for emergency repair teams was constructed;

[0036] Construct an objective function based on the tasks of restoring the load at the distribution network nodes and repairing the damaged roads in the transportation network;

[0037] In the calculation process, constraints (34) and (35) are linearized by the big M method, and constraint (39) is linearized by the second-order cone relaxation method; constraints (46) and (47) of the emergency repair team scheduling problem are linearized by the McCormick envelope method;

[0038] The model driving part uses formula (51)-(53) as the objective function and formula (20)-(50) as the constraint to construct the mixed integer second-order cone programming model MISOCP, which is solved by Gurobi solver. The load of the distribution network node restored by the mobile energy storage vehicle is obtained, and the mobile energy storage vehicle reward function is calculated by formula (54):

[0039]

[0040] In the formula, represents the set of distribution network nodes supported by the mth mobile energy storage vehicle at time t; is the distribution network node at time t The load recovery amount; For distribution network nodes The load level weight coefficient; the reward value of the mobile energy storage vehicle during the movement between charging stations in the transportation network is 0;

[0041] By solving MISOCP, we can obtain the road maintenance strategy or dispatch strategy of the emergency repair team and calculate the reward function of the emergency repair team.

[0042] Further, in step S10, in the mobile energy storage vehicle image reinforcement learning memory unit D M In the importance sampling method, samples N are collected respectively s , and calculate the importance weight of each sample using formula (62):

[0043]

[0044] Where N D Represents memory unit D M The total number of samples in ; Indicates the priority of the sample; β is a hyperparameter;

[0045] Combined with the strategy of the Double DQN algorithm, formulas (63) and (64) are used to calculate the loss function:

[0046]

[0047] In the formula, and are the current network parameters and target network parameters of the mobile energy storage vehicle respectively; γ is the discount factor reflecting the impact of the future Q value on the current action, and its value is [0,1];

[0048] The neural network parameters of the mobile energy storage vehicle graph reinforcement learning algorithm are updated based on the stochastic gradient descent method and sample importance weights, which is specifically expressed as follows:

[0049]

[0050] In the formula, α M is the learning rate of the gradient descent algorithm of the current network of the mobile energy storage vehicle; and after a fixed number of steps N up Then, the parameters of the target network of the mobile energy storage vehicle are adjusted in the following way. To update:

[0051]

[0052] The TD-error value calculated according to formula (63) Calculate the new priority for each sample Among them, ∈ is a constant used to prevent the priority from being zero; in the mobile energy storage vehicle, the reinforcement learning memory unit D M Update the priority of each sample in;

[0053] Update the parameters of the neural network of the emergency repair team graph reinforcement learning algorithm.

[0054] And, an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that when the processor executes the program, the steps of optimizing the scheduling of multi-type mobile emergency resources and improving the resilience of the distribution network based on the data-model hybrid driving method as described above are implemented.

[0055] Compared with the prior art, the present invention and its preferred solution can solve the complex high-dimensional nonlinear multi-type mobile emergency resource random optimization scheduling problem, and its beneficial effects include at least:

[0056] Firstly, the present invention dispatches an emergency repair team to repair the faulty roads in the traffic network, taking into account the impact of the dynamic changes in the road status of the traffic network on the improvement of the resilience of the distribution network, avoiding the problems of ignoring road damage or assuming that the road is repaired in a fixed time in the prior art. Compared with the prior art, the present invention can be more effectively used in the actual post-disaster power-traffic coupling network fault repair, and obtain better distribution network resilience improvement effect.

[0057] Secondly, the present invention deeply studies the impact of dynamic interaction between multiple types of mobile emergency resources on their coordinated scheduling. All mobile emergency resources (i.e. mobile energy storage vehicles and emergency repair teams) are regarded as agents, and the dynamic adjacency relationships of all agents are abstracted as edges, and then the coordination and cooperation relationships of all mobile energy storage vehicles and emergency repair team agents are converted into a dynamic graph structure of mobile emergency resources; at the same time, effective feature information is mined from the changes in traffic flow caused by the movement of mobile emergency resources in the transportation network and the changes in adjacency relationships in the dynamic graph structure of mobile emergency resources to solve the optimal routing and scheduling strategies.

[0058] Finally, the present invention combines the high efficiency of the data-driven method with the accuracy of the model-driven method to propose a new data-model hybrid driven method. This method can more effectively solve the mathematical model of optimal scheduling of multi-type mobile emergency resources considering uncertainty factors such as uncertainty in traffic road repair time, dynamic changes in adjacency relationships, changes in traffic flow, and the time required to repair roads. In the data-driven part, in order to study the dynamic interactive relationship of mobile emergency resources, this method constructs a new graph attention neural network multi-agent reinforcement learning algorithm GATD3QN. Compared with the prior art, this algorithm abandons the use of traditional fully connected neural networks, and instead uses a graph neural network based on an attention mechanism to extract features from complex graph information and dynamic adjacency relationships. Compared with traditional fully connected neural networks, graph attention neural networks can focus on the most relevant adjacent nodes in dynamic graphs and effectively aggregate information by introducing node-level attention, which improves the model's ability to capture important local and global patterns, thereby enhancing the ability of mobile emergency resources to mine effective information from adjacent agents. In addition, in the face of the shortcomings of the traditional DQN algorithm such as over-estimation of Q value, training instability and low sample efficiency, the GATD3QN algorithm combines the advantages of the Double DQN algorithm and the Dueling DQN algorithm to optimize the feature extraction ability and Q value calculation of the traditional DQN algorithm, and adopts the Prioritized Experience Replay (PER) strategy to improve the sampling efficiency and training effect of the algorithm (corresponding to the design of the GATD3QN algorithm sampling and calculating TD-error from the sample experience pool for reverse updating of the neural network parameters). The GATD3QN algorithm used in the data-driven part can combine the excellent non-Euclidean data information extraction ability of the graph attention network and the powerful sequential stochastic optimization decision-making ability of the reinforcement learning algorithm, and effectively extract key information from the dynamic adjacency relationship between the mobile energy storage vehicle and the emergency repair team and the power-transportation coupling network graph data to formulate routing behavior strategies for the mobile energy storage vehicle and the emergency repair team; in the model-driven part, this method constructs the scheduling model of the mobile energy storage vehicle and the emergency repair team, the distribution network reconstruction model and the distribution network optimal power flow model into a MISOCP model and solves the optimal multi-type mobile emergency resource scheduling behavior strategy, including the power scheduling decision of the mobile energy storage vehicle, the road repair decision of the emergency repair team or the return to the warehouse decision.

[0059] In summary, the present invention can simultaneously consider the dynamic changes in the road status of the transportation network and the dynamic changes in the relationship between adjacent mobile emergency resources. The routing and scheduling strategies of mobile energy storage vehicles and emergency repair teams solved by the data-model hybrid driving method can restore the node load of the distribution network and repair damaged roads in the transportation network, thereby improving the resilience of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0061] Figure 1 This is a flow chart of the strategy implementation of an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to make the features and advantages of this patent more obvious and easy to understand, the following embodiments are specifically described in detail as follows:

[0063] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0064] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0065] like Figure 1 As shown, a multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on a data-model hybrid driving method provided by an embodiment of the present invention, and its implementation process includes the following steps:

[0066] S1: Initialize the power-traffic coupling network model, the mobile energy storage vehicle graph reinforcement learning model, and the emergency repair team graph reinforcement learning model;

[0067] S2: Simulate the power-transportation coupling network failure scenario after an extreme event occurs: set line disconnection in the distribution network and road damage in the transportation network;

[0068] S3: Initialization of the graph reinforcement learning algorithm environment and the fault scenario of the power-transportation coupling network;

[0069] S4: Based on the information of the power-traffic coupling network and the status information of the mobile emergency resources themselves, the mobile energy storage vehicle status of the graph neural network reinforcement learning algorithm is constructed and the status of the emergency repair team Get the state matrix o t , construct the adjacency matrix A according to the adjacency relationship of mobile emergency resources t ;

[0070] S5: Calculate the routing behavior strategy of the mobile energy storage vehicle based on the ε-Greedy strategy and the graph neural network reinforcement learning algorithm (data-driven method) and emergency repair team routing behavior strategy

[0071] S6: Execute the routing behavior strategy of mobile energy storage vehicles separately and emergency repair team routing behavior strategy And judge and update the status of mobile energy storage vehicles and emergency repair teams;

[0072] S7: Calculate the power dispatch strategy and reward function r of the mobile energy storage vehicle based on the mobile energy storage vehicle dispatch model combined with the distribution network reconstruction model and the distribution network optimal power flow model (model-driven method) t M , calculate the maintenance scheduling strategy and reward function r of the emergency repair team according to the emergency repair team scheduling model (model-driven method) t R ;

[0073] S8: Update the state of the mobile energy storage vehicle using the graph neural network reinforcement learning algorithm and the status of the emergency repair team Get the next state matrix o t+1 , update the adjacency matrix A t+1 ;

[0074] S9: The information of the current step and initial priority Stored in the mobile energy storage vehicle Figure reinforcement learning memory unit D M ,Will and initial priority Stored in the emergency repair team map reinforcement learning memory unit D R ;

[0075] S10: The priority experience replay strategy is used to sample the importance of the memory unit, and the weights of the graph neural network of the mobile energy storage vehicle and the emergency repair team are updated based on the stochastic gradient descent method;

[0076] S11: Determine whether the end time T has been reached End , if not, execute steps S3 to S10;

[0077] S12: Determine whether the training end scene number E has been reached End If not, execute steps S3 to S11; if yes, output the graph reinforcement learning algorithm parameters, routing and scheduling strategies of the mobile energy storage vehicle and the emergency repair team.

[0078] The following is a detailed introduction to this embodiment:

[0079] 1. Initialize the power-traffic coupling network model, the mobile energy storage vehicle graph reinforcement learning model, and the emergency repair team graph reinforcement learning model.

[0080] (1) Initialize the number of training scenes to e = 0;

[0081] (2) Initialization of the power-transportation coupling network model, including: voltage limits at distribution network nodes, line and transformer parameter settings, upper and lower constraints on voltage and output of distributed generators, output of renewable energy, load rate of distribution network, and the access location of charging stations in the distribution network; transportation network nodes, road length, road capacity and free driving speed, the location of charging stations in the transportation network, and the location of emergency repair team warehouses;

[0082] (3) The nodes in the distribution network and the lines connecting the nodes are regarded as a distribution network graph based on points and edges; charging station nodes are added to it to form a distribution network graph model Where N PDS represents the set of distribution network nodes, L PDS represents the set of distribution network lines, Represents the charging station node set. Initialize the load rate of the distribution network node;

[0083] (4) Treat the nodes in the traffic network and the roads connecting the nodes as points and edges to form a traffic network graph; add charging station nodes and emergency repair team warehouses to form a traffic network graph model Where N TN represents the set of nodes in the transportation network; L TN Represents a collection of transportation network roads; Represents the set of charging station nodes in the transportation network; Represents the set of repair team warehouse nodes in the transportation network;

[0084] (5) All mobile energy storage vehicles and emergency repair teams are considered as agents, and each agent is considered as a node, and their connections are considered as edges to construct a dynamic network model of mobile emergency resources. Where M is the mobile energy storage vehicle set, W is the emergency repair team set, Mobile emergency resource adjacency The adjacency matrix formed;

[0085] (6) Initialization of the mobile energy storage vehicle graph reinforcement learning model, including: graph neural network layer number and neuron settings, weight W M and bias b M Initialization; learning rate α M , sampling batch size N s and memory unit D M Initialization of hyperparameters such as capacity;

[0086] (7) Initialization of the emergency repair team graph reinforcement learning model, including: graph neural network layer number and neuron settings, weight W R and bias b RInitialization; learning rate α R , sampling batch size N s and memory unit D R Initialization of hyperparameters such as capacity.

[0087] 2. Simulate the power-transportation coupled network failure scenario after an extreme event occurs.

[0088] In order to study the impact of extreme events on the power-transportation coupling network, some lines are randomly interrupted in the distribution network and a damaged line set L is constructed. PDS,off , simulating the interruption of weak lines in the distribution network in real life after an extreme event; and randomly setting some road damage in the traffic network to construct a damaged road set Simulate the situation in real life where some roads in the traffic network are blocked by obstacles caused by extreme events.

[0089] 3. Initialization of the Graph Reinforcement Learning Algorithm Environment and Power-Transportation Coupled Network Failure Scenario

[0090] (1) Initialize training time t = 0;

[0091] (2) Initialize various variables in the graph reinforcement learning algorithm environment, including: the initial position of the mobile energy storage vehicle (charging station node), the initial SOC value of the mobile energy storage vehicle, the initial speed of the mobile energy storage vehicle, the initial position of the emergency repair team (their respective repair team warehouses), the emergency repair team's maintenance materials, the initial speed of the emergency repair team, and the traffic flow and speed of each road in the traffic network;

[0092] (3) Initialize the fault scenario of the power-transportation coupling network, that is, initialize the damaged line set L constructed in step S2 PDS,off and damaged road collection

[0093] 4. Construct the mobile energy storage vehicle state of the graph neural network reinforcement learning algorithm based on the power-traffic coupling network information and the mobile emergency resource status information and the status of the emergency repair team Get the state matrix o t , construct the adjacency matrix A according to the adjacency relationship of mobile emergency resources t .

[0094] (1) Collect information from the power-traffic coupling network graph model and the mobile emergency resource dynamic graph model to construct the state of the mobile energy storage vehicle graph neural network reinforcement learning algorithm Including the mobile energy storage vehicle's own status information Adjacent mobile emergency resource status information Traffic network information of your location Distribution network node load data P t ED 、New Energy Output P t WT , specifically expressed as:

[0095]

[0096] In the formula, M is the set of mobile energy storage vehicles; T is the time set; m is the number of the mobile energy storage vehicle; Represents the location information of the mobile energy storage vehicle in the traffic network at time t; Indicates the moving speed of the mobile energy storage vehicle in the traffic network; S m,t Indicates the remaining SOC value of the mobile energy storage vehicle; and They are the information of the mobile energy storage vehicles or emergency repair teams adjacent to the mth mobile energy storage vehicle at time t; N r is the number of adjacent emergency resources of the mobile energy storage vehicle; J is the number of adjacent roads of the mobile energy storage vehicle; represents the number of the jth road adjacent to the mth mobile energy storage vehicle at time t, and Indicates the road The first node number and the last node number of and Respectively represent roads the status of the road, speed of traffic and number of vehicles on the road; Indicates the road Length; Indicates the current traffic network node location number, Indicates the position number of the next node to be selected, Indicates the previous node position number, Indicates the road number to be entered and Indicates the previous road number, RT w,t It indicates the number of repair materials remaining in RC;

[0097] (2) Collect information from the power-traffic coupling network graph model and the mobile emergency resource dynamic graph model to construct the state of the emergency repair team graph neural network reinforcement learning algorithm Including the mobile energy storage vehicle's own status information Adjacent mobile emergency resource status information Traffic network information of your location Traffic network road fault information and the length of time it takes to repair damaged roads Specifically expressed as:

[0098]

[0099]

[0100] (3) Constructing the state matrix

[0101] (4) Collect information from the dynamic graph model of mobile emergency resources to construct the adjacency matrix A t , specifically expressed as:

[0102]

[0103] In the formula, N is the adjacency relationship formed by all mobile emergency resources at time t; v The number of mobile emergency resources can be expressed by N v =|M|+|W| calculation; is a binary variable, indicating the adjacency relationship between mobile emergency resources i and j. If they are adjacent, it is 1, otherwise it is 0. The adjacency matrix A t It will be input into the graph reinforcement learning algorithm to participate in the calculation of state-action value.

[0104] 5. Calculate the routing behavior strategy of mobile energy storage vehicles based on the ε-Greedy strategy and graph neural network reinforcement learning algorithm (data-driven method) and emergency repair team routing behavior strategy

[0105] (1) Using the ε-Greedy strategy: First, calculate the random number p at time t t , compare the current ε t value and random number p t , if p t <ε t , a random method is used to generate the mobile energy storage vehicle routing behavior strategy and emergency repair team routing behavior strategy Specifically, as shown in formula (150):

[0106]

[0107] In the formula, Represents the routing action set of the mobile energy storage vehicle routing and the emergency repair team routing; when the selected action a i,t =0 means stopping at the original location, and also means that the mobile emergency resource route reaches the destination location (the mobile energy storage vehicle is the charging station, and the emergency repair team is the damaged road or warehouse node), ready to accept the subsequent model scheduling decision; when the selected action a i,t ≠0 means that the mobile energy storage vehicle and the emergency repair team choose to enter the next traffic road to their destination according to the routing strategy;

[0108] (2) If p t ≥ε t , the routing behavior strategy of the mobile energy storage vehicle is generated through the graph neural network multi-agent reinforcement learning algorithm and emergency repair team routing behavior strategy

[0109]

[0110] In the formula, argmax(·) represents the parameter corresponding to the maximum value; The neural network parameters at time t for the mobile energy storage vehicle graph reinforcement learning algorithm According to the mobile emergency resource status set o t , Routing Action and the mobile emergency resource adjacency matrix A t The calculated action-value function; The neural network parameters at time t for the emergency repair team graph reinforcement learning algorithm According to the mobile emergency resource status set o t , Routing Action and the mobile emergency resource adjacency matrix A t The calculated action-value function;

[0111] The neural network structure used in the mobile energy storage vehicle graph reinforcement learning algorithm and the emergency repair team graph reinforcement learning algorithm mentioned in this example is the same, that is, the same neural network is used to control multiple mobile energy storage vehicles, and another neural network is used to control multiple emergency repair teams; the specific neural network structure is: first, a fully connected input layer is used to input the mobile emergency resource state set o t Perform preliminary feature extraction and dimension conversion, output This is followed by two layers of graph attention network layers. These two layers of networks use the attention mechanism to and the mobile emergency resource adjacency matrix A t Calculate the attention coefficient and extract features based on the attention matrix. Output is based on adjacency relationship. Features The last layer adopts the strategy of Dueling DQN algorithm to split the feature information into two fully connected branches: the first branch outputs the state function scalar value V(o t ,A t ); The second branch output status The action advantage value function vector under Therefore, the calculation of the Q value in equations (151) and (152) can be expressed as:

[0112]

[0113] In the formula, ψ t V(o t ,A t ) The fully connected neural network parameters of the branch where it is located; for The fully connected neural network parameters of the branch. The mobile energy storage vehicle and the emergency repair team select the routing action according to the action value function calculated by their respective neural networks. and

[0114] 6. Execute the mobile energy storage vehicle routing behavior strategy separately and emergency repair team routing behavior strategy And judge and update the status of mobile energy storage vehicles and emergency repair teams.

[0115] 7. Calculate the power dispatch strategy and reward function r of the mobile energy storage vehicle based on the mobile energy storage vehicle dispatch model, emergency repair team dispatch model, distribution network reconstruction model and distribution network optimal power flow model (model-driven method) t M and the maintenance scheduling strategy and reward function r of the emergency repair team t R .

[0116] (1) Construct a mobile energy storage vehicle dispatching model, a distribution network reconstruction model, and a distribution network optimal power flow model, as shown below:

[0117]

[0118]

[0119] Where, formulas (154)-(159) are the dispatching models of mobile energy storage vehicles; Represents the set of charging station nodes in the transportation network; is a binary variable, indicating whether the mobile energy storage vehicle is located at the traffic network node n tn , if it is at the node, it is 1, otherwise it is 0; and are binary variables, representing the charging and discharging status of the mth mobile energy storage vehicle at time t; and They represent the charging active power, discharging active power and reactive power at time t respectively; and Respectively represent the upper limit of active power and the upper limit of reactive power; S m,t Indicates the battery state-of-charge (SOC) at time t; and are the charging efficiency coefficient and the discharging efficiency coefficient respectively;S m and are the upper and lower limits of the SOC value respectively; formulas (154)-(156) restrict the charging and discharging power and reactive power of the mobile energy storage vehicle; formula (157) restricts the charging and discharging behavior of the mobile energy storage vehicle, that is, the mobile energy storage vehicle can only be in the charging state or the discharging state. At the same time, if the current location of the mobile energy storage vehicle is not located at the charging station node of the transportation network, it cannot be charged or discharged; formulas (158)-(159) constrain the SOC value of the mobile energy storage vehicle;

[0120] Where, formulas (160)-(165) are distribution network reconstruction models; L PDS , L PDS,off , L PDS,switch and They represent the set of distribution network lines, the set of damaged lines, the set of lines equipped with tie switches, and the set of distribution network source nodes respectively; ρ(·) and φ(·) represent the set of parent nodes and the set of child nodes respectively; α ij,t Indicates the connection status of the distribution network line (i, j) at time t, if connected, it is 1, and if disconnected, it is 0; N PDS Indicates the number of distribution network nodes; is a binary variable. If the distribution network node n at time t pds If it is a source node, it is 1, otherwise it is 0; f ij,t represents the virtual power flow into the distribution network node j at time t; M is a constant with a large value. Constraint (160) sets the line state constraint based on the distribution network line fault in this example three; Constraint (161) is the distribution network radiation constraint, which describes the relationship between the distribution network line state (normal line, damaged line and line equipped with remote switch) and the distribution network source node; Constraints (162)-(165) describe the virtual power flow balance constraint of the distribution network;

[0121] Where, formulas (166)-(176) are the optimal power flow models of distribution networks; and They represent the active power and reactive power flowing through the distribution network line (i, j) at time t respectively; and They represent the active and reactive output of the generator at the node j of the distribution network respectively; and Respectively represent the active and reactive outputs of the wind turbine WT; and They represent the active power exchange amount and reactive power exchange amount between the mobile energy storage vehicle and node j respectively; and Both represent the load loss of node j; and They represent the load demand of node j respectively; and They represent the maximum active power and the maximum reactive power of the distribution network line (i, j) respectively; represents the upper capacity limit of the distribution network line (i, j); represents the square value of the voltage at the node j of the distribution network at time t; r ij and x ij Respectively represent the line resistance and reactance values ​​of the distribution network line (i, j); and They represent the upper and lower limits of the square value of the voltage of the node j in the distribution network at time t. Constraints (166) and (167) describe the active power balance and reactive power balance of the distribution network line; constraints (168) and (169) describe the active power and reactive power of the mobile energy storage vehicle charging (or discharging) at the charging station node at time t; constraint (170) limits the active power and reactive power of the mobile energy storage vehicle when it is not at the charging station node to 0; constraints (171) and (172) limit the active power and reactive power flowing through the distribution network line at time t; constraint (173) limits the thermal capacity of the line (i, j); constraints (174)-(176) describe the voltage relationship between adjacent nodes of the distribution network.

[0122] (2) Construct an emergency repair team dispatch model, as shown below:

[0123]

[0124]

[0125] In the formula, W is the emergency repair team set; and are binary variables. If at time t the wth emergency repair team is located at node n of the traffic network, tn Or road tn If yes, it is 1, otherwise it is 0; Represents the set of damaged roads in the transportation network; is a binary variable, indicating the wth emergency repair team's response to the damaged road at time t. The repair decision is 1 if it is repaired, otherwise it is 0; is a binary variable, indicating the road Status; To repair damaged roads Time required; w,t is a binary variable used to determine whether the emergency repair team has reached the warehouse node; β w,tBinary variable used to determine whether the emergency repair team needs to be dispatched to the warehouse. Constraints (177)-(182) represent the emergency repair team's decision on road maintenance, while constraints (183)-(184) refer to the formulas here to make a decision on whether the emergency repair team stays in the warehouse. Constraint (177) indicates that the emergency repair team has not reached the damaged road in the traffic network. , its road repair decision variable will be restricted to 0; constraint (178) describes the relationship between the emergency repair team's repair decision variable and the state of damaged roads in the traffic network. If the road in the traffic network at time t If the road is still damaged, the emergency repair team can repair it; Constraint (179) constrains the road status of the traffic network based on the traffic network road fault setting in Example 3; Constraint (180) indicates that each damaged road in the traffic network Only one emergency repair team can repair the damaged road. Constraint (181) indicates that the emergency repair team can repair the damaged road. Repaired After a while, the road Traffic can be restored at time t; constraint (182) assumes that damaged roads in the traffic network will not be damaged again after being repaired; constraint (183) calculates whether the emergency repair team has arrived at the warehouse node, and constraint (184) constrains the subsequent dispatch of the emergency repair team through the status of all damaged roads and the location of the emergency repair team at time t, that is, calculates the dispatch status of the emergency repair team;

[0126] (3) The objective function is constructed based on the distribution network node load recovery and traffic network damaged road repair tasks, as shown below:

[0127]

[0128] In the formula, c pds and c tn They are represented as distribution network nodes n pds Unit load loss cost and transportation network road Damage cost per unit time; Represents node n pds The importance coefficient of is: the objective function (185) represents minimizing the loss cost of the power-transportation coupling network; formula (186) calculates the load loss cost of the distribution network node; formula (187) calculates the road damage cost of the transportation network and the dispatch cost of the emergency repair team;

[0129] (4) Constraints (168) and (169) are linearized using the big-M method, and constraint (173) is linearized using the second-order cone relaxation method; while constraints (180) and (181) of the emergency repair team scheduling problem are linearized using the McCormick envelope method.

[0130] (5) (Model driving part) The mixed integer second-order cone programming model MISOCP is constructed with formulas (185)-(187) as the objective function and formulas (154)-(176) as the constraints, and the Gurobi commercial solver is used to solve it. The load of the distribution network node restored by the mobile energy storage vehicle is obtained, and the mobile energy storage vehicle reward function is calculated using formula (188)

[0131]

[0132] In the formula, represents the set of distribution network nodes supported by the mth mobile energy storage vehicle at time t; is the distribution network node at time t The load recovery amount; For distribution network nodes The load level weight coefficient. It should be noted here that the dispatch cost of the mobile energy storage vehicle is not taken into account in this example, so the reward value of the mobile energy storage vehicle is 0 when it moves between charging stations in the transportation network;

[0133] At the same time, the road maintenance strategy or dispatching strategy of the emergency repair team is obtained by solving MISOCP, and the reward function of the emergency repair team is calculated by formulas (189)-(195):

[0134]

[0135]

[0136] In the formula, It represents the travel time from the starting node b to the target node p in the transportation network; For road tn The free passage time of tn and They are road delay coefficients, usually with values ​​of 0.15 and 4; The reward value obtained by the emergency repair team for completing the task at time t; The road was repaired by the emergency repair team Contribution reward value; is the penalty value; and All are weight coefficients; At time t, the emergency repair team completes the repair of damaged roads in the traffic network. The reward obtained by repairing is set to a constant; The reward value for the emergency repair team to return to the initial warehouse node at time t after repairing all damaged roads is set as a constant; is a binary variable, which means that when the road selected at time t is the same as that at time t-1, it is 1, otherwise it is 0; c1 and c2 are both constants. Formula (193) calculates the number of damaged roads in the traffic network repaired by the wth emergency repair team at time t. The ratio of the number of vehicles to the total number of vehicles; the three penalty items in the formula (194) represent: the penalty for the emergency repair team to choose the road passed at the previous moment as the next road to be entered, the penalty for the emergency repair team to repair the road but fail to return to its maintenance station node at the last moment, and the penalty for failing to repair the road and fail to return to the warehouse at the last moment;

[0137] 8. Update the state of the mobile energy storage vehicle using the graph neural network reinforcement learning algorithm and the status of the emergency repair team Get the next state matrix o t+1 , update the adjacency matrix A t+1 ;

[0138] 9. The information of the current step and initial priority Stored in the mobile energy storage vehicle Figure reinforcement learning memory unit D M ,Will and initial priority Stored in the emergency repair team map reinforcement learning memory unit D R .

[0139] 10. The priority experience replay strategy is used to perform sample importance sampling on the memory unit, and the weights of the graph neural network of the mobile energy storage vehicle and the emergency repair team are updated respectively based on the stochastic gradient descent method.

[0140] (1) In the mobile energy storage vehicle, the image reinforcement learning memory unit D M In the importance sampling method, samples N are collected respectively s , and calculate the importance weight of each sample using formula (62):

[0141]

[0142] Where N D Represents memory unit D M The total number of samples in ; Indicates the priority of the sample; β is a hyperparameter;

[0143] (2) Combining the strategy of the Double DQN algorithm, the loss function is calculated using formulas (197) and (198):

[0144]

[0145] In the formula, and are the current network parameters and target network parameters of the mobile energy storage vehicle respectively; γ is the discount factor reflecting the impact of the future Q value on the current action, and its value is [0,1];

[0146] (3) The neural network parameters of the mobile energy storage vehicle graph reinforcement learning algorithm are updated based on the stochastic gradient descent method and sample importance weight, which is specifically expressed as:

[0147]

[0148] In the formula, α M is the learning rate of the gradient descent algorithm of the current network of the mobile energy storage vehicle. And after a fixed number of steps N up Then, the parameters of the target network of the mobile energy storage vehicle are adjusted in the following way. To update:

[0149]

[0150] (4) TD-error value calculated according to formula (63) Calculate the new priority for each sample Where ∈ is a small constant to prevent the priority from being zero. M Update the priority of each sample in;

[0151] (5) Update the parameters of the neural network of the emergency repair team graph reinforcement learning algorithm according to methods (1)-(4).

[0152] 11. Determine whether the end time T has been reached End , if not, execute steps 3 to 10.

[0153] 12. Determine whether the training end episode number E has been reached End If not, execute steps 3 to 11; if so, output the graph reinforcement learning algorithm parameters, routing and scheduling strategies of the mobile energy storage vehicle and the emergency repair team.

[0154] The present invention provides a multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on a data-model hybrid drive method. The method first regards mobile emergency resources (i.e., mobile energy storage vehicles and emergency repair teams) as agents, and abstracts the dynamic adjacency relationships of all agents as edges, and then converts the coordination and cooperation relationships of all mobile energy storage vehicles and emergency repair team agents into a dynamic graph structure. In the data-driven part, the method adopts a graph neural network reinforcement learning algorithm based on an attention mechanism, which can combine the excellent non-Euclidean data information extraction ability of the graph attention network and the powerful sequential random optimization decision-making ability of the reinforcement learning algorithm, and effectively extract key information from the dynamic adjacency relationship formed by the mobile energy storage vehicle and the emergency repair team and the power-transportation coupling network graph data to formulate the routing behavior strategy of the mobile energy storage vehicle and the emergency repair team; in the model-driven part, the scheduling model of the mobile energy storage vehicle and the emergency repair team, the distribution network reconstruction model and the distribution network optimal power flow model are constructed as a MISOCP model and the optimal multi-type mobile emergency resource scheduling behavior strategy is solved to achieve the improvement of the distribution network resilience. In summary, the present invention can simultaneously consider the dynamic changes in the road status of the transportation network and the dynamic changes in the relationship between adjacent mobile emergency resources. The routing and scheduling strategies of mobile energy storage vehicles and emergency repair teams solved by the data-model hybrid driving method can restore the node load of the distribution network and repair damaged roads in the transportation network, thereby improving the resilience of the distribution network.

[0155] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0156] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0157] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0158] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0159] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0160] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

[0161] This patent is not limited to the above-mentioned optimal implementation mode. Anyone can derive other various forms of a multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on a data-model hybrid driven method under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of this invention should be covered by this patent.

Claims

1. A multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on a data-model hybrid drive method, characterized by: Starting from the dynamic interactive relationship between mobile energy storage vehicles and emergency repair teams, the two teams are coordinated to restore load and repair roads. The emergency repair team is dispatched to repair damaged roads in the traffic network under the uncertainty of the time required to repair the roads. Mobile energy storage vehicles are dispatched to restore loads on distribution network nodes under the dynamic changes in the road status of the traffic network. Data methods and model methods are combined to solve the above multi-type mobile emergency resource scheduling problems. The data-driven part is supported by the graph neural network multi-agent reinforcement learning algorithm GATD3QN based on the graph attention mechanism. The GATD3QN algorithm combines the graph neural network based on the attention mechanism, Double The DQN algorithm and DuelingDQN algorithm optimize the feature extraction ability and Q-value calculation of the DQN algorithm, and adopt the priority experience replay strategy to improve the sampling efficiency and training effect of the algorithm, so as to solve the optimal multi-type mobile emergency resource routing behavior strategy; the model-driven part constructs the scheduling model of the mobile energy storage vehicle and the emergency repair team, the distribution network reconstruction model and the distribution network optimal power flow model into a MISOCP model and solves the optimal multi-type mobile emergency resource scheduling behavior strategy, including the power scheduling decision of the mobile energy storage vehicle, the road repair decision of the emergency repair team or the return to the warehouse decision, so as to improve the resilience of the distribution network.

2. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid driving method according to claim 1 is characterized by: The following steps are involved: Step S1: construct and initialize the power-traffic coupling network model, the mobile energy storage vehicle graph reinforcement learning model and the emergency repair team graph reinforcement learning model; Step S2: Simulate the power-transportation coupling network failure scenario after the extreme event occurs, and set line disconnection in the distribution network and road damage in the transportation network; Step S3: Initialize the graph reinforcement learning algorithm environment and the power-transportation coupling network fault scenario; Step S4: Construct the mobile energy storage vehicle state of the graph neural network reinforcement learning algorithm based on the power-traffic coupling network information and the mobile emergency resource state information and the status of the emergency repair team Get the state matrix o t , construct the adjacency matrix A according to the adjacency relationship of mobile emergency resources t ; Step S5: Calculate the mobile energy storage vehicle routing behavior strategy based on the ε-Greedy strategy and the graph neural network reinforcement learning algorithm as the data-driven part and emergency repair team routing behavior strategy Step S6: Execute the mobile energy storage vehicle routing behavior strategy respectively and emergency repair team routing behavior strategy And judge and update the status of mobile energy storage vehicles and emergency repair teams; Step S7: The model driving part calculates the power dispatching strategy and reward function of the mobile energy storage vehicle according to the mobile energy storage vehicle dispatching model combined with the distribution network reconstruction model and the distribution network optimal power flow model. Calculate the maintenance scheduling strategy and reward function of the emergency repair team based on the emergency repair team scheduling model Step S8: Update the state of the mobile energy storage vehicle of the graph neural network reinforcement learning algorithm and the status of the emergency repair team Get the next state matrix o t+1 , update the adjacency matrix A t+1 ; Step S9: The information of the current step and initial priority Stored in the mobile energy storage vehicle Figure reinforcement learning memory unit D M ,Will and initial priority Stored in the emergency repair team map reinforcement learning memory unit D R ; Step S10: The priority experience replay strategy is used to sample the importance of the memory unit, and the weights of the graph neural network of the mobile energy storage vehicle and the emergency repair team are updated based on the stochastic gradient descent method; Step S11: Determine whether the end time T has been reached End , if not, execute steps S3 to S10; Step S12: Determine whether the training end episode number E has been reached End If not, execute steps S3 to S11; if yes, output the graph reinforcement learning algorithm parameters, routing and scheduling strategies of the mobile energy storage vehicle and the emergency repair team.

3. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid driving method according to claim 2 is characterized by: In step S1, the power-traffic coupling network model, the mobile energy storage vehicle graph reinforcement learning model and the emergency repair team graph reinforcement learning model are initialized, and the number of training episodes is initialized. The nodes in the distribution network and the lines connecting the nodes are regarded as points and edges, forming a distribution network graph. Charging station nodes are added to form a distribution network graph model. Where N PDS represents the set of distribution network nodes, L PDS represents the set of distribution network lines, Represents a collection of charging station nodes; The nodes in the traffic network and the roads connecting the nodes are regarded as points and edges to form a basic traffic network graph; charging station nodes and emergency repair team warehouses are added to form a traffic network graph model Where N TN represents the set of nodes in the transportation network; L TN Represents a collection of transportation network roads; Represents the set of charging station nodes in the transportation network; Represents the set of repair team warehouse nodes in the transportation network; All mobile energy storage vehicles and emergency repair teams are regarded as agents, and each agent is regarded as a node, and their connections are regarded as edges to construct a dynamic network model of mobile emergency resources. Where M is the mobile energy storage vehicle set, W is the emergency repair team set, Mobile emergency resource adjacency The adjacency matrix constructed.

4. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid drive method according to claim 3 is characterized by: In step S2, some lines are randomly interrupted in the distribution network to construct a damaged line set L PDS,off , which is used to simulate the interruption of weak lines in the distribution network after an extreme event; and randomly set some road damage in the traffic network to construct a damaged road set This is to simulate the situation where some roads in the transportation network are blocked by obstacles caused by extreme events.

5. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid drive method according to claim 2 is characterized by: Step S3 specifically includes: initializing various variables in the graph reinforcement learning algorithm environment, initializing the power-transportation coupling network fault scenario, and initializing the training time.

6. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid drive method according to claim 4 is characterized by: In step S4: According to the information of the power-traffic coupling network and the state information of the mobile emergency resources themselves, the state of the mobile energy storage vehicle graph neural network reinforcement learning algorithm is constructed Including the mobile energy storage vehicle's own status information Adjacent mobile emergency resource status information Traffic network information of your location Distribution network node load data P t ED 、New Energy Output P t WT , specifically expressed as: In the formula, M is the set of mobile energy storage vehicles; T is the time set; m is the number of the mobile energy storage vehicle; Represents the location information of the mobile energy storage vehicle in the traffic network at time t; Indicates the moving speed of the mobile energy storage vehicle in the traffic network; S m,t Indicates the remaining SOC value of the mobile energy storage vehicle; and They are the information of the mobile energy storage vehicles or emergency repair teams adjacent to the mth mobile energy storage vehicle at time t; N r is the number of adjacent emergency resources of the mobile energy storage vehicle; J is the number of adjacent roads of the mobile energy storage vehicle; represents the number of the jth road adjacent to the mth mobile energy storage vehicle at time t, and Indicates the road The first node number and the last node number of and Respectively represent roads the status of the road, speed of traffic and number of vehicles on the road; Indicates the road Length; Indicates the current traffic network node location number, Indicates the position number of the next node to be selected, Indicates the previous node position number, Indicates the road number to be entered and Indicates the previous road number, RT w,t It indicates the number of repair materials remaining in RC; According to the information of the power-traffic coupling network and the status information of the mobile emergency resources themselves, the state of the emergency repair team graph neural network reinforcement learning algorithm is constructed Including the mobile energy storage vehicle's own status information Adjacent mobile emergency resource status information Traffic network information of your location Traffic network road fault information E t DL and the length of time it takes to repair damaged roads Specifically expressed as: In the formula, W is the emergency repair team set; The state matrix is ​​constructed from Then, the adjacency matrix A is constructed based on the adjacency relationship of mobile emergency resources. t , specifically expressed as: In the formula, N is the adjacency relationship formed by all mobile emergency resources at time t; v is the number of mobile emergency resources, expressed as N v =|M|+|W| calculation; is a binary variable, indicating the adjacency relationship between mobile emergency resources i and j. If they are adjacent, it is 1, otherwise it is 0. Adjacency Matrix A t Used to input into the graph reinforcement learning algorithm to participate in the calculation of state-action values.

7. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid drive method according to claim 6 is characterized by: In step S5, first, the random number p at time t is calculated t , whose value range is [0,1), and the ε-Greedy strategy is used to compare the current ε t value and random number p t , if p t <ε t , a random method is used to generate the mobile energy storage vehicle routing behavior strategy and emergency repair team routing behavior strategy Specifically, as shown in formula (16): In the formula, Represents the routing action set of the mobile energy storage vehicle routing and the emergency repair team routing; when the selected action a i,t = 0 means stopping at the original location, corresponding to the mobile emergency resource route reaching the destination location, ready to accept the subsequent model scheduling decision; when the selected action a i,t ≠0 means that the mobile energy storage vehicle and the emergency repair team choose to enter the next traffic road to their destination according to the routing strategy; If there is p t ≥ε t , the routing behavior strategy of the mobile energy storage vehicle is generated through the graph neural network multi-agent reinforcement learning algorithm and the emergency repair team routing behavior strategy a t R , as shown in formulas (17) and (18): In the formula, argmax(·) represents the parameter corresponding to the maximum value; The neural network parameters at time t for the mobile energy storage vehicle graph reinforcement learning algorithm According to the mobile emergency resource status set o t , Routing Action and the mobile emergency resource adjacency matrix A t The calculated action-value function; The neural network parameters at time t for the emergency repair team graph reinforcement learning algorithm According to the mobile emergency resource status set o t , Routing Action and the mobile emergency resource adjacency matrix A t The calculated action-value function; The neural network structure used in the mobile energy storage vehicle graph reinforcement learning algorithm and the emergency repair team graph reinforcement learning algorithm is the same: the same neural network is used to control multiple mobile energy storage vehicles, and another neural network is used to control multiple emergency repair teams; the neural network structure is: first, a fully connected input layer is used to input the mobile emergency resource state set o t Perform preliminary feature extraction and dimension conversion, output This is followed by two layers of graph attention network layers. These two layers of networks use the attention mechanism to and the mobile emergency resource adjacency matrix A t Calculate the attention coefficient and extract features based on the attention matrix. Output is based on adjacency relationship. Features The last layer adopts the strategy of Dueling DQN algorithm to split the feature information into two fully connected branches: the first branch outputs the state function scalar value V(o t ,A t ); The second branch output status The action advantage value function vector under The calculation of the Q value in formulas (17) and (18) is expressed as: In the formula, ψ t V(o t ,A t ) is the fully connected neural network parameter of the branch; θ t for The fully connected neural network parameters of the branch; the mobile energy storage vehicle and the emergency repair team select the routing action according to the action value function calculated by their respective neural networks and 8. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid driving method according to claim 7 is characterized by: In step S7, the mobile energy storage vehicle dispatching model, distribution network reconstruction model and distribution network optimal power flow model are constructed as follows: Among them, formulas (20)-(25) are the dispatching models of mobile energy storage vehicles; Represents the set of charging station nodes in the transportation network; is a binary variable, indicating whether the mobile energy storage vehicle is located at the traffic network node n tn , if it is at the node, it is 1, otherwise it is 0; and are binary variables, representing the charging and discharging status of the mth mobile energy storage vehicle at time t; and They represent the charging active power, discharging active power and reactive power at time t respectively; and Respectively represent the upper limit of active power and the upper limit of reactive power; S m,t Indicates the battery state of charge SOC at time t; and are the charging efficiency coefficient and the discharging efficiency coefficient respectively; S m and are the upper and lower limits of the SOC value respectively; formulas (20)-(22) limit the charging and discharging power and reactive power of the mobile energy storage vehicle; formula (23) limits the charging and discharging behavior of the mobile energy storage vehicle, that is, the mobile energy storage vehicle can only be in the charging state or the discharging state. At the same time, if the current location of the mobile energy storage vehicle is not located at the charging station node of the transportation network, it cannot be charged or discharged; formulas (24)-(25) constrain the SOC value of the mobile energy storage vehicle; Formulas (26)-(31) are the distribution network reconstruction model; L PDS , L PDS,off , L PDS,switch and They represent the set of distribution network lines, the set of damaged lines, the set of lines equipped with tie switches, and the set of distribution network source nodes respectively; ρ(·) and φ(·) represent the set of parent nodes and the set of child nodes respectively; α ij,t Indicates the connection status of the distribution network line (i, j) at time t, if connected, it is 1, and if disconnected, it is 0; N PDS Indicates the number of distribution network nodes; is a binary variable. If the distribution network node n at time t pds If it is a source node, it is 1, otherwise it is 0; f ij,t represents the virtual power flow into the distribution network node j at time t; M is a constant with a large value; constraint (26) is a line state constraint set based on the distribution network line fault; constraint (27) is a distribution network radiation constraint, which describes the relationship between the distribution network line state and the distribution network source node; constraints (28)-(31) describe the virtual power flow balance constraints of the distribution network; Formulas (32)-(42) are the optimal power flow models of distribution networks; and They represent the active power and reactive power flowing through the distribution network line (i, j) at time t respectively; and They represent the active and reactive output of the generator at the node j of the distribution network respectively; and Respectively represent the active and reactive outputs of the wind turbine WT; and They represent the active power exchange amount and reactive power exchange amount between the mobile energy storage vehicle and node j respectively; and Both represent the load loss of node j; and They represent the load demand of node j respectively; and They represent the maximum active power and the maximum reactive power of the distribution network line (i, j) respectively; represents the upper capacity limit of the distribution network line (i, j); represents the square value of the voltage at the node j of the distribution network at time t; r ij and x ij Respectively represent the line resistance and reactance values ​​of the distribution network line (i, j); and They represent the upper and lower limits of the square value of the voltage of the node j in the distribution network at time t respectively; constraints (32) and (33) describe the active power balance and reactive power balance of the distribution network line; constraints (34) and (35) describe the active power and reactive power of the mobile energy storage vehicle charging or discharging at the charging station node at time t; constraint (36) limits the active power and reactive power of the mobile energy storage vehicle when it is not at the charging station node to 0; constraints (37) and (38) limit the active power and reactive power flowing through the distribution network line at time t; constraint (39) limits the thermal capacity of the line (i, j); constraints (40)-(42) describe the voltage relationship between adjacent nodes of the distribution network; The emergency repair team dispatch model constructed is as follows: In the formula, W is the emergency repair team set; and are binary variables. If at time t the wth emergency repair team is located at node n of the traffic network, tn Or road tn If yes, it is 1, otherwise it is 0; Represents the set of damaged roads in the transportation network; is a binary variable, indicating the wth emergency repair team's response to the damaged road at time t. The repair decision is 1 if it is repaired, otherwise it is 0; is a binary variable, indicating the road Status; To repair damaged roads Time required; w,t is a binary variable used to determine whether the emergency repair team has reached the warehouse node; β w,t Binary variable used to determine whether it is necessary to continue dispatching the emergency repair team to the warehouse. Constraints (43)-(48) represent the emergency repair team's decision on road maintenance, and constraints (49)-(50) make decisions on whether the emergency repair team stays in the warehouse. Constraint (43) indicates that the emergency repair team has not reached the damaged road in the traffic network. , its road repair decision variable will be restricted to 0; Constraint (44) describes the relationship between the emergency repair team's repair decision variable and the state of damaged roads in the traffic network. If the road in the traffic network at time t If the road is still damaged, the emergency repair team will repair it; Constraint (45) constrains the road status of the traffic network based on the traffic network road fault setting; Constraint (46) indicates that each damaged road in the traffic network Only one emergency repair team can repair the damaged road. Constraint (47) indicates that the emergency repair team can only repair the damaged road. Repaired After a while, the road Traffic is restored at time t; Constraint (48) assumes that damaged roads in the traffic network will not be damaged again after being repaired; Constraint (49) calculates whether the emergency repair team has arrived at the warehouse node, and Constraint (50) constrains the subsequent dispatch of the emergency repair team through the status of all damaged roads and the location of the emergency repair team at time t, that is, calculates the dispatch status of the emergency repair team; The objective function is constructed based on the distribution network node load recovery and traffic network damaged road repair tasks, as shown below: In the formula, c pds and c tn They are represented as distribution network nodes n pds Unit load loss cost and transportation network road Damage cost per unit time; Represents node n pds The importance coefficient of ; the objective function (51) represents minimizing the loss cost of the power-traffic coupling network; formula (52) is used to calculate the load loss cost of the distribution network node; formula (53) is used to calculate the road damage cost of the traffic network and the dispatch cost of the emergency repair team; In the calculation process, constraints (34) and (35) are linearized by the big M method, and constraint (39) is linearized by the second-order cone relaxation method; constraints (46) and (47) of the emergency repair team scheduling problem are linearized by the McCormickenvelope method; The model driving part uses formula (51)-(53) as the objective function and formula (20)-(50) as the constraint to construct the mixed integer second-order cone programming model MISOCP, which is solved by Gurobi solver. The load of the distribution network node restored by the mobile energy storage vehicle is obtained, and the mobile energy storage vehicle reward function is calculated by formula (54): In the formula, represents the set of distribution network nodes supported by the mth mobile energy storage vehicle at time t; is the distribution network node at time t The load recovery amount; For distribution network nodes The load level weight coefficient; the reward value of the mobile energy storage vehicle during the movement between charging stations in the transportation network is 0; By solving MISOCP, we can obtain the road maintenance strategy or dispatch strategy of the emergency repair team, and calculate the reward function of the emergency repair team through formulas (55)-(61): In the formula, It represents the travel time from the starting node b to the target node p in the transportation network; For road tn The free passage time of tn and All are road delay coefficients; The reward value obtained by the emergency repair team for completing the task at time t; The road was repaired by the emergency repair team Contribution reward value; is the penalty value; and All are weight coefficients; At time t, the emergency repair team completes the repair of damaged roads in the traffic network. The reward obtained by repairing is set to a constant; The reward value for the emergency repair team to return to the initial warehouse node at time t after repairing all damaged roads is set as a constant; is a binary variable, which means that when the road selected at time t is the same as that at time t-1, it is 1, otherwise it is 0; c1 and c2 are both constants; Formula (59) calculates the number of damaged roads in the traffic network repaired by the wth emergency repair team at time t. The three penalty items in formula (60) represent: the penalty for the emergency repair team to choose the road it passed at the previous moment as the next road to be entered, the penalty for the emergency repair team to repair the road but fail to return to its maintenance station node at the last moment, and the penalty for failing to repair the road and fail to return to the warehouse at the last moment.

9. The multi-type mobile emergency resource optimization scheduling and distribution network resilience improvement strategy based on the data-model hybrid driving method according to claim 8 is characterized by: In step S10, in the mobile energy storage vehicle graph reinforcement learning memory unit D M In the importance sampling method, samples N are collected respectively s , and calculate the importance weight of each sample using formula (62): Where N D Represents memory unit D M The total number of samples in ; Indicates the priority of the sample; β is a hyperparameter; Combined with the strategy of the Double DQN algorithm, formulas (63) and (64) are used to calculate the loss function: In the formula, and are the current network parameters and target network parameters of the mobile energy storage vehicle respectively; γ is the discount factor reflecting the impact of the future Q value on the current action, and its value is [0,1]; The neural network parameters of the mobile energy storage vehicle graph reinforcement learning algorithm are updated based on the stochastic gradient descent method and sample importance weights, which is specifically expressed as follows: In the formula, α M is the learning rate of the gradient descent algorithm of the current network of the mobile energy storage vehicle; and after a fixed number of steps N up Then, the parameters of the target network of the mobile energy storage vehicle are adjusted in the following way. To update: The TD-error value calculated according to formula (63) Calculate the new priority for each sample Among them, ∈ is a constant used to prevent the priority from being zero; in the mobile energy storage vehicle, the reinforcement learning memory unit D M Update the priority of each sample in; Update the parameters of the neural network of the emergency repair team graph reinforcement learning algorithm.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps of optimizing the scheduling of multi-type mobile emergency resources and improving the resilience of the distribution network based on the data-model hybrid driving method as described in any one of claims 2-9.

Images