Power distribution network post-disaster intelligent recovery method and device based on RainbowDQN algorithm

Abstract

The invention provides a power distribution network post-disaster intelligent recovery method and device based on a RainbowDQN algorithm. The method comprises the steps of obtaining a power distribution network model and a preset N-3 fault combination, and constructing a Markov decision environment model of a power distribution network load recovery task when a line has an N-3 fault; setting a target function and a constraint condition of the Markov decision environment model; training an intelligent agent in the Markov decision environment model by adopting a RainbowDQN algorithm added with a self-adaptive action mask to obtain a trained intelligent agent; and parameter information of the power distribution network with the N-3 line fault is transmitted to the trained intelligent agent, and the intelligent agent outputs a load recovery strategy of the power distribution network with the N-3 line fault. According to the method, the flexible soft switching technology is utilized to quickly isolate an affected area when a fault is caused by a natural disaster, so that the risk of fault diffusion is reduced, the power supply grid is quickly subjected to necessary topological adjustment after the disaster, and efficient distribution of power support after the disaster is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault recovery, and in particular to a method and device for intelligent post-disaster recovery of a distribution network based on a RainbowDQN algorithm. Background Art

[0002] With the widespread integration of distributed renewable energy, the rapid development of load-side diversification, and the continued advancement of distribution automation technology, distribution networks are gradually evolving into highly complex, multi-energy interconnected systems. However, in response to emergencies such as natural disasters (such as typhoons, rainstorms, and earthquakes), equipment aging, and operational errors, the vulnerability of distribution networks has become increasingly prominent. Line failures, node voltage loss, and switch refusal to operate are frequent, easily causing large-scale power outages. This is particularly true in core urban areas and critical load zones, such as hospitals, communication hubs, and transportation nodes. Failure to promptly restore power after a disaster can have serious impacts on social security and economic order.

[0003] Therefore, how to diagnose the network immediately after a fault occurs and develop an efficient and feasible load restoration strategy has become a core issue for the restorative operation of distribution systems. Distribution network load restoration aims to maximize user load recovery by rationally allocating available resources, optimizing switching operation sequences, and reconstructing network power supply paths, while satisfying multiple constraints such as radiation structure, voltage stability, power flow, and equipment capacity. This process not only reflects the resilience and intelligence level of the power system but also directly affects the efficiency of disaster emergency response and the user's electricity experience. It is one of the key supporting technologies for realizing a "self-healing" power grid and "active distribution management."

[0004] Currently, there is no effective post-disaster intelligent recovery method for distribution networks in the existing technology. Summary of the Invention

[0005] The embodiments of the present invention provide a method and device for intelligent post-disaster recovery of a distribution network based on the RainbowDQN algorithm, so as to effectively improve the resilience and sustainability of the distribution network.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0007] According to one aspect of the present invention, a method for intelligent post-disaster recovery of a distribution network based on the RainbowDQN algorithm is provided, comprising:

[0008] Obtain the distribution network model and the preset N-3 fault combination, and construct a Markov decision environment model for the distribution network load restoration task when N-3 faults occur on the line;

[0009] Setting the objective function and constraints of the Markov decision environment model;

[0010] The RainbowDQN algorithm with adaptive action mask is used to train the intelligent agent in the Markov decision environment model to obtain a trained intelligent agent;

[0011] The parameter information of the distribution network with N-3 line failure is transmitted to the trained intelligent agent, and the intelligent agent outputs the load recovery strategy of the distribution network with N-3 line failure.

[0012] Preferably, the step of acquiring a distribution network model and a preset N-3 fault combination and constructing a Markov decision environment model for a distribution network load restoration task when an N-3 fault occurs on a line includes:

[0013] A modified IEEE123 node system is set as the distribution network model, which includes node data, line data, transformer data, load data, and switch data. A certain number of random N-3 line fault data are preset as training data and test data respectively.

[0014] A Markov decision environment model for the load restoration task of the distribution network with N-3 line faults is constructed. The Markov decision model is defined as:

[0015] M=<S,A,P,R,γ> (1)

[0016] Wherein, the Markov decision process consists of state space S, action space A, state transition probability P, reward function R and discount factor γ;

[0017] The state space S is defined as:

[0018] S t ={N t ,L t ,W t} (2)

[0019] Where S t is the state of the environment at step t, N t is the node voltage state at step t. If the node loses voltage, N t If the node is not depressurized, N t is 1; L t is the state of the line at step t. If the line fails, L t If the line is not faulty, L t =1; W t is the switch state at step t. If the switch is closed, W t If the switch is open, W t is 0;

[0020] The action space A is defined as:

[0021] A t={a1,a2,a3,…a n} (3)

[0022] Where A t is the action selected by the agent at step t, a i Indicates that the corresponding i-th switch is in action, that is, if the corresponding switch is open, the switch is closed to put the line into operation; if the corresponding switch is closed, the switch is opened to disconnect the line from operation;

[0023] Reward function R e,t Defined as:

[0024] R e,t ={R nt +R b} (4)

[0025]

[0026] Where R e,t is the reward obtained at step t, R nt R is the number of nodes newly restored after the corresponding switch takes action compared to the last action. b The reward is to get the agent closer to the goal of fully restoring the load;

[0027] The corresponding penalties for negative actions taken by the agent are as follows:

[0028]

[0029] The formula indicates that when the agent takes an action, if no node is restored, it means that the action is invalid and a corresponding penalty is given.

[0030] Preferably, the objective function and constraint conditions of the Markov decision environment model are set, including:

[0031] The task of post-disaster load restoration in the distribution network is to restore power to the nodes that have lost power due to a fault. The objective function of the Markov decision environment model is to restore power to all nodes, that is, to maximize the value of the cumulative total reward.

[0032] The node voltage constraints and line current constraints of the Markov decision environment model are set as follows:

[0033] U i,min ≤U i ≤U i,max (8)

[0034] I ij,min ≤I ij ≤I ij,max (9)

[0035] Where U i represents the voltage amplitude at node i, U i,min and U i,max Indicates the minimum and maximum values ​​allowed; I ij Indicates the current on line ij, and I ij,min and I ij,max Indicates the minimum and maximum values ​​allowed;

[0036] The power balance constraint equation of the Markov decision environment model is set as follows:

[0037]

[0038] Where, P i , Q i and V i represents the active power, reactive power and voltage of node i; B ij , G ij and θ ij represents the susceptance, conductance and phase difference of the branch; l represents the set of child nodes of node i; and represents the active power and reactive power exchanged between node i and the upstream power grid; and represents the active power and reactive power requirements of node i.

[0039] Preferably, the method of using the RainbowDQN algorithm with an adaptive action mask to train the agent in the Markov decision environment model to obtain a trained agent includes:

[0040] Based on the environmental model of the distribution network model, the RainbowDQN algorithm with an adaptive action mask is used to train the intelligent agent in the Markov decision environment model. At each moment, the intelligent agent dynamically adjusts the action selection mechanism according to the action mask returned by the environment. The action mask is a Boolean vector, where: True indicates that the current action meets the specifications and can be executed; False indicates that the current action does not meet the specifications or has expired and should be prohibited from selection;

[0041] In the action selection phase, the agent calculates the Q-values ​​of all actions based on the current state and then performs the following processing:

[0042]

[0043] In each round, the agent receives the initial state of the distribution network and uses the noise network to output a Q value with random perturbations based on the current state. The Q value of the invalid action is set to negative infinity based on the adaptive action masking technology. The agent then performs the action with the largest Q value and completes a switching operation. After each step, the action is applied to the environment model. The environment updates the state according to the current topology structure and feedbacks the new system state, the immediate reward value, and the flag bit indicating whether the recovery task is completed. The reward function evaluates the contribution of the action to load recovery and encourages the agent to restore more load. In this process, the agent packages the state-action-reward-new state information generated by each interaction into experience samples and stores them in the priority experience replay pool. When the accumulated data in the experience pool reaches a predetermined amount, the agent begins to sample a batch of samples from the experience pool according to the priority experience replay strategy every certain number of steps to update the online network parameters. The target network parameters are updated slightly each time according to the soft update coefficient.

[0044] After the agent executes the action, the topology of the distribution network changes. The environment performs a power flow calculation based on the new network structure to obtain updated information on the recovery of voltage, current, and node load, and feeds the new status back to the agent. At the same time, the environment calculates the immediate reward corresponding to the action of this step: if the action is legal and successfully restores part of the load, a reward is given according to the reward function; if the action has no recovery contribution or causes the constraint to be violated, a negative penalty is imposed; by continuously perceiving the environment, selecting actions, receiving feedback, and updating strategies, the agent gradually learns how to make the correct switching operation sequence under different regional faults and different load conditions, and realizes the adaptive recovery of multi-region distribution networks. When the fluctuation of the reward function during training is less than the set numerical range, the training ends and a trained agent is obtained.

[0045] According to another aspect of the present invention, a distribution network post-disaster intelligent recovery device based on the RainbowDQN algorithm is provided, comprising:

[0046] The Markov decision environment model construction module is used to obtain the distribution network model and the preset N-3 fault combination, and construct a Markov decision environment model for the distribution network load restoration task when the line has N-3 faults;

[0047] An objective function and constraint condition setting module, used to set the objective function and constraint conditions of the Markov decision environment model;

[0048] An agent training module is used to train the agent in the Markov decision environment model using the RainbowDQN algorithm with an adaptive action mask to obtain a trained agent;

[0049] The load recovery strategy acquisition module is used to transmit parameter information of the distribution network where the N-3 line fault occurs to the trained intelligent agent, and the intelligent agent outputs the load recovery strategy of the distribution network where the N-3 line fault occurs.

[0050] Preferably, the Markov decision environment model building module is used to set a modified IEEE123 node system as a distribution network model, the distribution network model includes node data, line data, transformer data, load data and switch data, and a certain number of random N-3 line fault data are preset as training data and test data respectively;

[0051] A Markov decision environment model for the load restoration task of the distribution network with N-3 line faults is constructed. The Markov decision model is defined as:

[0052] M=<S,A,P,R,γ> (1)

[0053] Wherein, the Markov decision process consists of state space S, action space A, state transition probability P, reward function R and discount factor γ;

[0054] The state space S is defined as:

[0055] S t ={N t ,L t ,W t} (2)

[0056] Where S t is the state of the environment at step t, N t is the node voltage state at step t. If the node loses voltage, N t If the node is not depressurized, N t is 1; L t is the state of the line at step t. If the line fails, L t If the line is not faulty, L t =1; W t is the switch state at step t. If the switch is closed, W t If the switch is open, W t is 0;

[0057] The action space A is defined as:

[0058] A t ={a1,a2,a3,…a n} (3)

[0059] Where A t is the action selected by the agent at step t, a iIndicates that the corresponding i-th switch is in action, that is, if the corresponding switch is open, the switch is closed to put the line into operation; if the corresponding switch is closed, the switch is opened to disconnect the line from operation;

[0060] Reward function R e,t Defined as:

[0061] R e,t ={R nt +R b} (4)

[0062]

[0063] Where R e,t is the reward obtained at step t, R nt R is the number of nodes newly restored after the corresponding switch takes action compared to the last action. b The reward is to get the agent closer to the goal of fully restoring the load;

[0064] The corresponding penalties for negative actions taken by the agent are as follows:

[0065]

[0066] The formula indicates that when the agent takes an action, if no node is restored, it means that the action is invalid and a corresponding penalty is given.

[0067] Preferably, the objective function and constraint condition setting module is specifically used for the task of post-disaster load recovery of the distribution network to restore power supply to the nodes that have lost power due to the fault, and the objective function of the Markov decision environment model is set to restore power supply to all nodes, that is, to maximize the value of the cumulative total reward;

[0068] The node voltage constraints and line current constraints of the Markov decision environment model are set as follows:

[0069] U i,min ≤U i ≤U i,max (8)

[0070] I ij,min ≤I ij ≤I ij,max (9)

[0071] Where U i represents the voltage amplitude at node i, U i,min and U i,max Indicates the minimum and maximum values ​​allowed; I ij Indicates the current on line ij, and I ij,min and I ij,maxIndicates the minimum and maximum values ​​allowed;

[0072] The power balance constraint equation of the Markov decision environment model is set as follows:

[0073]

[0074] Where, P i , Q i and V i represents the active power, reactive power and voltage of node i; B ij , G ij and θ ij represents the susceptance, conductance and phase difference of the branch; l represents the set of child nodes of node i; and represents the active power and reactive power exchanged between node i and the upstream power grid; and represents the active power and reactive power requirements of node i.

[0075] Preferably, the agent training module is specifically used to train the agent in the Markov decision environment model based on the environment model of the distribution network model, using the RainbowDQN algorithm with an adaptive action mask. At each moment, the agent dynamically adjusts the action selection mechanism according to the action mask returned by the environment. The action mask is a Boolean vector, where: True indicates that the current action complies with the specification and is executable; False indicates that the current action does not comply with the specification or has expired and should be prohibited from selection;

[0076] In the action selection phase, the agent calculates the Q-values ​​of all actions based on the current state and then performs the following processing:

[0077]

[0078] In each round, the agent receives the initial state of the distribution network and uses the noise network to output a Q value with random perturbations based on the current state. The Q value of the invalid action is set to negative infinity based on the adaptive action masking technology. The agent then performs the action with the largest Q value and completes a switching operation. After each step, the action is applied to the environment model. The environment updates the state according to the current topology structure and feedbacks the new system state, the immediate reward value, and the flag bit indicating whether the recovery task is completed. The reward function evaluates the contribution of the action to load recovery and encourages the agent to restore more load. In this process, the agent packages the state-action-reward-new state information generated by each interaction into experience samples and stores them in the priority experience replay pool. When the accumulated data in the experience pool reaches a predetermined amount, the agent begins to sample a batch of samples from the experience pool according to the priority experience replay strategy every certain number of steps to update the online network parameters. The target network parameters are updated slightly each time according to the soft update coefficient.

[0079] After the agent executes the action, the topology of the distribution network changes. The environment performs a power flow calculation based on the new network structure to obtain updated information on the recovery of voltage, current, and node load, and feeds the new status back to the agent. At the same time, the environment calculates the immediate reward corresponding to the action of this step: if the action is legal and successfully restores part of the load, a reward is given according to the reward function; if the action has no recovery contribution or causes the constraint to be violated, a negative penalty is imposed; by continuously perceiving the environment, selecting actions, receiving feedback, and updating strategies, the agent gradually learns how to make the correct switching operation sequence under different regional faults and different load conditions, and realizes the adaptive recovery of multi-region distribution networks. When the fluctuation of the reward function during training is less than the set numerical range, the training ends and a trained agent is obtained.

[0080] A non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the distribution network post-disaster intelligent recovery method based on the RainbowDQN algorithm is implemented.

[0081] A computer program product, characterized in that it includes a computer program, which, when running on one or more processors, is used to implement the distribution network post-disaster intelligent recovery method based on the RainbowDQN algorithm.

[0082] It can be seen from the technical solutions provided by the above-mentioned embodiments of the present invention that the method of the present invention uses flexible soft switching technology to quickly isolate the affected area when a fault is caused by a natural disaster, reduce the risk of fault spread, and enable the power supply grid to quickly make necessary topological adjustments after the disaster, thereby achieving efficient distribution of post-disaster power support, thereby effectively improving the resilience and sustainability of the power supply grid after the disaster.

[0083] Additional aspects and advantages of the present invention will be set forth in part in the following description, will become apparent from the following description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0085] Figure 1 A processing flow chart of a method for intelligent post-disaster recovery of a distribution network based on the RainbowDQN (Deep Q Network) algorithm provided by an embodiment of the present invention;

[0086] Figure 2 A topological structure diagram of a modified IEEE123 node distribution network example provided in an embodiment of the present invention.

[0087] Figure 3 A schematic diagram of a RainbowDQN-based agent training recovery strategy reward provided in an embodiment of the present invention;

[0088] Figure 4 A schematic diagram of a RainbowDQN-based agent test recovery strategy reward provided in an embodiment of the present invention;

[0089] Figure 5 A structural diagram of a distribution network post-disaster intelligent recovery device based on the RainbowDQN algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0090] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0091] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0092] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention pertains. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such herein, will not be interpreted in an idealized or overly formal sense.

[0093] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0094] Distributed power sources (such as photovoltaics, energy storage, and microgrids) are widely distributed in distribution networks, and there is an urgent need to introduce intelligent algorithms with autonomous learning and high-dimensional decision-making capabilities to support fast, flexible, and efficient load recovery in uncertain fault scenarios.

[0095] The present invention proposes a method for intelligent post-disaster recovery of distribution networks based on the RainbowDQN algorithm. This method integrates multiple deep reinforcement learning optimization mechanisms, such as dual-Q networks, prioritized experience replay, noise networks, and multi-step rewards, to construct a recovery agent with efficient convergence and policy generalization capabilities, enabling autonomous recovery strategy learning under multi-scenario fault conditions. At the same time, the present invention constructs an interactive simulation environment for a modified IEEE123 node system. By randomly injecting N-3 line faults, dynamically detecting node de-voltage conditions, and combining an adaptive action mask mechanism with power flow feedback calculation, the present invention implements full-process intelligent interactive training from state perception, action selection, to environment transfer, ensuring that the learned strategy is feasible, safe, and engineering deployable.

[0096] The embodiment of the present invention proposes a flow chart of a method for post-disaster distribution network load recovery based on RainbowDQN. Figure 1As shown, the processing steps include the following:

[0097] Step S1: Obtain the distribution network model and the preset N-3 fault combination, and construct a Markov decision environment model for the distribution network load restoration task when N-3 faults occur on the line.

[0098] A modified IEEE123 node system is set as the distribution network model, which includes node data, line data, transformer data, load data, and switch data. A certain number (e.g., 500 groups) of random N-3 line faults are preset as training data, and a certain number (e.g., 100 groups) of random N-3 line faults are preset as test data. The above IEEE123 node system contains 123 nodes, 114 lines, and 11 switches, and divides the distribution network into five functional areas to improve the controllability and flexibility of fault recovery and reconstruction strategies. Each area is connected to the boundary through a tie switch.

[0099] A Markov decision environment model for the load restoration task of the distribution network with N-3 line faults is constructed. The Markov decision model is defined as:

[0100] M=<S,A,P,R,γ> (1)

[0101] Wherein, the Markov decision process consists of state space S, action space A, state transition probability P, reward R and discount factor γ.

[0102] The state space S is defined as:

[0103] S t ={N t ,L t ,W t} (2)

[0104] Where S t is the state of the environment at step t, N t is the node voltage state at step t. If the node loses voltage, N t If the node is not depressurized, N t is 1; L t is the state of the line at step t. If the line fails, L t If the line is not faulty, L t =1; W t is the switch state at step t. If the switch is closed, W t If the switch is open, W t is 0.

[0105] The action space A is defined as:

[0106] A t ={a1,a2,a3,…an} (3)

[0107] Where A t is the action that the agent can choose at step t, a i Indicates that the corresponding i-th switch is in action, that is, if the corresponding switch is open, the switch is closed to put the line into operation; if the corresponding switch is closed, the switch is opened to disconnect the line from operation.

[0108] The established environment is a load recovery model, so the state transfer function is deterministic, that is, the state will give the node voltage state, line state and switch state of the corresponding topology after the action according to the action taken by the intelligent agent.

[0109] Reward function R e,t Defined as:

[0110] R e,t ={R nt +R b} (4)

[0111]

[0112] Where R e,t is the reward obtained at step t, R nt R is the number of nodes newly restored after the corresponding switch takes action compared to the last action. b The purpose is to get the agent closer to the goal of fully restoring the load, that is, to give it a larger reward when it has restored all the loads.

[0113] In order to prevent the agent from making invalid or even wrong actions, it is necessary to punish the agent for its negative actions. The penalties are as follows:

[0114]

[0115] The formula indicates that when the agent takes an action, if no node is restored, it means that the action is invalid. In order to improve the decision-making efficiency of the agent and avoid invalid actions, it is punished accordingly.

[0116] Taking into account the above reward function and penalty function, the total reward function of the present invention is:

[0117] R=R e,t +R p,t (7)

[0118] In a constructed Markov decision environment, the intelligent agent continuously learns switching strategies to achieve efficient load restoration by interacting with the distribution network post-disaster recovery model. At each moment, the agent first perceives the current environmental state, including the power supply status of each region, whether a node has lost voltage, and the overall load recovery progress of the system. Based on this perception, the agent selects an action and applies it to the environmental model. The environment updates its state based on the current topology and provides feedback on the new system state, an immediate reward, and a flag indicating whether the recovery task has been completed. This reward function evaluates the contribution of each action to load recovery, incentivizing the agent to restore more load as quickly and safely as possible. During this process, the agent packages the state-action-reward-new-state information generated by each interaction into experience samples and stores them in a prioritized experience replay pool. During the training phase, the agent regularly draws samples from the experience pool and updates the policy network to reduce Q-value estimation errors and accelerate convergence. By continuously perceiving the environment, selecting actions, receiving feedback, and updating its policy, the agent gradually learns how to make the correct switching operation sequence under different regional faults and load conditions, achieving rapid, safe, and adaptive recovery of multi-region distribution networks.

[0119] Step S2: Based on the distribution network post-disaster load recovery task, set the objective function of the above Markov decision environment model, node voltage constraints, power balance constraints and other constraints.

[0120] The task of post-disaster load recovery in the distribution network is to restore power to the nodes that have lost power due to a fault. The objective function of the above Markov decision environment model is to restore power to all nodes, that is, to maximize the value of the cumulative total reward.

[0121] The node voltage constraints and line current constraints of the above Markov decision environment model are set as follows:

[0122] U i,min ≤U i ≤U i,max (8)

[0123] I ij,min ≤I ij ≤I ij,max (9)

[0124] Where U i represents the voltage amplitude at node i, U i,min and U i,max Indicates the minimum and maximum values ​​allowed; I ij Indicates the current on line ij, and I ij,min and I ij,max Indicates the minimum and maximum values ​​allowed.

[0125] The power balance constraint equation of the above Markov decision environment model is set as follows:

[0126]

[0127] Where, P i , Q i and V i represents the active power, reactive power and voltage of node i; B ij , G ij and θ ij represents the susceptance, conductance and phase difference of the branch; l represents the set of child nodes of node i; and represents the active power and reactive power exchanged between node i and the upstream power grid; and represents the active power and reactive power requirements of node i.

[0128] Step S3: Based on the environment model, the RainbowDQN algorithm with adaptive action mask is used to train the agent in the Markov decision environment model to obtain a trained Markov decision environment model.

[0129] The present invention uses the RainbowDQN algorithm with an adaptive action mask to train the intelligent agent in the Markov decision environment model. During the load recovery process of the distribution network, the executable switching action at each step is subject to multiple operating constraints, resulting in dynamic changes in the action space under different states. In order to prevent the intelligent agent from selecting invalid or illegal actions during the training and reasoning process, the present invention introduces invalid action masking technology in the algorithm implementation. At each moment, the intelligent agent dynamically adjusts the action selection mechanism according to the Action Mask returned by the environment. The action mask is a Boolean vector, where: True indicates that the current action meets the specifications and can be executed; False indicates that the current action does not meet the specifications or has expired, and selection should be prohibited. In the action selection stage, the intelligent agent first calculates the Q value of all actions based on the current state, and then performs the following processing:

[0130]

[0131] That is, the Q value corresponding to the illegal action is set to negative infinity to ensure that these actions will not be selected during the action sampling process. Ensure that the agent outputs valid actions.

[0132] When training an agent using the RainbowDQN algorithm, at the beginning of each round, the agent receives the initial state of the distribution network under a specific fault scenario. Using a policy network with an integrated noise structure, the agent outputs a Q-value vector with random perturbations to enhance its exploration capabilities. To prevent the agent from taking illegal or invalid switching operations, the system incorporates an adaptive action masking technique at each decision step. This forcibly sets the Q-values ​​corresponding to actions that violate the distribution network's operating constraints to negative infinity, thereby excluding these actions from the policy output. The agent then selects the action with the highest expected reward based on these filtered Q-values, executing the switching operation with the highest Q-value. After executing this action, the distribution network topology changes, and the environment performs a power flow calculation based on the new network structure to obtain updated information such as voltage, current, and node load recovery. The new state is then fed back to the agent. Simultaneously, the environment calculates the immediate reward for this action: if the action is legal and successfully restores some load, a reward is awarded according to the reward function; if the action does not contribute to recovery or violates a constraint, a negative penalty is applied.

[0133] At the same time, the agent packages the current step's state, action, reward, and next state into an experience unit and stores it in the experience pool. When the number of samples stored in the experience pool exceeds a preset threshold, the agent begins updating the policy network at fixed intervals. Specifically, the agent uses a prioritized experience replay (PER) mechanism to sample high-value samples from the experience pool based on the TD error to form a mini-batch dataset for gradient backpropagation training of the Q network. Simultaneously, the target network parameters are smoothly replaced using a soft update.

[0134] Throughout the training process, the goal of the agent is to maximize the objective function, that is, to maximize the total reward value of the agent.

[0135] like Figure 2 A modified IEEE123 node distribution network topology diagram provided by an embodiment of the present invention. Figure 3 This figure illustrates the rewards for a RainbowDQN-based agent training recovery strategy, as provided in an embodiment of the present invention. Initially, the rewards are low and fluctuate widely. With experience, they rapidly increase and stabilize, indicating that the agent has mastered an effective load recovery strategy. Later, the sliding average reward stabilizes and fluctuations decrease, validating the effectiveness and convergence of RainbowDQN in post-disaster recovery tasks. Training terminates when the fluctuations in the reward function during training fall within the set range, resulting in a trained agent.

[0136] The trained agent demonstrates efficient convergence and policy generalization capabilities. After training, its network parameters are retained, effectively saving the agent. The agent is then tested using a test set containing data from 100 N-3 fault scenarios. Both the test and training models use Markov decision models, but the training and test data differ. The agent's solution follows the same steps as training: detecting states, deriving actions, and calculating rewards, but without parameter updates. Subsequently, the parameters of the distribution network experiencing the N-3 line fault are transmitted to the qualified agent, which then outputs a load recovery strategy for the distribution network.

[0137] Step S4: Use the N-3 line fault in the distribution network to test the intelligent agent and derive the corresponding load recovery strategy, as follows:

[0138] In order to verify the effectiveness of the technical solution of the present invention, a modified IEEE123 node distribution network example is used for analysis and verification.

[0139] like Figure 4 A diagram illustrating the rewards for testing recovery strategies for a RainbowDQN-based agent in an embodiment of the present invention. The agent no longer experiences negative total rewards under various fault scenarios, demonstrating its ability to reliably restore loads lost due to N–3 line faults. These results demonstrate that the trained agent exhibits strong generalization and robustness under various random N–3 fault conditions, reliably developing high-quality load recovery scheduling plans within a limited number of steps.

[0140] Table 1

[0141] Line failure 0,49,64 9,110,17 103,42,23 Switch action 5,8,2 2,10,8 2,9,10

[0142] Table 1 shows the specific results obtained from three selected test environments. When faults occurred on lines 0, 49, and 64, the agent first operated switches 5 and 8 to restore power to the loads at nodes 52–83; then, it operated switch 2 to restore power to the loads at nodes 1–17. When faults occurred on lines 9, 110, and 17, the agent first operated switch 2 to reconnect the load area downstream of node 13, which had been disconnected by the fault, with the healthy feeder segment at the parent node 18, forming a new feeder branch without closing any loops. Next, it operated switch 10 to restore power to the isolated load area on the left. Finally, it operated switch 8 to reconnect the isolated, deenergized area in the middle with the area at node 117, restoring the middle fault. The load restoration process for the third fault combination follows the same principle.

[0143] The structure of a distribution network post-disaster intelligent recovery device based on the RainbowDQN algorithm provided by the embodiment of the present invention is as follows: Figure 5 As shown, it includes the following modules:

[0144] A Markov decision environment model building module 51 is used to obtain a distribution network model and a preset N-3 fault combination, and to build a Markov decision environment model for a distribution network load restoration task when an N-3 fault occurs on a line;

[0145] An objective function and constraint condition setting module 52 is used to set the objective function and constraint conditions of the Markov decision environment model;

[0146] An agent training module 53 is configured to train an agent in the Markov decision environment model using a RainbowDQN algorithm with an adaptive action mask to obtain a trained agent;

[0147] The load recovery strategy acquisition module 54 is used to transmit parameter information of the distribution network with N-3 line failure to the trained intelligent agent, and the intelligent agent outputs the load recovery strategy of the distribution network with N-3 line failure.

[0148] Specifically, the Markov decision environment model building module is used to set a modified IEEE123 node system as a distribution network model, which includes node data, line data, transformer data, load data, and switch data, and preset a certain number of random N-3 line fault data as training data and test data respectively;

[0149] A Markov decision environment model for the load restoration task of the distribution network with N-3 line faults is constructed. The Markov decision model is defined as:

[0150] M=<S,A,P,R,γ> (1)

[0151] Wherein, the Markov decision process consists of state space S, action space A, state transition probability P, reward function R and discount factor γ;

[0152] The state space S is defined as:

[0153] S t ={N t ,L t ,W t} (2)

[0154] Where S t is the state of the environment at step t, N t is the node voltage state at step t. If the node loses voltage, N t If the node is not depressurized, N t is 1; L t is the state of the line at step t. If the line fails, L t If the line is not faulty, Lt is 1; W t is the switch state at the t-th step, if the switch is closed then W t is 1, if the switch is open then W t is 0;

[0155] The action space A is defined as:

[0156] A t = {a1, a2, a3, … a n} (3)

[0157] In the formula, A t is the action selected by the agent at the t-th step, a i indicates that the corresponding i-th switch is operated, that is, if the corresponding switch is open, the switch is closed to put the line into operation, and if the corresponding switch is closed, the switch is opened to disconnect the line from operation;

[0158] The reward function R e,t is defined as:

[0159] R e,t = {R nt + R b} (4)

[0160]

[0161] In the formula, R e,t is the reward obtained at the t-th step, R nt is the number of newly recovered nodes after the corresponding switch is operated compared with the last operation, and R b is the reward for the agent to approach the target of completely restoring the load;

[0162] The corresponding punishment for the negative behavior of the agent is as follows:

[0163]

[0164] In the formula, if the agent takes an action and does not recover any node, it indicates that the action is invalid, and a corresponding punishment is given.

[0165] Specifically, the above-mentioned objective function and constraint setting module, which is specifically used for the task of post-disaster load restoration of a power distribution network, is to restore power supply to voltage loss nodes caused by failures, and the objective function of the Markov decision environment model is set to restore power supply to all nodes, that is, to maximize the value of the cumulative total reward;

[0166] The node voltage constraint and line current constraint of the Markov decision environment model are set as follows:

[0167] U i,min ≤ Ui ≤U i,max (8)

[0168] I ij,min ≤I ij ≤I ij,max (9)

[0169] Where U i represents the voltage amplitude at node i, U i,min and U i,max Indicates the minimum and maximum values ​​allowed; I ij Indicates the current on line ij, and I ij,min and I ij,max Indicates the minimum and maximum values ​​allowed;

[0170] The power balance constraint equation of the Markov decision environment model is set as follows:

[0171]

[0172] Where, P i , Q i and V i represents the active power, reactive power and voltage of node i; B ij , G ij and θ ij represents the susceptance, conductance and phase difference of the branch; l represents the set of child nodes of node i; and represents the active power and reactive power exchanged between node i and the upstream power grid; and represents the active power and reactive power requirements of node i.

[0173] Specifically, the above-mentioned intelligent agent training module is specifically used to train the intelligent agent in the Markov decision environment model based on the environmental model of the distribution network model, using the RainbowDQN algorithm with an adaptive action mask. At each moment, the intelligent agent dynamically adjusts the action selection mechanism according to the action mask returned by the environment. The action mask is a Boolean vector, where: True indicates that the current action meets the specifications and is executable; False indicates that the current action does not meet the specifications or has expired and should be prohibited from selection;

[0174] In the action selection phase, the agent calculates the Q-values ​​of all actions based on the current state and then performs the following processing:

[0175]

[0176] In each round, the agent receives the initial state of the power distribution network, outputs the Q value with random disturbance according to the current state using the noise network, sets the Q value of invalid actions to negative infinity according to the adaptive action mask technology, then the agent executes the action with the maximum Q value, completes a switch operation, after each action, the action is applied to the environment model, the environment updates the state according to the current topology structure, and feeds back the new system state, the immediate reward value and the flag indicating whether the restoration task is completed, the reward function evaluates the contribution of the current action to the load restoration, and encourages the agent to restore more loads, in this process, the agent packs the state-action-reward-new state information generated in each interaction into experience samples and stores them in the priority experience replay pool, when the amount of data accumulated in the experience pool reaches a predetermined amount, the agent starts to sample a batch of samples from the experience pool according to the priority experience replay strategy to update the online network parameters, and the target network parameters are updated with small amplitude network parameters each time according to the soft update coefficient;

[0177] After the agent executes the action, the topology structure of the power distribution network changes, the environment performs a power flow calculation based on the new network structure to obtain the voltage, current, node load restoration update information, and feeds back the new state to the agent, at the same time, the environment calculates the immediate reward corresponding to the current action: if the action is legal and successfully restores part of the load, the reward function is given a reward; if the action has no restoration contribution or causes the constraint to be broken, a negative penalty will be applied; through continuous perception of the environment, selection of action, reception of feedback and update of strategy, the agent gradually learns how to make correct switch operation sequences in different regional faults and different load states, and realizes the adaptive restoration of the multi-regional power distribution network, when the fluctuation of the reward function in the training process is less than a set value range, the training is ended, and the trained agent is obtained.

[0178] The specific process of the device for post-disaster intelligent restoration of the power distribution network based on the RainbowDQN algorithm is similar to the foregoing method embodiment, and will not be described here.

[0179] In summary, the embodiment of the present application constructs a post-disaster load restoration method for power distribution networks based on RainbowDQN reinforcement learning, which significantly improves the restoration speed and reliability in complex multi-fault scenarios. The method deeply integrates the N-3 fault combination of the modified IEEE123 node power distribution network and the voltage loss node detection process, constructs a true circuit topology semantic environment, and provides accurate state information for the agent; the dual network and Dueling structure, priority experience replay technology, noise network and multi-step replay technology are introduced, which effectively enhances the exploration efficiency and experience utilization rate of the high-dimensional action space; combined with the adaptive action mask mechanism and the soft update strategy, the dynamic shielding and smooth parameter updating of the executed or safety constraint outside actions are realized, ensuring that the strategy is always iterated within the legal action subset.

[0180] Experimental results demonstrate that the Q-value decomposition and masking mechanism improves decision interpretability and robustness. Compared with existing technologies, this method not only significantly reduces the reliance on manual rules and parameter adjustment costs, but also effectively suppresses the risk of policy "jumps." This provides highly reliable, controllable, and easily deployable online decision support for post-disaster intelligent dispatching of distribution networks, significantly shortening power outage recovery time and improving system resilience and power supply reliability.

[0181] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0182] From the above description of the embodiments, it can be seen that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus the necessary general-purpose hardware platform. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0183] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without making any creative efforts.

[0184] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A distribution network post-disaster intelligent recovery method based on the RainbowDQN algorithm, characterized in that: include: Obtain the distribution network model and the preset N-3 fault combination, and construct a Markov decision environment model for the distribution network load restoration task when N-3 faults occur on the line; Setting the objective function and constraints of the Markov decision environment model; The RainbowDQN algorithm with adaptive action mask is used to train the intelligent agent in the Markov decision environment model to obtain a trained intelligent agent; The parameter information of the distribution network with N-3 line failure is transmitted to the trained intelligent agent, and the intelligent agent outputs the load recovery strategy of the distribution network with N-3 line failure.

2. The method according to claim 1, characterized in that The method of obtaining a distribution network model and a preset N-3 fault combination and constructing a Markov decision environment model for a distribution network load restoration task when an N-3 fault occurs on a line includes: A modified IEEE123 node system is set as the distribution network model, which includes node data, line data, transformer data, load data, and switch data. A certain number of random N-3 line fault data are preset as training data and test data respectively. A Markov decision environment model for the load restoration task of the distribution network with N-3 line faults is constructed. The Markov decision model is defined as: M=<S,A,P,R,γ> (1) Wherein, the Markov decision process consists of state space S, action space A, state transition probability P, reward function R and discount factor γ; The state space S is defined as: S t ={N t ,L t ,W t } (2) Where S t is the state of the environment at step t, N t is the node voltage state at step t. If the node loses voltage, N t If the node is not depressurized, N t is 1; L t is the state of the line at step t. If the line fails, L t If the line is not faulty, L t =1; W t is the switch state at step t. If the switch is closed, W t If the switch is open, W t is 0; The action space A is defined as: A t ={a1,a2,a3,…a n } (3) Where A t is the action selected by the agent at step t, a i Indicates that the corresponding i-th switch is in action, that is, if the corresponding switch is open, the switch is closed to put the line into operation; if the corresponding switch is closed, the switch is opened to disconnect the line from operation; Reward function R e,t Defined as: R e,t ={R nt +R b } (4) Where R e,t is the reward obtained at step t, R nt R is the number of nodes newly restored after the corresponding switch takes action compared to the last action. b The reward is to get the agent closer to the goal of fully restoring the load; The corresponding penalties for negative actions taken by the agent are as follows: The formula indicates that when the agent takes an action, if no node is restored, it means that the action is invalid and a corresponding penalty is given.

3. The method according to claim 2, characterized in that The objective function and constraint conditions of the Markov decision environment model are set as follows: The task of post-disaster load restoration in the distribution network is to restore power to the nodes that have lost power due to a fault. The objective function of the Markov decision environment model is to restore power to all nodes, that is, to maximize the value of the cumulative total reward. The node voltage constraints and line current constraints of the Markov decision environment model are set as follows: IN i,min ≤U i ≤U i,max (8) I ij,min ≤I ij ≤I ij,max (9) Where U i represents the voltage amplitude at node i, U i,min and U i,max Indicates the minimum and maximum values ​​allowed; I ij Indicates the current on line ij, and I ij,min and I ij,max Indicates the minimum and maximum values ​​allowed; The power balance constraint equation of the Markov decision environment model is set as follows: P i =P i grid -P i load (12) Where, P i , Q i and V i represents the active power, reactive power and voltage of node i; B ij , G ij and θ ij represents the susceptance, conductance and phase difference of the branch; l represents the set of child nodes of node i; P i grid and Q i grid represents the active power and reactive power exchanged between node i and the upstream power grid; P i load and Q i load represents the active power and reactive power requirements of node i.

4. The method according to claim 3, characterized in that The RainbowDQN algorithm with adaptive action mask is used to train the agent in the Markov decision environment model to obtain a trained agent, including: Based on the environmental model of the distribution network model, the RainbowDQN algorithm with an adaptive action mask is used to train the intelligent agent in the Markov decision environment model. At each moment, the intelligent agent dynamically adjusts the action selection mechanism according to the action mask returned by the environment. The action mask is a Boolean vector, where: True indicates that the current action meets the specifications and is executable; False indicates that the current action does not meet the specifications or has expired and should be prohibited from selection; In the action selection phase, the agent calculates the Q-values ​​of all actions based on the current state and then performs the following processing: In each round, the agent receives the initial state of the distribution network and uses the noise network to output a Q value with random perturbations based on the current state. The Q value of the invalid action is set to negative infinity based on the adaptive action masking technology. The agent then performs the action with the largest Q value and completes a switching operation. After each step, the action is applied to the environment model. The environment updates the state according to the current topology structure and feedbacks the new system state, the immediate reward value, and the flag bit indicating whether the recovery task is completed. The reward function evaluates the contribution of the action to load recovery and encourages the agent to restore more load. In this process, the agent packages the state-action-reward-new state information generated by each interaction into experience samples and stores them in the priority experience replay pool. When the accumulated data in the experience pool reaches a predetermined amount, the agent begins to sample a batch of samples from the experience pool according to the priority experience replay strategy every certain number of steps to update the online network parameters. The target network parameters are updated slightly each time according to the soft update coefficient. After the agent executes the action, the topology of the distribution network changes. The environment performs a power flow calculation based on the new network structure to obtain updated information on the recovery of voltage, current, and node load, and feeds the new status back to the agent. At the same time, the environment calculates the immediate reward corresponding to the action of this step: if the action is legal and successfully restores part of the load, a reward is given according to the reward function; if the action has no recovery contribution or causes the constraint to be violated, a negative penalty is imposed; by continuously perceiving the environment, selecting actions, receiving feedback, and updating strategies, the agent gradually learns how to make the correct switching operation sequence under different regional faults and different load conditions, and realizes the adaptive recovery of multi-region distribution networks. When the fluctuation of the reward function during training is less than the set numerical range, the training ends and a trained agent is obtained.

5. A distribution network post-disaster intelligent recovery device based on the RainbowDQN algorithm, characterized in that: include: The Markov decision environment model construction module is used to obtain the distribution network model and the preset N-3 fault combination, and construct a Markov decision environment model for the distribution network load restoration task when the line has N-3 faults; An objective function and constraint condition setting module, used to set the objective function and constraint conditions of the Markov decision environment model; An agent training module is used to train the agent in the Markov decision environment model using the RainbowDQN algorithm with an adaptive action mask to obtain a trained agent; The load recovery strategy acquisition module is used to transmit parameter information of the distribution network where the N-3 line fault occurs to the trained intelligent agent, and the intelligent agent outputs the load recovery strategy of the distribution network where the N-3 line fault occurs.

6. The device according to claim 5, characterized in that: The Markov decision environment model building module is used to set a modified IEEE123 node system as a distribution network model, which includes node data, line data, transformer data, load data and switch data, and preset a certain number of random N-3 line fault data as training data and test data respectively; A Markov decision environment model for the load restoration task of the distribution network with N-3 line faults is constructed. The Markov decision model is defined as: M=<S,A,P,R,γ> (1) Wherein, the Markov decision process consists of state space S, action space A, state transition probability P, reward function R and discount factor γ; The state space S is defined as: S t ={N t ,L t ,W t } (2) Where S t is the state of the environment at step t, N t is the node voltage state at step t. If the node loses voltage, N t If the node is not depressurized, N t is 1; L t is the state of the line at step t. If the line fails, L t If the line is not faulty, L t =1; W t is the switch state at step t. If the switch is closed, W t If the switch is open, W t is 0; The action space A is defined as: A t ={a1,a2,a3,…a n } (3) Where A t is the action selected by the agent at step t, a i Indicates that the corresponding i-th switch is in action, that is, if the corresponding switch is open, the switch is closed to put the line into operation; if the corresponding switch is closed, the switch is opened to disconnect the line from operation; Reward function R e,t Defined as: R e,t ={R nt +R b } (4) Where R e,t is the reward obtained at step t, R nt R is the number of nodes newly restored after the corresponding switch takes action compared to the last action. b The reward is to get the agent closer to the goal of fully restoring the load; The corresponding penalties for negative actions taken by the agent are as follows: The formula indicates that when the agent takes an action, if no node is restored, it means that the action is invalid and a corresponding penalty is given.

7. The device according to claim 6, characterized in that: The objective function and constraint condition setting module is specifically used for the task of post-disaster load recovery of the distribution network, which is to restore power to the nodes that have lost power due to the fault. The objective function of the Markov decision environment model is set to restore power to all nodes, that is, to maximize the value of the cumulative total reward; The node voltage constraints and line current constraints of the Markov decision environment model are set as follows: IN i,min ≤U i ≤U i,max (8) I ij,min ≤I ij ≤I ij,max (9) Where U i represents the voltage amplitude at node i, U i,min and U i,max Indicates the minimum and maximum values ​​allowed; I ij Indicates the current on line ij, and I ij,min and I ij,max Indicates the minimum and maximum values ​​allowed; The power balance constraint equation of the Markov decision environment model is set as follows: P i =P i grid -P i load (12) Q i =Q i grid -Q i load (13) Where, P i , Q i and V i represents the active power, reactive power and voltage of node i; B ij , G ij and θ ij represents the susceptance, conductance and phase difference of the branch; l represents the set of child nodes of node i; P i grid and Q i grid represents the active power and reactive power exchanged between node i and the upstream power grid; P i load and Q i load represents the active power and reactive power requirements of node i.

8. The device according to claim 7, characterized in that: The agent training module is specifically used to train the agent in the Markov decision environment model based on the environment model of the distribution network model, using the RainbowDQN algorithm with an adaptive action mask. At each moment, the agent dynamically adjusts the action selection mechanism according to the action mask returned by the environment. The action mask is a Boolean vector, where: True indicates that the current action meets the specifications and is executable; False indicates that the current action does not meet the specifications or has expired and should be prohibited from selection; In the action selection phase, the agent calculates the Q-values ​​of all actions based on the current state and then performs the following processing: In each round, the agent receives the initial state of the distribution network and uses the noise network to output a Q value with random perturbations based on the current state. The Q value of the invalid action is set to negative infinity based on the adaptive action masking technology. The agent then performs the action with the largest Q value and completes a switching operation. After each step, the action is applied to the environment model. The environment updates the state according to the current topology structure and feedbacks the new system state, the immediate reward value, and the flag bit indicating whether the recovery task is completed. The reward function evaluates the contribution of the action to load recovery and encourages the agent to restore more load. In this process, the agent packages the state-action-reward-new state information generated by each interaction into experience samples and stores them in the priority experience replay pool. When the accumulated data in the experience pool reaches a predetermined amount, the agent begins to sample a batch of samples from the experience pool according to the priority experience replay strategy every certain number of steps to update the online network parameters. The target network parameters are updated slightly each time according to the soft update coefficient. After the agent executes the action, the topology of the distribution network changes. The environment performs a power flow calculation based on the new network structure to obtain updated information on the recovery of voltage, current, and node load, and feeds the new status back to the agent. At the same time, the environment calculates the immediate reward corresponding to the action of this step: if the action is legal and successfully restores part of the load, a reward is given according to the reward function; if the action has no recovery contribution or causes the constraint to be violated, a negative penalty is imposed; by continuously perceiving the environment, selecting actions, receiving feedback, and updating strategies, the agent gradually learns how to make the correct switching operation sequence under different regional faults and different load conditions, and realizes the adaptive recovery of multi-region distribution networks. When the fluctuation of the reward function during training is less than the set numerical range, the training ends and a trained agent is obtained.

9. A non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the distribution network post-disaster intelligent recovery method based on the RainbowDQN algorithm as described in any one of claims 1 to 4 is implemented.

10. A computer program product, characterized in that The invention comprises a computer program, which, when running on one or more processors, is used to implement the distribution network post-disaster intelligent recovery method based on the RainbowDQN algorithm as described in any one of claims 1 to 4.

Images