A method and system for optimizing the resilience of campus distribution network based on reinforcement learning

Abstract

The present invention provides a method and system for optimizing the resilience of a park distribution network based on reinforcement learning, the method comprising obtaining a historical operating status data set of a target park distribution network; based on the idea of ​​reinforcement learning, performing interactive training of power grid scheduling simulation according to the historical operating status data set, and constructing a dynamic optimization scheduling model for the resilience of the distribution network; obtaining the real-time operating status data of the target park distribution network, inputting the real-time operating status data into the dynamic optimization scheduling model for the resilience of the distribution network for optimization scheduling analysis, and generating a corresponding real-time optimal scheduling strategy for the distribution network; and optimizing the resilience of the target park distribution network according to the real-time optimal scheduling strategy for the distribution network. The present invention realizes real-time status monitoring and analysis of the park power grid and intelligent power grid resilience optimization scheduling, can promptly and effectively respond to emergencies and nonlinear impact loads, and provide reliable protection for the power supply quality and power supply safety of the park distribution network.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and in particular to a method and system for optimizing the resilience of a campus distribution network based on reinforcement learning. Background Art

[0002] With the rapid development of social economy, various park users such as industrial technology parks, ecological parks and real estate parks are expanding rapidly. At the same time, various electrical equipment are constantly being put into use. In particular, the sharp increase in high-power and nonlinear impact loads in industrial parks has brought severe challenges to the power supply quality and power supply safety of the park's distribution network.

[0003] The research on improving the ability of existing park distribution networks to resist impact loads mainly includes: 1) using distributed controllable resources such as distributed power sources, energy storage and controllable loads to complement each other, realize peak shaving and valley filling of distribution networks, and provide backup protection for scenarios subject to large load impacts; 2) In view of the uncertainty and complexity of fault scenarios caused by multiple impact loads, Monte Carlo simulation and information entropy method are used to extract fault scenario sets, and resilience assessment under impact loads is performed in the form of expected values ​​and probability distributions, providing a basis for subsequent resilience optimization and improvement. Although the above methods can meet the needs of improving the resilience of park distribution networks to a certain extent, they are limited to post-analysis optimization for large load impact scenarios, and cannot make full use of multi-source data to carry out real-time monitoring and timely dispatch optimization of the resilience status of distribution networks. There is a lack of intelligent and automated means to improve the resilience of power grids, making it difficult to effectively respond to emergencies and nonlinear impact loads, and unable to ensure the power supply quality and power supply safety of park distribution networks. Summary of the invention

[0004] The purpose of the present invention is to provide a method for optimizing the resilience of a park distribution network based on reinforcement learning. By collecting high-efficiency and multi-dimensional park distribution network operation data, an intelligent power grid resilience dynamic optimization scheduling model is constructed based on nonlinear optimization theory and reinforcement learning ideas to perform real-time status monitoring, analysis and scheduling optimization on the park distribution network. While effectively responding to emergencies and the impact of nonlinear shock loads, it can also improve the operating efficiency and response speed of the distribution network, significantly improve the resilience and stability of the distribution network, and provide reliable protection for the power supply quality and power supply safety of the park distribution network.

[0005] In order to achieve the above objectives, it is necessary to provide a reinforcement learning-based campus distribution network resilience optimization scheduling method and system to address the above technical problems.

[0006] In a first aspect, an embodiment of the present invention provides a method for optimizing the resilience of a campus distribution network based on reinforcement learning, the method comprising the following steps:

[0007] Obtain a historical operating status data set of the target park distribution network; the historical operating status data set includes the load conditions of power equipment, the level of harmonic distortion of the power grid, the amplitude of voltage fluctuation of the power grid, the change in resilience, external environmental conditions and power grid fault records;

[0008] Based on the idea of ​​reinforcement learning, simulated interactive training between historical operating status and real-time operating status is carried out according to the historical operating status data set to build a dynamic optimization scheduling model for distribution network resilience;

[0009] Acquire the real-time operating status data of the target park distribution network, input the real-time operating status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis, and generate the corresponding distribution network real-time optimal scheduling strategy;

[0010] According to the real-time optimal dispatching strategy of the distribution network, the distribution network of the target park is dispatched for grid resilience optimization.

[0011] Furthermore, the step of performing interactive training of power grid dispatch simulation based on the historical operation status data set and constructing a distribution network resilience dynamic optimization dispatch model based on the reinforcement learning idea includes:

[0012] Based on the reinforcement learning idea, a dynamic optimization reinforcement learning model is constructed according to the dispatching process of the target park distribution network; the state space and action space of the dynamic optimization reinforcement learning model are respectively defined according to the operating state and dispatching strategy of the target park distribution network;

[0013] The dynamic optimization reinforcement learning model is trained according to the historical operating status data set to obtain the distribution network resilience dynamic optimization scheduling model.

[0014] Furthermore, the reward function of the dynamic optimization reinforcement learning model is constructed according to a preset indicator system; the preset indicator system includes changes in load stability, changes in power dispatching costs, changes in equipment reliability, changes in harmonic distortion levels, changes in voltage fluctuation amplitudes, and changes in grid resilience.

[0015] Furthermore, the dynamic optimization reinforcement learning model is expressed as:

[0016]

[0017]

[0018]

[0019] in, and Respectively The state space and action space at each moment; , , , , and Respectively The load of electrical equipment at all times, external environmental conditions, grid harmonic distortion level, grid voltage fluctuation amplitude, grid fault records, and changes in grid resilience; Represents a set of scheduling strategies; express The first a scheduling strategy; Indicates the number of scheduling strategies; , , , and They represent the changes in load stability, power dispatching cost, equipment reliability, harmonic distortion level and voltage fluctuation amplitude at time t respectively; is the weight coefficient; express Status at all times Next action The corresponding reward value.

[0020] Furthermore, the value function network of the dynamic optimization reinforcement learning model is a deep neural network; the deep neural network includes a state input layer, a feature extraction layer, a feature fusion layer and a strategy value output layer connected in sequence; the feature extraction layer includes a parallel load feature extraction layer, a harmonic feature extraction layer, a voltage feature extraction layer, a toughness feature extraction layer, an environmental feature extraction layer and a fault feature extraction layer; the output of the strategy value output layer is expressed as:

[0021]

[0022] In the formula, Represents the output value of the strategy value output layer; , and Represent the weight matrix, bias vector and activation function of the strategy value output layer respectively; and Represent the weight matrix and bias vector of the feature fusion layer respectively; Represents the network parameters of the value function network.

[0023] Furthermore, the step of training the dynamic optimization reinforcement learning model according to the historical operation status data set to obtain the distribution network resilience dynamic optimization scheduling model includes:

[0024] Acquire each state to be analyzed from the historical operation state data set according to a preset time step, and obtain a corresponding action to be executed according to the state to be analyzed and the dynamic optimization reinforcement learning model;

[0025] Simulate and execute the action to be executed, and obtain the corresponding next running state and action execution reward;

[0026] Create a corresponding state transition experience pool based on each state to be analyzed, the corresponding action to be executed, the action execution reward and the next running state;

[0027] A preset number of batch samples are randomly extracted from the state transition experience pool, and the dynamic optimization reinforcement learning model is trained based on a preset loss function to obtain the distribution network resilience dynamic optimization scheduling model; the preset loss function is expressed as:

[0028]

[0029] in, Indicates status Next action Enter the state And get the reward value Expected value; Represents the state transition experience pool; represents the discount factor; and Represent the current value function network parameters and the target value function network parameters respectively; Represents the current value function network parameters Next, status Execute an action The current strategy value at time ; Represents the current value function network parameters Next, status Execute an action The target strategy value at the time Represents the current value function network parameters Next, status Execute all actions The maximum target strategy value; , and Respectively Changes in load stability, harmonic distortion level, voltage fluctuation amplitude and grid resilience at each moment; Represents the weight coefficient.

[0030] Furthermore, the step of inputting the real-time operating status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis to generate the corresponding distribution network real-time optimal scheduling strategy also includes:

[0031] Extracting features from the real-time operating status data according to the distribution network resilience dynamic optimization scheduling model to obtain corresponding comprehensive feature values;

[0032] The comprehensive characteristic value is compared with a preset fault characteristic threshold and a preset event characteristic threshold respectively to obtain a corresponding fault identification result and event identification result.

[0033] Furthermore, the method further comprises:

[0034] Obtain the new operating status data and new status execution reward value of the target park distribution network after the grid resilience optimization scheduling, and dynamically optimize the scheduling strategy set of the distribution network resilience dynamic optimization scheduling model based on the new operating status data and the new status execution reward value.

[0035] In a second aspect, an embodiment of the present invention provides a park distribution network resilience optimization scheduling system based on reinforcement learning, the system comprising:

[0036] A data set acquisition module is used to acquire a historical operating status data set of the target park distribution network; the historical operating status data set includes the load of electrical equipment, the level of harmonic distortion of the power grid, the amplitude of voltage fluctuation of the power grid, the change in the resilience of the power grid, external environmental conditions and power grid fault records;

[0037] A model building module is used to perform interactive training of power grid dispatch simulation based on the historical operation status data set based on the idea of ​​reinforcement learning, and to build a dynamic optimization dispatch model for distribution network resilience;

[0038] A strategy generation module is used to obtain the real-time operating status data of the target park distribution network, input the real-time operating status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis, and generate the corresponding distribution network real-time optimal scheduling strategy;

[0039] The optimization scheduling module is used to perform grid resilience optimization scheduling on the target park distribution network according to the real-time optimal scheduling strategy of the distribution network.

[0040] Furthermore, the system further comprises:

[0041] The strategy optimization module is used to obtain the new operating status data and new status execution reward value of the target park distribution network after the grid resilience optimization scheduling, and dynamically optimize the scheduling strategy set of the distribution network resilience dynamic optimization scheduling model based on the new operating status data and the new status execution reward value.

[0042] The above application provides a method and system for optimizing the resiliency of a park distribution network based on reinforcement learning. The method can obtain the historical operating status data set of the target park distribution network, conduct interactive training of power grid scheduling simulation based on the historical operating status data set based on the reinforcement learning idea, and construct a dynamic optimization scheduling model for the resilience of the distribution network. Then, the real-time operating status data of the target park distribution network is obtained, and the real-time operating status data is input into the dynamic optimization scheduling model for the resilience of the distribution network for optimization scheduling analysis to generate the corresponding real-time optimal scheduling strategy for the distribution network, and the technical solution for optimizing the resiliency of the power grid for the target park distribution network according to the real-time optimal scheduling strategy for the distribution network. Compared with the prior art, the method for optimizing the resiliency of a park distribution network based on reinforcement learning collects high-efficiency and multi-dimensional park distribution network operation data, and constructs an intelligent power grid resilience dynamic optimization scheduling model based on nonlinear optimization theory to perform real-time status monitoring and analysis and scheduling optimization on the park distribution network. While effectively responding to emergencies and nonlinear impact loads, it can also improve the operating efficiency and response speed of the distribution network, significantly improve the resilience and stability of the distribution network, and provide reliable protection for the power supply quality and power supply safety of the park distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of a method for optimizing the resilience of a campus distribution network based on reinforcement learning in an embodiment of the present invention;

[0044] Figure 2 is a schematic diagram of the structure of a value function network of a dynamic optimization reinforcement learning model in an embodiment of the present invention;

[0045] Figure 3 It is another flow chart of the method for optimizing the resilience of a campus distribution network based on reinforcement learning in an embodiment of the present invention;

[0046] Figure 4 It is a structural schematic diagram of a park distribution network resilience optimization dispatching system based on reinforcement learning in an embodiment of the present invention;

[0047] Figure 5 It is another structural diagram of the campus distribution network resilience optimization and dispatching system based on reinforcement learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and beneficial effects of the present application clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the embodiments described below are part of the embodiments of the present invention and are only used to illustrate the present invention, but are not used to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] The method for optimizing the resilience of the campus distribution network based on reinforcement learning provided by the present invention can be understood as a method based on the current application status that the existing technology cannot respond to the emergencies and nonlinear impact loads of the campus distribution network in real time. It is based on high-efficiency and multi-dimensional park distribution network operation data, and constructs an intelligent power grid resilience dynamic optimization scheduling model based on nonlinear optimization theory combined with reinforcement learning ideas to perform real-time state monitoring and analysis and scheduling optimization on the campus distribution network, effectively respond to emergencies and nonlinear impact loads during the operation of the power grid, so as to improve the operation efficiency and response speed of the distribution network, and enhance the resilience and stability of the distribution network. The following embodiments will explain in detail the method for optimizing the resilience of the campus distribution network based on reinforcement learning of the present invention.

[0050] In one embodiment, Figure 1 As shown, a method for optimizing the resilience of a campus distribution network based on reinforcement learning is provided, comprising the following steps:

[0051] S11. Obtain a historical operating status data set of the target park distribution network; wherein, the historical operating status data set can be understood as the target park distribution network operating status data obtained for a preset time period; the corresponding preset time period and collection frequency can be set according to actual application requirements, and are not specifically limited here.

[0052] In order to improve the operation efficiency and response speed of the power grid and realize real-time, comprehensive and reliable monitoring and analysis of nonlinear power grid load impact, this embodiment preferably collects multi-source data of the park distribution network for analysis and research; specifically, the historical operation status data set includes the load conditions of power equipment, the level of power grid harmonic distortion, the amplitude of power grid voltage fluctuation, the change in power grid resilience, external environmental conditions and power grid fault records; wherein, the load conditions of power equipment can be understood as the power consumption conditions of each power equipment in the distribution network obtained through smart meters or existing load monitoring equipment; the level of power grid harmonic distortion can be understood as the total harmonic distortion THD (Total Harmonic Distortion) of the power grid obtained through power quality monitoring equipment; the amplitude of power grid voltage fluctuation can be understood as the amplitude of distribution network voltage fluctuation obtained through voltage monitoring equipment; external environmental conditions can be understood as data such as temperature, humidity and weather conditions of the target park obtained by deploying environmental sensors; power grid fault records can be understood as the fault occurrence time and corresponding fault type of each power equipment obtained through the existing power equipment fault diagnosis system.

[0053] It should be noted that the historical operating status data set in this embodiment can be understood as a data set obtained by cleaning, deduplicating, processing missing exceptions and standardizing the actually collected park distribution network operating data set in order to ensure the reliability and effectiveness of the subsequent training to build a dynamic optimization scheduling model for distribution network resilience; that is, in actual applications, after obtaining the operating data set of the target park distribution network in the above manner, each type of operating data is cleaned, deduplicated and processed for missing exceptions, and then standardized, and each type of data is normalized to the same dimensional range to eliminate the dimensional differences between different data sources, and obtain the required historical operating status data set. The specific method of standardizing various types of operating data can be selected according to actual application requirements, and is not specifically limited here.

[0054] S12. Based on the idea of ​​reinforcement learning, interactive training of power grid dispatching simulation is performed according to the historical operating status data set to construct a dynamic optimization dispatching model of distribution network resilience; wherein, the dynamic optimization dispatching model of distribution network resilience can be understood as a nonlinear power grid impact dispatching dynamic optimization model obtained by simulation training based on nonlinear optimization theory combined with reinforcement learning ideas in a pre-constructed simulation environment according to the park distribution network dispatching process, which can directly give the corresponding optimization dispatching strategy according to the input real-time operating status data; specifically, the steps of constructing a dynamic optimization dispatching model of distribution network resilience by performing interactive training of power grid dispatching simulation based on the historical operating status data set according to the idea of ​​reinforcement learning include:

[0055] Based on the reinforcement learning idea, a dynamic optimization reinforcement learning model is constructed according to the dispatching process of the target park distribution network; the state space and action space of the dynamic optimization reinforcement learning model are respectively defined according to the operating state and dispatching strategy of the target park distribution network; that is, the state space is defined according to the load conditions of power equipment, external environmental conditions, grid harmonic distortion levels, grid voltage fluctuation amplitudes, grid fault records and grid resilience changes during the operation of the distribution network; at the same time, the action space is defined according to the preset dispatching strategy set used in the distribution network operation process, wherein the dispatching strategy set includes load transfer strategy, renewable energy dispatching strategy, energy storage system use strategy, dynamic adjustment of power factor correction equipment configuration strategy and backup power supply activation sequence adjustment strategy, etc., and different operating states correspond to different dispatching strategies. It should be noted that in actual applications, the formulation of various strategies in the scheduling strategy set can be set according to actual needs. For example, the load transfer strategy may include transferring part of the load from peak hours to off-peak hours according to the real-time load of power users to reduce load peaks; the renewable energy scheduling strategy may include maximizing the use of renewable energy by adjusting the output power of renewable energy such as photovoltaic power generation and wind power generation; the energy storage system use strategy may include controlling the charging and discharging of the energy storage system to balance the grid load according to the grid load; the dynamic adjustment of the power factor correction device configuration strategy may include optimizing the power factor and reducing reactive power loss by adjusting the compensation capacity of the power factor correction device in real time; the strategy for adjusting the order of enabling the backup power supply may include determining the order and time of enabling the backup power supply according to the severity of the fault or emergency to ensure the continuity of power supply, etc.

[0056] In order to ensure the reliability of the reward function setting, this embodiment preferably sets the reward function of the dynamic optimization reinforcement learning model to be constructed according to a preset indicator system; wherein the preset indicator system includes the change in load stability, the change in power dispatching cost, the change in equipment reliability, the change in harmonic distortion level, the change in voltage fluctuation amplitude and the change in grid resilience; correspondingly, the load stability indicator is the load fluctuation standard deviation or the load fluctuation mean square deviation; the power dispatching cost change indicator is the cost difference before and after power dispatching, and the relevant costs include electricity charges and equipment loss costs, etc.; the equipment reliability change indicator includes the change in reliability indicators before and after equipment failure, such as the change in mean time between failures MTBF (Mean Time Between Failure) and mean repair time MTTR (Mean Time To The harmonic distortion level change index is the difference in the harmonic distortion level of the power grid before and after power dispatch; the voltage fluctuation amplitude change index is the difference in the voltage fluctuation amplitude of the power grid before and after power dispatch, which can be the standard deviation or mean square error of the voltage fluctuation before and after power dispatch; the power grid resilience change index is the difference in the power grid recovery capacity before and after power dispatch, and the change in the power grid recovery capacity before and after power dispatch can be calculated by obtaining the recovery time and recovery effect of the power grid after the emergency.

[0057] Based on the above definition, it can be seen that the dynamic optimization reinforcement learning model in this embodiment is preferably expressed as:

[0058]

[0059]

[0060]

[0061] in, and Respectively The state space and action space at each moment; , , , , and Respectively The load of electrical equipment at all times, external environmental conditions, grid harmonic distortion level, grid voltage fluctuation amplitude, grid fault records, and changes in grid resilience; Represents a set of scheduling strategies; express The first a scheduling strategy; Indicates the number of scheduling strategies; , , , and They represent the changes in load stability, power dispatching cost, equipment reliability, harmonic distortion level and voltage fluctuation amplitude at time t respectively; is the weight coefficient, which can be adjusted according to actual needs; express Status at all times Next action The corresponding reward value.

[0062] It should be noted that in practical applications, in order to improve the efficiency of calculating the reward value after the scheduling strategy is changed, the above reward function can be optimized by pre-setting the contribution values ​​of different scheduling strategies corresponding to the preset actions executed in different states to the comprehensive reward value, so as to quickly obtain the full representation value corresponding to the reward function in different states. For example, in practical applications, assuming that a certain action only adjusts the load stability or the power scheduling cost, the corresponding reward function value can be quickly obtained according to the contribution value of the scheduling strategy corresponding to the load stability or the power scheduling cost, thereby effectively improving the training efficiency of the dynamic optimization reinforcement learning model; correspondingly, the improved reward function can be expressed as:

[0063]

[0064] in, Indicates status Execute an action The contribution value of the i-th dispatching strategy to the comprehensive reward value, including the contribution values ​​of load stability control, power dispatching cost control, equipment reliability control, harmonic distortion level control, voltage fluctuation amplitude control and grid resilience improvement; The weight coefficient representing the contribution of the i-th scheduling strategy to the comprehensive reward value.

[0065] In addition, in order to improve the adaptive learning ability of the dynamic optimization reinforcement learning model, effectively respond to emergencies and nonlinear impact loads, and significantly improve the resilience and stability of the power grid, this embodiment preferably introduces a nonlinear power grid value approximation function to improve the iteration method of the reinforcement learning model, so as to ensure that the dispatching strategy can be automatically adjusted based on the dynamic optimization reinforcement learning model to ensure the stable operation of the power grid. Specifically, the value function network of the dynamic optimization reinforcement learning model is a deep neural network; the deep neural network is as follows Figure 2As shown, it includes a state input layer, a feature extraction layer, a feature fusion layer and a strategy value output layer connected in sequence; the feature extraction layer includes a parallel load feature extraction layer, a harmonic feature extraction layer, a voltage feature extraction layer, a toughness feature extraction layer, an environmental feature extraction layer and a fault feature extraction layer.

[0066] The state input layer in this embodiment can be understood as receiving the operating state data such as the load condition of the power equipment in the park distribution network, the harmonic distortion level of the power grid, the voltage fluctuation amplitude of the power grid, the change in the resilience of the power grid, the external environmental conditions and the fault records of the power grid. The load feature extraction layer can be understood as a processing layer that extracts load features based on the load condition of the power equipment by performing nonlinear transformation through a number of hidden units and activation functions. In practical applications, the load features can be effectively extracted and processed by sequentially using a hidden unit containing 128 neurons and 1 activation function ReLU and a hidden unit containing 64 neurons and 1 activation function ReLU, and by using the ReLU activation function to capture complex patterns in the load condition at a higher dimension, the accuracy of load prediction and management can be improved. The corresponding output can be expressed as:

[0067]

[0068] in, and is the weight matrix of the load feature extraction layer, and is the bias vector, is the activation function of the load feature extraction layer.

[0069] The harmonic feature extraction layer can be understood as a processing layer that extracts harmonic features based on the harmonic distortion level of the power grid by performing nonlinear transformations through several hidden units and activation functions. In practical applications, the harmonic features can be effectively extracted and processed by sequentially using hidden units containing 64 neurons and 1 activation function ReLU and hidden units containing 32 neurons and 1 activation function ReLU. The nonlinear relationship in the harmonic features is captured by using the ReLU activation function to improve the model's sensitivity and processing capability to harmonic changes. The corresponding output can be expressed as:

[0070]

[0071] in, and is the weight matrix of the harmonic feature extraction layer, and is the bias vector, is the activation function of the harmonic feature extraction layer.

[0072] The voltage feature extraction layer can be understood as a processing layer that extracts voltage features based on the voltage fluctuation amplitude of the power grid by performing nonlinear transformation through several hidden units and activation functions. In practical applications, the voltage features can be effectively extracted and processed by sequentially using a hidden unit containing 64 neurons and 1 activation function ReLU and a hidden unit containing 32 neurons and 1 activation function ReLU. The corresponding output can be expressed as:

[0073]

[0074] in, and is the weight matrix of the voltage feature extraction layer, and is the bias vector, is the activation function of the voltage feature extraction layer.

[0075] The environmental feature extraction layer can be understood as a processing layer that extracts environmental features based on external environmental conditions through nonlinear transformations through a number of hidden units and activation functions. In practical applications, the environmental condition data can be effectively processed by sequentially including a hidden unit containing 32 neurons and 1 activation function ReLU and a hidden unit containing 16 neurons and 1 activation function ReLU, thereby improving the model's adaptability to environmental influences and prediction accuracy. The corresponding output can be expressed as:

[0076]

[0077] in, and is the weight matrix of the environment feature extraction layer, and is the bias vector, is the activation function of the environment feature extraction layer.

[0078] Similarly, the resilience feature extraction layer and the fault feature extraction layer can be understood as processing layers that perform nonlinear transformations through a number of hidden units and activation functions, and extract corresponding features based on the changes in grid resilience and grid fault records respectively. The specific hidden unit settings can also be set according to actual application requirements.

[0079] The feature fusion layer can be understood as integrating the load characteristics, harmonic characteristics, voltage characteristics, resilience characteristics, environmental characteristics and fault characteristics extracted by the above-mentioned load feature extraction layer, harmonic feature extraction layer, voltage feature extraction layer, resilience feature extraction layer, environmental feature extraction layer and fault feature extraction layer, and obtaining a processing layer of multi-dimensional fusion features through comprehensive processing by hidden units and activation functions.

[0080] The strategy value output layer can be understood as calculating the value of each scheduling strategy based on the fusion features output by the feature fusion layer. , in order to select the model result output layer of the optimal scheduling strategy with the highest output value. That is, the output of the strategy value output layer is expressed as:

[0081]

[0082] In the formula, Represents the output value of the strategy value output layer; , and Represent the weight matrix, bias vector and activation function of the strategy value output layer respectively; and Represent the weight matrix and bias vector of the feature fusion layer respectively; Represents the network parameters of the value function network.

[0083] It should be noted that the number of hidden units used in the above-mentioned feature extraction layers and the number of neurons in each hidden unit are only exemplary and can be optimized and adjusted according to the specific task requirements in practical applications to improve the performance and accuracy of the overall model. This embodiment can ensure that the model can show higher efficiency and accuracy when dealing with complex power grid scheduling problems by refining the feature extraction method.

[0084] The dynamic optimization reinforcement learning model obtained through the above method steps can be understood as a model framework that can be used for optimal scheduling of the park distribution network. It is necessary to optimize the model parameters of the dynamic optimization reinforcement learning model based on the operating status data of the actual park distribution network in order to obtain a network model that can be applied to the target park distribution network scenario.

[0085] According to the historical operating status data set, the dynamic optimization reinforcement learning model is trained to obtain the distribution network resilience dynamic optimization scheduling model; specifically, the step of training the dynamic optimization reinforcement learning model according to the historical operating status data set to obtain the distribution network resilience dynamic optimization scheduling model includes:

[0086] Acquire each state to be analyzed from the historical operation state data set according to a preset time step, and obtain a corresponding action to be executed according to the state to be analyzed and the dynamic optimization reinforcement learning model;

[0087] Simulate and execute the action to be executed, and obtain the corresponding next running state and action execution reward;

[0088] Create a corresponding state transition experience pool based on each state to be analyzed, the corresponding action to be executed, the action execution reward and the next running state;

[0089] A preset number of batch samples are randomly extracted from the state transition experience pool, and the dynamic optimization reinforcement learning model is trained based on a preset loss function to obtain the distribution network resilience dynamic optimization scheduling model.

[0090] In practical applications, the dynamic optimization reinforcement learning model is first initialized in a simulation environment to set the initial state and initial action ,in is the state space, As the action space, the historical operation status data set is used to train the dynamic optimization reinforcement learning model to obtain the distribution network resilience dynamic optimization scheduling model with converged model parameters. The detailed process is as follows:

[0091] 1) At a preset time step Get the current status , according to the current state and dynamically optimize the reinforcement learning model to select the corresponding action ;

[0092] 2) Execute the action , get the next state And the corresponding action instant reward ;

[0093] 3) The quadruple Stored in the state transfer experience pool middle;

[0094] 4) Transfer experience pool from state Randomly draw small batches of samples from , training is performed according to the preset loss function until the performance index converges or the preset number of training times is reached to obtain a distribution network resilience dynamic optimization scheduling model; wherein, It can be understood as the number of randomly selected small batch samples.

[0095] In order to improve the stability and convergence speed of the above training process, this embodiment preferably sets the preset loss function as:

[0096]

[0097] in, Indicates status Next action Enter the state And get the reward value Expected value; Represents the state transition experience pool; represents the discount factor; and Represent the current value function network parameters and the target value function network parameters respectively; Represents the current value function network parameters Next, status Execute an action The current strategy value at time ; Represents the current value function network parameters Next, status Execute an action The target strategy value at the time Represents the current value function network parameters Next, status Execute all actions The maximum target strategy value; , and Respectively Changes in load stability, harmonic distortion level, voltage fluctuation amplitude and grid resilience at each moment; Represents the weight coefficient.

[0098] This embodiment uses samples in the state transition experience pool for training by minimizing the above loss function Dynamically optimize the network parameters of the reinforcement learning model , the trained model can be applied to the actual power grid operation. By dynamically adjusting the dispatching strategy, real-time optimization and emergency response can be performed on large-scale nonlinear power grid shocks, and the model can also adapt to the power grid dispatching needs under different load modes.

[0099] S13. Acquire the real-time operating status data of the target park distribution network, input the real-time operating status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis, and generate the corresponding real-time optimal scheduling strategy for the distribution network.

[0100] The real-time optimal dispatching strategy of the distribution network in this embodiment can be understood as directly deploying the trained distribution network resilience dynamic optimization dispatching model into the actual power grid dispatching system, and receiving the real-time operating status data of the target park distribution network system for analysis and processing to obtain the optimal dispatching action under the current state, which can be expressed as:

[0101]

[0102] in, is the value function of the dynamic optimization dispatch model for distribution network resilience, are the parameters of the neural network; Represents the real-time operating status of the distribution network resilience dynamic optimization dispatch model output The corresponding real-time optimal dispatching strategy for the distribution network is used to dynamically adjust the operating status of the park distribution network, which may include load transfer, renewable energy dispatching, use of energy storage systems, dynamic adjustment of the settings of power factor correction equipment, and adjustment of the activation order of backup power sources.

[0103] In order to facilitate the timely perception of emergencies and equipment failures based on the real-time collected operating status data of the target park distribution network, this embodiment preferably analyzes the real-time operating status data to generate the real-time optimal dispatching strategy for the distribution network, and also performs fault identification and emergency identification. Specifically, the step of inputting the real-time operating status data into the distribution network resilience dynamic optimization dispatching model for optimization dispatching analysis and generating the corresponding real-time optimal dispatching strategy for the distribution network also includes:

[0104] According to the distribution network resilience dynamic optimization scheduling model, features are extracted from the real-time operating status data to obtain corresponding comprehensive feature values; wherein, the acquisition of comprehensive feature values ​​can be understood as the fused feature values ​​output by the feature fusion layer after the real-time operating status data is input into the distribution network resilience dynamic optimization scheduling model. The specific process can be found in the description of the value function network in the aforementioned distribution network resilience dynamic optimization scheduling model, which will not be repeated here.

[0105] The comprehensive characteristic value is compared with the preset fault characteristic threshold and the preset event characteristic threshold respectively to obtain corresponding fault identification results and event identification results; wherein, the preset fault characteristic threshold and the preset event characteristic threshold can be understood as thresholds used for fault identification and event classification, and can be set according to the actual application scenario requirements. For example, multiple fault characteristic thresholds are set according to different fault types or fault degrees, and the actual comprehensive characteristic value is compared with the fault characteristic thresholds of different levels to quickly identify the fault type and fault level; similarly, multiple event characteristic thresholds can also be set according to different event types or event levels, and the actual comprehensive characteristic value is compared with the event characteristic thresholds of different levels to quickly identify the event type and event level. No specific limitation is made here.

[0106] It should be noted that if a fault or emergency is identified through the above method, the real-time optimal dispatching strategy of the distribution network generated based on the dynamic optimization dispatching model of the distribution network resilience can be understood as the emergency dispatching strategy corresponding to the fault / emergency. The emergency dispatching strategy also includes load transfer, enabling backup power supply, adjusting renewable energy output, dispatching energy storage system, and adjusting power factor correction equipment. Finally, the operation state of the power grid is adjusted in real time by executing the emergency dispatching strategy. That is, when an emergency or fault occurs, the above-mentioned dynamic optimization dispatching model of the distribution network resilience can quickly analyze and identify the fault or emergency, and can also adjust the dispatching strategy in real time for emergency response. It can also monitor the operation state of the power grid after dispatch in real time, detect load changes and fault changes, and dynamically adjust the dispatching strategy of the power grid to ensure that the reward value of the model can accurately reflect the recovery effect and response speed of the power grid, optimize the response capability to similar emergencies or faults in the future, and thus ensure the stable operation and rapid recovery of the power grid under extreme conditions.

[0107] S14. According to the real-time optimal dispatching strategy of the distribution network, the distribution network of the target park is dispatched for grid resilience optimization. The specific content of the real-time optimal dispatching strategy of the distribution network may vary depending on the actual collected operating status data, and may include one or more of the load transfer strategy, renewable energy dispatching strategy, energy storage system use strategy, dynamic adjustment of power factor correction equipment configuration strategy and backup power supply activation sequence strategy; the corresponding grid resilience optimization dispatching can be understood as executing the corresponding grid dispatching optimization action based on the actual acquired real-time optimal dispatching strategy of the distribution network.

[0108] The present embodiment provides a technical solution for obtaining a historical operating status data set of a target park distribution network, performing grid dispatching simulation interactive training based on the historical operating status data set based on the idea of ​​reinforcement learning, and constructing a distribution network resilience dynamic optimization dispatching model. Then, the real-time operating status data of the target park distribution network is obtained, and the real-time operating status data is input into the distribution network resilience dynamic optimization dispatching model for optimization dispatching analysis to generate a corresponding distribution network real-time optimal dispatching strategy, and performing grid resilience optimization dispatching of the target park distribution network according to the distribution network real-time optimal dispatching strategy. By collecting high-efficiency and multi-dimensional park distribution network operation data, an intelligent grid resilience dynamic optimization dispatching model is constructed based on nonlinear optimization theory to perform real-time status monitoring and analysis and dispatch optimization of the park distribution network, which can effectively respond to emergencies and the impact of nonlinear shock loads while improving the operating efficiency and response speed of the distribution network, and significantly improving the resilience and stability of the distribution network.

[0109] In addition, in order to ensure that the dynamic optimization scheduling model of distribution network resilience can be applicable to the real-time changing grid operation status and to ensure the rationality and accuracy of the generation of the scheduling strategy, this embodiment preferably introduces an adaptive training framework to optimize and adjust the dynamic optimization scheduling model of distribution network resilience based on the real-time operation status data collected in the actual grid system application; specifically, Figure 3 As shown, the method also includes:

[0110] S15. Obtain the new operating status data and new status execution reward value of the target park distribution network after the grid resilience optimization scheduling, and dynamically optimize the scheduling strategy set of the distribution network resilience dynamic optimization scheduling model based on the new operating status data and the new status execution reward value; the implementation method of the specific scheduling strategy set dynamic optimization can be set according to the actual application scenario requirements and is not specifically limited here.

[0111] The embodiment of the present application obtains the historical operating status data set of the target park distribution network, and based on the idea of ​​reinforcement learning, performs grid dispatching simulation interactive training according to the historical operating status data set, constructs the distribution network resilience dynamic optimization dispatching model, obtains the real-time operating status data of the target park distribution network, inputs the real-time operating status data into the distribution network resilience dynamic optimization dispatching model for optimization dispatching analysis to generate the corresponding distribution network real-time optimal dispatching strategy, and performs grid resilience optimization dispatching on the target park distribution network according to the real-time optimal dispatching strategy of the distribution network, and dynamically optimizes the dispatching strategy set of the distribution network resilience dynamic optimization dispatching model according to the new operating status data and new status execution reward value after the target park distribution network is obtained for grid resilience optimization dispatching. By collecting high-efficiency and multi-dimensional park distribution network operation data, it can not only construct an intelligent grid resilience dynamic optimization dispatching model based on nonlinear optimization theory to perform real-time state monitoring and analysis and dispatch optimization on the park distribution network, effectively respond to emergencies and nonlinear impact loads, but also improve the operation efficiency and response speed of the distribution network, significantly improve the resilience and stability of the distribution network, and ensure that the grid resilience dynamic optimization dispatching model can adapt to changes in the grid state, ensure the accuracy and reliability of grid dispatching, and thus provide reliable protection for the power supply quality and power supply safety of the park distribution network.

[0112] In order to further demonstrate the application effect of the park distribution network resilience optimization scheduling method based on reinforcement learning proposed in the present invention, the power grid scheduling of a modern industrial park is taken as an example for verification and explanation:

[0113] Industrial Zone A is a typical modern industrial park, which includes several major building groups of Company A, such as large-volume production workshops, large-header production workshops, pipe production workshops, assembly workshops, finished product workshops, general company factories and several equipment rooms. The annual electricity consumption of the park is about 16 million kWh. There are a large number of nonlinear loads in the equipment in the park, which leads to power quality problems such as harmonics, voltage fluctuations, power factor and three-phase imbalance. These problems have a significant impact on the safe operation of power distribution facilities. Especially during the high temperatures in summer, the demand for cooling loads increases dramatically, which seriously aggravates the peak-to-valley difference of the load and poses a safety hazard to the power grid. In order to cope with these problems, the park power grid resilience optimization scheduling method based on reinforcement learning provided by the present invention is adopted.

[0114] First, the historical and real-time operating status data of Industrial Zone A were collected, including the load of power equipment, the level of harmonic distortion of the power grid, the amplitude of voltage fluctuation of the power grid, the change of power grid resilience, the external environmental conditions and the power grid fault records, and the data were cleaned and standardized. Then, the power grid dispatching process of the park was modeled as a dynamic optimization problem, the state space, action space and reward function were determined, and a deep neural network was introduced as a value function approximator to design a dynamic optimization reinforcement learning model suitable for power grid dispatching. The model was trained with the preprocessed data set in a simulation environment so that it could learn and optimize the dispatching strategy, and finally the trained distribution network resilience dynamic optimization dispatching model was applied to the actual power grid operation.

[0115] On a July day of a certain year, Industrial Zone A experienced several consecutive days of high temperature weather, and the cooling equipment in the park was turned on on a large scale, causing a sharp increase in the load on the power grid. According to real-time data monitoring, the peak load in the park reached 8,000 kilowatts, and the harmonic level and voltage fluctuations also increased significantly. Traditional power grid dispatching methods are difficult to cope with such high load fluctuations, resulting in a series of power quality problems, and some equipment even overheated and failed. When the method provided by the present invention is used for real-time dispatch optimization, the model receives the current power grid operation status data for rapid analysis and selects the optimal dispatching strategy. For the load peak, the model recommends load transfer, transferring some non-critical loads to the backup power supply, while increasing the output of the photovoltaic power station and the energy storage power station to balance the power grid load. For harmonics and voltage fluctuations, the model recommends adjusting the settings of the power factor correction equipment and starting the backup power supply for voltage regulation. In this scenario, after the implementation of the above dispatching strategy, the power grid load quickly stabilized at around 6,500 kilowatts, and the harmonic level and voltage fluctuations were also effectively controlled. Compared with the traditional method, the power quality of the method of the present invention is significantly improved under load fluctuation conditions. The specific comparison data results are shown in Table 1:

[0116] Table 1 Comparison of the effects of the park power grid resilience optimization dispatching methods

[0117]

[0118] It can be seen from the data in Table 1 that the method of the present invention significantly reduces the load fluctuation of the power grid under high load conditions, reduces the harmonic level and voltage fluctuation, improves the stability and reliability of the power grid through real-time scheduling optimization, and can quickly analyze and adjust the scheduling strategy in real time when a fault occurs to ensure the safe operation of the power grid.

[0119] It should be noted that although the steps in the above flowchart are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders.

[0120] In one embodiment, Figure 4 As shown, a park distribution network resilience optimization scheduling system based on reinforcement learning is provided, and the system includes:

[0121] The data set acquisition module 1 is used to acquire the historical operation status data set of the target park distribution network; the historical operation status data set includes the load of power equipment, the harmonic distortion level of the power grid, the voltage fluctuation amplitude of the power grid, the change of the power grid resilience, the external environmental conditions and the power grid fault records;

[0122] Model building module 2, used to perform interactive training of power grid dispatch simulation based on the historical operation status data set based on the reinforcement learning idea, and build a distribution network resilience dynamic optimization dispatch model;

[0123] Strategy generation module 3, used to obtain the real-time operation status data of the target park distribution network, input the real-time operation status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis, and generate the corresponding distribution network real-time optimal scheduling strategy;

[0124] The optimization scheduling module 4 is used to perform power grid resilience optimization scheduling on the target park distribution network according to the real-time optimal scheduling strategy of the distribution network.

[0125] In one embodiment, Figure 5 As shown, a park distribution network resilience optimization scheduling system based on reinforcement learning is provided, and the system also includes:

[0126] The strategy optimization module 5 is used to obtain the new operating status data and new status execution reward value of the target park distribution network after the grid resilience optimization scheduling, and dynamically optimize the scheduling strategy set of the distribution network resilience dynamic optimization scheduling model based on the new operating status data and the new status execution reward value.

[0127] For the specific limitations of the campus distribution network resilience optimization and dispatching system based on reinforcement learning, please refer to the limitations of the campus distribution network resilience optimization and dispatching method based on reinforcement learning above. The corresponding technical effects can also be equivalently obtained, which will not be repeated here. Each module in the above-mentioned campus distribution network resilience optimization and dispatching system based on reinforcement learning can be fully or partially implemented by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the corresponding operations of each of the above modules.

[0128] In summary, the embodiments of the present invention provide a method and system for optimizing the resilience of a park distribution network based on reinforcement learning. The method and system for optimizing the resilience of a park distribution network based on reinforcement learning can obtain a historical operating status data set of a target park distribution network, perform grid scheduling simulation interactive training based on the historical operating status data set based on the reinforcement learning idea, and construct a distribution network resilience dynamic optimization scheduling model. Then, the real-time operating status data of the target park distribution network is obtained, and the real-time operating status data is input into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis to generate a corresponding distribution network real-time optimal scheduling strategy, and according to the distribution network real-time optimal scheduling strategy, a technical solution for optimizing the resilience of the target park distribution network is performed. This method collects high-efficiency and multi-dimensional park distribution network operation data, constructs an intelligent grid resilience dynamic optimization scheduling model based on nonlinear optimization theory, performs real-time status monitoring and analysis and scheduling optimization on the park distribution network, effectively responds to emergencies and nonlinear shock loads, and can also improve the operation efficiency and response speed of the distribution network, significantly improve the resilience and stability of the distribution network, and provide reliable protection for the power supply quality and power supply safety of the park distribution network.

[0129] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The above-mentioned embodiments only express several preferred implementation modes of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in the technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be based on the protection scope of the claims.

Claims

1. A method for optimizing the resilience of a campus distribution network based on reinforcement learning, characterized in that: The method comprises the following steps: Obtain a historical operating status data set of the target park distribution network; the historical operating status data set includes the load of power equipment, the level of power grid harmonic distortion, the amplitude of power grid voltage fluctuation, the change in power grid resilience, external environmental conditions and power grid fault records; Based on the idea of ​​reinforcement learning, interactive training of power grid dispatch simulation is carried out according to the historical operation status data set to build a dynamic optimization dispatch model for distribution network resilience; Acquire the real-time operating status data of the target park distribution network, input the real-time operating status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis, and generate the corresponding distribution network real-time optimal scheduling strategy; According to the real-time optimal dispatching strategy of the distribution network, the distribution network of the target park is dispatched for optimal resilience of the distribution network; Among them, the steps of conducting interactive training of power grid dispatch simulation based on the historical operation status data set and constructing a dynamic optimization dispatch model for distribution network resilience based on the idea of ​​reinforcement learning include: Based on the reinforcement learning idea, a dynamic optimization reinforcement learning model is constructed according to the dispatching process of the target park distribution network; the state space and action space of the dynamic optimization reinforcement learning model are respectively defined according to the operating state and dispatching strategy of the target park distribution network; the reward function of the dynamic optimization reinforcement learning model is constructed according to a preset indicator system; the preset indicator system includes the change in load stability, the change in power dispatching cost, the change in equipment reliability, the change in harmonic distortion level, the change in voltage fluctuation amplitude and the change in grid resilience; the dynamic optimization reinforcement learning model is expressed as: in, and Respectively The state space and action space at each moment; , , , , and Respectively The load of electrical equipment at all times, external environmental conditions, grid harmonic distortion level, grid voltage fluctuation amplitude, grid fault records, and changes in grid resilience; Represents a set of scheduling strategies; express The first a scheduling strategy; Indicates the number of scheduling strategies; , , , and They represent the changes in load stability, power dispatching cost, equipment reliability, harmonic distortion level and voltage fluctuation amplitude at time t respectively; is the weight coefficient; express Status at all times Next action The corresponding reward value; The dynamic optimization reinforcement learning model is trained according to the historical operating status data set to obtain the distribution network resilience dynamic optimization scheduling model.

2. The method for optimizing the resilience of a campus distribution network based on reinforcement learning as claimed in claim 1, characterized in that: The value function network of the dynamic optimization reinforcement learning model is a deep neural network; the deep neural network includes a state input layer, a feature extraction layer, a feature fusion layer and a strategy value output layer connected in sequence; the feature extraction layer includes a parallel load feature extraction layer, a harmonic feature extraction layer, a voltage feature extraction layer, a toughness feature extraction layer, an environmental feature extraction layer and a fault feature extraction layer; the output of the strategy value output layer is expressed as: In the formula, Represents the output value of the strategy value output layer; , and Represent the weight matrix, bias vector and activation function of the strategy value output layer respectively; and Represent the weight matrix and bias vector of the feature fusion layer respectively; Network parameters representing the value function network; , , , , and They represent the outputs of the load feature extraction layer, harmonic feature extraction layer, voltage feature extraction layer, toughness feature extraction layer, environment feature extraction layer and fault feature extraction layer respectively.

3. The method for optimizing the resilience of a campus distribution network based on reinforcement learning as claimed in claim 1, characterized in that: The step of training the dynamic optimization reinforcement learning model according to the historical operation status data set to obtain the distribution network resilience dynamic optimization scheduling model includes: Acquire each state to be analyzed from the historical operation state data set according to a preset time step, and obtain a corresponding action to be executed according to the state to be analyzed and the dynamic optimization reinforcement learning model; Simulate and execute the action to be executed, and obtain the corresponding next running state and action execution reward; Create a corresponding state transition experience pool based on each state to be analyzed, the corresponding action to be executed, the action execution reward and the next running state; A preset number of batch samples are randomly extracted from the state transition experience pool, and the dynamic optimization reinforcement learning model is trained based on a preset loss function to obtain the distribution network resilience dynamic optimization scheduling model; the preset loss function is expressed as: in, Indicates status Next action Enter the state And get the reward value Expected value; Represents the state transition experience pool; represents the discount factor; and Represent the current value function network parameters and the target value function network parameters respectively; Represents the current value function network parameters Next, status Execute an action The current strategy value at time ; Represents the current value function network parameters Next, status Execute an action The target strategy value at the time; Represents the current value function network parameters Next, status Execute all actions The maximum target strategy value; , and Respectively Changes in load stability, harmonic distortion level, voltage fluctuation amplitude and grid resilience at each moment; Represents the weight coefficient.

4. The method for optimizing the resilience of a campus distribution network based on reinforcement learning as claimed in claim 1, characterized in that: The step of inputting the real-time operating status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis to generate the corresponding distribution network real-time optimal scheduling strategy also includes: Extracting features from the real-time operating status data according to the distribution network resilience dynamic optimization scheduling model to obtain corresponding comprehensive feature values; The comprehensive characteristic value is compared with a preset fault characteristic threshold and a preset event characteristic threshold respectively to obtain a corresponding fault identification result and event identification result.

5. The method for optimizing the resilience of a campus distribution network based on reinforcement learning as claimed in claim 1, characterized in that: The method further comprises: Obtain the new operating status data and new status execution reward value of the target park distribution network after the grid resilience optimization scheduling, and dynamically optimize the scheduling strategy set of the distribution network resilience dynamic optimization scheduling model based on the new operating status data and the new status execution reward value.

6. A park distribution network resilience optimization dispatching system based on reinforcement learning, characterized in that: The system comprises: A data set acquisition module is used to acquire a historical operating status data set of the target park distribution network; the historical operating status data set includes the load of electrical equipment, the level of harmonic distortion of the power grid, the amplitude of voltage fluctuation of the power grid, the change in the resilience of the power grid, external environmental conditions and power grid fault records; A model building module is used to perform interactive training of power grid dispatch simulation based on the historical operation status data set based on the idea of ​​reinforcement learning, and to build a dynamic optimization dispatch model for distribution network resilience; A strategy generation module is used to obtain the real-time operating status data of the target park distribution network, input the real-time operating status data into the distribution network resilience dynamic optimization scheduling model for optimization scheduling analysis, and generate the corresponding distribution network real-time optimal scheduling strategy; An optimization scheduling module, used to perform power grid resilience optimization scheduling on the target park distribution network according to the real-time optimal scheduling strategy of the distribution network; Among them, based on the reinforcement learning idea, the power grid dispatching simulation interactive training is carried out according to the historical operation status data set to build a distribution network resilience dynamic optimization dispatching model, including: Based on the reinforcement learning idea, a dynamic optimization reinforcement learning model is constructed according to the dispatching process of the target park distribution network; the state space and action space of the dynamic optimization reinforcement learning model are respectively defined according to the operating state and dispatching strategy of the target park distribution network; the reward function of the dynamic optimization reinforcement learning model is constructed according to a preset indicator system; the preset indicator system includes the change in load stability, the change in power dispatching cost, the change in equipment reliability, the change in harmonic distortion level, the change in voltage fluctuation amplitude and the change in grid resilience; the dynamic optimization reinforcement learning model is expressed as: in, and Respectively The state space and action space at each moment; , , , , and Respectively The load of electrical equipment at all times, external environmental conditions, grid harmonic distortion level, grid voltage fluctuation amplitude, grid fault records, and changes in grid resilience; Represents a set of scheduling strategies; express The first a scheduling strategy; Indicates the number of scheduling strategies; , , , and They represent the changes in load stability, power dispatching cost, equipment reliability, harmonic distortion level and voltage fluctuation amplitude at time t respectively; is the weight coefficient; express Status at all times Next action The corresponding reward value; The dynamic optimization reinforcement learning model is trained according to the historical operating status data set to obtain the distribution network resilience dynamic optimization scheduling model.

7. The park distribution network resilience optimization dispatching system based on reinforcement learning as claimed in claim 6 is characterized in that: The system further comprises: The strategy optimization module is used to obtain the new operating status data and new status execution reward value of the target park distribution network after the grid resilience optimization scheduling, and dynamically optimize the scheduling strategy set of the distribution network resilience dynamic optimization scheduling model based on the new operating status data and the new status execution reward value.

Images