A method and related device for source-network-load-storage collaborative partitioning and collaborative optimization

Abstract

The application belongs to the technical field of new energy power generation, and discloses a source network load storage collaborative partitioning and collaborative optimization method and related device; wherein the source network load storage collaborative partitioning and collaborative optimization method comprises the following steps: based on the selected distribution network, the historical data of the power grid load is obtained; based on the obtained historical data of the power grid load, the trained reinforcement learning model is used for prediction to obtain the control strategy prediction result of the control device in the selected distribution network. In the training process of the reinforcement learning model, the multi-agent reinforcement learning algorithm is adopted, which has better efficiency and performance compared with the single-agent architecture reinforcement learning. In addition, the reinforcement learning model adopts a topological change graph convolution network model, and the trained reinforcement learning model can be applied to different topological structures, which can reduce the training cost required for model transplantation.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation technology, and specifically relates to a method and related apparatus for coordinated optimization of power generation, grid, load and storage in different regions. Background Technology

[0002] With the increasing global pursuit of sustainable energy, the installed capacity of new energy sources in power distribution networks has shown a significant upward trend. Against the backdrop of today's energy transition, new energy sources such as solar and wind power, with their clean and renewable characteristics, have become important choices for meeting ever-growing energy demands. As the proportion of renewable energy in the energy structure increases, its volatility and intermittency pose significant challenges to the stable operation of the power grid. Simultaneously, the widening peak-to-valley difference in electricity load and the increasing demands from users for power supply reliability and power quality make coordinated optimization among power generation, grid, load, and storage inevitable. Explained, "source" mainly refers to various power generation sources, including traditional thermal and hydropower, as well as new energy power generation (such as photovoltaic and wind power); "grid" represents the transmission and distribution network, responsible for the transmission and distribution of electricity; "load" refers to the electricity load, which refers to the equipment or systems that consume electricity at the user end; and "storage" refers to energy storage systems, such as battery energy storage, thermal storage, and cold storage.

[0003] The topology of power grids is complex, especially at the distribution network level, where the integration of distributed generation and energy storage systems further complicates power flow distribution. Excessive injection of active power can cause voltage fluctuations to exceed grid standard thresholds. Voltage fluctuations can be mitigated by utilizing the control flexibility of photovoltaic inverters themselves and other controllable devices (exemplarily, static var compensators (SVCs) and on-load tap changers (OLTCs)). However, a carefully designed scheme is needed to coordinate these devices on a large scale to regulate the voltage of the entire network with limited local information. Furthermore, the computational time cost of Optimal Power Flow (OPF) algorithms is high, making it difficult to effectively address voltage fluctuations in a short period.

[0004] Currently, some studies attempt to use single-agent deep reinforcement learning (DRL) models to model the reactive power optimization problem of distribution networks as a Markovian decision process (MDP). For example, they employ actor-critic (AC) algorithms to minimize network losses and equipment operating costs as optimization objectives, using the switching commands of discrete reactive power regulators as control variables. Other studies use deterministic policy gradient (DPG) algorithms, treating the reactive power of inverters or static var compensators as control actions. However, these existing studies only use grid node data as features of the same dimension for reinforcement learning training, neglecting the impact of grid topology modeling on the model. Reinforcement learning models trained using these methods often fail to achieve good results and are difficult to apply to grid scenarios with different topologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method and related apparatus for coordinated optimization of power generation, grid, load, and storage in different zones, in order to solve one or more of the aforementioned technical problems. The technical solution provided by this invention utilizes a trained reinforcement learning model for prediction, obtaining the predicted control strategies of selected control devices in the distribution network. During the training process of the reinforcement learning model, a multi-agent reinforcement learning algorithm is employed, which has better efficiency and performance compared to single-agent reinforcement learning architectures. Furthermore, the reinforcement learning model uses a topology-transformation graph convolutional network model, and the trained reinforcement learning model can be applied to different topologies, reducing the training cost required for model porting.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for coordinated optimization of source-grid-load-storage partitioning, comprising the following steps:

[0008] Based on the selected distribution network, obtain historical data on the grid load;

[0009] Based on the acquired historical data of power grid load, a trained reinforcement learning model is used to make predictions and obtain the prediction results of the control strategy of the selected control device in the distribution network.

[0010] The training steps of the reinforcement learning model include: acquiring distribution network operation data and topology information based on distribution network samples, and constructing a mathematical model for voltage and reactive power optimization of the distribution network; converting the constructed mathematical model into a multi-agent Markov decision process model; constructing a topology change graph convolutional network model based on distribution network operation data and topology information; training the topology change graph convolutional network model using a multi-agent reinforcement learning algorithm based on the multi-agent Markov decision process model to obtain a reinforcement learning model that can dynamically adapt to distribution network topology changes and optimize voltage and reactive power control; and offline training and verification of the obtained reinforcement learning model using actual grid data acquired based on distribution network samples to obtain a trained reinforcement learning model.

[0011] A further improvement of the present invention is that,

[0012] The topology transformation graph convolutional network model includes:

[0013] The input layer is used to input power distribution network operation data and topology information;

[0014] The learnable spectral convolutional modules with residual maps, which are repeatedly stacked, are used to extract features from the input distribution network operation data and topology information to obtain high-dimensional features of the distribution network operation data and topology information. Each learnable spectral convolutional module with residual maps includes a learnable spectral convolutional layer with residual maps, a first normalization layer, and a pooling layer.

[0015] The second normalization layer is used to normalize the obtained high-dimensional features to obtain distribution network optimization prediction information.

[0016] The output layer is used to output the obtained distribution network optimization prediction information.

[0017] A further improvement of the present invention is that,

[0018] The topology transformation graph convolutional network model also includes:

[0019] A bilateral filter layer is disposed between a learnable spectral convolution module with residual maps that is repeatedly stacked and a second normalization layer. It is used to filter the high-dimensional features of the obtained distribution network operation data and topology information, obtain the filtered high-dimensional features, and input them into the second normalization layer.

[0020] A further improvement of the present invention is that,

[0021] The learnable spectral convolutional layer with residual maps is represented as follows:

[0022] ;

[0023] In the formula, The output of a learnable spectral convolutional layer with residual maps; It is an activation function; It is a learnable spectral convolution kernel with residual maps. The result of the operation is a learnable spectral convolution matrix with residual maps; It is the feature vector corresponding to each vertex in the graph; is a trainable weight matrix; These are trainable weight vectors;

[0024] ;

[0025] In the formula, It is a learnable Laplace matrix with residuals; For about of k polynomial of order 1; k The summation function iterates through the indexes; K This represents the total number of nodes in the graph; for Parameters of the polynomial;

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, L The normalized graph Laplacian matrix; For a given constant value; It is the identity matrix; The distance metric matrix is ​​based on the generalized Mahalanobis distance. A It is an adjacency matrix; , These are the randomly initialized distance metric matrix and adjacency matrix, respectively. The residual Laplacian matrix obtained during training when the distance metric and adjacency relationship are randomly initialized.

[0030] A further improvement of the present invention is that,

[0031] The expression for calculating the distance metric matrix using the generalized Mahalanobis distance is as follows:

[0032] ;

[0033] In the formula, It is the vertex corresponding to the graph eigenvectors; , These are trainable weights.

[0034] A further improvement of the present invention is that,

[0035] In the multi-agent Markov decision process model mentioned above

[0036] Control devices with independent operating space in the power distribution network are defined as intelligent agents;

[0037] The state observation space is defined as: ;

[0038] In the formula, It is a set of active and reactive power loads; It is a set of active power generated by power generation equipment; It is a set of reactive power generated by power generation equipment; It is a set of voltage amplitudes Voltage phase in radians ; These are the active power, reactive power, and phase angle of the load, respectively. These refer to the active power and reactive power generated by the power generation equipment, respectively.

[0039] The action space of an agent is defined as:

[0040] For continuously controlled devices, the agent is equipped with a continuous set of actions, represented as:

[0041] ;

[0042] In the formula, To regulate the output of the equipment; These represent the lower and upper limits of the output of the continuously variable control device, respectively.

[0043] For discrete control devices, the agent is equipped with a discrete set of actions, represented as:

[0044] ;

[0045] In the formula, To regulate the output of the equipment; This indicates all the actions the device is capable of performing;

[0046] reward function for:

[0047] ;

[0048] In the formula, This is a penalty item for network damage; This is a voltage deviation penalty term;

[0049] ;

[0050] In the formula, It is a positive weighting coefficient; Indicates network loss;

[0051] ;

[0052] In the formula, This refers to the number of nodes in the distribution network. A positive weighting coefficient; Represents a node The voltage amplitude; Represents a node The reference voltage amplitude.

[0053] A further improvement of the present invention is that the control device in the power distribution network includes a switch, an inverter, and an energy storage device.

[0054] In a second aspect, the present invention provides a system for coordinated optimization of source-grid-load-storage zones, comprising:

[0055] The data acquisition module is used to acquire historical data of the power grid load based on the selected distribution network.

[0056] The prediction module is used to make predictions based on the acquired historical data of the power grid load using a trained reinforcement learning model, and to obtain the prediction results of the control strategy of the selected control device in the distribution network.

[0057] The training steps of the reinforcement learning model include: acquiring distribution network operation data and topology information based on distribution network samples, and constructing a mathematical model for voltage and reactive power optimization of the distribution network; converting the constructed mathematical model into a multi-agent Markov decision process model; constructing a topology change graph convolutional network model based on distribution network operation data and topology information; training the topology change graph convolutional network model using a multi-agent reinforcement learning algorithm based on the multi-agent Markov decision process model to obtain a reinforcement learning model that can dynamically adapt to distribution network topology changes and optimize voltage and reactive power control; and offline training and verification of the obtained reinforcement learning model using actual grid data acquired based on distribution network samples to obtain a trained reinforcement learning model.

[0058] A further improvement of the present invention is that,

[0059] The topology transformation graph convolutional network model includes:

[0060] The input layer is used to input power distribution network operation data and topology information;

[0061] The learnable spectral convolutional modules with residual maps, which are repeatedly stacked, are used to extract features from the input distribution network operation data and topology information to obtain high-dimensional features of the distribution network operation data and topology information. Each learnable spectral convolutional module with residual maps includes a learnable spectral convolutional layer with residual maps, a first normalization layer, and a pooling layer.

[0062] The second normalization layer is used to normalize the obtained high-dimensional features to obtain distribution network optimization prediction information.

[0063] The output layer is used to output the obtained distribution network optimization prediction information.

[0064] A further improvement of the present invention is that,

[0065] The topology transformation graph convolutional network model also includes:

[0066] A bilateral filter layer is disposed between a learnable spectral convolution module with residual maps that is repeatedly stacked and a second normalization layer. It is used to filter the high-dimensional features of the obtained distribution network operation data and topology information, obtain the filtered high-dimensional features, and input them into the second normalization layer.

[0067] A further improvement of the present invention is that,

[0068] The learnable spectral convolutional layer with residual maps is represented as follows:

[0069] ;

[0070] In the formula, The output of a learnable spectral convolutional layer with residual maps; It is an activation function; It is a learnable spectral convolution kernel with residual maps. The result of the operation is a learnable spectral convolution matrix with residual maps; It is the feature vector corresponding to each vertex in the graph; is a trainable weight matrix; These are trainable weight vectors;

[0071] ;

[0072] In the formula, It is a learnable Laplace matrix with residuals; For about of k polynomial of order 1; k The summation function iterates through the indexes; K This represents the total number of nodes in the graph; for Parameters of the polynomial;

[0073] ;

[0074] ;

[0075] ;

[0076] In the formula, L The normalized graph Laplacian matrix; For a given constant value; It is the identity matrix; The distance metric matrix is ​​based on the generalized Mahalanobis distance. A It is an adjacency matrix; , These are the randomly initialized distance metric matrix and adjacency matrix, respectively. The residual Laplacian matrix obtained during training when the distance metric and adjacency relationship are randomly initialized.

[0077] A further improvement of the present invention is that,

[0078] The expression for calculating the distance metric matrix using the generalized Mahalanobis distance is as follows:

[0079] ;

[0080] In the formula, It is the vertex corresponding to the graph eigenvectors; , These are trainable weights.

[0081] A further improvement of the present invention is that,

[0082] In the multi-agent Markov decision process model mentioned above

[0083] Control devices with independent operating space in the power distribution network are defined as intelligent agents;

[0084] The state observation space is defined as: ;

[0085] In the formula, It is a set of active and reactive power loads; It is a set of active power generated by power generation equipment; It is a set of reactive power generated by power generation equipment; It is a set of voltage amplitudes Voltage phase in radians ; These are the active power, reactive power, and phase angle of the load, respectively. These refer to the active power and reactive power generated by the power generation equipment, respectively.

[0086] The action space of an agent is defined as:

[0087] For continuously controlled devices, the agent is equipped with a continuous set of actions, represented as:

[0088] ;

[0089] In the formula, To regulate the output of the equipment; These represent the lower and upper limits of the output of the continuously variable control device, respectively.

[0090] For discrete control devices, the agent is equipped with a discrete set of actions, represented as:

[0091] ;

[0092] In the formula, To regulate the output of the equipment; This indicates all the actions the device is capable of performing;

[0093] reward function for:

[0094] ;

[0095] In the formula, This is a penalty item for network damage; This is a voltage deviation penalty term;

[0096] ;

[0097] In the formula, It is a positive weighting coefficient; Indicates network loss;

[0098] ;

[0099] In the formula, This refers to the number of nodes in the distribution network. A positive weighting coefficient; Represents a node The voltage amplitude; Represents a node The reference voltage amplitude.

[0100] A further improvement of the present invention is that the control device in the power distribution network includes a switch, an inverter, and an energy storage device.

[0101] In a third aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for source-grid-load-storage collaborative partitioning and collaborative optimization as described in any one of the first aspects of the present invention.

[0102] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for source-grid-load-storage collaborative partitioning and collaborative optimization as described in any one of the first aspects of the present invention.

[0103] Compared with the prior art, the present invention has the following beneficial effects:

[0104] This invention provides a method for coordinated optimization of power generation, grid, load, and storage systems in different zones. It utilizes a trained reinforcement learning model to predict the control strategies of selected control devices in a distribution network. During the training of the reinforcement learning model, a multi-agent reinforcement learning algorithm is employed, which offers better efficiency and performance compared to single-agent architectures. Furthermore, the reinforcement learning model uses a topology-transformation graph convolutional network model, allowing the trained model to be applied to different topologies and reducing the training costs required for model porting. Further explanation is provided: the multi-agent reinforcement learning algorithm (MARL) performs exceptionally well when centralized control mechanisms cannot accurately predict the next action of each individual agent. Distribution power networks involve numerous issues requiring cooperation or competition among a large number of agents, making MARL a suitable training framework for power network optimization models.

[0105] This invention improves the spectral convolution kernel in reinforcement learning models by employing a learnable Laplacian matrix. This allows the spectral convolution kernel to be applied to data with diverse graph topologies, improving the network model's topology generalization ability. The trained model can be adapted to different topologies, reducing the training cost required for model portability. This invention also proposes a residual Laplacian matrix to create graph structures in an unsupervised manner, improving the model's perception of the internal relationships between nodes in the graph and enabling dynamic adjustments to adapt to changes in power grid topology (such as switching and load changes).

[0106] In the reinforcement learning model proposed in this invention, the generalized Mahalanobis distance is used instead of Euclidean distance as the distance metric when measuring the relationship between nodes. The generalized Mahalanobis distance is used as a parameter for training, which can be adaptive according to the topology and the relationship between nodes, so that the constructed Laplacian matrix can more reasonably predict the loss in the environment of changing topology. Attached Figure Description

[0107] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0108] Figure 1 This is a flowchart illustrating a method for coordinated optimization of source-grid-load-storage zones in an embodiment of the present invention.

[0109] Figure 2 This is a flowchart illustrating a method for coordinated optimization of source-grid-load-storage zones in a specific embodiment of the present invention.

[0110] Figure 3 This is a schematic diagram of a topology transformation graph convolutional network model in an embodiment of the present invention;

[0111] Figure 4 This is a schematic diagram of a source-grid-load-storage collaborative partitioned collaborative optimization system in an embodiment of the present invention. Detailed Implementation

[0112] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention; obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0113] Based on the technical solutions disclosed in the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0114] Please see Figure 1 The present invention provides a method for source-grid-load-storage collaborative partitioning optimization, comprising the following steps:

[0115] Step 1: Based on the selected distribution network, obtain historical data on the power grid load;

[0116] Step 2: Based on the historical power grid load data obtained in Step 1, use the trained reinforcement learning model to make predictions and obtain the prediction results of the control strategies of the selected control devices in the distribution network; in the exemplary technical solution, the control devices in the distribution network may include switches, inverters and energy storage devices, etc.

[0117] The training steps of the reinforcement learning model include:

[0118] Step (1): Based on the distribution network sample, obtain the distribution network operation data and topology information, and construct a mathematical model for voltage and reactive power optimization of the distribution network;

[0119] Step (2) transforms the mathematical model constructed in step (1) into a Markov decision process model based on multi-agent agents;

[0120] Step (3): Based on the power distribution network operation data and topology information, construct a topology change graph convolutional network model;

[0121] Step (4): Based on the Markov decision process model based on multi-agent obtained in step (2), a multi-agent reinforcement learning algorithm (exemplary, such as Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments, Group Relative Policy Optimization, etc.) is used to train the topology change graph convolutional network model obtained in step (3) to obtain a reinforcement learning model that can dynamically adapt to the topology changes of the distribution network and optimize voltage and reactive power control.

[0122] Step (5) uses the actual (i.e. real) power grid data obtained based on the distribution network sample to train and validate the reinforcement learning model obtained in step (4) offline, and obtain the trained reinforcement learning model; further interpretably, the trained reinforcement learning model can be saved, and then its effectiveness and scalability under distribution network topology changes can be verified.

[0123] The technical solution provided in this invention achieves intelligent decision-making for voltage and reactive power control in distribution networks by constructing a Markov decision process model based on multi-agent agents and training it using a multi-agent reinforcement learning algorithm. The reinforcement learning model can automatically predict and optimize the control strategies of control devices based on the real-time status and historical data of the distribution network, without manual intervention. Furthermore, the trained reinforcement learning model can automatically optimize the control of voltage and reactive power in the distribution network, improving the timeliness and accuracy of control and reducing errors and delays caused by human factors. In this invention, by designing a topology change graph convolutional network model and training it using a multi-agent reinforcement learning algorithm, the reinforcement learning model can dynamically adapt to changes in the distribution network topology. When lines, loads, or power generation facilities in the distribution network change, the reinforcement learning model can quickly adjust its strategy to maintain optimized performance. Moreover, this technical solution is not only applicable to specific distribution networks but can also be extended to other similar distribution network environments by adjusting model parameters and structure. This scalability gives this technical solution broad application prospects. This invention, through the prediction and optimization of reinforcement learning models, can significantly reduce voltage deviation and network losses in distribution networks, thereby improving power supply quality. This not only helps improve the stability and reliability of the power system but also reduces energy waste and lowers operating costs. The invention obtains a trained reinforcement learning model through offline training, enabling rapid deployment and use in practical applications. This offline training combined with real-time application improves model training efficiency and reduces the complexity of real-time computation. In summary, the technical solutions of this invention apply improved techniques such as multi-agent reinforcement learning algorithms and topology transformation graph convolutional network models to the field of voltage and reactive power optimization in distribution networks. This not only improves the intelligence and automation level of voltage and reactive power control in distribution networks but also enhances dynamic adaptability and scalability, providing new ideas and methods for the optimized control of distribution networks.

[0124] Please see Figure 2 and Figure 3 In one specific embodiment of the present invention, the training steps of the reinforcement learning model specifically include the following steps:

[0125] Step S101: Construct a mathematical model for voltage and reactive power optimization of the distribution network; wherein, the mathematical model includes: optimization objective, equality constraints and inequality constraints for power flow calculation;

[0126] In a specific exemplary technical solution, considering medium-voltage (10-24kV) and low-voltage (0.23-1kV) distribution networks, the distribution network is modeled as follows: , and ) represent the set of nodes (buses) and the set of edges (branches), respectively; among them, node (bus) 0 is regarded as the connection point with the main power grid, balancing the active and reactive power in the distribution network;

[0127] By coordinating various control devices, the voltage of the entire distribution network is regulated. Voltage quality is an important indicator for measuring power supply quality. Reactive power has a significant impact on voltage amplitude. Reasonable reactive power optimization can maintain the node (bus) voltage within a specified range. Voltage deviation can be used to measure voltage quality.

[0128] For safety and optimal operation, a 5% voltage deviation is typically allowed, and the target voltage formula is:

[0129] ;

[0130] in, For nodes The per-unit voltage value; Represents a per-unit value, with no dimensionless unit;

[0131] Minimizing network losses is an important objective in reactive power optimization of distribution networks. Explained, network losses are mainly due to power loss caused by line resistance, and the rational allocation of reactive power can effectively reduce network losses.

[0132] The total active power loss of a distribution network can be expressed by the following formula:

[0133] ;

[0134] in, It is the total active power loss of the distribution network; It refers to the number of lines; It is the first The resistance of the line; It is the first The voltage at both ends of the line, It is the first Active and reactive power on each line;

[0135] The optimization goal is to make Minimum, that is .

[0136] For each busbar ,make and The magnitude and phase angle of the complex voltage, For complex power injection, active and reactive power injection can be defined as follows:

[0137] ;

[0138] ;

[0139] in, It is connected to the bus. The index set of the busbar; and It is a side road The conductivity and susceptance on the surface; It is a busbar and The phase difference between them; and It is a busbar The active and reactive power of the power generation facilities (interpretatively, zero if there is no power on the bus). and It is a busbar The active and reactive power at load (interpretatively, zero if there is no load on the bus); this formula can represent the dynamics of the power system and is crucial for solving power flow problems and active voltage control problems.

[0140] Step S102: Transform the mathematical model constructed in step S101 into a Markov decision process model based on multi-agent agents.

[0141] In an exemplary technical solution, the mathematical model is converted into a Markov Decision Process (MDP) model to describe the specific solution process of reinforcement learning; wherein, a partially observable multi-agent Markov game is considered, and it is assumed that communication between all agents is available; at time step intelligent agent Observe your own observation It is the global environment state. Part of it, based on my own observations and the message from my previous time step In addition to messages received from other intelligent agents Select Action and information Assuming an intelligent agent News It is sent to all other agents and in the next time step (i.e., time step 1) ), available to other intelligent agents; in The combined effect produces the next environmental state. And based on the transition probability and reward function respectively, and These are the environment state space and the joint action space, respectively. (Intelligent agent) The goal is to find a strategy Maximize its discount benefits Therefore, intelligent agents The objective function is defined as ,in and These are the common strategy and the discount factor, respectively.

[0142] Step S103: Based on the topology characteristics of the distribution network, design a topology transformation graph convolutional network model; where, for example, such as... Figure 3 As shown, the structure of the topology transformation graph convolutional network model includes:

[0143] The input layer is used to input power distribution network operation data and topology information;

[0144] Repeated stacking N The spectral convolutional module with residual maps is used to extract features from the input distribution network operation data and topology information to obtain high-dimensional features of the distribution network operation data and topology information. Each spectral convolutional module with residual maps includes a spectral convolutional layer with residual maps, a normalization layer, and a pooling layer. N These are adjustable parameters;

[0145] In a preferred embodiment, a bilateral filter layer may be included to filter the high-dimensional features of the obtained distribution network operation data and topology information to obtain filtered high-dimensional features.

[0146] The normalization layer is used to normalize the high-dimensional features or filtered high-dimensional features of the obtained distribution network operation data and topology information to obtain distribution network optimization prediction information.

[0147] The output layer is used to output the obtained distribution network optimization prediction information.

[0148] Step S104: Design a power distribution network optimization reinforcement learning framework based on multi-agent reinforcement learning, and train a topology transformation graph convolutional network model based on the multi-agent reinforcement learning MARL algorithm.

[0149] Step S105: Train the reinforcement learning model offline using real distribution network data, save the trained reinforcement learning model, and then verify its effectiveness and scalability when the distribution network topology changes.

[0150] The application scenario of the technical solution of this invention is to predict, based on historical data of the power grid load, how to adjust the control strategies of control devices such as switches, inverters and energy storage devices in the next period of time, so as to maximize the reduction of active power loss in the power grid and meet safety requirements.

[0151] The core improvement of the technical solution in this invention is to transform the power grid topology into a graph network that conforms to the power flow calculation equality constraint, describing a Markov decision process. Based on the Multi-Agent Reinforcement Learning (MARL) algorithm, an Actor-Critic (AC) algorithm is used to minimize network losses and equipment operating costs as the optimization objective. A reward function is designed to evaluate the power grid operating status during operation. The optimized graph convolutional network is used to predict the operation of the next time period and train the reinforcement learning model. A further improvement is that the main module of the graph convolutional network is a spectral convolutional layer with residuals, using generalized Mahalanobis distance as the distance metric and a learnable Laplacian matrix as the topological representation matrix. Simultaneously, the Laplacian matrix with residuals is trained to perceive the intrinsic connections between nodes. Feature transformation and reparameterization are used to construct intra-node and inter-node mappings, thus forming the overall part of the spectral convolutional layer with residuals.

[0152] In this invention, a learnable Laplacian matrix is ​​employed, enabling the spectral convolution kernel to be applicable to data with diverse graph topologies. This improves the topology generalization of the network model, allowing the trained model to be adapted to different topologies and reducing the training cost required for model portability. The reinforcement learning model proposed in this invention uses generalized Mahalanobis distance instead of Euclidean distance as the distance metric when measuring relationships between nodes. Generalized Mahalanobis distance is used as a parameter for training, allowing for adaptive adaptation based on the topology and relationships between nodes. This enables the constructed Laplacian matrix to more reasonably predict losses in environments with changing topologies. The influence relationships between power grid nodes are distance-dependent; closer nodes have a greater influence. This invention reparameterizes the feature transformation in the convolutional network to construct a mapping between intra-node and inter-node features in the topology, improving model accuracy. This invention proposes a residual Laplacian matrix to create a graph structure in an unsupervised manner, improving the model's perception of intra-node relationships in the graph and enabling dynamic adjustments to adapt to changes in power grid topology (such as switching and load changes).

[0153] In one specific embodiment of the present invention, a spectral convolutional layer with residual maps is proposed, which is an improvement upon the spectral convolutional layer; wherein,

[0154] This invention improves the spectral convolution kernel in reinforcement learning models by incorporating a learnable Laplacian matrix into the kernel parameter learning process, enabling the model to adapt to more diverse power grid topologies. Furthermore, this invention proposes a learnable Laplacian matrix with residuals. Considering the existence of discrete variables such as switches in the power grid topology, the spectral convolutional layer with residual graphs can explore networks with features similar to the original topology by training the residual graph. Moreover, this invention uses generalized Mahalanobis distance as the distance metric in the Laplacian matrix, which can adapt to the topology and relationships between nodes, providing a more reasonable representation of the power grid node topology. Finally, this invention introduces a transformation matrix and a bias vector applied to the output features in the topology-optimized spectral convolutional layer to construct intra-node and inter-node mappings.

[0155] Specifically, the formula for the spectral convolutional layer is:

[0156] ;

[0157] In the formula, This is the output of the spectral convolutional layer; It is an activation function; It is a spectral convolution kernel; The result of the operation is a spectral convolution matrix; It is the feature vector corresponding to each vertex in the graph;

[0158] Spectral convolution kernel It can be represented as:

[0159] ;

[0160] In the formula, These are trainable arbitrary parameters that can be initialized and then adjusted using backpropagation of the error. The characteristic matrix of the Laplace matrix is, i.e. .

[0161] In this embodiment of the invention, the description of the spectral convolutional layer with residual maps is as follows: given a graph and its adjacency matrix Distance metric matrix Learnable Laplace matrix with residuals The formula is as follows:

[0162] ;

[0163] ;

[0164] ;

[0165] In the formula, The specified value, typically 0.1; a symmetric positive semi-definite matrix. , These are trainable weight parameters. It is the feature vector corresponding to each vertex in the graph.

[0166] In this embodiment of the invention, the distance metric matrix of generalized Mahalanobis distance is used. The calculation method is as follows:

[0167] ;

[0168] In the formula, It is the vertex corresponding to the graph eigenvectors.

[0169] In this embodiment of the invention, the formula for the spectral convolutional layer with residuals is:

[0170] ;

[0171] In the formula, The output of a learnable spectral convolutional layer with residual maps; It is an activation function; It is a learnable spectral convolution kernel with residual maps. The result of the operation is a learnable spectral convolution matrix with residual maps; It is the feature vector corresponding to each vertex in the graph; is a trainable weight matrix; These are trainable weight vectors;

[0172] Given the initial Laplace matrix ,feature and parameters ,function Output a learnable Laplacian matrix Spectrum, convolution kernel Represented as:

[0173] ;

[0174] In the formula, It is a learnable Laplace matrix with residuals. It is the feature vector corresponding to each vertex in the graph.

[0175] In this embodiment of the invention, given a diagram and its adjacency matrix and distance metric matrix Normalized graph Laplace matrix It can be obtained in the following ways:

[0176] ;

[0177] In the formula, the matrix Here, is the distance metric between nodes, specifically the generalized Mahalanobis distance; matrix A is the adjacency matrix.

[0178] Explained, the Laplace matrix The Laplace matrix determines the connectivity between nodes and the degree of vertices. Unique expression The topological structure. This study proposes a learnable Laplacian matrix, which... Parameterization is performed given the initial Laplacian matrix. ,feature and parameters ,function Output a learnable Laplacian matrix The score.

[0179] The convolution kernel will be represented as:

[0180] ;

[0181] The main description of the improved spectral convolutional layer is as follows:

[0182] ;

[0183] Due to dense matrix multiplication The complexity of this formula is O(n). ;if Approximated as Given a polynomial function that can be recursively computed, then due to the Laplace matrix... The sparsity of the expression will reduce the complexity to [value missing]. Chebyshev expansion is used to calculate polynomial of order .

[0184] This invention proposes a learnable Laplacian matrix with residuals. Considering the discrete variables such as switches in the power grid topology, the spectral graph convolutional layer with residual graphs can explore networks with features similar to the original topology graph by training the residual graph. Since there is no prior knowledge about the distance metric, the metric... Since it is randomly initialized, it may take a long time to converge. To accelerate training and increase the stability of the learned graph topology, a learnable graph Laplacian operator with residuals is used. In the Laplace matrix Make a slight offset based on the existing structure:

[0185] ;

[0186] In the formula, This is a configurable parameter, typically set to 0.1; The Laplacian matrix obtained during training when M is randomly initialized is used as the metric.

[0187] Laplace matrix A great deal of useful graph structure information has been revealed, in addition to those substructures consisting of virtual vertex connections that cannot be directly learned on the intrinsic graph; therefore, the residual Laplace matrix The impact on the final graph topology is determined by control.

[0188] This invention employs generalized Mahalanobis distance as the distance metric in the Laplacian matrix, which can adaptively adapt to the topology and relationships between nodes, providing a more reasonable representation of the topology of power grid nodes. For power grid topologies, with discretized data forming a graph structure, Euclidean distance is no longer a good measure of vertex similarity. Therefore, during training, the distance metric needs to adapt to the task and features. In metric learning, algorithms are divided into supervised and unsupervised learning. The optimal metric obtained in an unsupervised manner minimizes intra-class distance while maximizing inter-class distance. For labeled datasets, the learning objective is to find the metric that minimizes the loss. and The generalized Mahalanobis distance formula between them is:

[0189] ;

[0190] if The formula then degenerates into Euclidean distance. In the model of this invention, the symmetric positive semi-definite matrix... ,in These are trainable weights.

[0191] It is to measure and The transformation basis of the space between the Euclidean distances is then used to calculate the Gaussian kernel:

[0192] ;

[0193] right After normalization, a dense adjacency matrix is ​​obtained. In the model of this invention, the optimal metric It is to construct a set of graph Laplace's that minimizes the prediction loss. That measurement.

[0194] In this invention, a transformation matrix and a bias vector are introduced into the topology-optimized spectral convolutional layer to construct intra-node and inter-node mappings.

[0195] In classic convolutional neural networks, the output feature of a convolutional layer is the sum of all feature maps from the previous layer. This means that new features are constructed not only based on neighboring vertices but also depend on features within other vertices. In a power grid topology environment, creating and training separate topologies for different vertex features on the same graph is difficult to interpret. To construct intra-node and inter-node mappings, a transformation matrix and a bias vector applied to the output features are introduced in the topology-optimized spectral convolutional layer. The reparameterization formula for the output features is:

[0196] ;

[0197] In the Layer, transformation matrix ,deviation With measurement Training together, among them It's the feature dimension. In each spectral convolutional layer, we have a total of... Parameters of learning complexity This is independent of the size or degree of the input graph. In the next spectral convolutional layer, the spectral convolution kernel will be constructed in another feature domain with a different metric.

[0198] The following is a further explanation of the multi-agent reinforcement learning algorithm used in this embodiment of the invention:

[0199] 1) Problem Description

[0200] The multi-agent reinforcement learning algorithm proposed in this invention can be applied to large-scale power networks containing control devices such as inverters, energy storage devices, and switches. Each device can be controlled in a distributed manner and can therefore be considered as an agent. To ensure the security of the distribution network, all agents within a region share observations of that region.

[0201] 2) Mathematical modeling

[0202] Since each agent can only observe a portion of the entire power grid, and ensuring the security of the power grid is a common goal among the agents, this problem is modeled as a distributed partially observable Markov decision process (Dec-POMDP) ​​that can be described as a 10-tuple, where... It is a collection of intelligent agents. It is a set of states. It is a set of actions. It is a reward function. It is the set of observations. It is a state transition function. It is the probability of observation. It is the initial state probability distribution. It is a discount factor.

[0203] Intelligent agents: There exists a group of intelligent agents that control a group of control devices (that is, a control device with an independent action space is defined as an intelligent agent), denoted as Each agent resides at a node in the graph (i.e., the power network). In this embodiment of the invention, a function is defined. This indicates the node where each agent is located.

[0204] Region: The entire power grid is divided into The union of the regions is denoted as . ,in and (if We define a function. To map nodes to their respective regions.

[0205] The state observation space of a multi-agent Markov decision process model is defined as follows:

[0206]

[0207] in, It is a set of active and reactive power loads; It is a set of active power generated by power generation equipment; It is a set of reactive power generated by power generation equipment; It is a set of voltage amplitudes Voltage phase in radians ; These are the active power, reactive power, and phase angle of the load, respectively. These refer to the active power and reactive power generated by the power generation equipment, respectively.

[0208] In this embodiment of the invention, a function is defined.

[0209] ;

[0210] The function defined above will A subset is mapped to its associated metric, where Represents a power set;

[0211] Observation is defined as ,in Indicating in intelligent agents Measurements within the region.

[0212] Action space of an agent:

[0213] For continuously controlled equipment (such as inverters), each The intelligent agent is equipped with a continuous set of actions:

[0214] ;

[0215] In the formula, To regulate the output of the equipment; These represent the lower and upper limits of the output of the continuously variable control device, respectively.

[0216] Interpretatively, continuous action represents the maximum reactive power it produces, i.e., the power generated by the first continuous action. The reactive power generated by each inverter is ,in It is the first Each device has the maximum apparent power that it can generate based on its physical capacity. If This means that reactive power is being output to the distribution network. If This means absorbing reactive power from the distribution network. The value is usually selected based on the capacity of the distribution network, which is for safety operation considerations.

[0217] For discrete control devices (such as switches), each The intelligent agent is equipped with a discrete set of actions:

[0218] ;

[0219] In the formula, To regulate the output of the equipment; This represents all actions that the device can perform; combined actions are defined as... .

[0220] State transition probability function:

[0221] Since the state includes the previous action, and the load change is random (theoretically, it can be modeled as any probability distribution), the state transition probability function can be defined as:

[0222] ;

[0223] It follows a Markov process, specifically... ;

[0224] in, and ;

[0225] Define and understand space, and The changes in load are described. It is an extremely short time interval, much shorter than the time interval between two control operations (i.e., one time step), as shown in the example of this invention. .

[0226] Observation probability function:

[0227] In the context of power grids, it describes the measurement errors that may occur in sensors. Mathematically, it can be defined as... Specifically, ;in, It is a multivariate isotropic Gaussian distribution. It depends on the physical characteristics of the sensor.

[0228] Reward function for a multi-agent Markov decision process model:

[0229] First, calculate the network loss penalty; the greater the network loss, the greater the penalty.

[0230] Define network loss penalty items: ;

[0231] in, It is a positive weighting coefficient used to adjust the importance of network loss in the reward function.

[0232] For voltage stability, calculate the voltage deviation penalty term;

[0233] Voltage deviation can be measured using the root mean square error (RMSE):

[0234] ;

[0235] in, It is another positive weighting coefficient;

[0236] Ultimately, the reward function was designed as follows:

[0237] ;

[0238] in, Represents the number of nodes in the distribution network. Indicates network loss. Represents a node voltage amplitude, Represents a node The reference voltage amplitude.

[0239] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0240] Please see Figure 4In this embodiment of the invention, a system for source-grid-load-storage coordinated partitioning and coordinated optimization is provided, comprising:

[0241] The data acquisition module is used to acquire historical data of the power grid load based on the selected distribution network.

[0242] The prediction module is used to make predictions based on the acquired historical data of the power grid load using a trained reinforcement learning model, and to obtain the prediction results of the control strategy of the selected control device in the distribution network.

[0243] The training steps of the reinforcement learning model include: acquiring distribution network operation data and topology information based on distribution network samples, and constructing a mathematical model for voltage and reactive power optimization of the distribution network; converting the constructed mathematical model into a multi-agent Markov decision process model; constructing a topology change graph convolutional network model based on distribution network operation data and topology information; training the topology change graph convolutional network model using a multi-agent reinforcement learning algorithm based on the multi-agent Markov decision process model to obtain a reinforcement learning model that can dynamically adapt to distribution network topology changes and optimize voltage and reactive power control; and offline training and verification of the obtained reinforcement learning model using actual grid data acquired based on distribution network samples to obtain a trained reinforcement learning model.

[0244] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used to execute the operation of a source-grid-load-storage collaborative partitioning and collaborative optimization method.

[0245] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM (Random Access Memory) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the source-network-load-storage collaborative partitioning and collaborative optimization method in the above embodiments.

[0246] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0247] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0248] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0249] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0250] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for source-network-load-balance collaborative partitioning and optimization, characterized in that, The method comprises the following steps: Based on the selected power distribution network, the historical data of the power grid load is obtained; Based on the obtained historical data of the power grid load, the trained reinforcement learning model is used for prediction to obtain the prediction result of the control strategy of the control device in the selected power distribution network; The training steps of the reinforcement learning model include: based on the power distribution network sample, the power distribution network operation data and topology information are obtained, and a mathematical model of power distribution network voltage reactive power optimization is constructed; the constructed mathematical model is converted into a multi-agent based Markov decision process model; according to the power distribution network operation data and topology information, a topology change graph convolution network model is constructed; based on the multi-agent based Markov decision process model, the topology change graph convolution network model is trained by using a multi-agent reinforcement learning algorithm to obtain a reinforcement learning model capable of dynamically adapting to the topology change of the power distribution network and optimizing the voltage reactive power control; the obtained reinforcement learning model is trained and verified offline using the actual power grid data obtained based on the power distribution network sample, and a trained reinforcement learning model is obtained.

2. The method of claim 1, wherein the topology change graph convolution network model comprises: an input layer for inputting power distribution network operation data and topology information; a repeatedly stacked learnable spectral graph convolution module with residual graph for feature extraction of the input power distribution network operation data and topology information to obtain high-dimensional features of the power distribution network operation data and topology information; wherein each learnable spectral graph convolution module with residual graph comprises a learnable spectral graph convolution layer with residual graph, a first normalization layer, and a pooling layer; a second normalization layer for normalizing the obtained high-dimensional features to obtain power distribution network optimization prediction information; and an output layer for outputting the obtained power distribution network optimization prediction information.

3. The method of claim 2, wherein the topology change graph convolution network model further comprises: a bilateral filter layer disposed between the repeatedly stacked learnable spectral graph convolution module with residual graph and the second normalization layer, for filtering the obtained high-dimensional features of the power distribution network operation data and topology information to obtain filtered high-dimensional features and input the second normalization layer.

4. The method of claim 2, wherein the learnable spectral graph convolution layer with residual graph is represented as:

5. The method of claim 4, wherein the calculation expression of the distance measurement matrix using the generalized Mahalanobis distance is:

6. The method of claim 1, wherein in the multi-agent based Markov decision process model, the control devices in the power distribution network with independent action space are defined as agents; ； wherein is the output of the learnable spectral graph convolution layer with residual graph; is an activation function; is a learnable spectral convolution kernel with residual graph, is the operation result of the learnable spectral convolution matrix with residual graph; is the feature vector corresponding to each vertex on the graph; is a trainable weight matrix; is a trainable weight vector; ； wherein is a learnable Laplacian matrix with residuals; is a polynomial of degree is a k polynomial of degree k is a summation function index; K is the total number of nodes in the graph; is a parameter of the polynomial in ； ； ； wherein, L is a normalized graph Laplacian matrix; is a given constant value; is an identity matrix; is a distance metric matrix using generalized Mahalanobis distance; A is an adjacency matrix; , are respectively a randomly initialized distance metric matrix and an adjacency matrix; is a residual Laplacian matrix trained when the distance metric and the adjacency relationship are randomly initialized. the action space of the agent is defined as: for continuous control devices, the agent is equipped with a continuous action set, represented as: ； wherein is the eigenvector on the graph corresponding to the vertex of the graph G; , are trainable weights. ​ ​ ​ The state observation space is defined as: ; wherein is a set of active and reactive power loads; is a set of active power generated by power plants; is a set of reactive power generated by power plants; is a set of voltage magnitudes and voltage phases in radians ; are active power, reactive power and phase angle of loads, respectively; are active power, reactive power generated by power plants, respectively; ​ ​ ； In the formula, to regulate the output of the device; respectively represent the lower limit and the upper limit of the output of the continuous regulation device. For discrete control devices, the agent is equipped with a discrete action set, denoted as: ； wherein to regulate the output of the device; represents all actions the device is capable of performing; Reward function is: ； In the formula, is a network loss penalty term; is a voltage deviation penalty term; ； In the formula, is a positive weight coefficient; represents the network loss; ； In the formula, This refers to the number of nodes in the distribution network. A positive weighting coefficient; Represents a node The voltage amplitude; Represents a node The reference voltage amplitude.

7. The method of source network load storage collaborative partitioning and collaborative optimization according to claim 1, characterized in that, The control devices in the power distribution network include switches, inverters and energy storage devices.

8. A system for source-network-load-balance coordination partitioning and optimization, characterized in that, Comprise: The data acquisition module is used for acquiring the historical data of the power grid load based on the selected power distribution network; The prediction module is used for predicting based on the acquired historical data of the power grid load, using the trained reinforcement learning model to obtain the control strategy prediction result of the control device in the selected power distribution network; Wherein, the training steps of the reinforcement learning model include: based on the power distribution network sample, the operation data and topology information of the power distribution network are acquired, and a mathematical model of power distribution network voltage reactive power optimization is constructed; the constructed mathematical model is converted into a multi-agent based Markov decision process model; according to the operation data and topology information of the power distribution network, a topology change graph convolution network model is constructed; based on the multi-agent based Markov decision process model, the topology change graph convolution network model is trained by using multi-agent reinforcement learning algorithm, so as to obtain a reinforcement learning model which can dynamically adapt to the topology change of the power distribution network and optimize the voltage reactive power control; the obtained reinforcement learning model is trained and verified offline using the actual power grid data obtained based on the power distribution network sample, to obtain the trained reinforcement learning model.

9. The system for source-grid-load-storage collaborative zoning and collaborative optimization according to claim 8, wherein The topology change graph convolution network model comprises: An input layer for inputting the operation data and topology information of the power distribution network; A repeatedly stacked learnable spectral graph convolution module with residual graph for feature extraction of the input operation data and topology information of the power distribution network to obtain high-dimensional features of the operation data and topology information of the power distribution network; wherein each learnable spectral graph convolution module with residual graph comprises a learnable spectral graph convolution layer with residual graph, a first normalization layer and a pooling layer; A second normalization layer for normalizing the obtained high-dimensional features to obtain power distribution network optimization prediction information; An output layer for outputting the obtained power distribution network optimization prediction information.

10. The system for source-grid-load-storage collaborative zoning and collaborative optimization according to claim 9, wherein The topology change graph convolution network model further comprises: A bilateral filter layer, which is arranged between the repeatedly stacked learnable spectral graph convolution module with residual graph and the second normalization layer, for filtering the obtained high-dimensional features of the operation data and topology information of the power distribution network to obtain filtered high-dimensional features and input the second normalization layer.

11. The system for source-grid-load-storage collaborative zoning and collaborative optimization according to claim 9, wherein The learnable spectral graph convolution layer with residual graph is denoted as: ； wherein is the output of the learnable spectral graph convolution layer with residual graph; is an activation function; is a learnable spectral convolution kernel with residual graph, is the operation result of the learnable spectral convolution matrix with residual graph; is the feature vector corresponding to each vertex on the graph; is a trainable weight matrix; is a trainable weight vector; ； wherein is a learnable Laplacian matrix with residuals; is a polynomial of degree is a k polynomial of degree k is a summation function index; K is the total number of nodes in the graph; is a parameter of the polynomial in ； ； ； wherein, L is a normalized graph Laplacian matrix; is a given constant value; is an identity matrix; is a distance metric matrix using generalized Mahalanobis distance; A is an adjacency matrix; , are respectively a randomly initialized distance metric matrix and an adjacency matrix; is a residual Laplacian matrix trained when the distance metric and the adjacency relationship are randomly initialized.

12. The system for source-grid-load-storage collaborative zoning and collaborative optimization according to claim 11, wherein The calculation expression of the distance measurement matrix using the generalized Mahalanobis distance is: ； wherein is the eigenvector on the graph corresponding to the vertex ; , are trainable weights.

13. The system for source-grid-load-storage collaborative zoning and collaborative optimization according to claim 8, wherein In the multi-agent based Markov decision process model, The control device with independent action space in the power distribution network is defined as an agent; The state observation space is defined as: ; wherein is a set of active and reactive power loads; is a set of active power generated by power plants; is a set of reactive power generated by power plants; is a set of voltage magnitudes and voltage phases in radians ; are active power, reactive power and phase angle of the loads, respectively; are active power, reactive power generated by power plants, respectively; The action space of the agent is defined as: For continuous control equipment, the agent is equipped with a continuous action set, denoted as: ； In the formula, to regulate the output of the device; respectively represent the lower limit and the upper limit of the output of the continuous regulation device. For discrete control equipment, the agent is equipped with a discrete action set, denoted as: ； wherein to regulate the output of the device; represents all actions the device is capable of performing; Reward function is: ； In the formula, is a network loss penalty term; is a voltage deviation penalty term; ； In the formula, is a positive weight coefficient; represents the network loss; ； In the formula, is the number of nodes in the power distribution network; is a positive weight coefficient; represents the voltage amplitude of the node ; and represents the reference voltage amplitude of the node .

14. The system for source-network-load-bunker collaborative zoning and collaborative optimization according to claim 8, characterized in that, The control device in the power distribution network includes a switch, an inverter, and an energy storage device.

15. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, The processor implements the source network load storage collaborative partitioning and collaborative optimization method of any one of claims 1-7 when executing the program. 16.A non-transitory computer-readable storage medium having stored thereon a computer program. The computer program implements the source network load storage collaborative partitioning and collaborative optimization method of any one of claims 1-7 when executed by the processor.

Images