Layered cooperative control method and system for multi-port converter

Abstract

The invention relates to a hierarchical cooperative control method and system for a multi-port converter, and the method comprises the steps: constructing a self-adaptive weighted directed graph model which reflects an electrical and communication coupling relation, and carrying out the fusion of a system state matrix and a weighted graph model through a graph convolution network, extracting graph embedding features representing spatial coupling and state dependence between nodes, generating a local instruction according to node features and local states by using a local controller based on a policy network, and performing global aggregation on the node features by using a graph attention network to generate an energy scheduling matrix; and finally, generating a unified control instruction through weighted fusion of the local instruction and global scheduling. According to the method, accurate power distribution and efficient energy scheduling of the multi-port converter cluster in complex environments such as communication limitation, topology change and power fluctuation are realized, and the robustness and the operation efficiency of an energy supply system are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of power grid control, specifically relating to a hierarchical collaborative control method and system for multi-port converters. Background Technology

[0002] Multiport converters, as key power electronic interface devices connecting power sources, grids, loads, and energy storage, are widely used in microgrids, distributed energy systems, and AC / DC hybrid distribution networks. With the increasing penetration of distributed energy, systems often contain multiple multiport converter nodes. These nodes not only have complex electrical connections but also require information exchange through communication networks to achieve coordinated control. Existing multiport converter control strategies mainly include centralized control and distributed control. While centralized control can achieve global optimization, it places extremely high demands on the reliability and bandwidth of the communication network and is susceptible to single-point failures, making it difficult to adapt to dynamic scenarios with frequent topology changes. Distributed control, while possessing strong robustness and flexibility, lacks a global perspective because each node makes decisions based solely on local information, making it difficult to achieve optimal energy scheduling at the system level.

[0003] To balance global optimization and local response, hierarchical cooperative control architecture has gradually become a research hotspot. However, existing hierarchical control methods often treat the physical topology of the electrical network and the information topology of the communication network separately, or treat them only as static connections. In actual operation, the electrical connection impedance between converter nodes may change due to line switching, and communication quality may fluctuate due to network latency or packet loss. Traditional control methods struggle to simultaneously quantify and map these dynamically changing electrical connection strengths and communication quality into the control model, resulting in the control system's inability to accurately perceive the spatial coupling characteristics and state dependencies between nodes. Furthermore, existing cooperative control often uses fixed rules or simplified linear models for power allocation. When the system faces multiple disturbances or nonlinear load changes, conflicts can easily arise between local adaptive adjustment and global energy dispatch, leading to decreased power allocation accuracy, bus voltage fluctuations, and even system instability.

[0004] Therefore, existing technologies urgently need a control scheme that can integrate network topology characteristics and operating status, and achieve local and global coordination and unification in a dynamic environment. Summary of the Invention

[0005] One objective of this invention is to at least solve one or more of the aforementioned problems existing in the prior art. In other words, one objective of this invention is to provide a hierarchical cooperative control method for a multi-port converter that satisfies one or more of the aforementioned requirements, comprising:

[0006] A weighted directed graph model is generated based on the architecture of the power supply system, which includes multi-port converters. The nodes of the weighted directed graph model represent the various multi-port converters in the power supply system, and the edge weights are quantitative parameters that integrate the electrical connection strength and communication quality between nodes. Obtain the operating parameters of the multi-port converter, generate the state vector of each multi-port converter based on the operating parameters, and further generate the system state matrix of the power supply system; The weighted directed graph model and the system state matrix are fused using a graph convolutional network to generate a graph embedding matrix that reflects the spatial coupling relationship and state dependency characteristics between multi-port converters. The policy network generates local control commands based on the state vector and the corresponding node features in the graph embedding matrix. The policy network is configured in the local controller of each multiport converter. A graph attention network is used to perform global aggregation of node features in the graph embedding matrix to generate an energy scheduling matrix containing the active power target value and reactive power target value of each multi-port converter. Based on the preset fusion weights, the local control commands and energy scheduling matrix are weighted and fused to generate unified control commands, which are then synchronously sent to each multi-port converter.

[0007] As a preferred implementation, the edge weights are calculated using the following method: An electrical connection matrix is ​​generated based on the equivalent electrical impedance between each multi-port converter. A communication connection matrix is ​​generated based on the communication delay and data packet loss rate between each multi-port converter. The electrical connection matrix and the communication connection matrix are summed by weight to generate a weighted adjacency matrix; Generate a weighted directed graph model of the power supply system based on the weighted adjacency matrix.

[0008] As a further preferred implementation, the weighting coefficients of the weighted summation are generated based on the relative proportions of electrical connection strength and communication connection strength, such that the weight of the communication connection matrix is ​​increased when the electrical connection strength decreases, or the weight of the electrical connection matrix is ​​increased when the communication connection strength decreases.

[0009] As a preferred implementation, the operating parameters of the multi-port converter include the active power, reactive power, port voltage, device temperature, and operating mode identifier of the multi-port converter.

[0010] As a preferred implementation, a graph convolutional network is used to fuse a weighted directed graph model and a system state matrix to generate a graph embedding matrix that reflects the spatial coupling relationships and state dependency characteristics between multi-port converters, including: The adjacency matrix of the weighted directed graph model is normalized to obtain the normalized Laplace matrix. The system state matrix is ​​used as the initial feature input to the graph convolutional network; Graph convolutional networks use multi-order polynomial filters to perform convolution calculations on input features, and then use residual connection mechanisms to superimpose the convolution results with the input features. After multiple rounds of iteration to aggregate multi-order neighborhood features, a graph embedding matrix is ​​generated.

[0011] In a preferred embodiment, the local controller is also equipped with an evaluation network that evaluates the value of local control commands based on a preset reward function and updates the parameters of the policy network.

[0012] As a preferred implementation, a graph attention network is used to globally aggregate the node features in the graph embedding matrix to generate an energy dispatch matrix containing the active power target values ​​and reactive power target values ​​of each multi-port converter, including: A graph attention network is used to linearly map the node features in the graph embedding matrix; Calculate the attention coefficient between nodes, which represents the degree of influence of neighboring nodes on the central node; The features of neighboring nodes are weighted and aggregated based on the attention coefficient to generate an energy scheduling matrix containing the active power target value and reactive power target value of each multi-port converter.

[0013] As a preferred embodiment, the method further includes: Real-time monitoring of disturbance indicators in the power supply system; when the disturbance indicators exceed the preset threshold, triggering the operation mode switching mechanism to generate operation mode instructions. The unified control commands also include operating mode commands.

[0014] As a further preferred implementation, the disturbance indicators include active power fluctuation rate, reactive power fluctuation rate, voltage fluctuation rate, and communication interruption duration.

[0015] On the other hand, the present invention also provides a hierarchical cooperative control system for multi-port converters, comprising: The graph model building module is used to generate a weighted directed graph model based on the architecture of the power supply system, which includes multi-port converters. The nodes of the weighted directed graph model represent the various multi-port converters in the power supply system, and the edge weights are quantitative parameters that integrate the electrical connection strength and communication quality between nodes. The state awareness module is used to acquire the operating parameters of the multi-port converter, generate the state vector of each multi-port converter based on the operating parameters, and further generate the system state matrix of the power supply system. The feature extraction module is used to fuse the weighted directed graph model and the system state matrix using a graph convolutional network to generate a graph embedding matrix that reflects the spatial coupling relationship and state dependency characteristics between multi-port converters. A local control module is used to generate local control commands using a policy network based on the state vector and the corresponding node features in the graph embedding matrix. The policy network is configured in the local controller of each multi-port converter. The global coordination module is used to perform global aggregation processing on the node features in the graph embedding matrix using a graph attention network to generate an energy scheduling matrix containing the active power target value and reactive power target value of each multi-port converter. The instruction fusion module is used to perform weighted fusion of local control instructions and energy scheduling matrix according to preset fusion weights, generate unified control instructions, and synchronously send the unified control instructions to each multi-port converter.

[0016] Compared with the prior art, the multi-port converter hierarchical cooperative control method and system provided by the present invention have the following beneficial effects: The multi-port converter hierarchical collaborative control method of this invention constructs a weighted directed graph model that integrates electrical connection strength and communication quality, and utilizes a graph convolutional network to deeply fuse the architectural features of the power supply system with its real-time operating status, generating a graph embedding matrix that can characterize the spatial coupling relationship and state dependency characteristics between multi-port converters. This approach overcomes the limitation of traditional methods that separate operating conditions from system information, enabling the control system to maintain accurate perception of the global topology and operating status of the system even when local communication quality deteriorates or electrical parameters are disturbed, thereby achieving a combination of local control and global scheduling.

[0017] At the local level, the policy network configured for each converter in this invention can utilize the node features in the graph embedding matrix, enabling the local controller to have global awareness when making decisions, ensuring rapid response while also taking into account coordination with neighboring nodes; at the global level, this invention uses a graph attention network to globally aggregate node features to identify key energy interaction paths and generate a globally optimal energy scheduling matrix.

[0018] Ultimately, by weighting and fusing local control commands and the global energy dispatch matrix, the system can generate unified control commands according to preset weights. This mechanism effectively solves the problem that a single control mode is difficult to balance steady-state accuracy and dynamic performance under dynamic operating conditions, and realizes accurate power allocation and efficient energy dispatch of multi-port converter clusters in complex environments, significantly improving the robustness and operating efficiency of the power supply system. Attached Figure Description

[0019] Figure 1This is a flowchart of the multi-port converter hierarchical collaborative control method of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the described elements without departing from the scope of the invention. Various processes or components may be appropriately omitted, substituted, or added to the various examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined into other examples.

[0022] Embodiments of the present invention provide a hierarchical cooperative control method for multi-port converters, such as... Figure 1 As shown, the method in this embodiment mainly includes the following steps S100-S600.

[0023] Step S100: Generate a weighted directed graph model based on the architecture of the power supply system including the multi-port converter.

[0024] Step S100 defines the weighted directed graph model of the energy supply system as follows: , where the set of nodes Each node in Represents a multiport converter unit in an energy supply system; edge set Represents the connection relationships between nodes; adjacency matrix Used to store connection weights between nodes.

[0025] In this model, the set of nodes in the graph represents the various multi-port converter devices in the system, while the edges between nodes represent the connections between the converters. To accurately reflect the actual operating environment of the system, the connection weights between nodes do not depend on a single factor, but rather incorporate quantitative parameters that integrate the electrical connection strength and communication quality between nodes.

[0026] This embodiment also provides a specific method for generating the weighted directed graph model in step S100, assuming that the system contains A set of multi-port converter nodes: .

[0027] in, Indicates the first A multi-port converter .

[0028] For multi-port converters with electrical connections Its connection strength is determined by electrical impedance.

[0029] First calculate and Equivalent impedance between two nodes ,in For resistance, Let the reactance be defined. Define the admittance. The calculation of the elements of the electrical connection matrix follows the formula below:

[0030] in, These are the elements of the normalized electrical connection matrix. The admittance magnitude, It is the maximum value of the magnitude of the admittance of all connections in the system.

[0031] Simultaneously, a communication connection matrix is ​​constructed, and the communication connection values ​​comprehensively reflect the information exchange capabilities between nodes. This embodiment uses the following formula to calculate the communication connection strength. : .

[0032] In the formula, For nodes With nodes The actual communication delay between them; The reference delay is used to measure the system's tolerance for delay. Let be the data packet loss rate. The above formula for calculating communication connection strength ensures that the lower the communication latency and the lower the packet loss rate, the closer the communication connection strength is to 1.

[0033] In a further embodiment, to fully characterize the connection relationships between different multi-port converters, an adaptive weight fusion method is used to construct the final weighted adjacency matrix. .

[0034] First, calculate the adaptive weighting coefficients. , .

[0035] .

[0036] The above coefficients The dynamics reflect the proportion of electrical connection strength in the total connection strength. Dynamically reflects the proportion of communication connection strength in the total connection strength, by adjusting , It can adjust the proportion of the influence of electrical connections and communication connections on the overall adjacency matrix.

[0037] Through the above adaptive mechanism, when the communication connection strength When reduced (e.g., during communication congestion). The value will automatically increase, allowing the system to rely more on electrical connection information when making control decisions, thus enhancing the system's robustness in communication-constrained environments.

[0038] Subsequently, the electrical connection matrix and the communication connection matrix are weighted and summed using this coefficient to obtain the final adjacency matrix elements. :

[0039] in, Self-connection strength; It is a unit array.

[0040] Final adoption constitute Represents the adjacency matrix, used for quantization. and The connection weights between them.

[0041] This completes the weighted directed graph model. The construction.

[0042] Step S200: Obtain the operating parameters of the multi-port converter, generate the state vector of each multi-port converter based on the operating parameters, and further generate the system state matrix of the power supply system.

[0043] In this step, the system collects the operating status data for each multi-port converter node. The collected operating parameters include active power. reactive power Output port voltage amplitude Output port voltage phase angle Device surface temperature and operating mode identifier .

[0044] Based on the above parameters, construct a state vector with nodes as the unit. : .

[0045] in, For runtime; , These are the active power and reactive power output by the multi-port converter, respectively, which directly reflect the energy interaction status of the multi-port converter. This refers to the voltage amplitude at the output port. The phase angle of the output port voltage; Maintaining the surface temperature of the main control power module or device is a key safety constraint to ensure that the multi-port converter is not damaged by overheating, and is crucial for achieving long-term reliable operation.

[0046] Constituting the first Multiport converter The operating state can be represented as: .

[0047] Among them, the operating mode identifier Represented using one-hot encoding, for example, grid-connected mode is represented as Offline mode is represented as . This is an identifier variable for the operating mode. When the system is running in grid-connected mode, it keeps synchronized with the main grid. In constant voltage control mode, the goal is to maintain the port voltage of the multi-port converter. In constant power control mode, the goal is to track the load. In protection mode, the system will limit the output of the multi-port converter and enter a safety protection state.

[0048] Specifically, this embodiment also provides an operating mode identifier. An encoding example is shown below: .

[0049] All of the energy supply system The state vectors of the multiport converter nodes at time 1 Superimpose the data to form the system state matrix. :

[0050] This state matrix is ​​used to describe the global operating status of the system in the current control cycle.

[0051] Step S300: Use a graph convolutional network to fuse the weighted directed graph model and the system state matrix to generate a graph embedding matrix that reflects the spatial coupling relationship and state dependency characteristics between multi-port converters.

[0052] This step utilizes Graph Convolutional Networks (GCNs) to extract structure-state features.

[0053] In a preferred embodiment, the GCN network structure consists of one preprocessing layer and three graph convolutional layers, with the output being a graph embedding matrix.

[0054] First, the weighted adjacency matrix generated in step S100 is processed in the preprocessing layer. After performing symmetric normalization, the normalized Laplace matrix is ​​obtained. The specific calculations are as follows: .

[0055] In the formula, This is the self-connection enhancement factor, with a value of 0.1; A degree matrix, a diagonal matrix; To ensure numerical stability, the spectral radius is normalized to form the symmetric normalized Laplacian matrix used for graph convolution operations. .

[0056] In the formula, for The largest eigenvalue.

[0057] Subsequently, the system state matrix obtained in step S200 is... as input features To address the gradient vanishing problem in deep networks and achieve multi-scale feature fusion, this embodiment employs a GCN layer structure that includes residual connections and multinomial filters.

[0058] To capture the coupling effects of neighbors with different hop counts and achieve multi-scale feature fusion, this embodiment employs polynomial filters of different orders in the convolution operation. For the... Layers, each with a different order. Filtered output And then splice them together: .

[0059] In the formula, This is the maximum order of the polynomial filter, typically taken as 2 to 4; For the first Layered Graph Convolutional Networks The output of an order polynomial filter; For the first Layered Graph Convolutional Networks The first order polynomial filter Several coefficients. The filtering results of different orders are concatenated. The model can simultaneously aggregate local (low-order) and global (high-order) spatial structure information.

[0060] No. The feature update formula for the layer is as follows: .

[0061] In the formula, For the first The hidden state matrix of a layered graph convolutional network; For the first The learnable weight matrix of a layered graph convolutional network; For the first The residual connection weight matrix of the layer is used for gradient flow optimization; This is the activation function. The structure is achieved by directly superimposing the input outside the activation function. This ensures the effective transmission of original information and gradients in deep networks.

[0062] Furthermore, attention enhancement and adaptive Dropout operations are performed at the output of each layer to improve feature selection effectiveness and prevent overfitting. Specifically, attention enhancement calculates and weights the fused features of this layer, improving feature selection effectiveness, and places the weights at the output of this layer; adaptive Dropout applies weights to the attention-enhanced features... Apply adaptive Dropout to suppress overfitting, placing it at the output of this layer before feeding it into the next layer.

[0063] go through The node features obtained after layer graph convolution are represented as follows: .

[0064] After iterative computation through a 3-layer graph convolutional network, the output of the 3rd layer is finally denoted as the graph embedding matrix. each of the lines A high-dimensional embedding feature representing a node reflects the spatial coupling relationship and state dependency characteristics of the node and its neighborhood.

[0065] Step S400: Use the policy network to generate local control commands based on the state vector and the corresponding node features in the graph embedding matrix. The policy network is configured in the local controller of each multi-port converter. This step employs a local controller based on the Deep Deterministic Policy Gradient (DDPG) algorithm. Before constructing the specific network structure, the input state model and action space of the controller are first defined.

[0066] Definition of the first Local controller input status of a multi-port converter for: .

[0067] In the formula, The local running state vector obtained in step S200; The graph embedding matrix output by the graph convolutional network in step S300 Corresponding node The feature vectors. This joint input ensures that the controller can both perceive the accurate local state and utilize the neighborhood coupling information extracted by the graph network.

[0068] Define the action vector of the local controller Its constraints are:

[0069] In the formula, , , These are active power regulation, reactive power regulation, and voltage regulation, and all regulation quantities must meet the action constraints.

[0070] The policy network of the local controller adopts an actor architecture; specifically, each multiport converter... The policy network takes the input state as its basis. Generate action vectors .

[0071] The specific model structure of the Actor network is as follows:

[0072] In the formula, For input state transition; , These are the hidden states of layers 1 and 2 of the Actor network; , These are the weight matrices of the Actor network; , These are the bias vectors of the Actor network.

[0073] The final output action vector As an operating mode command, it includes active power regulation. Reactive power regulation Voltage regulation and operating mode identifier .

[0074] In some further embodiments of the present invention, in order to optimize the policy network, the local controller is also configured with an evaluation network, which evaluates the value of local control instructions based on a preset reward function and updates the parameters of the policy network.

[0075] Specifically, the evaluation network adopts the dual-Critic structure of TD3. Compared with the single-Critic structure, this structure can effectively reduce error propagation and improve training convergence and policy performance.

[0076]

[0077] In the formula, , For two Critic networks, the current state-action pair Long-term return estimates under the current strategy; , , This is the weight matrix of the first Critic network; , , This is the weight matrix of the first Critic network.

[0078] Training based on reward function : .

[0079] In the formula, For the reward function; Represents a reference value; , , , , These are the weighting coefficients; To constrain the penalties for violations.

[0080] The evaluation network updates the policy network using the following method: Set the target function: .

[0081] In the formula, and These are the model parameters.

[0082] Set the loss function for the Critic network and update the Critic network parameters by minimizing the mean squared error between the predicted and target values: .

[0083] Set up a policy network update function and use the deterministic policy gradient theorem to update the Actor network parameters. The goal is to maximize the value score of the Critic network. .

[0084] Through the gradient backpropagation described above, the policy network can continuously evolve in the direction of obtaining higher rewards.

[0085] Step S500: Use a graph attention network to perform global aggregation processing on the node features in the graph embedding matrix to generate an energy scheduling matrix containing the active power target value and reactive power target value of each multi-port converter.

[0086] This step utilizes a Graph Attention Network (GAT) for global coordination. First, the multiport converter nodes are computed. For multiport converter nodes Attention coefficient between Characterizing multi-port converter nodes For multiport converter nodes Attention Score:

[0087] in, The feature mapping matrix is ​​obtained through linear transformation of node features: .

[0088] In the formula For the multi-port converter node output in step S300 High-dimensional embedding features , Let be the step weight matrix.

[0089] We perform weighted aggregation of the features of neighboring nodes to generate a feature representation that includes global scheduling information: .

[0090] In the formula, For adjacent multiport converter nodes For the central multi-port converter node Normalized attention weights; It is a non-linear activation function; For multiport converter nodes Globally perceived feature vectors; For multi-attention fusion, It is a linear transformation matrix for multi-head output.

[0091] Furthermore, the feature representation is mapped into an energy dispatch matrix containing the active power target value and reactive power target value of each multi-port converter: .

[0092] In the formula, This is the global energy scheduling output model; , , These are the globally suggested output values ​​after active power, reactive power, and voltage scheduling allocation, respectively.

[0093] Step S600: According to the preset fusion weight, the local control command and energy scheduling matrix are weighted and fused to generate a unified control command, which is then synchronously sent to each multi-port converter.

[0094] Specifically, step S600 integrates the adjustment amount output by the local DDPG controller with the scheduling result output by the global GAT module, and generates the final unified control command according to the set weight coefficients.

[0095] Specifically, the fusion formula includes: . In the formula, , , These are the active, reactive, and voltage control commands for the multi-port converter. , , Taken from the output of step S300; , , Taken from the output of step S500; , , The weighting coefficients are in the range [0,1].

[0096] Some embodiments of the present invention also add a disturbance response and operation mode switching link. When disturbance events such as power imbalance, topology changes or local node failures are detected during system operation, the operation mode switching mechanism is automatically triggered. According to the control strategy, the multi-port converter node is guided to intelligently switch between multiple operation modes such as grid-connected, off-grid, voltage control, and power priority, thereby enhancing the system's anti-disturbance capability and continuous power supply capability.

[0097] In these embodiments, step S600 further includes S610: The system monitors disturbance indicators of the power supply system in real time. When the disturbance indicators exceed the preset threshold, it triggers the operation mode switching mechanism to generate operation mode instructions.

[0098] Specifically, the disturbance metrics include the fluctuation rates of active and reactive power, voltage, and the time interval since the last communication at each multi-port converter port: .

[0099] In the formula, , , These represent the fluctuation rates of active power, reactive power, and voltage at each port of the multi-port converter. This is the time interval between the current communication and the most recent communication. For fixed communication cycles.

[0100] The triggering criteria for disturbance indicators exceeding preset thresholds include: .

[0101] In the formula, , , These are the threshold values ​​for active power, reactive power, and voltage fluctuations during the operation of a multi-port converter. This is a threshold for the communication time interval. When any disturbance indicator exceeds the corresponding threshold, the operation mode switch is triggered.

[0102] Specifically, when , When the system's active and reactive power exceed the threshold, the target multi-port converter switches to power priority mode to adjust the power in a timely manner, aiming to stabilize the port output power. When the target multi-port converter switches to voltage priority mode; when When the operating mode is switched, the multi-port converter switches to off-grid or protection mode. After the operating mode switch, a new output state is displayed. .

[0103] In some further preferred embodiments, to prevent the system from being impacted by a sudden mode switch, the operating mode switching mechanism employs a smooth transition strategy. Specifically, it follows the formula below: .

[0104] In the formula, For the target operating mode; The smoothing coefficient is set to [0.85, 0.9] in this embodiment.

[0105] For step S600, which includes the disturbance response and operating mode switching mechanism S610, the unified control command for the multi-port converter may also include an operating mode switching command when generating the command. .

[0106] Specifically, the unified control commands for each multi-port converter include: .

[0107] in, .

[0108] In the formula, , , The target output of the target multiport converter node.

[0109] The unified control commands obtained above will be synchronously sent to each multi-port converter to solve the problem that a single control mode is difficult to balance steady-state accuracy and dynamic performance under dynamic conditions. This enables precise power allocation and efficient energy scheduling of the multi-port converter cluster in complex environments, significantly improving the robustness and operating efficiency of the power supply system.

[0110] Some embodiments of the present invention also provide state feedback and iterative update steps, using the following formula for iterative updates: .

[0111] In the formula, This is the system's dynamic response function; The system control command matrix at time t contains the control commands for each multiport converter node. Unified control commands; Let be the external disturbance vector of the system at time t; The system's state matrix at time t contains the state of each multiport converter node. The state vector contains operating parameters including active power, reactive power, output port voltage amplitude, output port voltage phase angle, device surface temperature, and operating mode identifier.

[0112] After the instruction is issued, the system updates the status X(t+1) for the next control cycle.

[0113] Another embodiment of the present invention provides a hierarchical collaborative control system for a multi-port converter, which mainly includes a graph model construction module 100, a state perception module 200, a feature extraction module 300, a local control module 400, a global coordination module 500, an instruction fusion module, and a disturbance processing module 600.

[0114] The graph model construction module 100 is configured to generate a weighted directed graph model based on the architecture of a power supply system including a multi-port converter. Specifically, this module is used to perform step S100 in the above method, including calculating the electrical connection matrix and the communication connection matrix, generating adaptive weight coefficients based on the relative proportions of the electrical connection strength and the communication connection strength, and then weighted summing to obtain a weighted adjacency matrix reflecting the coupling relationship.

[0115] The state awareness module 200 is configured to acquire the operating parameters of the multi-port converter, generate a state vector for each multi-port converter based on the operating parameters, and further generate a system state matrix for the power supply system. This module is used to execute step S200 in the above method, and to collect key parameters in real time, including active power, reactive power, port voltage, device temperature, and operating mode identifier, to construct a state matrix that characterizes the global operating status of the system.

[0116] The feature extraction module 300 is configured to use a graph convolutional network to fuse the weighted directed graph model and the system state matrix, generating a graph embedding matrix that reflects the spatial coupling relationship and state dependency characteristics between multi-port converters. This module is used to execute step S300 in the above method, extracting node embedding features containing multi-order neighborhood information through normalized Laplacian matrix construction, multi-order polynomial filtering, and a dual residual connection mechanism.

[0117] The local control module 400 is configured to generate local control commands using a policy network based on the state vector and the corresponding node features in the graph embedding matrix. The policy network is configured in the local controller of each multi-port converter. This module executes step S400 of the above method. Its built-in local controller is constructed based on the DDPG algorithm, utilizes the policy network to generate actions containing active / reactive power regulation and operating mode commands, and achieves adaptive optimization of the policy through an evaluation network and reward function mechanism.

[0118] The global coordination module 500 is configured to use a graph attention network to perform global aggregation processing on the node features in the graph embedding matrix, generating an energy scheduling matrix containing the active power target values ​​and reactive power target values ​​of each multi-port converter. This module is used to execute step S500 in the above method, identifying key energy interaction paths by calculating the attention coefficients between nodes, and realizing power optimization allocation from a global perspective.

[0119] The instruction fusion module 600 is configured to perform weighted fusion of local control instructions and energy scheduling matrices according to preset fusion weights, generate unified control instructions, and synchronously distribute the unified control instructions to each multi-port converter. This module is used to execute step S600 in the above method, balancing the control objectives of local adaptive adjustment and global optimal scheduling through a weighted fusion algorithm, and completing the state update of closed-loop control.

[0120] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This invention is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A hierarchical cooperative control method for a multi-port converter, characterized in that, include: A weighted directed graph model is generated based on the architecture of the power supply system including multi-port converters; the nodes of the weighted directed graph model represent the various multi-port converters in the power supply system, and the edge weights are quantitative parameters that integrate the electrical connection strength and communication quality between nodes. Obtain the operating parameters of the multi-port converter, generate the state vector of each multi-port converter based on the operating parameters, and further generate the system state matrix of the power supply system. The weighted directed graph model and the system state matrix are fused using a graph convolutional network to generate a graph embedding matrix that reflects the spatial coupling relationship and state dependency characteristics between multi-port converters. The policy network generates local control commands based on the state vector and the corresponding node features in the graph embedding matrix. The policy network is configured in the local controller of each multiport converter. A graph attention network is used to perform global aggregation processing on the node features in the graph embedding matrix to generate an energy scheduling matrix containing the active power target value and reactive power target value of each multi-port converter. According to the preset fusion weight, the local control command and the energy scheduling matrix are weighted and fused to generate a unified control command, which is then synchronously sent to each multi-port converter.

2. The hierarchical cooperative control method for a multi-port converter as described in claim 1, characterized in that, The edge weights are calculated using the following method: An electrical connection matrix is ​​generated based on the equivalent electrical impedance between each multi-port converter. A communication connection matrix is ​​generated based on the communication delay and data packet loss rate between each multi-port converter. The electrical connection matrix and the communication connection matrix are weighted and summed to generate a weighted adjacency matrix; A weighted directed graph model of the power supply system is generated based on the weighted adjacency matrix.

3. The hierarchical cooperative control method for a multi-port converter as described in claim 2, characterized in that, The weighting coefficients of the weighted summation are generated based on the relative proportions of electrical connection strength and communication connection strength, such that the weight of the communication connection matrix is ​​increased when the electrical connection strength decreases, or the weight of the electrical connection matrix is ​​increased when the communication connection strength decreases.

4. The hierarchical cooperative control method for a multi-port converter as described in claim 1, characterized in that, The operating parameters of a multi-port converter include its active power, reactive power, port voltage, device temperature, and operating mode indicator.

5. The hierarchical cooperative control method for a multi-port converter as described in claim 1, characterized in that, The weighted directed graph model and the system state matrix are fused using a graph convolutional network to generate a graph embedding matrix that reflects the spatial coupling relationship and state dependency characteristics between multi-port converters, including: The adjacency matrix of the weighted directed graph model is normalized to obtain the normalized Laplace matrix. The system state matrix is ​​used as the initial feature input to the graph convolutional network; The graph convolutional network uses multi-order polynomial filters to perform convolution calculations on the input features, and superimposes the convolution results with the input features through a residual connection mechanism. After multiple rounds of iteration to aggregate multi-order neighborhood features, the graph embedding matrix is ​​generated.

6. The hierarchical cooperative control method for a multi-port converter as described in claim 1, characterized in that, The local controller is also configured with an evaluation network, which evaluates the value of the local control command based on a preset reward function and updates the parameters of the policy network.

7. The hierarchical cooperative control method for a multi-port converter as described in claim 1, characterized in that, A graph attention network is used to globally aggregate the node features in the graph embedding matrix to generate an energy dispatch matrix containing the active power target value and reactive power target value of each multi-port converter, including: A graph attention network is used to linearly map the node features in the graph embedding matrix; Calculate the attention coefficient between nodes, whereby the attention coefficient characterizes the degree of influence of neighboring nodes on the central node; The features of neighboring nodes are weighted and aggregated based on the attention coefficient to generate an energy scheduling matrix containing the active power target value and reactive power target value of each multi-port converter.

8. The hierarchical cooperative control method for a multi-port converter as described in claim 1, characterized in that, The method also includes: The disturbance index of the power supply system is monitored in real time. When the disturbance index exceeds a preset threshold, the operation mode switching mechanism is triggered to generate an operation mode command. The unified control command also includes the operating mode command.

9. The hierarchical cooperative control method for a multi-port converter as described in claim 8, characterized in that, The disturbance indicators include active power fluctuation rate, reactive power fluctuation rate, voltage fluctuation rate, and communication interruption duration.

10. A hierarchical collaborative control system for a multi-port converter, characterized in that, include: The graph model construction module is used to generate a weighted directed graph model based on the architecture of the power supply system including multi-port converters; the nodes of the weighted directed graph model represent the various multi-port converters in the power supply system, and the edge weights are quantitative parameters that integrate the electrical connection strength and communication quality between nodes. The state awareness module is used to acquire the operating parameters of the multi-port converter, generate the state vector of each multi-port converter based on the operating parameters, and further generate the system state matrix of the power supply system. The feature extraction module is used to fuse the weighted directed graph model and the system state matrix using a graph convolutional network to generate a graph embedding matrix that reflects the spatial coupling relationship and state dependency characteristics between multi-port converters. A local control module is used to generate local control commands using a policy network based on the state vector and the corresponding node features in the graph embedding matrix. The policy network is configured in the local controller of each multi-port converter. The global coordination module is used to perform global aggregation processing on the node features in the graph embedding matrix using a graph attention network to generate an energy scheduling matrix containing the active power target value and reactive power target value of each multi-port converter. The instruction fusion module is used to perform weighted fusion of the local control instructions and the energy scheduling matrix according to the preset fusion weights, generate unified control instructions, and synchronously send the unified control instructions to each multi-port converter.

Images