Self-adaptive collaborative decision-making method and device for power distribution network, computer equipment, readable storage medium and program product

Abstract

The invention relates to a self-adaptive collaborative decision-making method and device for a power distribution network, computer equipment, a computer readable storage medium and a computer program product. The method comprises the steps of dividing a distributed power supply cluster according to access positions and parameters of distributed power supplies in a target power distribution network, and obtaining power grid historical operation data of the target power distribution network based on the divided distributed power supply cluster; performing model construction processing to obtain a power distribution network adaptive collaborative decision model and a training set; splitting the model based on splitting learning to obtain an edge-side lightweight model and a cloud-side large-volume model; performing model training processing through the training set, performing parameter fusion updating based on federated learning in the model training process, and performing loss function weight adjustment based on an adaptive reward mechanism to obtain a trained model; and finally carrying out collaborative decision-making processing on the target power distribution network to obtain a collaborative decision-making result. According to the invention, the voltage control effect of the collaborative decision-making process can be effectively ensured.

Description

Technical Field

[0001] The present application relates to the technical field of distribution network, and in particular to a distribution network adaptive collaborative decision-making method, device, computer equipment, computer-readable storage medium and computer program product. Background Art

[0002] With the development of distribution network technology, distributed generators (DG) have also begun to be connected to the distribution network on a large scale and unevenly, which has brought huge challenges to the operation and control of the distribution network. The output of distributed generators is characterized by strong uncertainty, which can easily cause voltage crossing and large voltage fluctuations in the distribution network. Traditional distribution network voltage regulation equipment mainly includes on-load tap changer (OLTC) and capacitor bank (CB), etc., which have slow regulation speeds and are often difficult to adapt to the scenario of fast fluctuations in the output of distributed generators in the distribution network; energy storage systems can use their own capacity reserves to achieve the effect of peak shaving and valley filling in voltage control, but the economy is poor; while distributed generators can use their remaining capacity to provide voltage support for the system, with fast regulation speed, which is an efficient and flexible means of voltage regulation in distribution networks.

[0003] The traditional centralized distributed power supply voltage control method needs to centrally integrate all the data of the distribution network, and then make control decisions on the distribution network based on the global data. However, the traditional centralized method ignores the issue of data ownership in the distribution network and cannot meet the privacy protection needs of multiple stakeholders in the distribution network. At present, the privacy protection problem of multiple stakeholders in the distribution network can be solved through the framework of federated learning. However, due to the limited computing power resources of the edge computing device of the distribution network, it is difficult to meet the resource requirements of deep learning model training, resulting in poor voltage control effect of the collaborative decision-making method based on the federated learning framework. Summary of the invention

[0004] Based on this, it is necessary to provide a distribution network adaptive collaborative decision-making method, device, computer equipment, computer-readable storage medium and computer program product that can effectively improve the voltage control effect in response to the above-mentioned technical problems.

[0005] In a first aspect, the present application provides a distribution network adaptive collaborative decision-making method, comprising:

[0006] Divide the distributed power supply clusters according to the access locations and parameters of the distributed power supply in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters;

[0007] Model building processing is performed based on preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network;

[0008] The distribution network adaptive collaborative decision-making model is split based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side;

[0009] The edge-side lightweight model and the cloud-side large-volume model are subjected to model training processing by using the training set, and during the model training process, the cloud-side large-volume model is subjected to parameter fusion update based on federated learning, and the loss function weight of the cloud-side large-volume model is adjusted based on an adaptive reward mechanism, so as to obtain a target edge model and a target cloud-side model after training;

[0010] Based on the target edge model and the target cloud-side model, collaborative decision processing of the target distribution network is performed to obtain a collaborative decision result.

[0011] In one of the embodiments, the training set includes a node privacy data matrix and a label matrix for each training sample;

[0012] The node privacy data matrix is ​​constructed based on the node active power data and node reactive power data in the historical operation data of the distribution network. The node privacy data matrix is ​​expressed as:

[0013]

[0014] In the above formula, X n,g represents the node privacy data matrix of cluster n at time h, and They represent the active power data and reactive power data of node i of cluster n at time h, respectively. i represents the node number, and I represents the total number of nodes in the distribution network.

[0015] The label matrix is ​​constructed based on the reactive output data of distributed power sources in the historical operation data of the distribution network, and the label matrix is ​​expressed as:

[0016]

[0017] In the above formula, Y n,h represents the label matrix of cluster n at time h, It represents the reactive power output data of the rth distributed generation in cluster n at time h, r represents the sequence number of the distributed generation, and R represents the total number of distributed generation in the distribution network.

[0018] In one embodiment, the performing model training processing on the edge-side lightweight model and the cloud-side large-volume model through the training set includes:

[0019] Inputting the node privacy data matrix and label matrix of the training samples in the training set into the corresponding edge-side lightweight model, performing forward propagation processing through the edge-side lightweight model to obtain cutting layer data;

[0020] The cutting layer data is sent to the cloud-side large-volume model, and the cutting layer data is forward propagated through the cloud-side large-volume model to obtain the output layer loss and the output layer gradient, and the network parameters of the cloud-side large-volume model are updated;

[0021] Back-propagating the output layer loss to the cutting layer through the cloud-side large-volume model to generate a cutting layer gradient, and sending the cutting layer gradient to the edge-side lightweight model;

[0022] The cutting layer gradient is back-propagated through the edge-side lightweight model to update the model parameters of the edge-side lightweight model.

[0023] In one of the embodiments, performing parameter fusion updating on the cloud-side large-volume model based on federated learning during the model training process includes:

[0024] After each round of model training, the network parameters of the cloud-side large-scale model of all clusters are integrated based on the federated average algorithm to obtain the global parameters.

[0025] The network parameters of the cloud-side large-volume model are updated according to the global parameters.

[0026] In one embodiment, adjusting the loss function of the cloud-side large-volume model based on the adaptive reward mechanism includes:

[0027] Obtaining an average voltage deviation index of the training samples in the training set at each moment;

[0028] Based on the average voltage deviation index, determining a weighted value of the loss function at each moment;

[0029] Based on the weighted values ​​of the loss functions at each moment, the cloud-side large-volume model is subjected to loss function adjustment processing to obtain the cloud-side large-volume model after the loss function is adjusted.

[0030] In one of the embodiments, the collaborative decision processing of the target distribution network is performed based on the target edge model and the target cloud-side model to obtain a collaborative decision result, including:

[0031] Performing collaborative decision processing on the target distribution network based on the target edge model and the target cloud-side model to obtain distributed power supply decision data;

[0032] Mask processing is performed on the distributed power source decision data to obtain a collaborative decision result.

[0033] In a second aspect, the present application also provides a distribution network adaptive collaborative decision-making device, comprising:

[0034] A data acquisition module, used to divide the distributed power supply clusters according to the access locations and parameters of the distributed power supplies in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters;

[0035] A model building module, used to perform model building processing based on preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network, to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network;

[0036] A model splitting module, used for splitting the distribution network adaptive collaborative decision-making model based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side;

[0037] A model training module, used to perform model training processing on the edge-side lightweight model and the cloud-side large-volume model through the training set, perform parameter fusion update on the cloud-side large-volume model based on federated learning during the model training process, and adjust the loss function weight of the cloud-side large-volume model based on an adaptive reward mechanism to obtain a target edge model and a target cloud-side model after training;

[0038] The collaborative decision-making module is used to perform collaborative decision processing on the target distribution network based on the target edge model and the target cloud-side model to obtain a collaborative decision result.

[0039] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0040] Divide the distributed power supply clusters according to the access locations and parameters of the distributed power supply in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters;

[0041] Model building processing is performed based on preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network;

[0042] The distribution network adaptive collaborative decision-making model is split based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side;

[0043] The edge-side lightweight model and the cloud-side large-volume model are subjected to model training processing by using the training set, and during the model training process, the cloud-side large-volume model is subjected to parameter fusion update based on federated learning, and the loss function weight of the cloud-side large-volume model is adjusted based on an adaptive reward mechanism, so as to obtain a target edge model and a target cloud-side model after training;

[0044] Based on the target edge model and the target cloud side model, collaborative decision processing of the target distribution network is performed to obtain a collaborative decision result

[0045] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0046] Divide the distributed power supply clusters according to the access locations and parameters of the distributed power supply in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters;

[0047] Model building processing is performed based on preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network;

[0048] The distribution network adaptive collaborative decision-making model is split based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side;

[0049] The edge-side lightweight model and the cloud-side large-volume model are subjected to model training processing by using the training set, and during the model training process, the cloud-side large-volume model is subjected to parameter fusion update based on federated learning, and the loss function weight of the cloud-side large-volume model is adjusted based on an adaptive reward mechanism, so as to obtain a target edge model and a target cloud-side model after training;

[0050] Based on the target edge model and the target cloud side model, collaborative decision processing of the target distribution network is performed to obtain a collaborative decision result

[0051] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0052] Divide the distributed power supply clusters according to the access locations and parameters of the distributed power supply in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters;

[0053] Model building processing is performed based on preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network;

[0054] The distribution network adaptive collaborative decision-making model is split based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side;

[0055] The edge-side lightweight model and the cloud-side large-volume model are subjected to model training processing by using the training set, and during the model training process, the cloud-side large-volume model is subjected to parameter fusion update based on federated learning, and the loss function weight of the cloud-side large-volume model is adjusted based on an adaptive reward mechanism, so as to obtain a target edge model and a target cloud-side model after training;

[0056] Based on the target edge model and the target cloud side model, collaborative decision processing of the target distribution network is performed to obtain a collaborative decision result

[0057] The above-mentioned distribution network adaptive collaborative decision-making method, device, computer equipment, computer-readable storage medium and computer program product divide the distributed power supply clusters according to the access location and parameters of the distributed power supplies in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters; perform model construction processing based on preset model parameters, historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network; split the distribution network adaptive collaborative decision-making model based on split learning to obtain an edge-side lightweight model and a cloud-side large-volume model; perform model training processing on the edge-side lightweight model and the cloud-side large-volume model through the training set, perform parameter fusion update on the cloud-side large-volume model based on federated learning during the model training process, and adjust the loss function weight of the cloud-side large-volume model based on an adaptive reward mechanism to obtain the target edge model and target cloud-side model after training; perform collaborative decision processing on the target distribution network based on the target edge model and the target cloud-side model to obtain a collaborative decision result. This application constructs a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network by first acquiring various types of data on the target distribution network and distributed power sources. And through the framework of split federated learning, the training process of the distribution network adaptive collaborative decision-making model is completed through the training set data. The target edge model and target cloud-side model that can be used for adaptive collaborative decision-making of the distribution network are obtained, and then the collaborative decision-making process of the distribution network is completed. This application fully considers the data privacy protection performance among multiple stakeholders corresponding to distributed power sources through the control framework of split-federated learning, and adds an adaptive reward mechanism to the model to optimize the control strategy, which can effectively ensure the voltage control effect of the collaborative decision-making process. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0059] Figure 1 An application environment diagram of a distribution network adaptive collaborative decision-making method in an embodiment;

[0060] Figure 2 It is a flowchart of a method for adaptive collaborative decision-making in a distribution network in one embodiment;

[0061] Figure 3 A structural diagram of a certain node calculation example in an embodiment;

[0062] Figure 4A test day source load fluctuation curve in an embodiment;

[0063] Figure 5 A comparison diagram of voltage extreme value deviation curves of Schemes I and II in an embodiment;

[0064] Figure 6 A comparison diagram of voltage extreme value deviation curves of Schemes II and III in an embodiment;

[0065] Figure 7 A comparison diagram of voltage extreme value deviation curves of Schemes II and IV in an embodiment;

[0066] Figure 8 It is a curve of neural network convergence when cluster three-no reward mechanism is used in one embodiment;

[0067] Fig. 9 It is a curve of neural network convergence when cluster three has a reward mechanism in one embodiment;

[0068] Fig.10 It is a curve of neural network convergence when cluster 4 has no reward mechanism in one embodiment;

[0069] Fig.11 This is a curve of neural network convergence when cluster four has a reward mechanism in one embodiment.

[0070] Fig.12 It is a structural block diagram of a distribution network adaptive collaborative decision-making device in one embodiment;

[0071] Fig.13 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0073] The distribution network adaptive collaborative decision-making method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. When the staff of the terminal 102 needs to formulate a collaborative decision-making plan for the target distribution network connected to multiple distributed power sources to control the operation of the distribution network, the corresponding collaborative decision-making model can be generated by the distribution network adaptive collaborative decision-making method of the present application. First, the user specifies the target distribution network through the terminal 102, and the server 104 divides the distributed power supply clusters according to the access location and parameters of the distributed power supplies in the specified target distribution network, and obtains the historical operation data of the target distribution network based on the divided distributed power supply clusters; the model is constructed based on the preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain the distribution network adaptive collaborative decision-making model and training set based on the graph convolutional neural network; the distribution network adaptive collaborative decision-making model is split based on split learning to obtain the edge side lightweight model and the cloud side large volume model; the edge side lightweight model and the cloud side large volume model are trained through the training set, the loss function of the cloud side large volume model is adjusted based on the adaptive reward mechanism, and the cloud side large volume model is updated based on the federated learning during the model training process to obtain the target edge model and the target cloud side model after the training is completed; the target distribution network is collaboratively decided based on the target edge model and the target cloud side model to obtain the collaborative decision result. Among them, the terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, projection devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0074] In an exemplary embodiment, Figure 2 As shown in the figure, a distribution network adaptive collaborative decision-making method is provided, which is applied to Figure 1 The server 104 in the example is used as an example to illustrate the method, which includes the following steps 201 to 209. Among them:

[0075] Step 201 : divide the distributed power source clusters according to the access locations and parameters of the distributed power sources in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power source clusters.

[0076] Among them, the target distribution network refers to the target distribution network that needs to be processed by collaborative decision-making. The distribution network is connected to multiple distributed power sources, which belong to multiple stakeholders. The output of distributed power sources is characterized by strong uncertainty, which easily causes the distribution network to have voltage crossing and large voltage fluctuations. Therefore, it is necessary to process the target distribution network by collaborative decision-making to alleviate the voltage fluctuation problem. A distributed power source cluster refers to a cluster composed of multiple distributed power sources. The historical operation data of the power grid can be node active power data, node reactive power data, and distributed power source reactive output data. These data can be determined based on the divided distributed power source clusters after the distributed power source clusters are divided.

[0077] Exemplarily, when a user needs to perform efficient collaborative decision-making processing on a distribution network connected to multiple distributed power sources to adjust the voltage fluctuation of the distribution network, the corresponding decision-making scheme can be generated through the adaptive collaborative decision-making of the distribution network of the present application. First, the user can specify the target distribution network through the terminal 102, and the server 104 responsible for model generation and collaborative decision-making can collect relevant processing after determining the target distribution network. First, the target distribution network needs to be specified, the access location and parameters of the distributed power sources in the target distribution network are determined, the distributed power source clusters are divided, and the historical operation data of the power grid of the target distribution network is obtained based on the divided distributed power source clusters. In one embodiment, the user can select an active distribution network through the terminal, and input the distribution network topology connection relationship and the access location and parameters of the distributed power source to the server, and divide the distributed power source cluster; input the historical operation data of the distribution network, including node active power data, node reactive power data and distributed power source reactive output data.

[0078] Step 203, a model is constructed based on preset model parameters, historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network.

[0079] Among them, the preset model parameters refer to the parameters used to construct the initial model. The model constructed in this application is a distribution network adaptive collaborative decision-making model based on graph convolutional neural network. The preset model parameters specify the graph convolutional neural network parameters, including the number of network layers, learning rate, loss function, number of training rounds, and regularization coefficient, etc. At the same time, the construction of the model and the generation of the training set can also be combined with the historical operation data of the distribution network and the topological connection relationship of the target distribution network.

[0080] For example, after obtaining the historical operation data of the power grid, the model can be constructed and the training set can be generated based on these historical operation data of the power grid. At the same time, the model also needs to take into account the topological connection relationship of the target distribution network. When generating the adaptive collaborative decision-making model of the distribution network, it is necessary to design the loss function, activation function and inter-layer transfer structure of the adaptive collaborative decision-making model. For the training set, it is necessary to design the node privacy data matrix, label matrix and topology matrix of each training object in the training set.

[0081] The constructed node privacy data matrix is ​​constructed based on the node active power data and node reactive power data in the historical operation data of the distribution network. The node privacy data matrix is ​​expressed as:

[0082]

[0083] In the above formula, X n,h represents the node privacy data matrix of cluster n at time h, and They represent the active power data and reactive power data of node i of cluster n at time h, respectively. i represents the node number, and I represents the total number of nodes in the distribution network.

[0084] The label matrix is ​​constructed based on the reactive output data of distributed generation in the historical operation data of the distribution network. The label matrix is ​​expressed as:

[0085]

[0086] In the above formula, Y n,h represents the label matrix of cluster n at time h, It represents the reactive power output data of the rth distributed generation in cluster n at time h, r represents the sequence number of the distributed generation, and R represents the total number of distributed generation in the distribution network.

[0087] Step 205, splitting the distribution network adaptive collaborative decision-making model based on split learning to obtain a lightweight model on the edge side and a large-scale model on the cloud side.

[0088] Step 207, perform model training processing on the edge side lightweight model and the cloud side large-volume model through the training set, adjust the loss function of the cloud side large-volume model based on the adaptive reward mechanism, and perform parameter fusion update on the cloud side large-volume model based on federated learning during the model training process to obtain the target edge model and target cloud side model after training.

[0089] Among them, split learning refers to splitting the model into multiple parts for learning, and in the solution of this application, it refers to splitting the overall large model of the distribution network adaptive collaborative decision-making model into an edge-side lightweight model and a cloud-side large-volume model. The edge-side lightweight model is deployed to the edge side, and the cloud-side large-volume model is deployed to the cloud. The adaptive reward mechanism is used to optimize the control strategy, mainly used to adjust the loss function of the cloud-side large-volume model, and the split edge-side lightweight model and cloud-side large-volume model are essentially still a model, so it can also be regarded as optimizing the loss function of the overall distribution network adaptive collaborative decision-making model. Federated learning is a machine learning framework that can effectively help multiple institutions use data and perform machine learning modeling while meeting the requirements of user privacy protection, data security and government regulations. This application uses the framework of federated learning to complete the parameter update of the cloud-side large-volume model, thereby effectively protecting the privacy security of multiple stakeholders corresponding to distributed power sources.

[0090] Exemplarily, after obtaining the adaptive collaborative decision-making model of the distribution network, the training process of the model can be completed based on the framework of split-federated learning, so as to ensure the performance of the model. First, the adaptive collaborative decision-making model of the distribution network can be split based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side. In a specific embodiment, different distributed power clusters each correspond to a lightweight model on the edge side, and during model training, the training process of the corresponding model of the cluster can be completed based on the historical data of the distributed power cluster. During the training process, the loss function of the large-volume model on the cloud side is adjusted based on the adaptive reward mechanism, and the parameter fusion update of the large-volume model on the cloud side is performed based on federated learning during the model training process. It is not only based on the voltage control problem of high-proportion distributed power sources connected to the distribution network, but also considers the data privacy protection performance among multiple stakeholders. It also optimizes the control strategy by adding an adaptive reward mechanism, alleviates the voltage fluctuation problem, has a good voltage control effect, and the adaptive reward mechanism effectively improves the convergence of the model.

[0091] Step 209: perform collaborative decision processing on the target distribution network based on the target edge model and the target cloud-side model to obtain a collaborative decision result.

[0092] For example, after the trained target edge model and target cloud-side model are obtained, these models can be deployed to edge devices and the cloud. These models can collect various operating data of the target distribution network and distributed power sources during the operation of the distribution network, and perform adaptive collaborative decision-making based on real-time operating data to obtain collaborative decision results, and perform distributed control of the distribution network based on the results of the collaborative decision.

[0093] The above-mentioned distribution network adaptive collaborative decision-making method divides the distributed power supply clusters according to the access location and parameters of the distributed power supply in the target distribution network, and obtains the historical operation data of the target distribution network based on the divided distributed power supply clusters; the model is constructed based on the preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain the distribution network adaptive collaborative decision-making model and training set based on the graph convolutional neural network; the distribution network adaptive collaborative decision-making model is split based on split learning to obtain the edge side lightweight model and the cloud side large volume model; the edge side lightweight model and the cloud side large volume model are trained by the training set, and the cloud side large volume model is updated with parameters based on federated learning during the model training process, and the loss function weight of the cloud side large volume model is adjusted based on the adaptive reward mechanism to obtain the target edge model and the target cloud side model after training; the target distribution network is collaboratively decided based on the target edge model and the target cloud side model to obtain the collaborative decision result. This application constructs a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network by first obtaining various types of data of the target distribution network and distributed power supplies. And through the framework of split federated learning, the training process of the adaptive collaborative decision-making model of the distribution network is completed through the training set data. The target edge model and target cloud side model that can be used for adaptive collaborative decision-making of the distribution network are obtained, and then the collaborative decision-making process of the distribution network is completed. This application fully considers the data privacy protection performance among multiple stakeholders corresponding to distributed power sources through the control framework of split-federated learning, and adds an adaptive reward mechanism to the model to optimize the control strategy, which can effectively ensure the voltage control effect of the collaborative decision-making process.

[0094] In an exemplary embodiment, the model training processing of the edge-side lightweight model and the cloud-side large-volume model through the training set in step 207 includes: inputting the node privacy data matrix and label matrix of the training samples in the training set into the corresponding edge-side lightweight model, performing forward propagation processing through the edge-side lightweight model to obtain cutting layer data; sending the cutting layer data to the cloud-side large-volume model, performing forward propagation processing on the cutting layer data through the cloud-side large-volume model to obtain the output layer loss and output layer gradient, and updating the network parameters of the cloud-side large-volume model; back-propagating the output layer loss to the cutting layer through the cloud-side large-volume model to generate a cutting layer gradient, and sending the cutting layer gradient to the edge-side lightweight model; back-propagating the cutting layer gradient through the edge-side lightweight model to update the model parameters of the edge-side lightweight model.

[0095] Exemplarily, the training set contains training samples corresponding to different clusters. Therefore, during the model training process, the training samples can be input into the edge-side lightweight model, and forward propagation processing can be performed through the edge-side lightweight model to obtain the corresponding data of the forward propagation. These data will be transmitted to the cloud-side large-volume model, and the cloud-side large-volume model will be further forward-propagated to perform loss calculation and gradient calculation and other processing to update the network parameters of the cloud-side large-volume model. After the network parameters of the cloud-side large-volume model are updated, back propagation can be performed based on the calculated loss to determine the cutting layer gradient, and then the model parameters of the edge-side lightweight model can be updated to complete the parameter update process of the entire model. By iterating the model training process, the final target edge model and target cloud-side model can be obtained. In a specific embodiment, the training process of the edge-side lightweight model and the cloud-side large-volume model is expressed as:

[0096] a) First, the lightweight model on the edge side and the large-volume model on the cloud side initialize their respective network parameters and the number of training rounds k;

[0097] b) The nth cluster of the distribution network collects the node privacy data X during the operation of the distribution network n With label data Y n , and X n and Y n The data is packaged and transmitted to the input end of the lightweight model on the edge side corresponding to the nth cluster;

[0098] c) The edge side lightweight model receives X transmitted by the distribution network n and Y n , forward propagates to the cutting layer of the lightweight model on the edge side, and encrypts the cutting layer data S k,n With label Y n Transmit to the cloud side large volume model, encrypt data S k,n The generation process of is expressed as:

[0099]

[0100] In the formula, S k,n represents the encrypted data of the cutting layer during the k-th round of training of the n-th cluster, F a () represents the forward propagation process function of the edge-side lightweight model, is the network parameter in the edge-side lightweight model during the k-th round of training for the n-th cluster, X n Represents the node privacy data of the nth cluster;

[0101] d) The cloud-side large-volume model receives the encrypted data S of the cutting layer k,n and label Y nPerform forward propagation and generate a loss function at the output layer. The output layer loss function can be expressed as:

[0102]

[0103] Where, L k,n is the output layer loss function of the cloud-side large-volume model during the k-th round of training for the n-th cluster, F b () represents the forward propagation process function of the large-volume model on the cloud side, is the network parameter of the large-scale model on the cloud side during the k-th round of training of the n-th cluster, S k,n Y is the encrypted data transmitted by the cutting layer during the k-th round of training for the n-th cluster, n Represents the label data of the nth cluster;

[0104] e) The cloud-side large-volume model is based on the output layer loss function L k,n Calculate the output layer gradient M k,n , and update the network parameters of the large-scale model on the cloud side The cloud-side large-volume model is based on the output layer loss function L k,n Back propagation to the cutting layer generates the cutting layer gradient J k,n , and the cutting layer gradient J k,n Transmitted to the edge side lightweight model, this process can be expressed as:

[0105]

[0106]

[0107] In the formula, is the network parameter of the large-scale model on the cloud side during the k+1th round of training of the nth cluster, is the network parameter of the large-scale model on the cloud side during the k-th round of training for the n-th cluster, η represents the learning rate during training, and M k,n represents the output layer gradient of the nth cluster during the kth round of training, J k,n represents the cutting layer gradient of the kth round of training for the nth cluster, represents the back propagation process function of the cloud-side neural network, L k,n is the output layer loss function of the k-th round of training for the n-th cluster;

[0108] f) Edge-side lightweight model back-propagates the cutting layer gradient J k,n , update the network parameters of the lightweight model on the edge side. This process can be expressed as:

[0109]

[0110] In the formula, is the network parameter in the edge-side lightweight model during the k+1th round of training for the nth cluster, is the network parameter in the edge-side lightweight model during the k-th round of training for the n-th cluster, η represents the learning rate during training, and J k,n represents the cutting layer gradient of the kth round of training for the nth cluster;

[0111] g) Set the number of training rounds k = k + 1, and repeat steps b) - g) until the output layer loss function L k,n Convergence. In this embodiment, the model parameter update processing of the edge-side lightweight model and the cloud-side large-volume model is completed through forward propagation and back propagation processing, thereby effectively ensuring the availability of the edge-side lightweight model and the cloud-side large-volume model. The trained model can effectively perform collaborative decision-making processing on the target distribution network.

[0112] In an exemplary embodiment, the parameter fusion update of the cloud-side large-volume model based on federated learning during the model training process in step 207 includes: after each round of training in the model training process, integrating the network parameters of the cloud-side large-volume model of all clusters based on the federated averaging algorithm to obtain global parameters; updating the network parameters of the cloud-side large-volume model according to the global parameters.

[0113] Exemplarily, the present application can specifically complete the processing of model parameter update based on federated learning. In the process of model parameter update, after updating the model parameters of the edge-side lightweight model and the cloud-side large-volume model through back-propagation and forward-propagation processing, the network parameters of the cloud-side large-volume model of all clusters are integrated based on the federated average algorithm to obtain global parameters, and then the network parameters of the cloud-side large-volume model are updated through the global parameters, thereby effectively ensuring the privacy of the stakeholders corresponding to the distributed power supply in the cluster. The process can be specifically expressed as:

[0114]

[0115]

[0116] In the formula, represents the global parameters during the kth round of training, N represents the total number of clusters, is the network parameter of the large-scale model on the cloud side during the k-th round of training for the n-th cluster, is the network parameter in the cloud-side large-volume model during the k+1th round of training of the nth cluster. In this embodiment, the network parameter update of the cloud-side large-volume model is implemented through the federated averaging algorithm, thereby effectively ensuring the privacy security of the system.

[0117] In an exemplary embodiment, the loss function adjustment of the cloud-side large-volume model based on the adaptive reward mechanism in step 207 includes: obtaining the average voltage deviation index of the training samples in the training set at each moment; determining the weighted value of the loss function at each moment based on the average voltage deviation index; and adjusting the loss function of the cloud-side large-volume model based on the weighted value of the loss function at each moment to obtain the cloud-side large-volume model after the loss function is adjusted.

[0118] Exemplarily, for the adjustment process of the adaptive reward mechanism, since this application mainly relies on the adaptive reward mechanism to optimize the voltage control strategy and ensure the voltage control effect. Therefore, in this process, it is necessary to first obtain the average voltage deviation index of the training samples in the training set at each time. Then, based on the weighting of the loss function at different times of these average voltage deviation indicators, the loss function adjustment processing of the cloud-side large-volume model is completed to obtain the cloud-side large-volume model after the loss function is adjusted. In one embodiment, the process can be expressed as:

[0119] w h =f(δ n,h )

[0120]

[0121]

[0122] In the formula, w h represents the weighted value of the loss function of the model at time h, δ n,h represents the average voltage deviation index of cluster n at time h, and f() represents the weighted value of the loss function w h With the average voltage deviation index δ n,h The function of change, I n is the number of nodes in cluster n, represents the square value of the voltage at node i at time h, ω represents the model parameter, L(ω) represents the overall loss function, and L h (ω) represents the loss function at time h, and H represents the total number of moments;

[0123] w h =δ(δ n,h ) represents the weighted value w of the loss function h According to the average voltage deviation index δ n,h Adjustment, the adjustment method is: loss function weighted value w h With the average voltage deviation index δ n,hIn this embodiment, the optimization and adjustment of the model loss function is completed by the average voltage deviation index of the training samples at each moment, thereby effectively ensuring that the voltage fluctuation problem is alleviated, having a good voltage control effect, and the adaptive reward mechanism improves the convergence of the model.

[0124] In an exemplary embodiment, step 209 includes: performing collaborative decision processing on the target distribution network based on the target edge model and the target cloud side model to obtain distributed power supply decision data; and performing mask processing on the distributed power supply decision data to obtain a collaborative decision result.

[0125] Exemplarily, the solution of the present application can also complete the correction and update processing of the decision strategy based on mask processing, perform collaborative decision processing of the target distribution network based on the target edge model and the target cloud side model, and obtain the distributed power supply active output, reactive output and other data of the distributed power supply decision data. Based on these data, mask processing is performed to obtain the distributed power supply decision result that is most suitable for the scene. The process can be expressed as:

[0126]

[0127]

[0128] In the formula, Q r represents the reactive power output of the rth distributed generation, P r represents the active output of the rth distributed generation, S r represents the capacity of the rth distributed generation, Q r,min and Q r,max They respectively represent the minimum and maximum reactive output of the rth distributed generation.

[0129] The present application also provides a comparative embodiment, using a certain node system example, whose topology and partition structure are as follows Figure 3 The total active power of the rated load is 10.05MW, the total reactive power is 7.54Mvar, and the rated voltage is 10.5kV. Eight wind turbines and 16 photovoltaic systems are connected to the node system. The total connected rated active capacity is 12.5kW, and its specific connection location and capacity are shown in Table 1 below.

[0130] Table 1 Rated capacity and access location of distributed generation

[0131]

[0132] In order to fully verify the advancement of this application method, in this example, the following four schemes are adopted for comparative analysis:

[0133] Scheme I: The reactive output of distributed generation is 0, and the initial operation state of the distribution network is obtained;

[0134] Solution II: Using the application method, the reactive power output of distributed power sources is determined to obtain the operating status of the distribution network;

[0135] Solution III: Optimize the reactive power output of distributed power sources using a centralized optimization method;

[0136] Scheme IV: Using the MADDPG control method, the output of the controlled equipment is determined and the operating status of the distribution network is obtained.

[0137] The average voltage deviation index AVD is introduced to calculate the average voltage deviation of the entire network, which more intuitively reflects the voltage fluctuation under different schemes:

[0138]

[0139] In the formula, represents the voltage of node i at time h, H is the total number of time, and I is the total number of all cluster nodes in the distribution network. The partial neural network hyperparameter settings of the present application solution can be referred to as shown in Table 2.

[0140] Table 2 Neural network hyperparameter settings

[0141]

[0142] The comparison of voltage control results of schemes I, II, III, and IV is shown in Table 3, and the comparison of the method proposed in this application and the MADDPG method is shown in Table 4.

[0143] Table 3 Voltage control results of different schemes on the test day

[0144]

[0145] Table 4 Comparison between this application method and MADDPG method

[0146]

[0147] By comparing Scheme II with Scheme I, the AVD index of Scheme II dropped from 2.9159 to 1.2024 compared with Scheme I, proving the effectiveness of the method proposed in this application. When the distributed power source is not working, the distribution network will have severe voltage fluctuations and voltage over-limit phenomena; after using this method to control the reactive output of the distributed power source, the voltage control effect is greatly improved, solving the problem of large voltage fluctuations and voltage over-limit in the scenario where the distributed power source is not working, and by controlling the reactive output of the distributed power source, the system voltage is maintained at a safe operating level.

[0148] By comparing Scheme II with Scheme III, it is shown that this method and the centralized optimization method can achieve similar control effects in certain time periods.

[0149] By comparing Scheme II with Scheme IV, it is shown that this method can achieve similar effects as the MADDPG method in terms of control capability. Compared with the MADDPG method, this method has stronger privacy protection performance. This method uses the advantages of the split federated learning framework to achieve that the original data of the node is not uploaded to the cloud, and experience is not shared between clusters. At the same time, it also has adaptive capabilities.

[0150] Since the loss function calculation methods of the models with reward function and without reward function are different, the magnitude of the loss function in the two cases is different, so the convergence of the model is judged by analyzing the convergence curve trend of the same cluster neural network model. For specific situations, please refer to Figure 4-11 , Figure 4 It is the source load fluctuation curve on the test day; Figure 5 This is a comparison chart of voltage extreme value deviation curves of Schemes I and II; Figure 6 This is a comparison chart of voltage extreme value deviation curves of Schemes II and III; Figure 7 This is a comparison chart of voltage extreme value deviation curves of Schemes II and IV; Figure 8 It is the convergence curve of the neural network when there is no reward mechanism for clusters; Fig. 9 It is the convergence curve of the neural network when cluster three has a reward mechanism; Fig.10 It is the convergence curve of the neural network when cluster 4 has no reward mechanism; Fig.11 This is the convergence curve of the neural network when cluster 4 has a reward mechanism. Figure 8 and Fig. 9 When there is a reward function, the loss function after 1000 iterations is significantly lower than the initial loss function, which shows the effectiveness of the reward mechanism in improving the convergence of the model. Fig.10 and Fig.11 , the oscillation range of the loss function of cluster 4 is larger when it approaches 1000 rounds. When there is a reward function, the loss function value of the model after 1000 iterations is significantly lower than that of the initial model, and it shows a convergence trend; compared with the case of the same cluster without a reward function, the loss function value does not drop significantly and does not show a convergence trend. It can be seen that the convergence performance of the neural network with a reward mechanism is much better than that without a reward mechanism, which shows that the reward mechanism can, to a certain extent, screen the quality of data according to its value and improve the convergence performance of the neural network.

[0151] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0152] Based on the same inventive concept, the embodiment of the present application also provides a distribution network adaptive collaborative decision-making device for implementing the distribution network adaptive collaborative decision-making method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more distribution network adaptive collaborative decision-making device embodiments provided below can refer to the limitations of the distribution network adaptive collaborative decision-making method above, and will not be repeated here.

[0153] In an exemplary embodiment, Fig.12 As shown, a distribution network adaptive collaborative decision-making device is provided, comprising:

[0154] The data acquisition module 1201 is used to divide the distributed power source clusters according to the access locations and parameters of the distributed power sources in the target distribution network, and obtain the historical operation data of the target distribution network based on the divided distributed power source clusters.

[0155] The model construction module 1203 is used to perform model construction processing based on preset model parameters, historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network.

[0156] The model splitting module 1205 is used to split the distribution network adaptive collaborative decision-making model based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side.

[0157] The model training module 1207 is used to perform model training processing on the edge side lightweight model and the cloud side large-volume model through the training set. During the model training process, the parameter fusion update of the cloud side large-volume model is performed based on federated learning, and the loss function weight of the cloud side large-volume model is adjusted based on the adaptive reward mechanism to obtain the target edge model and target cloud side model after the training is completed.

[0158] The collaborative decision module 1209 is used to perform collaborative decision processing on the target distribution network based on the target edge model and the target cloud side model to obtain a collaborative decision result.

[0159] In one embodiment, the training set includes a node privacy data matrix and a label matrix for each training sample;

[0160] The node privacy data matrix is ​​constructed based on the node active power data and node reactive power data in the historical operation data of the distribution network. The node privacy data matrix is ​​expressed as:

[0161]

[0162] In the above formula, X n,h represents the node privacy data matrix of cluster n at time H, and They represent the active power data and reactive power data of node i of cluster n at time h, respectively. i represents the node number, and I represents the total number of nodes in the distribution network.

[0163] The label matrix is ​​constructed based on the reactive output data of distributed generation in the historical operation data of the distribution network. The label matrix is ​​expressed as:

[0164]

[0165] In the above formula, U n,h represents the label matrix of cluster n at time h, It represents the reactive power output data of the rth distributed generation in cluster n at time h, r represents the sequence number of the distributed generation, and R represents the total number of distributed generation in the distribution network.

[0166] In one embodiment, the model training module 1207 is specifically used to: input the node privacy data matrix and label matrix of the training samples in the training set into the corresponding edge-side lightweight model, perform forward propagation processing through the edge-side lightweight model to obtain cutting layer data; send the cutting layer data to the cloud-side large-volume model, perform forward propagation processing on the cutting layer data through the cloud-side large-volume model to obtain the output layer loss and output layer gradient, and update the network parameters of the cloud-side large-volume model; back-propagate the output layer loss to the cutting layer through the cloud-side large-volume model to generate a cutting layer gradient, and send the cutting layer gradient to the edge-side lightweight model; back-propagate the cutting layer gradient through the edge-side lightweight model to update the model parameters of the edge-side lightweight model.

[0167] In one embodiment, the model training module 1207 is specifically used to: after each round of training in the model training process, integrate the network parameters of the cloud-side large-volume model of all clusters based on the federated average algorithm to obtain global parameters; and update the network parameters of the cloud-side large-volume model according to the global parameters.

[0168] In one embodiment, the model training module 1207 is specifically used to: obtain the average voltage deviation index of the training samples in the training set at each moment; determine the weighted value of the loss function at each moment based on the average voltage deviation index; and perform loss function adjustment processing on the cloud-side large-volume model based on the weighted value of the loss function at each moment to obtain the cloud-side large-volume model after the loss function is adjusted.

[0169] In one embodiment, the collaborative decision module 1209 is specifically used to: perform collaborative decision processing on the target distribution network based on the target edge model and the target cloud side model to obtain distributed power supply decision data; and perform mask processing on the distributed power supply decision data to obtain collaborative decision results.

[0170] Each module in the above-mentioned distribution network adaptive collaborative decision-making device can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0171] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig.13 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data related to the adaptive collaborative decision-making of the distribution network. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for adaptive collaborative decision-making of a distribution network is implemented.

[0172] Those skilled in the art will understand that Fig.13 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0173] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0174] In one embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0175] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-mentioned method embodiments.

[0176] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0177] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0178] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0179] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the present application. It should be noted that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A distribution network adaptive collaborative decision-making method, characterized in that: The method comprises: Divide the distributed power supply clusters according to the access locations and parameters of the distributed power supply in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters; Model building processing is performed based on preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network; The distribution network adaptive collaborative decision-making model is split based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side; The edge-side lightweight model and the cloud-side large-volume model are subjected to model training processing by using the training set, the loss function of the cloud-side large-volume model is adjusted based on an adaptive reward mechanism, and the cloud-side large-volume model is subjected to parameter fusion update based on federated learning during the model training process to obtain a target edge model and a target cloud-side model after training; Based on the target edge model and the target cloud-side model, collaborative decision processing of the target distribution network is performed to obtain a collaborative decision result.

2. The method according to claim 1, characterized in that The training set includes a node privacy data matrix and a label matrix of each training sample; The node privacy data matrix is ​​constructed based on the node active power data and node reactive power data in the historical operation data of the distribution network. The node privacy data matrix is ​​expressed as: In the above formula, X n，h represents the node privacy data matrix of cluster n at time h, and They represent the active power data and reactive power data of node i of cluster n at time h, respectively, i represents the node number, and I represents the total number of nodes in the distribution network; The label matrix is ​​constructed based on the reactive output data of distributed power sources in the historical operation data of the distribution network, and the label matrix is ​​expressed as: In the above formula, Y n，h represents the label matrix of cluster n at time h, It represents the reactive power output data of the rth distributed generation in cluster n at time h, r represents the sequence number of the distributed generation, and R represents the total number of distributed generation in the distribution network.

3. The method according to claim 2, characterized in that The performing model training processing on the edge-side lightweight model and the cloud-side large-volume model through the training set includes: Inputting the node privacy data matrix and label matrix of the training samples in the training set into the corresponding edge-side lightweight model, performing forward propagation processing through the edge-side lightweight model to obtain cutting layer data; The cutting layer data is sent to the cloud-side large-volume model, and the cutting layer data is forward propagated through the cloud-side large-volume model to obtain the output layer loss and the output layer gradient, and the network parameters of the cloud-side large-volume model are updated; Back-propagating the output layer loss to the cutting layer through the cloud-side large-volume model to generate a cutting layer gradient, and sending the cutting layer gradient to the edge-side lightweight model; The cutting layer gradient is back-propagated through the edge-side lightweight model to update the model parameters of the edge-side lightweight model.

4. The method according to claim 3, characterized in that The updating of the parameters of the cloud-side large-volume model through fusion based on federated learning during the model training process includes: After each round of model training, the network parameters of the cloud-side large-scale model of all clusters are integrated based on the federated average algorithm to obtain the global parameters. The network parameters of the cloud-side large-volume model are updated according to the global parameters.

5. The method according to claim 3, characterized in that: The adjusting the loss function of the cloud-side large-volume model based on the adaptive reward mechanism includes: Obtaining an average voltage deviation index of the training samples in the training set at each moment; Based on the average voltage deviation index, determining a weighted value of the loss function at each moment; Based on the weighted values ​​of the loss functions at each moment, the cloud-side large-volume model is subjected to loss function adjustment processing to obtain the cloud-side large-volume model after the loss function is adjusted.

6. The method according to claim 1, characterized in that The collaborative decision-making process of the target distribution network is performed based on the target edge model and the target cloud-side model to obtain a collaborative decision result, including: Performing collaborative decision processing on the target distribution network based on the target edge model and the target cloud-side model to obtain distributed power supply decision data; Mask processing is performed on the distributed power source decision data to obtain a collaborative decision result.

7. A distribution network adaptive collaborative decision-making device, characterized in that: The device comprises: A data acquisition module is used to divide the distributed power supply clusters according to the access locations and parameters of the distributed power supplies in the target distribution network, and obtain the historical power grid operation data of the target distribution network based on the divided distributed power supply clusters; A model building module, used to perform model building processing based on preset model parameters, the historical operation data of the distribution network and the topological connection relationship of the target distribution network, to obtain a distribution network adaptive collaborative decision-making model and training set based on a graph convolutional neural network; A model splitting module, used for splitting the distribution network adaptive collaborative decision-making model based on split learning to obtain a lightweight model on the edge side and a large-volume model on the cloud side; A model training module, used to perform model training processing on the edge-side lightweight model and the cloud-side large-volume model through the training set, perform parameter fusion update on the cloud-side large-volume model based on federated learning during the model training process, and adjust the loss function weight of the cloud-side large-volume model based on an adaptive reward mechanism to obtain a target edge model and a target cloud-side model after training; The collaborative decision-making module is used to perform collaborative decision processing on the target distribution network based on the target edge model and the target cloud-side model to obtain a collaborative decision result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images