Virtual power plant cloud edge collaborative optimization method and system based on improved federated learning

Abstract

The invention discloses a virtual power plant cloud edge collaborative optimization method and system based on improved federal learning, and the method comprises the steps: a cloud node of a virtual power plant builds an initial model, and transmits model parameters to an edge node; the edge node carries out low-rank adaptive decomposition on the model parameters, local data training is utilized to obtain updated model parameters, and the updated model parameters are uploaded to the cloud node; and the cloud node performs weighted aggregation on the uploaded model parameters and issues the aggregated model parameters to the edge node. Based on the federated learning framework, local data of the edge node does not need to be massively transmitted to the cloud node, the data transmission amount is reduced, and the data privacy is protected; meanwhile, a low-rank adaptive decomposition matrix is adopted at an edge node to train and update model parameters, so that the data transmission quantity of cloud-edge collaboration is further reduced, and the security of a book transmission process is improved; the weighted aggregation mode of the model parameters at the cloud is improved, and the rationality of cloud node global model aggregation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of virtual power plants, and in particular to a cloud-edge collaborative optimization method and system for virtual power plants based on improved federated learning. Background Art

[0002] In the context of the construction of a new power system, virtual power plants (VPPs) have great practical needs as a key technology carrier and effective means to improve the energy efficiency management level of the power demand side, promote the consumption of new energy, and enhance the flexibility of system regulation. The existing VPP cloud-edge collaborative architecture is based on the Internet of Things. It uses advanced communication, measurement, control technology and software systems to aggregate one or more resources such as massive, heterogeneous, and decentralized renewable energy, controllable loads, and energy storage in the distribution network through the VPP cloud energy management system, participate in power auxiliary service transactions and power spot market transactions, and realize autonomous coordination and optimization control. The state information is mainly collected by edge node devices and uploaded to the cloud server. The algorithm model deployed in the cloud performs coordinated optimization or prediction, and the generated control strategy or prediction results are sent to each edge node. The large-scale transmission of edge node data leads to high communication costs, low security, and the computing power of edge nodes is not fully utilized.

[0003] The "Cloud-edge collaborative virtual power plant scheduling method, device and computer equipment" disclosed in the Chinese patent literature has a publication number of CN116207793A and a publication date of 2023-06-02. The virtual power plant control unit aggregates the operating data of the distributed power equipment to obtain the virtual power plant control unit operating data, and the virtual power plant edge cluster obtains the power adjustment limit parameters corresponding to the virtual power plant control unit. According to the operating data of the virtual power plant control unit and the power adjustment limit parameters, the power adjustment information of the virtual power plant control unit is determined, and the virtual power plant edge cluster forwards the power adjustment information to the virtual power plant control unit, so that the virtual power plant control unit adjusts the operating power of its corresponding distributed power equipment based on the power adjustment information. Compared with the traditional virtual power plant that cannot perform power scheduling on massive data, a system consisting of a cloud server, a virtual power plant edge cluster and a virtual power plant control unit is used to schedule the power of the virtual power plant. However, the large-scale transmission of virtual power plant edge cluster data in this technology will result in high communication costs and low security, and the cloud bears the main computing pressure, and the computing power of the virtual power plant edge cluster is not fully utilized. Summary of the invention

[0004] The present invention aims to overcome the problems in the prior art of large-scale transmission of edge node data resulting in high communication costs and low security, and the cloud side bears the main computing pressure while the computing power of edge nodes is not fully utilized. A virtual power plant cloud-edge collaborative optimization method and system based on improved federated learning is provided.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: The virtual power plant cloud-edge collaborative optimization method based on improved federated learning includes: The cloud node of the virtual power plant builds the initial model and sends the model parameters to the edge node; The edge node performs low-rank adaptive decomposition on the model parameters, uses local data training to obtain updated model parameters, and uploads them to the cloud node; The cloud node performs weighted aggregation on the uploaded model parameters and sends the aggregated model parameters to the edge node.

[0006] The present invention is based on a federated learning framework, so that the local data of edge nodes does not need to be transmitted to cloud nodes in large quantities, reducing the amount of data transmission and protecting data privacy, and performing model training at the edge nodes, so that the computing power of the edge nodes is effectively utilized; at the same time, a low-rank adaptive decomposition matrix is ​​used at the edge nodes to train and update the model parameters, further reducing the amount of data transmission in cloud-edge collaboration and improving the security of the book transmission process; the weighted aggregation method of model parameters at the cloud is improved, taking into account the number of samples of edge nodes and the size of model training loss, and improving the rationality of global model aggregation at cloud nodes.

[0007] Preferably, after the edge node obtains the updated model parameters through training, the parameter model is encrypted and uploaded to the cloud node; The cloud node decrypts the received encrypted parameter model and performs weighted aggregation, then re-encrypts and sends it to the edge node; the edge node decrypts and obtains the aggregated model parameters. (Partial homomorphic encryption, model parameters are homomorphically encrypted, loss function and edge node sample number are encrypted normally) Preferably, the edge node trains to obtain updated model parameters, calculates the optimization loss corresponding to the model parameters, and uploads the results to the cloud node; The cloud node uses the proportion of training samples of each edge node and the inverse of the optimization loss as weights to perform weighted aggregation on the model parameters uploaded by the edge nodes.

[0008] Preferably, the process of calculating the optimization loss corresponding to the model parameters includes: Calculate the loss value of model parameters for each sample of edge nodes; The loss values ​​of all samples of edge nodes are averaged to obtain the optimization loss corresponding to the model parameters.

[0009] Preferably, the process of the edge node performing low-rank adaptive decomposition on the model parameters comprises: Decompose the model parameters into pre-trained weights W that are not updated with training and fine-tuned incremental weights ΔW containing trainable parameters; represent the fine-tuned incremental weights ΔW by matrix multiplication of matrix B and matrix A; Where W and ΔW are d*d matrices, B and A are d*r and r*d matrices respectively, and r is less than d.

[0010] Preferably, when training starts after completing the low-rank adaptive decomposition, matrix A is initialized with a random Gaussian and matrix B is initialized with a collar.

[0011] Preferably, the edge node stores the model parameters obtained from the previous training; If the difference between the model parameters obtained after the current training and the model parameters obtained after the previous training is greater than the difference threshold, the model parameters obtained after the current training are uploaded to the cloud node for weighted aggregation; On the contrary, if it is not uploaded, the cloud node uses the model parameters uploaded by the edge node last time for weighted aggregation.

[0012] A virtual power plant cloud-edge collaborative optimization system based on improved federated learning, including a cloud node and several edge nodes connected to it; The cloud node establishes an initial model and sends the initial model parameters or the weighted aggregated model parameters to the edge node; the edge node uses local data training to update the model parameters and uploads them to the cloud node.

[0013] Preferably, the edge node comprises: The low-rank adaptive unit performs low-rank adaptive decomposition on the model parameters sent by the cloud node; The parameter training unit uses local data to train and update the model parameters after low-rank adaptive decomposition; The data upload unit processes the data during the training process and uploads the data that needs to be uploaded to the cloud node.

[0014] Preferably, the cloud node includes: A model building unit, building an initial model and generating initial model parameters; The weighted aggregation unit performs weighted aggregation on the model parameters uploaded by the edge nodes and re-sends them to the edge nodes.

[0015] The present invention has the following beneficial effects: based on the federated learning framework, the local data of the edge nodes does not need to be transmitted to the cloud nodes in large quantities, reducing the amount of data transmission and protecting data privacy; at the same time, a low-rank adaptive decomposition matrix is ​​used at the edge nodes to train and update the model parameters, further reducing the amount of data transmission in cloud-edge collaboration and improving the security of the book transmission process; the weighted aggregation method of the model parameters at the cloud is improved, thereby improving the rationality of the global model aggregation of the cloud nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the virtual power plant cloud-edge collaborative optimization method of the present invention.

[0017] Figure 2 It is a schematic diagram of the virtual power plant cloud-edge collaborative optimization system of the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0019] The existing VPP cloud-edge collaboration can be mainly divided into three layers of architecture: cloud, edge, and end in terms of hierarchical architecture, so as to realize vertical information interaction. Specifically, the terminal application is responsible for the aggregation of various distributed resources and access to the edge cluster, realizing the collection, monitoring, and control functions of the access equipment, and serving as the smallest control unit in the VPP architecture. As the control and management platform of the VPP terminal, the edge cluster has the functions of terminal information collection and reporting, control command decomposition and issuance, monitoring and alarm, etc. It is responsible for the secure access of various distributed resources, and has certain analysis and computing capabilities, and can be built-in to deploy control strategy algorithms and intelligent models. The cloud management and control platform is deployed on the cloud server and is the decision-making center of the VPP. It is responsible for collecting data from various clusters, deploying artificial intelligence, decision optimization and other models based on technologies such as big data and cloud computing, and accepting external interaction needs with the power market, dispatching center, etc., optimizing decisions internally, and generating the best control plan.

[0020] The existing VPP cloud-edge collaboration has the following three main problems: (1) The cost of data transmission is high. In actual operation, each terminal application generates a large amount of perception data. The communication cost of data transmission is huge, which puts great pressure on network channels and cloud resources. At the same time, the transmission of massive data will lead to an increase in data loss rate and an increase in abnormal data.

[0021] (2) High data transmission risk. During edge cluster data transmission, there is a risk of data attack or data leakage, which may lead to the theft of user privacy data or even grid operation parameters, which will have a serious impact on the stable operation of VPP.

[0022] (3) High pressure on cloud computing. The existing VPP cloud-edge collaborative architecture mainly uses cloud computing resources for global collaborative optimization. All kinds of advanced computing applications are undertaken by cloud computing power, and edge computing resources are not effectively utilized, resulting in a waste of computing resources and insufficient timeliness of computing analysis.

[0023] Based on the above problems, the present invention proposes a virtual power plant cloud-edge collaborative optimization method based on improved federated learning. The method introduces federated learning and a low-rank adaptive matrix to reduce the model parameters of cloud-edge collaboration, further reducing the data interaction cost and communication cost of VPP cloud-edge collaboration, improving the resource utilization efficiency of edge computing devices, and enhancing data security and privacy protection in the cloud-edge collaboration process.

[0024] like Figure 1 As shown in the figure, the virtual power plant cloud-edge collaborative optimization method based on improved federated learning includes: The cloud node of the virtual power plant builds the initial model and sends the model parameters to the edge node; The edge node performs low-rank adaptive decomposition on the model parameters, uses local data training to obtain updated model parameters, and uploads them to the cloud node; The cloud node performs weighted aggregation on the uploaded model parameters and sends the aggregated model parameters to the edge node.

[0025] It should be noted that the present invention is based on the federated learning framework, so that the local data of the edge nodes does not need to be transmitted to the cloud nodes in large quantities, reducing the amount of data transmission and protecting data privacy, and performing model training at the edge nodes, so that the computing power of the edge nodes is effectively utilized; at the same time, a low-rank adaptive decomposition matrix is ​​used at the edge nodes to train and update the model parameters, further reducing the amount of data transmission in cloud-edge collaboration and improving the security of the book transmission process; the weighted aggregation method of the model parameters at the cloud is improved, taking into account the number of samples of the edge nodes and the size of the model training loss, and improving the rationality of the global model aggregation of the cloud nodes.

[0026] It is worth noting that federated learning is a distributed machine learning framework with privacy protection and secure encryption technology, which aims to allow decentralized participants to collaborate on model training without disclosing private data to other participants. Applying the federated learning framework to the cloud-edge collaboration of virtual power plants, the cloud server or cloud management and control platform of the virtual power plant can be used as a cloud node, and several edge clients of the virtual power plant can be used as edge nodes, thus forming a cloud-edge collaboration system under the federated learning framework.

[0027] During the entire federated learning process of the virtual power plant cloud-edge collaboration, no local data of any edge node is transmitted to the cloud node, which can protect the data privacy of the edge node. At the same time, during the entire VPP cloud-edge collaborative interaction process, only the model structure and model parameters are sent and uploaded. Compared with the massive measurement data sent by the existing VPP cloud-edge collaboration, the data communication cost is greatly reduced. The core concept of federated learning is "data is available but not visible, and the model runs without data". While ensuring that the data does not leave the edge node, it takes into account the advantages of distributed computing and joint computing. Each edge node only needs to perform model training locally, the data volume does not increase, the computing cost pressure is small, and the computing power of the edge node is fully utilized.

[0028] Furthermore, after the edge node obtains the updated model parameters through training, it encrypts the parameter model and uploads it to the cloud node; The cloud node decrypts and weighted aggregates the received encrypted parameter model, then re-encrypts it and sends it to the edge node; the edge node decrypts it and obtains the aggregated model parameters.

[0029] It should be noted that in order to protect the security during data transmission, data can also be encrypted during data transmission between edge nodes and cloud nodes. The specific process is: the cloud node of the virtual power plant establishes an initial model, and encrypts the model parameters and sends them to the edge node; the edge node decrypts the encrypted model parameters from the cloud node, performs low-rank adaptive decomposition on the decrypted model parameters, and uses local data training to obtain updated model parameters, and then encrypts the updated model parameters and uploads them to the cloud node; the cloud node decrypts the uploaded encrypted model parameters, performs weighted aggregation on the decrypted model parameters, and encrypts the aggregated model parameters again and sends them to the edge node.

[0030] It is worth noting that in order to reduce the computational complexity of frequent encryption and decryption in cloud nodes, the model parameters can be homomorphically encrypted at the edge nodes. Then, there is no need to decrypt the model parameters at the cloud nodes. The homomorphically encrypted model parameters can be directly aggregated and then transmitted back to the edge nodes for decryption. The decrypted model parameters can then be updated again.

[0031] For the aggregation method of the commonly used federated average algorithm in existing federated learning, since it directly calculates the average value of the model parameters of each edge node, the additive homomorphic encryption in the partial homomorphic encryption can be directly used; and for the weighted aggregation method that considers the sample proportion and optimization loss of the edge nodes as weights in this application, the additive homomorphic encryption is still used for the model parameters, while the number of samples and the optimization loss of the edge nodes can be encrypted by conventional encryption means; when transmitted to the cloud node, the encrypted number of samples and the optimization loss are decrypted, and the model parameters do not need to be decrypted and weighted aggregation is performed directly, and then the results are sent to the edge node for decryption.

[0032] As a specific embodiment, the edge node trains to obtain updated model parameters, calculates the optimization loss corresponding to the model parameters, and uploads the results to the cloud node; The cloud node uses the proportion of training samples of each edge node and the inverse of the optimization loss as weights to perform weighted aggregation on the model parameters uploaded by the edge nodes.

[0033] It should be noted that after the cloud node sends the model to the edge node for training, the latest model parameters generated by each edge node based on local data can be obtained. However, how to perform weight aggregation on the VPP cloud node based on the model parameters of each node has become a new problem. The existing federated learning framework usually adopts the Federated Averaging (FedAvg) algorithm. FedAvg calculates the average value of all uploaded model parameters and sends this average value as the global model parameter for the next iteration to all edge nodes or performs model inference. Although this method is simple and easy to understand, FedAvg does not consider the model training effect of each edge node, but only considers the number of samples of the edge node. If the loss function of a node is too large, the model training performance is poor, but its sample number accounts for a high proportion, which will cause the parameters of the global model to be biased towards this node, resulting in poor performance of the aggregated model. To address this problem, an improved dynamic weighted model aggregation method that considers the training loss of edge node models is introduced.

[0034] Specifically, the process of calculating the optimization loss corresponding to the model parameters includes: Calculate the loss value of model parameters for each sample of edge nodes; The loss values ​​of all samples of edge nodes are averaged to obtain the optimization loss corresponding to the model parameters.

[0035] Assume that the number of edge nodes participating is N, the sum of the number of samples participating in the training of each edge node is S, and the number of samples of edge node k is D k , and its optimization loss function is defined as L(w k ), where w k is the model parameter of edge node k, Where L i (w) is the model parameter w k The loss value of the i-th sample of edge node k, that is, the optimization loss corresponding to the model parameter, is the average of the loss values ​​of each sample of the edge node.

[0036] During aggregation, the optimization loss of each edge node and the sample ratio of each edge node are considered. The calculation method of the model parameters after aggregation is: The product of the sample proportion of edge node k and the inverse of the optimization loss is used as the weight of the model parameters of the edge node k for weighted aggregation. When the optimization loss of the edge node is too large, the weight of the corresponding model parameter can be reduced, thereby avoiding the shortcomings of using the federated average algorithm for aggregation.

[0037] As an optional embodiment, the edge node stores the model parameters obtained from the previous training; If the difference between the model parameters obtained after the current training and the model parameters obtained after the previous training is greater than the difference threshold, the model parameters obtained after the current training are uploaded to the cloud node for weighted aggregation; On the contrary, if it is not uploaded, the cloud node uses the model parameters uploaded by the edge node last time for weighted aggregation.

[0038] It should be noted that as the cloud nodes and edge nodes continuously interact and train to update the model parameters, after the model accuracy reaches a certain level, the model parameters uploaded by some edge nodes after training and updating are not of much help to the global model update. This is to save the amount of data uploaded by the edge nodes. This can prevent these edge nodes from uploading the updated model parameters, and the cloud nodes can directly reuse the model parameters uploaded last time (this can be done if the number of samples of the edge nodes remains unchanged; if the number of samples of the edge nodes changes, it is necessary to upload the updated model parameters and the corresponding number of samples and optimization loss).

[0039] It is worth noting that the edge node saves the model parameters obtained from the previous training, and also trains and updates the aggregated model parameters currently received from the cloud node to obtain new model parameters. The norm of the difference between the previous and subsequent model parameters is used as the difference and difference threshold for comparison. If the difference is greater than the difference threshold, it means that the change of the model parameters cannot be ignored during aggregation, and the current trained model parameters need to be uploaded to the cloud node for weighted aggregation (while updating the model parameters saved by the edge node itself). If the difference is less than or equal to the difference threshold, it means that the change of the model parameters is not very helpful for the update of the global model, so there is no need to upload the model parameters (nor to update the model parameters saved by itself). The edge node can send information to the cloud node indicating the use of the last model parameters to save the amount of uploaded data. The setting of the difference threshold can be a fixed value, or it can be the result of integrating and summarizing the model parameters updated multiple times with the edge node.

[0040] If represents the model parameters of the edge node k after current training, represents the model parameters of the previous training of edge node k, then the difference threshold τ can be expressed as: Where N represents the number of edge nodes; m represents the number of interactive transmissions between cloud nodes and edge nodes, that is, how many times the cloud nodes send model parameters to edge nodes; ω represents the confidence of the threshold; σ is used to balance the noise generated by the update of model parameters.

[0041] The storage unit at the cloud node stores the model parameters, optimization loss and number of samples obtained from the previous training of the edge node. When receiving information uploaded by an edge node about the use of the previous model parameters, the data uploaded by the edge node last time is retrieved for weighted aggregation, and the new model parameters, optimization loss and number of samples uploaded by the edge node are saved and updated in the storage unit.

[0042] As a specific embodiment, the process of the edge node performing low-rank adaptive decomposition of the model parameters includes: decomposing the model parameters into pre-trained weights W that are not updated with training and fine-tuning incremental weights ΔW containing trainable parameters; representing the fine-tuning incremental weights ΔW by matrix multiplication of matrix B and matrix A; Where W and ΔW are d*d matrices, B and A are d*r and r*d matrices respectively, and r is less than d.

[0043] When training starts after completing the low-rank adaptive decomposition, matrix A is initialized with a random Gaussian and matrix B is initialized with a collar.

[0044] It should be noted that although in the federated learning framework, edge node data no longer needs to be uploaded to the cloud server node, it only needs to receive the model parameters sent from the cloud to train at the local edge node and upload the trained model parameters for cloud aggregation update. However, with the continuous deepening of research on artificial intelligence and scheduling collaborative optimization technology, the amount of communication data containing only model parameters has also shown explosive growth. In order to further alleviate the communication pressure of VPP cloud-edge interaction and reduce the computing pressure of edge nodes, a federated learning framework is further improved based on low-rank adaptive decomposition.

[0045] Existing cloud-edge collaboration based on federated learning usually updates the global parameters of the pre-trained model, that is, fine-tuning the whole amount, which increases the training cost and cloud-edge communication cost. Federated learning based on low-rank adaptive decomposition does not need to update every parameter of the model. It introduces a low-rank adaptive matrix to approximate fine-tune the weights to achieve local model parameter update.

[0046] Full parameter update fine-tuning usually divides the model parameters into two parts, the pre-trained weight W and the fine-tuning incremental weight ΔW. For the input feature vector x, the output h is: h=Wx+ΔWx where W∈R d×d , ΔW∈R d×d ; For low-rank adaptive decomposition, matrix A and matrix B are introduced to approximate ΔW, that is, ΔW = BA; where B∈R d×r , A∈R r×d , and r is less than d; calculation shows that the model fine-tuning parameter quantity is directly reduced from d*d to 2*r*d, and the dimension of the output data is not changed. Therefore, according to the low-rank adaptive decomposition, the output h at this time is: h=Wx+BA.

[0047] For the matrix A, use a random Gaussian to initialize A∈N(0,σ 2 ), we use zero initialization B=0 for the matrix B, so ΔW=BA is zero at the beginning of training. We then scale ΔWx by α / r; where α is a constant with respect to r, and when optimizing with Adam, adjusting α is roughly the same as adjusting the learning rate if we scale the initialization appropriately. Therefore, we simply set α to the first r we try without adjusting it. This scaling helps reduce the need to re-tune hyperparameters when changing r.

[0048] The virtual power plant cloud-edge collaborative optimization method of the present invention is provided to improve the virtual power plant cloud-edge collaborative system. Figure 2 The virtual power plant cloud-edge collaborative optimization system based on improved federated learning shown includes a cloud node and several edge nodes connected thereto; The cloud node builds the initial model and sends the initial model parameters or the weighted aggregated model parameters to the edge node; The edge nodes use local data training to update model parameters and upload them to the cloud nodes.

[0049] Edge nodes include: The low-rank adaptive unit performs low-rank adaptive decomposition on the model parameters sent by the cloud node; The parameter training unit uses local data to train and update the model parameters after low-rank adaptive decomposition; The data upload unit processes the data during the training process and uploads the data that needs to be uploaded to the cloud node; The edge node storage unit saves the model parameters, optimization loss and number of samples obtained from the previous training of the edge node.

[0050] Cloud nodes include: A model building unit, building an initial model and generating initial model parameters; The weighted aggregation unit performs weighted aggregation on the model parameters uploaded by the edge nodes and re-sends them to the edge nodes; The cloud node storage unit saves the model parameters, optimization loss and number of samples obtained from the previous training uploaded by all edge nodes.

[0051] The computing model of the traditional virtual power plant cloud-edge interactive architecture is placed on the cloud management and control platform, and massive measurement data needs to be uploaded to the cloud. The present invention proposes a virtual power plant cloud-edge collaborative optimization method based on federated learning, which can send the cloud model to the computing device of the edge node without uploading measurement data, reducing communication pressure, enhancing system security, and effectively utilizing the computing power of the edge node device. The federated learning method based on low-rank adaptive technology does not need to update every parameter of the model sent from the cloud. By introducing a low-rank adaptive matrix to approximate the model parameters, the local model and the global model parameters can be updated, which greatly reduces the number of parameters uploaded and sent by the model. The improved dynamic weighted model aggregation technology takes into account the number of samples of the edge nodes and the size of the model training loss, and improves the rationality and accuracy of the global model aggregation of the virtual power plant cloud-edge interaction.

[0052] The above embodiments are further elaborations and illustrations of the present invention for ease of understanding, and are not limitations of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A virtual power plant cloud-edge collaborative optimization method based on improved federated learning, characterized in that: include: The cloud node of the virtual power plant builds the initial model and sends the model parameters to the edge node; The edge node performs low-rank adaptive decomposition on the model parameters, uses local data training to obtain updated model parameters, and uploads them to the cloud node; The cloud node performs weighted aggregation on the uploaded model parameters and sends the aggregated model parameters to the edge node.

2. The virtual power plant cloud-edge collaborative optimization method based on improved federated learning according to claim 1 is characterized in that: After the edge node obtains the updated model parameters through training, it encrypts the parameter model and uploads it to the cloud node; The cloud node decrypts and weighted aggregates the received encrypted parameter model, then re-encrypts it and sends it to the edge node; the edge node decrypts it and obtains the aggregated model parameters.

3. The virtual power plant cloud-edge collaborative optimization method based on improved federated learning according to claim 1 or 2, characterized in that: The edge node training obtains updated model parameters, calculates the optimization loss corresponding to the model parameters, and uploads them to the cloud node; the cloud node uses the training sample ratio of each edge node and the inverse of the optimization loss as weights to perform weighted aggregation on the model parameters uploaded by the edge node.

4. The virtual power plant cloud-edge collaborative optimization method based on improved federated learning according to claim 3 is characterized in that: The process of calculating the optimization loss corresponding to the model parameters includes: Calculate the loss value of model parameters for each sample of edge nodes; The loss values ​​of all samples of edge nodes are averaged to obtain the optimization loss corresponding to the model parameters.

5. The virtual power plant cloud-edge collaborative optimization method based on improved federated learning according to claim 1, 2 or 4, characterized in that: The process of the edge node performing low-rank adaptive decomposition on the model parameters includes: Decompose the model parameters into pre-trained weights W that are not updated with training and fine-tuned incremental weights ΔW that contain trainable parameters; The fine-tuning incremental weight ΔW is represented by the matrix multiplication of the matrix B and the matrix A; Where W and ΔW are d*d matrices, B and A are d*r and r*d matrices respectively, and r is less than d.

6. The virtual power plant cloud-edge collaborative optimization method based on improved federated learning according to claim 5 is characterized in that: When training starts after completing the low-rank adaptive decomposition, matrix A is initialized with a random Gaussian and matrix B is initialized with a collar.

7. The virtual power plant cloud-edge collaborative optimization method based on improved federated learning according to claim 3 is characterized in that: The edge node stores the model parameters obtained from the previous training; If the difference between the model parameters obtained after the current training and the model parameters obtained after the previous training is greater than the difference threshold, the model parameters obtained after the current training are uploaded to the cloud node for weighted aggregation; On the contrary, if it is not uploaded, the cloud node uses the model parameters uploaded by the edge node last time for weighted aggregation.

8. A virtual power plant cloud-edge collaborative optimization system based on improved federated learning, applicable to the virtual power plant cloud-edge collaborative optimization method according to any one of claims 1 to 7, characterized in that: Includes cloud nodes and several edge nodes connected to them; The cloud node establishes an initial model and sends the initial model parameters or the weighted aggregated model parameters to the edge node; The edge node uses local data training to update model parameters and uploads them to the cloud node.

9. The virtual power plant cloud-edge collaborative optimization system based on improved federated learning according to claim 8 is characterized in that: The edge node includes: The low-rank adaptive unit performs low-rank adaptive decomposition on the model parameters sent by the cloud node; The parameter training unit uses local data to train and update the model parameters after low-rank adaptive decomposition; The data upload unit processes the data during the training process and uploads the data that needs to be uploaded to the cloud node.

10. The virtual power plant cloud-edge collaborative optimization system based on improved federated learning according to claim 8 or 9, characterized in that: The cloud node includes: A model building unit, building an initial model and generating initial model parameters; The weighted aggregation unit performs weighted aggregation on the model parameters uploaded by the edge nodes and re-sends them to the edge nodes.

Images