Power dispatching literature privacy training and collaborative optimization method based on decentralized federated learning
By employing decentralized federated learning and differential privacy technology, collaborative training of power dispatch data across autonomous regions was achieved, solving the problems of data privacy leakage and data silos, and improving the efficiency and accuracy of the power dispatch system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA SOUTHERN POWER GRID COMPANY
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-28
AI Technical Summary
In the field of power dispatching, there are risks of data privacy leaks and data silos, which make it difficult to achieve cross-domain collaborative training and optimization, especially the inefficiency caused by data sharing and heterogeneity between autonomous domains.
By employing decentralized federated learning and differential privacy technology, cross-regional collaborative training is achieved through local training, privacy protection, model parameter uploading and aggregation, global model distribution and iterative training. Weighted average is used to merge model parameters to ensure data privacy and model accuracy.
It improves the data utilization efficiency and model accuracy of the power dispatching system, enhances the system's robustness and security, avoids the leakage of sensitive information, and solves the problem of data silos.
Smart Images
Figure CN121935949A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing and privacy protection technology in intelligent power dispatching systems, and in particular to a method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning. Background Technology
[0002] The field of power dispatching involves massive amounts of literature and data, including management systems, official documents, and reports, which play a crucial role in power dispatching decisions. However, this literature and data often contain sensitive information, such as dispatching plans, system load, and equipment status, thus posing significant privacy protection pressures during data sharing among autonomous systems (AS). Furthermore, the heterogeneity of power dispatching literature and data across different ASs, with substantial differences in data volume and characteristics, leads to data silos, hindering effective cross-domain collaborative training and optimization. Traditional methods face privacy risks in power dispatching data sharing, and the differences in data distribution and characteristics across different regions make effective collaborative optimization difficult. Therefore, it is necessary to propose a method for privacy-preserving training and collaborative optimization of power dispatching literature that enables cross-regional data collaborative training and optimization of power dispatching models. Summary of the Invention
[0003] Purpose of the invention: The purpose of this invention is to provide a method for privacy-preserving training and collaborative optimization of power dispatch literature based on decentralized federated learning. This method enables cross-regional collaborative training of data while ensuring data privacy, thereby optimizing the accuracy and robustness of the power dispatch model and improving the overall efficiency and intelligence level of the power dispatch system.
[0004] Technical solution: The power dispatching literature privacy training and collaborative optimization method based on decentralized federated learning described in this invention includes the following steps:
[0005] (1) Data acquisition and preprocessing;
[0006] (2) Local training and privacy protection;
[0007] (3) Model parameter uploading and aggregation;
[0008] (4) Global model distribution and local update;
[0009] (5) Iterative training and model convergence.
[0010] Further, step (2) local model training includes: the autonomous region uses local power dispatch literature data for model training, and each autonomous region adjusts the model parameters according to the characteristics of the literature data. It is assumed that the dataset for each autonomous region is... Autonomous system i trains a local model ,in These are the parameters of the autonomous system model.
[0011] Furthermore, step (2) differential privacy protection includes: after local model training, the autonomous region applies differential privacy technology to protect the trained model parameters, assuming the model gradient is... By adding Gaussian noise The perturbated model parameters are obtained as follows:
[0012]
[0013] Even if an external attacker obtains the perturbed model parameters, they will not be able to deduce the specific content of the local data.
[0014] Furthermore, step (3) local model upload includes uploading the locally trained model parameters to a decentralized storage platform.
[0015] Further, step (3) global model merging includes the global server receiving local model parameters from all autonomous systems, merging these parameters using a decentralized federated learning algorithm, and merging the model parameters of each autonomous system using a weighted average method, as shown in the formula:
[0016]
[0017] in, For the global model in the first Wheel parameters, It is the first The amount of data per autonomous region It represents the total number of all autonomous regions.
[0018] Furthermore, step (4) global model distribution includes distribution through a decentralized storage platform, where clients in each autonomous region download the latest global model parameters from the platform. Assuming the global model parameters are... Each autonomous system's client downloads this parameter and merges it with the local model parameters to update the local model.
[0019] Furthermore, the local model update process in step (4) is implemented using gradient descent, and the update formula is:
[0020]
[0021] in, These are the local model parameters updated in round t+1. These are the local model parameters for round t; For learning rate, To use local datasets The calculated gradient represents the error during local training. It is the influence factor of the global model on the local model update. These are global model parameters.
[0022] Furthermore, in step (5), the model training iteration involves multiple iterations, alternating between local training and global model merging. The autonomous region (AVM) adjusts its training based on the current global model parameters. The local model is trained and updated, and the updated parameters of the local model are uploaded. After each local training, all autonomous systems upload the updated model parameters to a decentralized storage platform for merging. In each round of local training, the loss function of the local model is... To minimize, the goal of each autonomous system is to reduce its local loss:
[0023]
[0024] in, It is the i-th autonomous region based on local data and model parameters The calculated loss function.
[0025] Furthermore, the following criteria are used to determine the convergence of the model in step (5):
[0026]
[0027] in, It is the loss of the global model in round t. It is a set convergence threshold. When the change in the model's loss is less than this threshold, the model is considered to have converged.
[0028] Furthermore, after convergence, step (5) yields the final global model. It is applied to power dispatching systems to provide optimal dispatching strategies.
[0029] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: By introducing decentralized federated learning and differential privacy technology, this invention solves the problems of data privacy leakage and data silos in the field of power dispatching; through a decentralized approach, autonomous regions can collaboratively train without sharing sensitive data, improving data utilization efficiency; simultaneously, the weighted merging algorithm addresses the data heterogeneity between different autonomous regions, ensuring that the global model accurately reflects the characteristics of each region, thereby improving the accuracy and robustness of the power dispatching system. Furthermore, the privacy protection mechanism during model training effectively avoids the leakage of sensitive information, ensuring data security and demonstrating significant security and reliability advantages. Attached Figure Description
[0030] Figure 1 This is a flowchart of the method described in this invention. Detailed Implementation
[0031] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0032] The decentralized federated learning-based method for privacy-preserving training and collaborative optimization of power dispatching literature, as described in this invention, effectively protects the privacy of literature data in the field of power dispatching through a decentralized training architecture and differential privacy technology. It also solves the data silo problem and achieves cross-autonomous region collaborative training and global model optimization. Specific steps are as follows: Figure 1 As shown:
[0033] (1) Data acquisition and preprocessing
[0034] (1.1) Data Acquisition:
[0035] Each power dispatching region (autonomous region) collects power dispatching documents and related data from its system. These documents include, but are not limited to, management regulations, official documents, and reports. These documents contain important and sensitive information regarding power grid operation, dispatching management, equipment status, and dispatching instructions. Because much of the content in these documents is confidential, it cannot be directly shared and requires protective measures to ensure data privacy.
[0036] (1.2) Data preprocessing:
[0037] To transform power dispatching literature into a format suitable for training, the literature content is first structured. Specific steps include:
[0038] Text cleaning: This involves removing unstructured parts of the text, such as redundant punctuation marks, spaces, irregular characters, or inconsistent formatting. This process ensures the consistency of the text data structure and provides a clean data source for subsequent analysis and processing.
[0039] Keyword extraction: Natural Language Processing (NLP) techniques are used to extract the core keywords of each document. For example, the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm is used to extract important information related to power dispatch based on the ratio of term frequency to inverse document frequency, helping the model quickly identify the core content and themes in the document.
[0040] Converting text data into numerical form facilitates model processing. BERT can be used to convert document content into word vectors, allowing for further numerical processing of the text and facilitating the application of subsequent machine learning algorithms.
[0041] The processed data (including text vectors and keywords) will be stored in a local database or a decentralized storage platform (such as IPFS) to ensure data security and traceability, and to prevent the leakage of sensitive data.
[0042] (2) Local training and privacy protection:
[0043] (2.1) Local model training:
[0044] Each autonomous region (AUR) uses local power dispatching literature data for model training. Each AUR adjusts its model parameters based on the characteristics of the literature data (such as content summaries and keyword frequencies). Assume each AUR's dataset is... Autonomous system i will train a local model ,in These are the parameters of the autonomous system model.
[0045] (2.2) Differential privacy protection:
[0046] To protect the privacy of power dispatching literature data, differential privacy techniques are applied to protect the trained model parameters after local model training. Specifically, assuming the model's gradient is... By adding Gaussian noise The perturbated model parameters are obtained as follows:
[0047]
[0048] This privacy protection mechanism ensures that even if external attackers obtain the perturbed model parameters, they cannot deduce the specific content of the local data, thus protecting the privacy and security of power dispatching literature data.
[0049] (3) Model parameter uploading and aggregation
[0050] (3.1) Uploading local models:
[0051] Clients in each autonomous system upload their locally trained model parameters (including gradients and weights) to a decentralized storage platform (such as IPFS). This decentralized storage platform ensures data security while avoiding the problems associated with relying on centralized servers.
[0052] (3.2) Global model merging:
[0053] The global server (decentralized learning platform) receives local model parameters from all autonomous systems (AS), and then merges these parameters using a decentralized federated learning algorithm. To address data heterogeneity between different ASs (such as differences in the number of data labels), a weighted average is used to merge the model parameters from each AS. The specific formula is as follows:
[0054]
[0055] in, For the global model in the first Wheel parameters, It is the first The amount of data per autonomous region This represents the total number of autonomous systems. The weighted average ensures that autonomous systems with larger datasets have a greater weight in the model merging process.
[0056] (4) Global model distribution and local update:
[0057] (4.1) Global model distribution:
[0058] The merged global model will be distributed through a decentralized storage platform (such as IPFS). Clients in each autonomous system download the latest global model parameters from the platform. Assume the global model parameters are... Each autonomous system client downloads this parameter and merges it with the local model parameters to update the local model.
[0059] (4.2) Local model update:
[0060] Each autonomous system (AAS) client fine-tunes its local model based on the downloaded global model. The goal of local model updates is to ensure that the local model parameters fully utilize information from the global optimization, guaranteeing that the global model's influence is reflected in the local training. The model update process can be implemented using gradient descent, with the update formula as follows:
[0061]
[0062] in, These are the local model parameters updated in round t+1. These are the local model parameters for round t. The learning rate controls the update step size. To use local datasets The calculated gradient represents the error during local training. It is the influence factor of the global model on the local model update. These are global model parameters, representing the global model after merging all autonomous domains in decentralized federated learning.
[0063] (5) Iterative training and model convergence:
[0064] (5.1) Model training iteration:
[0065] The training process involves multiple iterations, alternating between local training and global model merging. In each iteration, the autonomous region (AfD) adjusts its training based on the current global model parameters. The local model is trained and updated, and the updated parameters (such as gradients and weights) are uploaded. After each local training, all autonomous systems upload the updated model parameters to a decentralized storage platform (such as IPFS) for merging.
[0066] In each round of local training, the loss function of the local model... This will be minimized to better fit the local data. The goal for each autonomous system is to reduce its local loss:
[0067]
[0068] in, It is the i-th autonomous region based on local data and model parameters The calculated loss function.
[0069] (5.2) Model convergence and validation:
[0070] After multiple iterations, the system evaluates the performance of the global model to ensure its effectiveness in power dispatching tasks. Evaluation typically involves calculating the loss or accuracy of the global model on the validation set. If the loss does not change significantly, the model has reached convergence. Convergence can be determined using the following criteria:
[0071]
[0072] in, It is the loss of the global model in round t. It is a set convergence threshold. When the change in the model's loss is less than this threshold, the model is considered to have converged.
[0073] (5.3) Model Application:
[0074] After convergence, the final global model It will be applied to the power dispatching system to provide the optimal dispatching strategy.
Claims
1. A method for privacy-preserving training and collaborative optimization of power dispatch literature based on decentralized federated learning, characterized in that, Includes the following steps: (1) Data acquisition and preprocessing; (2) Local training and privacy protection; (3) Model parameter uploading and aggregation; (4) Global model distribution and local update; (5) Iterative training and model convergence.
2. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, Step (2) local model training includes: each autonomous region uses local power dispatch literature data for model training, and each autonomous region adjusts the model parameters according to the characteristics of the literature data. It is assumed that the dataset for each autonomous region is... Autonomous system i trains a local model ,in These are the parameters of the autonomous system model.
3. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, Step (2) differential privacy protection includes: after local model training, the autonomous region applies differential privacy technology to protect the trained model parameters, assuming the model gradient is... By adding Gaussian noise The perturbated model parameters are obtained as follows: ， Even if an external attacker obtains the perturbed model parameters, they will not be able to deduce the specific content of the local data.
4. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, Step (3) Local model upload includes uploading the locally trained model parameters to a decentralized storage platform.
5. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, Step (3) global model merging includes the global server receiving local model parameters from all autonomous systems, merging these parameters using a decentralized federated learning algorithm, and using a weighted average method to merge the model parameters of each autonomous system, as shown in the formula: ， in, For the global model in the first Wheel parameters, It is the first The amount of data per autonomous region It represents the total number of all autonomous regions.
6. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, Step (4) global model distribution includes distribution via a decentralized storage platform. Clients in each autonomous region download the latest global model parameters from the platform. Assume the global model parameters are... Each autonomous system's client downloads this parameter and merges it with the local model parameters to update the local model.
7. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, The local model update process in step (4) is implemented using gradient descent, and the update formula is: ， in, These are the local model parameters updated in round t+1. These are the local model parameters for round t; For learning rate, To use local datasets The calculated gradient represents the error during local training. It is the influence factor of the global model on the local model update. These are global model parameters.
8. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, The model training iteration in step (5) involves multiple iterations, alternating between local training and global model merging. The autonomous region (AVM) adjusts its training based on the current global model parameters. The local model is trained and updated, and the updated parameters of the local model are uploaded. After each local training, all autonomous systems upload the updated model parameters to a decentralized storage platform for merging. In each round of local training, the loss function of the local model is... To minimize, the goal of each autonomous system is to reduce its local loss: ， in, It is the i-th autonomous region based on local data and model parameters The calculated loss function.
9. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, The following criteria are used to determine the convergence of the model in step (5): ， in, It is the loss of the global model in round t. It is a set convergence threshold. When the change in the model's loss is less than this threshold, the model is considered to have converged.
10. The method for privacy-preserving training and collaborative optimization of power dispatching literature based on decentralized federated learning according to claim 1, characterized in that, After convergence, step (5) yields the final global model. It is applied to power dispatching systems to provide optimal dispatching strategies.