A virtual power plant collaborative scheduling system and method based on federated learning and digital twinning

Abstract

The application relates to a virtual power plant collaborative scheduling system and method based on federated learning and digital twinning, and belongs to the technical field of virtual power plants. Technical scheme: a high-fidelity digital twin of local equipment and operation model is constructed for each data-sensitive participant in the virtual power plant; local model training and data simulation are performed in the respective twin; a federated learning framework is adopted, so that the participants exchange only encrypted model parameters or gradients without sharing original local data, and a globally optimized collaborative scheduling decision model is generated by aggregation in the cloud or a designated coordinator; finally, the global model is used to drive the overall digital twin of the virtual power plant, and efficient, safe and self-adaptive collaborative operation optimization and real-time scheduling are performed. The application effectively protects data privacy and security, promotes extensive access to resources, realizes the balance between individualization and generalization, and improves the precision and adaptability of the scheduling model.

Description

Technical Field

[0001] This invention relates to a virtual power plant collaborative scheduling system and method based on federated learning and digital twins, belonging to the field of collaborative scheduling technology for internal resources of virtual power plants. Background Technology

[0002] To achieve the "dual carbon" goal, distributed energy sources, represented by photovoltaics and wind power, are developing rapidly. However, these resources are intermittent, volatile, and geographically dispersed, posing challenges to the stable operation of the power grid. "Virtual power plant" technology has emerged to address this, intelligently aggregating these scattered resources to provide stable and controllable power services similar to traditional power plants. One of the core functions of a virtual power plant is to achieve coordinated scheduling of internal resources. This typically relies on powerful data analysis and prediction models, such as load forecasting, renewable energy output forecasting, and energy storage charging and discharging strategy models. The traditional approach is to centrally upload the operational data of all participants to a central server for unified modeling. However, this faces two major challenges: First, data privacy and security issues. Energy consumption data from industrial and commercial users and residents, as well as core operational data of distributed power stations (such as battery health status), are sensitive information. Due to trade secrets, regulations (such as GDPR and data security laws), or user wishes, they are often unwilling or unable to share raw data, forming "data silos." Second, data heterogeneity and model generalization. The equipment models, operating environments, and energy consumption habits of different participants vary greatly, resulting in highly heterogeneous data collected centrally. A single model trained directly is difficult to adapt to all individuals, limiting the effectiveness of global optimization. Furthermore, with the addition of new members, the model needs frequent retraining, leading to poor scalability. Meanwhile, digital twin technology creates a computable and simulable virtual copy of each physical entity, making it possible to conduct extensive strategy deductions and "hypothesis analyses" in virtual space without interfering with physical devices.

[0003] In the prior art, the closest solution to this invention is "a virtual power plant optimization scheduling system based on centralized digital twins." This solution typically includes the following components: Centralized data acquisition module: Real-time acquisition of all data such as power, voltage, energy, and status of all participating members (distributed photovoltaic, energy storage, controllable load) in the virtual power plant through communication networks (such as 5G, fiber optics), and aggregation to the virtual power plant main station (central server); Centralized Digital Twin Building Module: On the main server, a digital twin of the entire virtual power plant aggregation system or key equipment is built using centralized full data. This twin serves as a simulation platform, reflecting the real-time status and topology of the virtual power plant. Centralized Optimized Scheduling Module: Based on the collected centralized data and twin model, a centralized optimized scheduling algorithm (such as mixed integer programming or model predictive control) is run. This algorithm calculates the optimal output / load curve for each member within a future scheduling cycle (e.g., 15 minutes, 1 hour) based on grid dispatch instructions and market price signals. Command issuance module: Decomposes the calculated optimal command and issues it to the local controller of each distributed energy source for execution.

[0004] While the aforementioned existing technologies employ digital twins for scheduling optimization, their centralized data and model architecture leads to the following inherent drawbacks: Drawback 1: High risk of data privacy breaches, limiting participation. This approach requires all members to share raw operational data. For industrial users, this data may contain production processes and trade secrets; for residential users, it involves personal privacy. Direct centralized data upload carries an extremely high risk of privacy breaches. Therefore, owners of many potentially high-quality resources (such as high-value adjustable loads) are unwilling to join due to security concerns, limiting the scale and flexibility of the virtual power plant. Drawback 2: Communication and computing loads are centralized, posing a single point of failure risk. All data needs to be uploaded to the central server, requiring high communication bandwidth. When the data volume is huge, the central server faces immense computational and storage pressure. Furthermore, if the central server fails, the entire virtual power plant's optimization and scheduling function will be paralyzed, resulting in poor system robustness. Drawback 3: The model is sensitive to data heterogeneity, making it difficult to balance personalization and generalization capabilities. The centralized optimization model attempts to fit all the highly diverse members with a single "general model." This can easily lead to poor model performance on specific individuals (insufficient personalization), or sacrificing the optimization potential of some members in pursuit of overall performance. When new members are added or equipment is updated, the centralized model needs to be retrained globally, resulting in poor adaptability and high update costs. Disadvantage 4: Difficulty in handling local constraints of members and rapidly changing operating conditions. Although the centralized model can handle global constraints (such as the total power limit), it often struggles to deeply and in real-time understand the complex local physical constraints of each member (such as energy storage charge and discharge rate limits, battery aging models, production line process constraints) and rapidly changing internal states. This may cause the issued scheduling instructions to be infeasible locally or damage the equipment, requiring frequent uplink and downlink communication for correction, thus reducing scheduling efficiency.

[0005] Federated learning, as a novel artificial intelligence technology where "data is available but not visible," offers a solution to the aforementioned data privacy and collaborative modeling issues. It allows multiple data owners to collaboratively build and share a more powerful machine learning model without exchanging data. Summary of the Invention

[0006] This invention proposes a virtual power plant collaborative scheduling system and method based on federated learning and digital twins. It effectively protects data privacy and security, promotes widespread resource access, achieves a balance between personalization and generalization, improves the accuracy and adaptability of the scheduling model, reduces communication and central computing pressure, enhances system robustness and scalability, and achieves efficient, closed-loop, and adaptive collaborative optimization. It solves the problems of data silos, privacy and security, and insufficient model generalization ability faced by virtual power plants when aggregating massive, heterogeneous, and distributed energy sources (such as photovoltaic, wind power, energy storage, and flexible loads) to participate in grid collaborative scheduling.

[0007] The technical solution of this invention is: A collaborative scheduling method for virtual power plants based on federated learning and digital twins is proposed. This method constructs high-fidelity digital twins of the local equipment and operational models of each data-sensitive participant in the virtual power plant. Localized model training and data simulation are performed within each twin. A federated learning framework is employed, enabling participants to exchange encrypted model parameters or gradients without sharing original local data. These parameters are then aggregated in the cloud or by a designated coordinator to generate a globally optimized collaborative scheduling decision model. Finally, this global model is used to drive the overall digital twin of the virtual power plant, enabling efficient, secure, and adaptive collaborative operation optimization and real-time scheduling.

[0008] The participating entities include industrial and commercial users and community energy storage users.

[0009] The method includes the following steps: S101, Constructing Local Digital Twins: Establish a high-fidelity local digital twin for each participating member in the virtual power plant, corresponding to its physical entity; each local twin runs within the member's local edge computing device or security domain; the twin integrates the member's physical device model, historical operating data, real-time sensor data, and local specific operating constraint rules; S102, Local Model Training and Data Augmentation: In each member's local digital twin environment, using its private historical operating data and twin simulation capabilities, the local scheduling strategy-related models are independently trained or fine-tuned, and the training data does not leave the local security boundary. S103, Federated aggregation of model parameters: a) Parameter upload: After each member completes model training or update locally, they upload the encrypted gradient or model parameters (not the original data) of the model to the virtual power plant main station or the server acting as the coordinator through a secure channel. b) Secure aggregation: After receiving the encrypted model parameters from all participating members, the main server uses a secure aggregation algorithm to aggregate and calculate these parameters, generating an updated global collaborative scheduling model; c) Parameter distribution: Distribute the aggregated global model parameters to all participating members; S104, Updating Local Models and Collaborative Inference: After receiving the global model parameters, each member merges them with its own existing model locally or directly replaces them, allowing the local model to absorb the "knowledge" of other members without exposing local data. Then, using the updated local model, it drives its local digital twin to perform collaborative scheduling strategy inference for the next scheduling cycle in the twin environment. During the inference, each local twin considers its own goals and indirectly "perceives" the global goals through the model. S105, Global Collaborative Decision Generation and Execution: a) Decision generation: The master station generates a global scheduling target based on the power grid scheduling needs or market information and broadcasts it to all members; each member calculates and feeds back a privacy-preserving, optimal local feasible scheduling solution commitment to the master station based on the updated model in S104 and the inference results in its own twin. b) Global Coordination and Confirmation: The main station gathers the commitments of all members, conducts rapid coordination and verification, and finally confirms a feasible global scheduling plan; the final scheduling instructions, broken down to each member, are then issued. c) Instruction execution and closed-loop feedback: After receiving instructions, each member executes them through the local controller; at the same time, real data during the execution process is fed back to the local digital twin in real time to calibrate the model and update the twin's state, forming a closed loop of "physical execution - digital mapping"; this closed loop also serves as the source of training data for the next round of federated learning.

[0010] Steps S101 to S105 constitute an iterative closed loop; step S101, local digital twin, provides a secure and private local simulation and training environment; steps S102-S103-S104, federated learning, are the key bridge connecting various privacy silos and achieving knowledge sharing; the final step S105, collaborative decision-making and execution, achieves the scheduling goal of the virtual power plant; the entire process realizes collaborative scheduling of "data not moving, model moving, privacy not leaked, and knowledge sharing".

[0011] The participants in the virtual power plant in step S101 include the flexible load of factory A, the energy storage power station of community B, and the photovoltaic system of building C.

[0012] In step S102, local model training and data augmentation involve training a local "output / load-benefit" model and a state prediction model (such as short-term power prediction), or simulating and generating "hypothetical" data under different scheduling strategies in a twin for data augmentation. Step S103 employs a secure aggregation algorithm that includes the FedAvg algorithm, which utilizes homomorphic encryption or secure multi-party computation techniques.

[0013] In the S104 simulation, the self-objective includes minimizing electricity costs, and the global objective is to minimize the virtual power plant tracking and dispatch instructions and the total cost.

[0014] The S105 generates a global scheduling target that includes the total adjustment power curve; the optimal local feasible scheduling scheme promises the range of power that can be increased / decreased and the cost promised to the master station during a certain period; the rapid coordination and verification of the quotation checks whether the total power meets the requirements.

[0015] A virtual power plant collaborative scheduling system based on federated learning and digital twins is disclosed. The system includes: a virtual power plant master station server, multiple distributed edge nodes, and a communication network; the edge nodes are equipped with local digital twin modules and model training modules; the virtual power plant master station server is equipped with a federated aggregation module, a global coordination module, and a communication encryption module to implement the above methods.

[0016] The innovation of this invention lies in the collaborative architecture driven by the dual engines of "local digital twin + federated learning": data privacy and local fine-grained modeling are guaranteed by building a local digital twin for each member, and knowledge fusion and collaborative optimization of the global model are achieved through federated learning without sharing data.

[0017] Localized, privacy-preserving model training and data augmentation in digital twins.

[0018] "Model parameters / gradients" serve as a secure medium to replace the original data transmission, enabling the construction of a collaborative scheduling model where "data is available but not visible".

[0019] Based on the updated model, a pre-scheduling co-inference mechanism in a twin environment improves the feasibility and economy of the final global scheduling instructions.

[0020] Compared with the prior art, the present invention has the following beneficial effects: Advantage 1: Effectively protects data privacy and security, and promotes widespread resource access. This invention requires members to upload only model parameters or gradients, rather than raw runtime data (S103), and allows for encrypted transmission and aggregation technologies. The direct technical effect of this is to fundamentally avoid the risk of sensitive raw data leakage. The ultimate result is that it alleviates the privacy and security concerns of industrial, commercial, and residential users, increasing their willingness to participate in virtual power plants. This enables virtual power plants to aggregate a wider range of diverse distributed energy resources, enhancing their dispatch potential and market competitiveness.

[0021] Advantage 2: Achieving a balance between personalization and generalization, improving the accuracy and adaptability of the scheduling model. This invention trains the model within a local digital twin (S102), enabling the model to fully learn local data patterns and physical constraints. Simultaneously, through federated learning and shared aggregation of global parameters (S103), each local model can integrate the "experience" of other members to address common challenges. Thus, the final model formed at each edge node is a hybrid model possessing both personalized adaptation and global generalization capabilities (S104). The direct effect is that during collaborative scheduling simulations, the model can more accurately describe and optimize the behavior of each specific member while still meeting the overall objective, thereby improving the accuracy and executability of the overall scheduling plan and demonstrating good adaptability to new members and devices.

[0022] Advantage 3: Reduced communication and central computing pressure, improving system robustness and scalability. Most data processing and model training computations are offloaded to the local digital twins of the edge nodes (S101, S102), with only a small number of encrypted model parameters needing to be transmitted between the main station and the edge (S103). This directly reduces the computational load on the central server and the bandwidth consumption of the backbone communication network. Therefore, the system can support larger-scale virtual power plants, and when the main station or some communication links fail, each edge node can still maintain basic optimal operation locally based on existing optimized models, significantly enhancing the overall robustness of the system.

[0023] Advantage 4: Achieves efficient, closed-loop, and adaptive collaborative optimization. This invention utilizes a local digital twin as a "sandbox," enabling extensive low-risk, low-cost collaborative simulations before physical execution (S104). Combined with a model updated through federated learning, the simulation results more closely approximate physical reality. From simulation to global decision generation (S105a-b), and then to physical execution and real-time feedback (S105c), a highly efficient "simulation-optimization-execution-feedback" closed loop is formed. Real-world feedback data can then be used for the next round of federated learning, continuously optimizing the model. Therefore, the system possesses self-iterative and self-optimizing capabilities, continuously adapting to changes in the internal and external environment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] A collaborative scheduling method for virtual power plants based on federated learning and digital twins is proposed. This method constructs high-fidelity digital twins of the local equipment and operational models of each data-sensitive participant in the virtual power plant. Localized model training and data simulation are performed within each twin. A federated learning framework is employed, enabling participants to exchange encrypted model parameters or gradients without sharing original local data. These parameters are then aggregated in the cloud or by a designated coordinator to generate a globally optimized collaborative scheduling decision model. Finally, this global model is used to drive the overall digital twin of the virtual power plant, enabling efficient, secure, and adaptive collaborative operation optimization and real-time scheduling.

[0026] The participating entities include industrial and commercial users and community energy storage users.

[0027] The method includes the following steps: S101, Constructing Local Digital Twins: Establish a high-fidelity local digital twin for each participating member in the virtual power plant, corresponding to its physical entity; each local twin runs within the member's local edge computing device or security domain; the twin integrates the member's physical device model, historical operating data, real-time sensor data, and local specific operating constraint rules; S102, Local Model Training and Data Augmentation: In each member's local digital twin environment, using its private historical operational data and twin simulation capabilities, independently train or fine-tune the local scheduling strategy-related model, with the training data remaining within the local security boundary;

[0028] S103, Federated aggregation of model parameters: a) Parameter upload: After each member completes model training or update locally, they upload the encrypted gradient or model parameters (not the original data) of the model to the virtual power plant main station or the server acting as the coordinator through a secure channel. b) Secure aggregation: After receiving the encrypted model parameters from all participating members, the main server uses a secure aggregation algorithm to aggregate and calculate these parameters, generating an updated global collaborative scheduling model; c) Parameter distribution: Distribute the aggregated global model parameters to all participating members; S104, Updating Local Models and Collaborative Inference: After receiving the global model parameters, each member merges them with its own existing model locally or directly replaces them, allowing the local model to absorb the "knowledge" of other members without exposing local data. Then, using the updated local model, it drives its local digital twin to perform collaborative scheduling strategy inference for the next scheduling cycle in the twin environment. During the inference, each local twin considers its own goals and indirectly "perceives" the global goals through the model. S105, Global Collaborative Decision Generation and Execution: a) Decision generation: The master station generates a global scheduling target based on the power grid scheduling needs or market information and broadcasts it to all members; each member calculates and feeds back a privacy-preserving, optimal local feasible scheduling solution commitment to the master station based on the updated model in S104 and the inference results in its own twin. b) Global Coordination and Confirmation: The main station gathers the commitments of all members, conducts rapid coordination and verification, and finally confirms a feasible global scheduling plan; the final scheduling instructions, broken down to each member, are then issued. c) Instruction execution and closed-loop feedback: After receiving instructions, each member executes them through the local controller; at the same time, real data during the execution process is fed back to the local digital twin in real time to calibrate the model and update the twin's state, forming a closed loop of "physical execution - digital mapping"; this closed loop also serves as the source of training data for the next round of federated learning.

[0029] Steps S101 to S105 constitute an iterative closed loop; step S101, local digital twin, provides a secure and private local simulation and training environment; steps S102-S103-S104, federated learning, are the key bridge connecting various privacy silos and achieving knowledge sharing; the final step S105, collaborative decision-making and execution, achieves the scheduling goal of the virtual power plant; the entire process realizes collaborative scheduling of "data not moving, model moving, privacy not leaked, and knowledge sharing".

[0030] The participants in the virtual power plant in step S101 include the flexible load of factory A, the energy storage power station of community B, and the photovoltaic system of building C.

[0031] In step S102, local model training and data augmentation involve training a local "output / load-benefit" model and a state prediction model (such as short-term power prediction), or simulating and generating "hypothetical" data under different scheduling strategies in a twin for data augmentation. Step S103 employs a secure aggregation algorithm that includes the FedAvg algorithm, which utilizes homomorphic encryption or secure multi-party computation techniques.

[0032] In the S104 simulation, the self-objective includes minimizing electricity costs, and the global objective is to minimize the virtual power plant tracking and dispatch instructions and the total cost.

[0033] The S105 generates a global scheduling target that includes the total adjustment power curve; the optimal local feasible scheduling scheme promises the range of power that can be increased / decreased and the cost promised to the master station during a certain period; the rapid coordination and verification of the quotation checks whether the total power meets the requirements.

[0034] A virtual power plant collaborative scheduling system based on federated learning and digital twins is disclosed. The system includes: a virtual power plant master station server, multiple distributed edge nodes, and a communication network; the edge nodes are equipped with local digital twin modules and model training modules; the virtual power plant master station server is equipped with a federated aggregation module, a global coordination module, and a communication encryption module to implement the above methods.

[0035] The innovation of this invention lies in the collaborative architecture driven by the dual engines of "local digital twin + federated learning": data privacy and local fine-grained modeling are guaranteed by building a local digital twin for each member, and knowledge fusion and collaborative optimization of the global model are achieved through federated learning without sharing data.

[0036] Localized, privacy-preserving model training and data augmentation in digital twins.

[0037] "Model parameters / gradients" serve as a secure medium to replace the original data transmission, enabling the construction of a collaborative scheduling model where "data is available but not visible".

[0038] Based on the updated model, a pre-scheduling co-inference mechanism in a twin environment improves the feasibility and economy of the final global scheduling instructions.

[0039] Explanation of related terms Virtual power plant: A coordinated management system that aggregates geographically dispersed distributed energy resources such as distributed power sources, energy storage systems, and controllable loads through advanced information and communication technologies and software systems, and participates in grid operation and electricity market transactions as a special "power plant".

[0040] Digital twins are full-element, dynamic, and bidirectional mappings of physical entities (such as an energy storage power station or a photovoltaic system) in virtual space. They integrate physical models, operational data, and domain knowledge, and can be used for simulation, analysis, prediction, and optimization.

[0041] Federated learning: a distributed machine learning paradigm. Its core involves multiple participants (clients) training models locally using their own data, then uploading only the model parameters (or model updates, such as gradients) to a central server for secure aggregation, forming a superior global model. The original data remains locally, effectively solving data privacy and security issues.

[0042] Coordinated dispatch: refers to the coordinated and optimized control of various distributed energy resources within a virtual power plant as a whole, in order to meet the dispatch instructions of the upper power grid (such as peak shaving and frequency regulation), participate in electricity market bidding, or achieve its own economic operation goals.

[0043] Alternative Solution 1 (Privacy Protection Technology Alternative): The federated learning framework in S103 can employ different combinations of privacy enhancement technologies. For example, in addition to homomorphic encryption, differential privacy technology can be combined to add an appropriate amount of random noise before uploading local model parameters, thus protecting individual data privacy while still ensuring the effectiveness of the aggregated model.

[0044] Alternative Solution 2 (Alternative for Twin Construction): When constructing a local digital twin in S101, for edge nodes with limited computing resources, instead of constructing a highly realistic physical model twin, a data-driven proxy model (such as a trained neural network) can be used to approximate the input-output response relationship of the local device. This proxy model can also support local extrapolation and training, and has lower computational overhead.

[0045] Alternative Solution 3 (Cooperative Optimization): The global coordination steps of S105 do not need to be entirely confirmed centrally by the main station. A distributed consensus mechanism can be designed, such as one based on blockchain smart contracts, where each member uploads their encrypted scheduling commitments to the chain. The consensus algorithm automatically verifies and generates the final, tamper-proof scheduling plan, achieving more decentralized trust and coordination.

[0046] Alternative Option 4 (Federated Learning Paradigm Replacement): Federated learning in S102-S103 can employ different paradigms. For example, federated transfer learning can be used, suitable for situations where participating members have significantly different data characteristics, accelerating learning by transferring shared model layers. Alternatively, hierarchical federated learning can be used, where a first federated aggregation is performed within multiple sub-regions (such as an industrial park), and then the aggregated model is uploaded to the central hub for a second aggregation, adapting to geographical or organizational hierarchical structures.

[0047] In addition to the basic FedAvg algorithm, federated learning aggregation algorithms can also employ weighted averaging aggregation strategies based on the unevenness of member data volume and quality, such as allocating aggregation weights according to the amount of member data or the quality of model updates.

[0048] Security technologies can employ secure multi-party computation techniques for model parameter aggregation, ensuring that even if the aggregation server is untrusted, it is impossible to deduce information about any individual member from the aggregation process.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A virtual power plant collaborative scheduling method based on federated learning and digital twins, characterized in that: For each data-sensitive participant in the virtual power plant, a high-fidelity digital twin of their local equipment and operating model is constructed. Localized model training and data simulation are performed in each twin. A federated learning framework is adopted, enabling each participant to exchange encrypted model parameters or gradients without sharing the original local data. These parameters are then aggregated in the cloud or by a designated coordinator to generate a globally optimized collaborative scheduling decision model. Finally, this global model is used to drive the overall digital twin of the virtual power plant for efficient, secure, and adaptive collaborative operation optimization and real-time scheduling.

2. The virtual power plant collaborative scheduling method based on federated learning and digital twins according to claim 1, characterized in that... The method includes the following steps: S101, Constructing Local Digital Twins: Establish a high-fidelity local digital twin for each participating member in the virtual power plant, corresponding to its physical entity; each local twin runs within the member's local edge computing device or security domain; the twin integrates the member's physical device model, historical operating data, real-time sensor data, and local specific operating constraint rules; S102, Local Model Training and Data Augmentation: In each member's local digital twin environment, using its private historical operating data and twin simulation capabilities, the local scheduling strategy-related models are independently trained or fine-tuned, and the training data does not leave the local security boundary. S103, Federated Aggregation of Model Parameters: a) Parameter Upload: After each member completes model training or update locally, they upload the encrypted gradient or model parameters of the model to the virtual power plant main station or the server acting as the coordinator through a secure channel. b) Secure aggregation: After receiving the encrypted model parameters from all participating members, the main server uses a secure aggregation algorithm to aggregate and calculate these parameters, generating an updated global collaborative scheduling model; c) Parameter distribution: Distribute the aggregated global model parameters to all participating members; S104, Updating Local Models and Collaborative Inference: After receiving the global model parameters, each member merges them with its own existing model locally or directly replaces them, allowing the local model to absorb the knowledge of other members without exposing local data. Then, using the updated local model, it drives its local digital twin to perform collaborative scheduling strategy inference for the next scheduling cycle in the twin environment. During the inference, each local twin considers its own goals and indirectly perceives the global goals through the model. S105, Global Collaborative Decision Generation and Execution: a) Decision generation: The master station generates a global scheduling target based on the power grid scheduling needs or market information and broadcasts it to all members; each member calculates and feeds back a privacy-preserving, optimal local feasible scheduling solution commitment to the master station based on the updated model in S104 and the inference results in its own twin. b) Global Coordination and Confirmation: The main station gathers the commitments of all members, conducts rapid coordination and verification, and finally confirms a feasible global scheduling plan; the final scheduling instructions, broken down to each member, are then issued. c) Instruction execution and closed-loop feedback: After receiving instructions, each member executes them through the local controller; at the same time, real data during the execution process is fed back to the local digital twin in real time to calibrate the model and update the twin state, forming a closed loop of physical execution-digital mapping; this closed loop also serves as the source of training data for the next round of federated learning.

3. The virtual power plant collaborative scheduling method based on federated learning and digital twins according to claim 2, characterized in that: The participants in the virtual power plant in step S101 include the flexible load of factory A, the energy storage power station of community B, and the photovoltaic system of building C.

4. The virtual power plant collaborative scheduling method based on federated learning and digital twins according to claim 2, characterized in that: In step S102, local model training and data augmentation involve training a local output / load-benefit model and a state prediction model, or simulating hypothetical data under different scheduling strategies in a twin for data augmentation.

5. The virtual power plant collaborative scheduling method based on federated learning and digital twins according to claim 2, characterized in that: Step S103 employs a secure aggregation algorithm that includes the FedAvg algorithm, which utilizes homomorphic encryption or secure multi-party computation techniques.

6. The virtual power plant collaborative scheduling method based on federated learning and digital twins according to claim 2, characterized in that: In the S104 simulation, the self-objective includes minimizing electricity costs, and the global objective is to minimize the virtual power plant tracking and dispatch instructions and the total cost.

7. The virtual power plant collaborative scheduling method based on federated learning and digital twins according to claim 2, characterized in that: The S105 generates a global scheduling target that includes the total adjustment power curve; the optimal local feasible scheduling scheme promises the range of power that can be increased / decreased and the cost promised to the master station during a certain period; the rapid coordination and verification of the quotation checks whether the total power meets the requirements.

8. A virtual power plant collaborative scheduling system based on federated learning and digital twins, characterized in that... The system includes: a virtual power plant master station server, multiple distributed edge nodes, and a communication network; the edge nodes are equipped with local digital twin modules and model training modules; the virtual power plant master station server is equipped with a federated aggregation module, a global coordination module, and a communication encryption module, for implementing the virtual power plant collaborative scheduling method based on federated learning and digital twins as described in any one of claims 1-7.