Island micro-grid virtual power plant collaborative scheduling method, device, equipment and medium

Abstract

This invention discloses a method, apparatus, equipment, and medium for collaborative scheduling of virtual power plants in isolated microgrids, which can resolve the contradiction between data privacy protection and scheduling optimization in isolated scenarios. The method includes: constructing a federated digital twin model through federated learning iterative training by a central node and multiple local nodes; expanding local training samples and simulating and verifying the local model using a local digital twin engine during the training process; simulating and verifying the aggregated global model using a central digital twin engine; generating source-load prediction results for each virtual power plant based on the federated digital twin model; obtaining a balanced strategy combination for collaborative scheduling of multiple virtual power plants through dynamic game optimization based on the source-load prediction results and the scheduling strategy space of each virtual power plant; and generating scheduling instructions for each virtual power plant according to the balanced strategy combination and issuing them to each virtual power plant to execute resource allocation.

Description

Technical Field

[0001] This invention relates to the field of microgrid dispatching technology, specifically to a method, apparatus, equipment, and medium for collaborative dispatching of virtual power plants in isolated microgrids. Background Technology

[0002] In recent years, microgrid dispatching technology has evolved from single-energy management to multi-energy complementary and coordinated operation. Virtual power plants, as an effective carrier for aggregating distributed resources, have improved energy utilization efficiency through market-based trading and optimized control. In typical isolated scenarios such as islands and industrial parks, virtual power plants have achieved coordinated dispatching of wind, solar, storage, and load. However, existing research mainly focuses on the optimization algorithm design under centralized architecture, and pays insufficient attention to the balance between multi-entity data privacy and dispatching autonomy.

[0003] Source-load forecasting in isolated microgrids faces three core challenges: First, data scarcity. In isolated scenarios, insufficient historical meteorological and load data leads to poor representativeness of training samples for forecasting models. Second, strong volatility. Renewable energy output is significantly affected by sudden weather changes, and load demand exhibits strong randomness due to factors such as production plans and resident behavior. Third, complex coupling relationships. Strong nonlinear coupling exists between the source and load sides, making it difficult for traditional univariate forecasting models to capture the interactive effects of multiple factors. Among existing forecasting algorithms, deep learning models such as LSTM and GRU can extract time-series features, but they are prone to overfitting in data-sparse scenarios. Physical mechanism models rely on precise equipment parameters and environmental data, resulting in poor adaptability in isolated scenarios. Furthermore, centralized forecasting models require collecting raw data from each node, increasing communication burden and raising data privacy risks, thus limiting the coordinated optimization of forecasting accuracy and scheduling decisions. Summary of the Invention

[0004] The purpose of this invention is to overcome the deficiencies in the existing technology and provide a method, device, equipment and medium for collaborative scheduling of virtual power plants in isolated microgrids that can solve the contradiction between data privacy protection and scheduling optimization in isolated scenarios.

[0005] To achieve the above objectives, the first aspect of the present invention provides a collaborative scheduling method for virtual power plants in an islanded microgrid, applied to a collaborative scheduling system including a central node and multiple local nodes deployed in each virtual power plant, the method comprising: Federated learning iterative training is performed through the central node and the multiple local nodes to construct a federated digital twin model, which includes a global model and local models. During the training process of the federated digital twin model, the local end digital twin engine of each local node is used to construct a corresponding local virtual image to expand the local training samples and perform simulation verification of the local model. The central end digital twin engine deployed on the central node is used to construct a global virtual scene to perform simulation verification of the aggregated global model. The source-load prediction results for each virtual power plant are generated based on the aforementioned federal digital twin model; Based on the source-load prediction results and the scheduling strategy space of each virtual power plant, a combination of equilibrium strategies for the coordinated scheduling of multiple virtual power plants is obtained through dynamic game optimization. Based on the aforementioned balance strategy combination, scheduling instructions for each virtual power plant are generated and issued to each virtual power plant to execute resource allocation for each virtual power plant.

[0006] Furthermore, federated learning iterative training is performed through the central node and the multiple local nodes to construct a federated digital twin model, including: A global initial model is deployed at the central node, and a local model is initialized on each local node based on local data. For each local node, a virtual sample is generated based on local data using a local digital twin engine. The local data and the virtual sample are then merged as local training data. The local model is trained using the local training data. The local data is the operating data of the virtual power plant collected by the local node. Each local node uploads the trained local model parameters to the central node; The central node aggregates the parameters of each local model to generate an updated global model, and then distributes the global model to each local node. Each local node merges its latest local data with the virtual samples to form local test data. The global model is then fine-tuned and tested using the local test data to obtain an updated local model, thus completing one round of iterative training.

[0007] Furthermore, based on the aforementioned federal digital twin model, the source-load prediction results for each virtual power plant are generated, including: For each local node, the historical data of the local node for a preset time period is input into the updated local model to generate preliminary prediction results; Based on the real-time status data of the virtual power plant collected by the local node, a dynamic correction factor is constructed through the local digital twin engine; The source load prediction result is generated by combining the preliminary prediction result with the dynamic correction factor, and then uploaded to the central node.

[0008] Furthermore, the dynamic correction factor is determined based on the weighted sum of the first correction component and the second correction component; the first correction component is determined based on the difference between the simulation output of the local digital twin engine and the preliminary prediction result, and the second correction component is determined based on the prediction error of the previous prediction result.

[0009] Furthermore, the method also includes: Based on the changing trend of the prediction error of multiple consecutive prediction results, adjust the first weight coefficient of the first correction component and / or the second weight coefficient of the second correction component. And / or, adjust the first weighting coefficient based on the deviation between the simulation output of the local digital twin engine and the actual operating data of the virtual power plant.

[0010] Furthermore, based on the source-load prediction results and the scheduling strategy space of each virtual power plant, a combination of equilibrium strategies for the coordinated scheduling of multiple virtual power plants is obtained through dynamic game optimization, including: Based on the source-load prediction results of each virtual power plant, a game model is constructed with each virtual power plant as the game participants. The boundary range of the strategy space of each virtual power plant is defined by the physical constraints of the schedulable resources of the virtual power plant, and the payoff function of each participant is determined based on the difference between independent scheduling cost and cooperative scheduling cost. By iteratively optimizing the Nash equilibrium point of the game model, a combination of equilibrium strategies is obtained that prevents all participants from increasing their own gains by unilaterally changing their strategies.

[0011] Furthermore, the method also includes: When the local node detects a device fault, it executes the corresponding fault response strategy according to the preset fault level. The fault levels include Level 1, Level 2, and Level 3 faults. The response strategy for Level 1 faults is for the local node to perform local fault clearing. The response strategy for Level 2 faults is for the local node to upload fault information to the central node to trigger the reallocation of resources from neighboring virtual power plants. The response strategy for Level 3 faults is for the local node to switch to independent scheduling mode and synchronize data with the central node after communication is restored.

[0012] A second aspect of the present invention provides a collaborative scheduling device for virtual power plants in an isolated microgrid, applied to a collaborative scheduling system including a central node and multiple local nodes deployed in each virtual power plant, the device comprising: The model building module is used to perform federated learning iterative training through the central node and the multiple local nodes to build a federated digital twin model. The federated digital twin model includes a global model and a local model. During the training process of the federated digital twin model, the local end digital twin engine of each local node is used to build a corresponding local virtual image to expand the local training samples and perform simulation verification of the local model. The central end digital twin engine deployed on the central node is used to build a global virtual scene to perform simulation verification of the aggregated global model. The prediction module is used to generate source-load prediction results for each virtual power plant based on the federated digital twin model. The game optimization module is used to obtain a combination of equilibrium strategies for the coordinated scheduling of multiple virtual power plants by dynamic game optimization based on the source-load prediction results and the scheduling strategy space of each virtual power plant. The scheduling execution module is used to generate scheduling instructions for each virtual power plant according to the combination of the balancing strategies and send them to each virtual power plant to execute the resource allocation of each virtual power plant.

[0013] Furthermore, when performing federated learning iterative training through the central node and the multiple local nodes to construct a federated digital twin model, the model building module is specifically used for: A global initial model is deployed at the central node, and a local model is initialized on each local node based on local data. For each local node, a virtual sample is generated based on local data using a local digital twin engine. The local data and the virtual sample are then merged as local training data. The local model is trained using the local training data. The local data is the operating data of the virtual power plant collected by the local node. Each local node uploads the trained local model parameters to the central node; The central node aggregates the parameters of each local model to generate an updated global model, and then distributes the global model to each local node. Each local node merges its latest local data with the virtual samples to form local test data. The global model is then fine-tuned and tested using the local test data to obtain an updated local model, thus completing one round of iterative training.

[0014] Furthermore, when generating source-load prediction results for each virtual power plant based on the aforementioned federal digital twin model, the prediction module is specifically used for: For each local node, the historical data of the local node for a preset time period is input into the updated local model to generate preliminary prediction results; Based on the real-time status data of the virtual power plant collected by the local node, a dynamic correction factor is constructed through the local digital twin engine; The source load prediction result is generated by combining the preliminary prediction result with the dynamic correction factor, and then uploaded to the central node.

[0015] Furthermore, the dynamic correction factor is determined based on the weighted sum of the first correction component and the second correction component; the first correction component is determined based on the difference between the simulation output of the local digital twin engine and the preliminary prediction result, and the second correction component is determined based on the prediction error of the previous prediction result.

[0016] Furthermore, the prediction module is also used for: Based on the changing trend of the prediction error of multiple consecutive prediction results, adjust the first weight coefficient of the first correction component and / or the second weight coefficient of the second correction component. And / or, adjust the first weighting coefficient based on the deviation between the simulation output of the local digital twin engine and the actual operating data of the virtual power plant.

[0017] Furthermore, when obtaining the equilibrium strategy combination for collaborative scheduling of multiple virtual power plants through dynamic game optimization based on the source-load prediction results and the scheduling strategy space of each virtual power plant, the game optimization module is specifically used for: Based on the source-load prediction results of each virtual power plant, a game model is constructed with each virtual power plant as the game participants. The boundary range of the strategy space of each virtual power plant is defined by the physical constraints of the schedulable resources of the virtual power plant, and the payoff function of each participant is determined based on the difference between independent scheduling cost and cooperative scheduling cost. By iteratively optimizing the Nash equilibrium point of the game model, a combination of equilibrium strategies is obtained that prevents all participants from increasing their own gains by unilaterally changing their strategies.

[0018] Furthermore, the device also includes a fault handling module for: When the local node detects a device fault, it executes the corresponding fault response strategy according to the preset fault level. The fault levels include Level 1, Level 2, and Level 3 faults. The response strategy for Level 1 faults is for the local node to perform local fault clearing. The response strategy for Level 2 faults is for the local node to upload fault information to the central node to trigger the reallocation of resources from neighboring virtual power plants. The response strategy for Level 3 faults is for the local node to switch to independent scheduling mode and synchronize data with the central node after communication is restored.

[0019] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the method described in the first aspect.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium having program instructions stored thereon, which, when executed, implement the method described in the first aspect.

[0021] The advantages and beneficial effects of this invention are as follows: Addressing the risk of user data leakage under centralized data processing models, this invention employs a federated learning framework. Each virtual power plant only uploads model parameters, not raw data, completing global model training without sharing private data, thus improving data security. By mapping the physical state of isolated microgrids in real time through digital twins, and combining the multi-source data feature fusion capabilities of federated learning, it overcomes the limitations of existing prediction algorithms based on single historical data, improving source-load prediction accuracy. Based on a dynamic game optimization strategy, it achieves second-level resource allocation, shortening scheduling response time and improving collaborative scheduling efficiency, thus solving the response lag problem in multi-virtual power plant collaborative scheduling. Attached Figure Description

[0022] Figure 1 This is a flowchart of the collaborative scheduling method for virtual power plants in isolated microgrids according to the present invention; Figure 2 This is a schematic diagram of the structure of the virtual power plant collaborative dispatching device for isolated microgrids of the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and examples. The following examples are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0024] The collaborative scheduling method for virtual power plants in isolated microgrids according to this invention is applied to a collaborative scheduling system including a central node and multiple local nodes deployed in each virtual power plant. The collaborative scheduling system architecture adopts a layered design, including a data layer, a model layer, and a scheduling layer. Each layer realizes data interaction and functional collaboration through standardized interfaces. The specific architecture modules and their functions are shown in the table below: The data layer collects local data through edge nodes and performs privacy protection processing; the model layer uses federated learning to train a global prediction model and combines it with a digital twin engine to realize real-time interaction between the physical system and the virtual model; the scheduling layer generates a collaborative scheduling strategy based on the prediction results and dynamic game algorithm, and continuously optimizes the decision through a feedback mechanism.

[0025] The hardware device of this invention must meet the requirements of edge computing, data acquisition, and real-time communication, and the software platform must support federated learning frameworks and digital twin simulation functions. The specific configuration list is shown in the table below: like Figure 1 As shown, the islanded microgrid virtual power plant collaborative scheduling method of the present invention includes: S101. Federated learning iterative training is performed through the central node and the multiple local nodes to construct a federated digital twin model. The federated digital twin model includes a global model and a local model. During the training process of the federated digital twin model, the local end digital twin engine of each local node is used to construct a corresponding local virtual image to expand the local training samples and perform simulation verification of the local model. The central end digital twin engine deployed on the central node is used to construct a global virtual scene to perform simulation verification of the aggregated global model. S102. Generate source-load prediction results for each virtual power plant based on the federated digital twin model; S103. Based on the source-load prediction results and the scheduling strategy space of each virtual power plant, the equilibrium strategy combination for the coordinated scheduling of multiple virtual power plants is obtained through dynamic game optimization. S104. Based on the combined balancing strategies, generate scheduling instructions for each virtual power plant and issue them to each virtual power plant to execute resource allocation for each virtual power plant.

[0026] The federated digital twin model is a composite model system integrating a federated learning framework and digital twin technology, comprising a global model and local models. The global model is maintained by a central node, while local models are maintained by individual local nodes. During model training, local nodes utilize their local digital twin engines to construct local virtual images. These images map the physical environment to local real data, generating virtual samples by simulating different operating conditions to expand local training samples and to perform simulation verification of the local model. Simultaneously, the central node utilizes its central digital twin engine to construct a global virtual scene. This scene integrates information from all local virtual images and is used to simulate and verify the aggregated global model, ensuring the overall effectiveness of the global model.

[0027] After the federated digital twin model is trained, each local node can use the updated local model to predict source load. The prediction process can combine local real-time data and dynamic correction from the digital twin engine to generate accurate future wind and solar power output and load demand curves.

[0028] The central node collects the source-load prediction results uploaded by each local node and, combined with the physical constraints of the dispatchable resources (such as energy storage, controllable loads, and standby units) of each virtual power plant, constructs a dynamic game model. By solving the Nash equilibrium, a set of equilibrium strategy combinations is obtained that prevents any virtual power plant from benefiting by unilaterally changing its own dispatch strategy, thereby achieving synergy between the interests of multiple stakeholders and overall efficiency.

[0029] The central node translates the balancing strategy into specific control commands, such as the energy storage charging and discharging power and controllable load adjustment amount sent to each local node, which are then executed by the local nodes to complete the dynamic allocation of resources.

[0030] This invention addresses the risk of user data leakage in centralized data processing models by employing a federated learning framework. Each virtual power plant uploads only model parameters, not raw data, completing global model training without sharing private data, thus improving data security. By mapping the physical state of isolated microgrids in real time through digital twins and combining the multi-source data feature fusion capabilities of federated learning, it overcomes the limitations of existing prediction algorithms based on single historical data, improving source-load prediction accuracy. Based on a dynamic game optimization strategy, it achieves second-level resource allocation, shortening scheduling response time and improving collaborative scheduling efficiency, thus solving the response lag problem in multi-virtual power plant collaborative scheduling.

[0031] To enable local models to inherit global commonalities while adapting to the latest local operating characteristics, thereby improving the model's generalization ability and prediction accuracy, a preferred embodiment of this invention involves constructing a federated digital twin model through federated learning iterative training performed by the central node and the multiple local nodes. This includes: deploying a global initial model at the central node and initializing local models on each local node based on local data; for each local node, generating virtual samples based on local data using a local digital twin engine, merging the local data and the virtual samples as local training data, and training the local model using the local training data, where the local data consists of operating data from a virtual power plant collected by the local node; each local node uploading the trained local model parameters to the central node; the central node aggregating the local model parameters to generate an updated global model and distributing the global model to each local node; each local node merging the latest local data with the virtual samples as local test data, fine-tuning and testing the global model using the local test data to obtain an updated local model, thus completing one round of iterative training.

[0032] Model initialization phase: A global initial model, such as an LSTM basic network, is deployed at the central node. Its structure can be preset to have 128 neurons in the input layer, 2 hidden layers with 64 neurons each, and 1 neuron in the output layer. Global parameters and biases are initialized. At the same time, the central digital twin engine imports the physical parameters of the isolated microgrid, such as 18% photovoltaic panel efficiency, 3m / s wind turbine cut-in wind speed, and 200kWh energy storage capacity, to construct an initial global virtual scene, including equipment parameters, topology, and historical operating data mapping.

[0033] Each local node initializes its local model based on local data distribution. Specifically, local nodes can fine-tune the global initial model based on 15-minute output / load data from the past three months to form a local initial model. Simultaneously, each local node's local digital twin engine constructs a local virtual image based on local device parameters and operational data for subsequent sample generation and simulation.

[0034] Local Training Phase: For each local node, virtual samples are generated based on real local data using a local digital twin engine. For example, 1000 sets of virtual samples are generated by simulating typical weather scenarios such as sunny, cloudy, and rainy days, as well as different equipment health levels. The real local data and virtual samples are then merged as local training data. Local training hyperparameters are set as follows: learning rate 0.001, batch size 32, number of iterations 50, using the Adam optimizer to minimize the MSE loss function, and training the local model using the local training data. During training, differential privacy techniques are used to add noise to the model gradient (e.g., noise intensity ε=1.0) to protect data privacy. Local data specifically refers to the virtual power plant operation data collected by the local nodes through the sensor network, including wind and solar power output, load demand, equipment status, and meteorological factors.

[0035] Parameter upload phase: Each local node uploads its trained local model parameters to the central node via an encrypted channel (such as TLS 1.3 protocol). The upload frequency can be set to once per hour. The central node performs integrity verification on the parameters (such as hash value comparison) and rejects abnormal parameters.

[0036] Global aggregation phase: The central node uses a weighted federated average algorithm (FedAvg) to aggregate the parameters of each local model, with weights ( The value is directly proportional to the amount of data in each node. ), ( (where is the number of data samples at node i, and j is the total number of nodes), generating the updated global model. The central digital twin engine loads the new global model parameters (where is the number of data samples at node i, and j is the total number of nodes). The global virtual scene is corrected in an integrated manner, and the prediction error of the global model in the virtual scene is verified by simulation. If the mean absolute error (MAE) is greater than 5%, a second aggregation is triggered.

[0037] The central node encrypts and distributes the updated global model to each local node. Each local node merges its latest locally collected data (such as the latest hourly data) with previously generated virtual samples as local test data. This test data is used to fine-tune and test the distributed global model, resulting in an updated local model. This completes one round of iterative training. The iterative process continues until the model converges; for example, 20 global iterations, each containing 5 local training iterations. The digital twin simulation step size is 1 minute, and the physical model update frequency is synchronized with the sensor sampling frequency (1Hz). The differential privacy noise intensity is dynamically adjusted based on data sensitivity (ε=1.2 for wind and solar data, ε=0.8 for load data).

[0038] In order to improve the accuracy of prediction by correcting the preliminary prediction results based on real-time status data, a preferred embodiment of the present invention is to generate source-load prediction results for each virtual power plant based on the federated digital twin model, including: for each local node, inputting the historical data of the local node for a preset period into the updated local model to generate a preliminary prediction result; constructing a dynamic correction factor through the local digital twin engine based on the status data of the virtual power plant collected in real time by the local node; fusing the preliminary prediction result and the dynamic correction factor to generate the source-load prediction result, and uploading the source-load prediction result to the central node.

[0039] After completing multiple rounds of iterative training, the system enters the online prediction phase. The system triggers predictions daily at preset times (e.g., 00:00, 06:00, 12:00, 18:00), with a prediction time domain of the next 24 hours and a time granularity of 15 minutes. Each local node executes the following process: First, historical data for a preset time period from the local node is input into the updated local model to generate preliminary prediction results. Historical data for the preset time period includes data on wind and solar power output, load, weather, and temperature for the previous 72 hours, as well as digital twin scenario parameters (such as equipment health and line loss rate). The updated local model then outputs preliminary predicted values ​​for wind and solar power output and load demand.

[0040] Then, based on the real-time status data of the virtual power plant collected by local nodes, a dynamic correction factor is constructed using the local digital twin engine. Dynamic correction factor ( The formula is as follows: in,( () is the output for twin simulation. The prediction error (mean absolute error) of the previous prediction result. These are preliminary predictions. As the first weighting coefficient, This is the second weighting coefficient. The initial weighting coefficients can be manually set with empirical values ​​(e.g., α=0.5, β=0.5), and then adaptively adjusted according to the operating status.

[0041] Finally, the preliminary prediction results are combined with the dynamic correction factor to generate the final source load prediction result, which is then uploaded to the central node. Final Source Load Prediction Result .

[0042] To achieve multi-dimensional correction of the prediction results and further improve the accuracy of the prediction, a preferred embodiment of the present invention is that the dynamic correction factor is determined by a weighted sum of a first correction component and a second correction component; the first correction component is determined based on the difference between the simulation output of the local digital twin engine and the preliminary prediction result, and the second correction component is determined based on the prediction error of the previous prediction result.

[0043] The first correction component is The second correction component is .

[0044] To enhance the model's adaptability to complex operating environments, a preferred embodiment of the present invention further includes: adjusting the first weight coefficient of the first correction component and / or the second weight coefficient of the second correction component based on the changing trend of the prediction error of multiple consecutive prediction results; and / or adjusting the first weight coefficient based on the deviation between the simulation output of the local digital twin engine and the actual operating data of the virtual power plant.

[0045] The local node records the prediction error (e.g., mean absolute error, MAE) of multiple consecutive predictions and analyzes its trend. Specifically, if the prediction error shows a continuous increasing trend (e.g., the MAE of three consecutive predictions increases sequentially, and the increase exceeds a preset threshold), it indicates that the current correction factor is insufficient to correct the prediction bias, and the first weighting coefficient needs to be increased. Second weighting coefficient (For example, increase each by 20%) to enhance the effectiveness of the correction factor. If the prediction error shows a continuous decreasing trend (e.g., the MAE of three consecutive predictions decreases sequentially), it indicates that the current correction effect is good, and the correction factor can be appropriately reduced. and (For example, reduce each by 10%) to avoid overcorrection. If the prediction error fluctuates smoothly (fluctuating within a preset range), the current weighting coefficients remain unchanged. Trend judgment can be performed using the moving average method or the linear regression slope method. For example, calculate the slope of the prediction error for the last three times; a positive slope indicates an increase in error, while a negative slope indicates a decrease in error. For example, in one implementation, regarding prediction deviation, when the deviation between the actual output / load and the predicted value exceeds 10% for three consecutive time periods, the local digital twin engine automatically adjusts the weighting coefficients. and (For example, increase the weights to 1.2 times), and trigger the next round of model fine-tuning to continuously optimize prediction accuracy.

[0046] When adjusting the first weighting coefficient based on the twin bias, the local node periodically compares the simulation output of the local digital twin engine (such as simulated wind and solar power output) with the actual collected operational data, and calculates the bias index (such as root mean square error RMSE). When the bias exceeds a preset threshold (such as 5%), it indicates that the mapping accuracy of the twin to the physical system has decreased, and the first weighting coefficient needs to be increased. (For example, adjust from 0.5 to 0.7) to enhance the weight of twin correction; when the bias is small, it can be reduced. .

[0047] and The core value selection logic is "scenario adaptation + dynamic adaptive adjustment". The initial value can be manually set based on experience, and then adaptive iterative optimization is performed based on the real-time operating status and prediction deviation characteristics of the isolated microgrid. The weight allocation of the two will be dynamically adjusted according to the matching degree between the physical system and the digital twin and the changing trend of prediction error. Ultimately, the dynamic correction factor can accurately correct the basic prediction deviation of the federated digital twin model and adapt to the characteristics of strong source load fluctuation in the isolated scenario.

[0048] In this embodiment of the invention, the weighting coefficients can be adaptively adjusted according to the prediction error trend and twin matching degree, so that the correction factor is always in the optimal state, thereby enhancing the model's adaptability to complex operating environments.

[0049] To avoid the imbalance of interests in traditional centralized scheduling, a preferred embodiment of this invention is to obtain an equilibrium strategy combination for collaborative scheduling of multiple virtual power plants through dynamic game optimization based on the source-load prediction results and the scheduling strategy space of each virtual power plant. This includes: constructing a game model with each virtual power plant as a player based on the source-load prediction results of each virtual power plant, wherein the boundary of the strategy space of each virtual power plant is defined by the physical constraints of the schedulable resources of the virtual power plant, and the payoff function of each player is determined based on the difference between independent scheduling cost and collaborative scheduling cost; and obtaining an equilibrium strategy combination that prevents all players from increasing their own payoff by unilaterally changing their strategies by iteratively optimizing the Nash equilibrium point of the game model.

[0050] The central node collects the source-load prediction results uploaded by all local nodes and constructs a game model with each virtual power plant as a player. Specifically, the strategy space S_i of each virtual power plant is determined, which is the set of all possible strategies for virtual power plant i, and its boundary is defined by the physical constraints of dispatchable resources. Dispatchable resources include energy storage systems (e.g., SOC range 20%-80%, charging and discharging power limit ±50kW), controllable loads (adjustment range ±15%), micro-turbine reserve capacity (e.g., 100kW), etc. Each specific strategy s_i in the strategy space represents a resource allocation scheme.

[0051] Let the set of virtual power plants be ( The strategy of participant (i) is ( ), the payoff function of each participant ( ),in( ( ) is the cost of independent scheduling, () represents the cost of coordinated scheduling.

[0052] The Nash equilibrium point is found through iterative optimization. For example, an iterative algorithm described in pseudocode might be used: initialize the policies of all participants, iterate through each participant, and find its optimal policy under the current policies of the other participants, until the maximum difference in policy updates is less than the convergence threshold, thus obtaining the equilibrium policy combination. This equilibrium strategy combination ensures that no participant can improve its own revenue by unilaterally changing the strategy. With the goal of minimizing scheduling costs, a Nash equilibrium strategy is solved, such as prioritizing the use of photovoltaic / wind power output (marginal cost 0 yuan / kWh), supplementing the shortfall with energy storage discharge (cost 0.3 yuan / kWh) or micro-turbines (cost 1.2 yuan / kWh), and absorbing excess output through energy storage charging (90% efficiency) or demand-side response (incentive 0.2 yuan / kWh).

[0053] When dispatching instructions are issued and executed, the balancing strategy combination is transformed into specific control instructions, such as: the output setpoint of wind and solar inverters, the energy storage charging and discharging power curve, the controllable load adjustment period, and the micro gas turbine start-up and shutdown plan; the instructions are sent to each device controller through the industrial bus (Modbus RTU protocol), with an execution delay requirement of ≤5 seconds; the edge nodes monitor the device response status in real time, and if the actual output of the inverter deviates from the instruction by more than 5%, local fine-tuning is immediately triggered.

[0054] To enable rapid and effective handling of equipment anomalies and ensure the stable operation of isolated microgrids, a preferred embodiment of the present invention further includes: when the local node detects a equipment fault, executing a corresponding fault response strategy according to a preset fault level; wherein, the fault levels include Level 1 faults, Level 2 faults, and Level 3 faults; the response strategy for Level 1 faults is for the local node to perform local fault clearing; the response strategy for Level 2 faults is for the local node to upload fault information to the central node to trigger the reallocation of resources from neighboring virtual power plants; and the response strategy for Level 3 faults is for the local node to switch to independent scheduling mode and synchronize data with the central node after communication is restored.

[0055] During the scheduling process, if a local node detects a device failure, it executes the corresponding strategy according to the preset failure level: Level 1 fault (such as a single photovoltaic string fault): The local controller automatically disconnects the faulty unit and adjusts the output of other strings to compensate for the power deficit.

[0056] Level 2 fault (such as energy storage system fault): The local node uploads fault information to the central node to trigger resource mutual assistance between virtual power plants. The central node redistributes resources of neighboring virtual power plants, such as scheduling energy storage resources of neighboring power plants to provide support.

[0057] Level 3 fault (such as communication interruption): The local node switches to independent scheduling mode and makes scheduling decisions autonomously based on local prediction data and local models; after communication is restored, the data from the independent scheduling period is synchronized to the central node.

[0058] The following is a description of the experimental design and dataset: 1. Experimental Scenario Setup The experiment was based on the simulation module of Acrel microgrid energy management platform. An island microgrid scenario containing three virtual power plants was built to simulate the energy supply system of an island industrial park, covering multiple entities such as photovoltaic, wind power, energy storage systems, industrial loads, and commercial loads.

[0059] 2. Dataset Source The training set was compiled from actual operational data collected from a coastal industrial park from January to December 2024, including: New energy output data: Daily power generation of photovoltaic modules and hourly power output curves of wind turbines (a total of 8760 samples). Load data: Hourly variation data of industrial production load and commercial lighting load (a total of 8760 samples). Equipment status data: Energy storage system SOC (State of Charge), inverter operating parameters, etc. (a total of 17,520 samples) 3. Evaluation Indicator Setting Source load prediction accuracy (MAE, RMSE), scheduling response time, and operating cost are selected as core evaluation indicators to compare the performance differences between the algorithm of this invention and the original scheduling algorithm.

[0060] The following are the experimental results and analysis: 1. Comparison of source load prediction performance In a simulation environment, the federated digital twin prediction model of this invention and the platform's original LSTM prediction model were tested respectively: Photovoltaic output prediction: The model of this invention has a MAE of 1.8% and an RMSE of 2.5%, which are 38% and 42% lower than the original model, respectively. The prediction stability is significantly improved in extreme weather scenarios (such as rainstorms and typhoons).

[0061] Load demand forecasting: The MAE of the model in this invention is 2.1% for industrial load and 2.5% for commercial load, which is 40% and 37% lower than the original model, respectively, effectively solving the forecasting bias problem caused by large load fluctuations in islanded scenarios.

[0062] 2. Comparison of Cooperative Scheduling Efficiency In the multi-virtual power plant collaborative scheduling experiment: The scheduling response time of the algorithm of this invention is 0.78 seconds on average, which is 63% shorter than the original scheduling algorithm of 2.1 seconds, thus meeting the real-time scheduling requirements of isolated microgrids.

[0063] Monthly operating cost experiments show that the average monthly operating cost of the islanded microgrid using the method of this invention is 126,000 yuan, a reduction of 18.2% compared to the original dispatch scheme of 154,000 yuan. The optimization of the energy storage system's charging and discharging contributed 70% of this cost reduction. The optimization formula is as follows: in, To optimize operating costs, Let t be the power purchased by the power grid. Let t be the electricity purchase price. Let t be the system loss cost at time t, and T be the scheduling cycle duration.

[0064] 3. Robustness Testing Even in extreme scenarios such as missing simulated data (missing rate of 20%) and equipment failure (shutdown of a photovoltaic system in a virtual power plant), the model of this invention can still maintain high prediction accuracy and scheduling stability, while the original algorithm's prediction error rises to 8.7% and scheduling cost increases by 25% in the case of missing data, verifying the strong robustness of the method of this invention.

[0065] Figure 2 The islanded microgrid virtual power plant collaborative dispatching device shown is applied to a collaborative dispatching system that includes a central node and multiple local nodes deployed in each virtual power plant, such as... Figure 2 As shown, the device includes: The model building module 201 is used to perform federated learning iterative training through the central node and the multiple local nodes to build a federated digital twin model. The federated digital twin model includes a global model and a local model. During the training process of the federated digital twin model, the local end digital twin engine of each local node is used to build a corresponding local virtual image to expand the local training samples and perform simulation verification of the local model. The central end digital twin engine deployed on the central node is used to build a global virtual scene to perform simulation verification of the aggregated global model. Prediction module 202 is used to generate source-load prediction results for each virtual power plant based on the federated digital twin model; The game optimization module 203 is used to obtain a combination of equilibrium strategies for the coordinated scheduling of multiple virtual power plants by dynamic game optimization based on the source load prediction results and the scheduling strategy space of each virtual power plant. The scheduling execution module 204 is used to generate scheduling instructions for each virtual power plant according to the balance strategy combination and send them to each virtual power plant to execute the resource allocation of each virtual power plant.

[0066] In some embodiments, when performing federated learning iterative training through the central node and the plurality of local nodes to construct a federated digital twin model, the model building module 201 is specifically used for: A global initial model is deployed at the central node, and a local model is initialized on each local node based on local data. For each local node, a virtual sample is generated based on local data using a local digital twin engine. The local data and the virtual sample are then merged as local training data. The local model is trained using the local training data. The local data is the operating data of the virtual power plant collected by the local node. Each local node uploads the trained local model parameters to the central node; The central node aggregates the parameters of each local model to generate an updated global model, and then distributes the global model to each local node. Each local node merges its latest local data with the virtual samples to form local test data. The global model is then fine-tuned and tested using the local test data to obtain an updated local model, thus completing one round of iterative training.

[0067] In some embodiments, when generating source-load prediction results for each virtual power plant based on the federated digital twin model, the prediction module 202 is specifically used for: For each local node, the historical data of the local node for a preset time period is input into the updated local model to generate preliminary prediction results; Based on the real-time status data of the virtual power plant collected by the local node, a dynamic correction factor is constructed through the local digital twin engine; The source load prediction result is generated by combining the preliminary prediction result with the dynamic correction factor, and then uploaded to the central node.

[0068] In some embodiments, the dynamic correction factor is determined based on a weighted sum of a first correction component and a second correction component; the first correction component is determined based on the difference between the simulation output of the local digital twin engine and the preliminary prediction result, and the second correction component is determined based on the prediction error of the previous prediction result.

[0069] In some embodiments, the prediction module 202 is further configured to: Based on the changing trend of the prediction error of multiple consecutive prediction results, adjust the first weight coefficient of the first correction component and / or the second weight coefficient of the second correction component. And / or, adjust the first weighting coefficient based on the deviation between the simulation output of the local digital twin engine and the actual operating data of the virtual power plant.

[0070] In some embodiments, when the equilibrium strategy combination for coordinated scheduling of multiple virtual power plants is obtained through dynamic game optimization based on the source-load prediction results and the scheduling strategy space of each virtual power plant, the game optimization module 203 is specifically used for: Based on the source-load prediction results of each virtual power plant, a game model is constructed with each virtual power plant as the game participants. The boundary range of the strategy space of each virtual power plant is defined by the physical constraints of the schedulable resources of the virtual power plant, and the payoff function of each participant is determined based on the difference between independent scheduling cost and cooperative scheduling cost. By iteratively optimizing the Nash equilibrium point of the game model, a combination of equilibrium strategies is obtained that prevents all participants from increasing their own gains by unilaterally changing their strategies.

[0071] In some embodiments, the apparatus further includes a fault handling module 205, configured to: When the local node detects a device fault, it executes the corresponding fault response strategy according to the preset fault level. The fault levels include Level 1, Level 2, and Level 3 faults. The response strategy for Level 1 faults is for the local node to perform local fault clearing. The response strategy for Level 2 faults is for the local node to upload fault information to the central node to trigger the reallocation of resources from neighboring virtual power plants. The response strategy for Level 3 faults is for the local node to switch to independent scheduling mode and synchronize data with the central node after communication is restored.

[0072] Figure 2 The islanded microgrid virtual power plant collaborative scheduling device in the illustrated embodiment can be used to execute the technical solution of the above method embodiment. Its implementation principle and technical effect are similar, and will not be repeated here.

[0073] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device can be a server as described in the above embodiment. The electronic device provided in this embodiment of the present invention can execute the processing flow provided in the embodiment of the collaborative scheduling method for virtual power plants in isolated microgrids, such as... Figure 3 As shown, the electronic device 1100 includes: a memory 1101, a processor 1102, a computer program, and a communication interface 1103; wherein, the computer program is stored in the memory 1101 and is configured to be executed by the processor 1102 to perform the islanded microgrid virtual power plant collaborative scheduling method as described above.

[0074] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, the computer program being executed by a processor to implement the islanded microgrid virtual power plant collaborative scheduling method described in the above embodiments.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0076] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A collaborative scheduling method for virtual power plants in an isolated microgrid, applied to a collaborative scheduling system including a central node and multiple local nodes deployed in each virtual power plant, characterized in that, The method includes: Federated learning iterative training is performed through the central node and the multiple local nodes to construct a federated digital twin model, which includes a global model and local models. During the training process of the federated digital twin model, the local end digital twin engine of each local node is used to construct a corresponding local virtual image to expand the local training samples and perform simulation verification of the local model. The central end digital twin engine deployed on the central node is used to construct a global virtual scene to perform simulation verification of the aggregated global model. The source-load prediction results for each virtual power plant are generated based on the aforementioned federal digital twin model; Based on the source-load prediction results and the scheduling strategy space of each virtual power plant, a combination of equilibrium strategies for the coordinated scheduling of multiple virtual power plants is obtained through dynamic game optimization. Based on the aforementioned balance strategy combination, scheduling instructions for each virtual power plant are generated and issued to each virtual power plant to execute resource allocation for each virtual power plant.

2. The method for collaborative scheduling of virtual power plants in isolated microgrids according to claim 1, characterized in that, Federated learning iterative training is performed through the central node and the multiple local nodes to construct a federated digital twin model, including: A global initial model is deployed at the central node, and a local model is initialized on each local node based on local data. For each local node, a virtual sample is generated based on local data using a local digital twin engine. The local data and the virtual sample are then merged as local training data. The local model is trained using the local training data. The local data is the operating data of the virtual power plant collected by the local node. Each local node uploads the trained local model parameters to the central node; The central node aggregates the parameters of each local model to generate an updated global model, and then distributes the global model to each local node. Each local node merges its latest local data with the virtual samples to form local test data. The global model is then fine-tuned and tested using the local test data to obtain an updated local model, thus completing one round of iterative training.

3. The method for collaborative scheduling of virtual power plants in isolated microgrids according to claim 2, characterized in that, The source-load prediction results for each virtual power plant are generated based on the aforementioned federated digital twin model, including: For each local node, the historical data of the local node for a preset time period is input into the updated local model to generate preliminary prediction results; Based on the real-time status data of the virtual power plant collected by the local node, a dynamic correction factor is constructed through the local digital twin engine; The source load prediction result is generated by combining the preliminary prediction result with the dynamic correction factor, and then uploaded to the central node.

4. The method for collaborative scheduling of virtual power plants in isolated microgrids according to claim 3, characterized in that, The dynamic correction factor is determined by a weighted sum of a first correction component and a second correction component; the first correction component is determined based on the difference between the simulation output of the local digital twin engine and the preliminary prediction result, and the second correction component is determined based on the prediction error of the previous prediction result.

5. The method for collaborative scheduling of virtual power plants in isolated microgrids according to claim 4, characterized in that, The method further includes: Based on the changing trend of the prediction error of multiple consecutive prediction results, adjust the first weight coefficient of the first correction component and / or the second weight coefficient of the second correction component. And / or, adjust the first weighting coefficient based on the deviation between the simulation output of the local digital twin engine and the actual operating data of the virtual power plant.

6. The method for collaborative scheduling of virtual power plants in isolated microgrids according to claim 1, characterized in that, Based on the source-load prediction results and the scheduling strategy space of each virtual power plant, a combination of equilibrium strategies for the coordinated scheduling of multiple virtual power plants is obtained through dynamic game optimization, including: Based on the source-load prediction results of each virtual power plant, a game model is constructed with each virtual power plant as the game participants. The boundary range of the strategy space of each virtual power plant is defined by the physical constraints of the schedulable resources of the virtual power plant, and the payoff function of each participant is determined based on the difference between independent scheduling cost and cooperative scheduling cost. By iteratively optimizing the Nash equilibrium point of the game model, a combination of equilibrium strategies is obtained that prevents all participants from increasing their own gains by unilaterally changing their strategies.

7. The method for collaborative scheduling of virtual power plants in isolated microgrids according to claim 1, characterized in that, The method further includes: When the local node detects a device fault, it executes the corresponding fault response strategy according to the preset fault level. The fault levels include Level 1, Level 2, and Level 3 faults. The response strategy for Level 1 faults is for the local node to perform local fault clearing. The response strategy for Level 2 faults is for the local node to upload fault information to the central node to trigger the reallocation of resources from neighboring virtual power plants. The response strategy for Level 3 faults is for the local node to switch to independent scheduling mode and synchronize data with the central node after communication is restored.

8. A collaborative scheduling device for virtual power plants in an isolated microgrid, applied to a collaborative scheduling system including a central node and multiple local nodes deployed in each virtual power plant, characterized in that, The device includes: The model building module is used to perform federated learning iterative training through the central node and the multiple local nodes to build a federated digital twin model. The federated digital twin model includes a global model and a local model. During the training process of the federated digital twin model, the local end digital twin engine of each local node is used to build a corresponding local virtual image to expand the local training samples and perform simulation verification of the local model. The central end digital twin engine deployed on the central node is used to build a global virtual scene to perform simulation verification of the aggregated global model. The prediction module is used to generate source-load prediction results for each virtual power plant based on the federated digital twin model. The game optimization module is used to obtain a combination of equilibrium strategies for the coordinated scheduling of multiple virtual power plants by dynamic game optimization based on the source-load prediction results and the scheduling strategy space of each virtual power plant. The scheduling execution module is used to generate scheduling instructions for each virtual power plant according to the combination of the balancing strategies and send them to each virtual power plant to execute the resource allocation of each virtual power plant.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores program instructions that, when executed, implement the method as described in any one of claims 1-7.

Images