Optimized scheduling method, system, medium and product of virtual power plant based on cloud-fog edge architecture

Abstract

The embodiment of the invention provides an optimal scheduling method and system for a virtual power plant based on a cloud and fog edge architecture, a computer readable storage medium and a computer program product. The method comprises the following steps: edge nodes realize local optimization scheduling of end side equipment, fog nodes realize coordinated aggregation of a plurality of edge nodes in a jurisdiction area, and a cloud platform realizes cross-area global optimization and knowledge sharing. According to the scheme, a cloud platform, fog node and edge node three-level collaborative optimal scheduling architecture is constructed, and local optimization capability, fog node area coordination capability and global optimization capability are given to the edge nodes, so that the inherent single-point fault risk of traditional centralized scheduling is overcome, and when the communication of the cloud platform is interrupted or invalid, the cloud platform can be effectively optimized. Each fog node can guide autonomous operation in the jurisdiction based on the latest regional model; and when the fog node hierarchy fails, the edge node can still independently carry out local optimization, so that the robustness and reliability of the system under the abnormal working condition are enhanced.

Description

Technical Field

[0001] This application relates to the field of power system dispatching technology, and in particular to an optimized dispatching method, system, equipment, computer-readable storage medium, and computer program product for a virtual power plant based on a cloud-fog-edge architecture. Background Technology

[0002] With the advancement of new power system construction, a massive amount of distributed resources, such as distributed photovoltaic power, energy storage, and flexible loads, have been connected to the distribution network. Virtual power plants, as a key technology for aggregating and coordinating these resources, are of great significance for improving grid flexibility and promoting the consumption of new energy sources.

[0003] Currently, the optimal scheduling of virtual power plants mainly relies on a centralized control model. This model communicates directly with all distributed resources and implements unified scheduling through a central control terminal. However, this architecture has inherent drawbacks: First, the central control terminal needs to process massive amounts of data and solve high-dimensional optimization problems, facing enormous computational pressure and making it difficult to meet real-time scheduling requirements; second, the entire system is highly dependent on continuous and reliable communication between the central control terminal and all distributed resources. Once communication is interrupted or the central server fails, the optimal scheduling function of the entire system will be paralyzed, resulting in poor robustness. Summary of the Invention

[0004] This application provides an optimized scheduling method, system, device, computer-readable storage medium, and computer program product for a virtual power plant based on a cloud-fog-edge architecture. It can achieve multi-level local area self-optimization control and global overall optimization control in a tiered and progressive manner. In the event of communication disconnection or partial disconnection, it can achieve partitioned and hierarchical self-optimization control, thereby improving the robustness of the entire system's optimized scheduling function.

[0005] In a first aspect, embodiments of this application provide an optimized scheduling method for a virtual power plant based on a cloud-fog-edge architecture, wherein the cloud-fog-edge architecture includes a cloud platform, fog nodes, and edge nodes that communicate and interact sequentially; the method includes:

[0006] The edge node receives the operating status information of the devices under its jurisdiction; using a local prediction model, it performs source-load prediction based on the operating status information to obtain the power output prediction curve and the load demand prediction curve; based on the local source-load prediction results, combined with electricity price, carbon emissions and grid operation constraints, it constructs and solves a local scheduling model, generates a first scheduling strategy and sends it to the devices under its jurisdiction for execution; and uploads the model parameters of the local prediction model and the local source-load prediction results to the fog node to which it belongs.

[0007] The fog node receives model parameters and local source load prediction results uploaded by each edge node within its jurisdiction; based on the received model parameters, it performs aggregation training using a federated averaging algorithm to generate a regional optimized prediction model; it receives a global optimized model from the cloud platform; it updates the regional optimized prediction model using the global optimized model; it distributes the updated regional optimized prediction model to each edge node under its jurisdiction so that each edge node updates its local prediction model; it receives a global coordination instruction from the cloud platform, the global coordination instruction containing boundary conditions generated based on the global optimized model to constrain the operation of the region; based on the received local source load prediction results of each edge node and the global coordination instruction, it performs regional-level collaborative optimization calculations to generate a second scheduling strategy for coordinating each edge node; it distributes the second scheduling strategy to each edge node under its jurisdiction to guide each edge node in generating the first scheduling strategy;

[0008] The cloud platform receives model parameters from the regional optimization prediction model of each fog node; based on the received model parameters, it performs global federated aggregation and training to generate a global optimization model; based on the global optimization model, it generates a global coordination instruction to be sent to each fog node; and it sends the global optimization model and the global coordination instruction to each fog node.

[0009] Secondly, embodiments of this application provide an optimized scheduling system for a virtual power plant based on a cloud-fog-edge architecture, which runs the optimized scheduling method for a virtual power plant based on a cloud-fog-edge architecture as described in the first aspect of this application. The system includes a cloud platform, fog nodes, and edge nodes that communicate and interact sequentially.

[0010] The edge node is used to receive the operating status information of the end-side equipment under its jurisdiction; using a local prediction model, it performs source-load prediction based on the operating status information to obtain the power output prediction curve and the load demand prediction curve; based on the local source-load prediction results, combined with electricity price, carbon emissions and grid operation constraints, it constructs and solves a local scheduling model, generates a first scheduling strategy and sends it to the end-side equipment under its jurisdiction for execution; and uploads the model parameters of the local prediction model and the local source-load prediction results to the fog node to which it belongs.

[0011] The fog node is used to receive model parameters and local source load prediction results uploaded by each edge node within its jurisdiction; based on the received model parameters, it performs aggregation training using a federated averaging algorithm to generate a regional optimized prediction model; it receives a global optimized model from the cloud platform; it updates the regional optimized prediction model using the global optimized model; it distributes the updated regional optimized prediction model to each edge node under its jurisdiction so that each edge node updates its local prediction model; it receives a global coordination instruction from the cloud platform, the global coordination instruction containing boundary conditions generated based on the global optimized model to constrain the operation of the region; based on the received local source load prediction results of each edge node and the global coordination instruction, it performs regional-level collaborative optimization calculations to generate a second scheduling strategy for coordinating each edge node; and it distributes the second scheduling strategy to each edge node under its jurisdiction to guide each edge node in generating the first scheduling strategy.

[0012] The cloud platform is used to receive model parameters from the regional optimization prediction model from each fog node; perform global federated aggregation and training based on the received model parameters to generate a global optimization model; generate a global coordination instruction to be sent to each fog node based on the global optimization model; and send the global optimization model and the global coordination instruction to each fog node.

[0013] Thirdly, embodiments of this application provide a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the optimized scheduling method for a virtual power plant based on a cloud-fog-edge architecture provided in the first aspect of this application.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the optimized scheduling method for a virtual power plant based on a cloud-fog-edge architecture provided in the first aspect of embodiments of this application.

[0015] Fifthly, embodiments of this application provide a computer program product on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the steps of the optimized scheduling method for a virtual power plant based on a cloud-fog edge architecture provided in the first aspect of embodiments of this application.

[0016] The technical solution provided in this application overcomes the inherent single-point failure risk of traditional centralized scheduling by constructing a three-level collaborative optimized scheduling architecture of cloud platform, fog nodes, and edge nodes, and by endowing edge nodes with local optimization capabilities, fog nodes with regional coordination capabilities, and the cloud platform with global optimization capabilities. When cloud platform communication is interrupted or fails, each fog node can guide autonomous operation within its jurisdiction based on the latest regional model; when the fog node level fails, the edge nodes can still independently perform local optimization, thus realizing a fundamental transformation from "total collapse due to central paralysis" to "local autonomy in the event of global failure," greatly enhancing the robustness of the system under abnormal operating conditions. This improves reliability; and decomposes the massive computing tasks of traditional centralized processing into three parallel processing layers: local optimization at edge nodes, regional collaboration at fog nodes, and global optimization at the cloud platform. This greatly reduces the computing load on single points and the communication pressure on the cloud, making it possible to schedule ultra-large-scale distributed power resources in real time and with fine precision, and significantly enhancing system scalability. The cloud platform sets constraints for the autonomous optimization of the lower layers by issuing global coordination commands, so that while fog nodes and edge nodes are autonomously optimizing and scheduling, their optimization and scheduling decisions can be guided to the most favorable direction for the whole system, ensuring the maximization of the overall optimization and scheduling benefits of the system. Attached Figure Description

[0017] Figure 1 A schematic diagram of a cloud edge architecture provided in an embodiment of this application;

[0018] Figure 2 A schematic diagram of a distributed or adjustable resource provided in an embodiment of this application;

[0019] Figure 3 A flowchart illustrating an optimized scheduling method for a virtual power plant based on a cloud-fog edge architecture, provided in an embodiment of this application.

[0020] Figure 4 A schematic diagram of the dynamic reconfiguration of a self-balancing unit provided in an embodiment of this application;

[0021] Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application. Those skilled in the art can make adjustments as needed to suit specific application scenarios. Furthermore, it should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not the entire structure.

[0023] Figure 1 This is a schematic diagram of a cloud-fog edge architecture provided in an embodiment of this application. Figure 1 As shown, this cloud-fog-edge architecture can include a cloud platform, fog nodes, and edge nodes that communicate and interact sequentially. The cloud platform communicates with one or more fog nodes within its jurisdiction, each fog node communicates with one or more edge nodes within its jurisdiction, and each edge node communicates with at least one end-side device. The edge nodes are primarily used for local optimization and scheduling of end-side devices, the fog nodes are primarily used for coordinated aggregation of multiple edge nodes within their jurisdiction, and the cloud platform is primarily used for cross-regional global optimization and knowledge sharing.

[0024] In this embodiment, as Figure 2 As shown, end-side equipment may include at least one of the following: various distributed power generation devices on the source side (such as waste heat, solar thermal, photovoltaic, hydrogen, etc.), multi-energy (cold-heat-electricity-gas-vapor-hydrogen) transmission and distribution networks, multi-energy load-side equipment (such as charging piles, lighting, air conditioning and other loads, etc.), and multi-element heterogeneous (cold-heat-electricity-gas-vapor-hydrogen) energy storage devices, etc.

[0025] Next, we will discuss the specific process of optimizing the scheduling of virtual power plants using the cloud-edge architecture described above:

[0026] Figure 3 This is a flowchart illustrating an optimized scheduling method for a virtual power plant based on a cloud-fog edge architecture, provided in an embodiment of this application. Figure 3 As shown, the method may include:

[0027] S301. The edge node receives the operating status information of the end-side equipment under its jurisdiction; using the local prediction model, it performs source-load prediction based on the operating status information to obtain the power output prediction curve and load demand prediction curve; based on the local source-load prediction results, combined with electricity price, carbon emission and grid operation constraints, it constructs and solves the local scheduling model, generates the first scheduling strategy and sends it to the end-side equipment under its jurisdiction for execution; and uploads the model parameters of the local prediction model and the local source-load prediction results to the fog node to which it belongs.

[0028] S302. The fog node receives model parameters and local source-load prediction results uploaded by each edge node within its jurisdiction; based on the received model parameters, it performs aggregation training using a federated averaging algorithm to generate a regional optimized prediction model; it receives a global optimized model from the cloud platform; it updates the regional optimized prediction model using the global optimized model; it distributes the updated regional optimized prediction model to each edge node under its jurisdiction so that each edge node updates its local prediction model; it receives a global coordination instruction from the cloud platform, which includes boundary conditions generated based on the global optimized model to constrain the operation of the region; based on the received local source-load prediction results of each edge node and the global coordination instruction, it performs regional-level collaborative optimization calculations to generate a second scheduling strategy for coordinating each edge node; it distributes the second scheduling strategy to each edge node under its jurisdiction to guide each edge node to update the first scheduling strategy.

[0029] S303. The cloud platform receives model parameters from the regional optimization prediction model of each fog node; based on the received model parameters, it performs global federated aggregation and training to generate a global optimization model; based on the global optimization model, it generates a global coordination instruction to be sent to each fog node; and it sends the global optimization model and the global coordination instruction to each fog node.

[0030] In this embodiment, the aforementioned cloud-fog edge architecture may further include a data acquisition module and a data processing and storage module. Figure 1 (Not shown in the image) The data acquisition module is used to collect operating parameters of various distributed power supply devices and mains power on the source side, operating parameters of multi-energy (cold-heat-electricity-gas-gas-hydrogen) transmission and distribution networks, operating parameters of multi-energy load-side equipment, and operating parameters of multi-element heterogeneous (electricity-heat-cold-gas-hydrogen) energy storage devices. The data processing and storage module is used to clean, process, and store the data collected by the data acquisition module. The data cleaning process includes analyzing the source, structure, and business attributes of the collected data; establishing backups and version records for the original data; employing deletion, interpolation, or predictive imputation methods based on machine learning models, depending on the missing data ratio and field importance; retaining data and adding missing data markers when necessary; identifying outliers and noise in the data by combining threshold judgment, statistical detection, and intelligent algorithms, and correcting them through smoothing or substitution; identifying and correcting duplicate records, multi-source conflicts, and redundant data; performing quality verification on the cleaned data; and marking the corrected or imputed data. The data governance process involves establishing unified data standards and metadata management mechanisms; standardizing the definition of data fields, units, sampling frequencies, and encodings; constructing a data quality monitoring system; monitoring data in real time through integrity, accuracy, timeliness, and consistency indicators; triggering alarms in abnormal situations; and using master data management methods to unify equipment, energy, and user identifiers to avoid redundancy and conflicts across systems.

[0031] Edge nodes can train local prediction models using historical operational data from their subordinate end-devices. The specific training process may include: using multi-source historical operational data from the data processing and storage module, constructing a local prediction model using a deep learning-based temporal prediction network (such as a recurrent neural network or a temporal convolutional network); normalizing the multi-source historical operational data in the data processing and storage module, and dividing it into training, validation, and test sets according to time order; constructing features from the multi-source historical operational data, including external variables such as trends, seasonality, weather, and operating conditions, and converting them into input tensors that the model can recognize; using the constructed local prediction model, iteratively updating model parameters on the training set through backpropagation and gradient descent, adjusting hyperparameters on the validation set, and evaluating the accuracy and stability of the local prediction model on the test set. After training, the trained local prediction model is deployed to the edge nodes.

[0032] Furthermore, when the optimized scheduling cycle arrives, the edge node acquires the operating status information of the devices under its jurisdiction and inputs this information into the local prediction model for source-load prediction, obtaining the power output prediction curve and load demand prediction curve. The edge node combines the local source-load prediction results with electricity prices, carbon emission constraints, and grid operating conditions to construct and solve the local scheduling model, obtaining a first scheduling strategy. This first scheduling strategy is then distributed to the devices under its jurisdiction, thereby controlling the actual operation of each device. For example, in the extreme case where both the fog node and the cloud platform are paralyzed, the edge node can use its self-generated first scheduling strategy to optimize the scheduling of its devices, enhancing the system's robustness and reliability under abnormal operating conditions.

[0033] Furthermore, optionally, edge nodes can upload their local source load prediction results to their respective fog nodes, which then perform fog node-level analysis based on the local source load prediction results. Fog nodes can combine the local source load prediction results of the edge nodes with global coordination instructions from the cloud platform to perform regional-level collaborative optimization calculations, generating a second scheduling strategy for coordinating all edge nodes under their jurisdiction. This second scheduling strategy is then distributed to all edge nodes to guide them in generating a first scheduling strategy. For example, edge nodes can update their first scheduling strategy based on the second scheduling strategy distributed from the upper layer, resulting in a more optimized final scheduling strategy.

[0034] Further, optionally, edge nodes can also upload model parameters of their local prediction models to their respective fog nodes. Fog nodes receive the model parameters from all their subordinate edge nodes, perform aggregation training using a federated averaging algorithm, generate a regional optimized prediction model, and distribute the regional optimized prediction model to all their subordinate edge nodes so that each edge node can update its local prediction model based on the regional optimized prediction model. Optionally, the cloud platform can also receive model parameters from the regional optimized prediction models of each fog node, perform global federated aggregation and training on the model parameters of each regional optimized prediction model, generate a global optimized model, and distribute the global optimized model to each fog node. Fog nodes can also receive a global optimized model from the cloud platform. The global optimized model is obtained by aggregating and training the regional optimized prediction models of each fog node. Fog nodes update their regional optimized prediction models using the global optimized model from the cloud platform and distribute the updated regional optimized prediction model to all their subordinate edge nodes so that each edge node can update its local prediction model based on the updated regional optimized prediction model.

[0035] In this embodiment, by having the lower layer report model parameters and the upper layer aggregate model parameters and optimize the model through an interaction mechanism, the local experience of edge nodes is aggregated into a more intelligent global optimization model at fog nodes and the cloud platform through federated learning, and then fed back to the lower layer for decision-making, thereby improving the prediction accuracy of the entire system and thus improving the accuracy of optimization scheduling.

[0036] Furthermore, optionally, the cloud platform can also combine the global optimization model, the real-time operating status information of all edge devices, and the optimization objectives to generate a global coordination command, and then distribute the global coordination command to each fog node. Each fog node receives the global coordination command from the cloud platform, which constrains the boundary conditions of the area under its jurisdiction. Each fog node combines the local source-load prediction results of its subordinate edge nodes with the global coordination command to perform regional-level collaborative optimization calculations, generating a second scheduling strategy for coordinating the edge nodes, and then distributes the second scheduling strategy to its subordinate edge nodes. Each edge node uses the second scheduling strategy as a reference, combining the local source-load prediction results with electricity prices, carbon emissions, and grid operation constraints to generate a first scheduling strategy. This approach allows edge nodes to autonomously optimize their scheduling while their optimization decisions are guided towards the most globally beneficial direction, ensuring the maximization of the overall system optimization scheduling benefits.

[0037] The optimized scheduling method for virtual power plants based on a cloud-fog-edge architecture provided in this application overcomes the inherent single-point failure risk of traditional centralized scheduling by constructing a three-level collaborative optimized scheduling architecture of cloud platform, fog nodes, and edge nodes, and endowing edge nodes with local optimization capabilities, fog nodes with regional coordination capabilities, and the cloud platform with global optimization capabilities. When cloud platform communication is interrupted or fails, each fog node can guide autonomous operation within its jurisdiction based on the latest regional model; when the fog node level fails, the edge nodes can still independently perform local optimization, thus achieving a fundamental shift from "total collapse due to central paralysis" to "local autonomy in the event of global failure," greatly enhancing the system's resilience in heterogeneous environments. The system exhibits robustness and reliability under normal operating conditions. Furthermore, it decomposes the massive computing tasks of traditional centralized processing into three parallel processing layers: local optimization at edge nodes, regional collaboration at fog nodes, and global optimization at the cloud platform. This significantly reduces the computing load on individual points and the communication pressure on the cloud, making real-time and refined scheduling of ultra-large-scale distributed power resources possible and significantly enhancing system scalability. By issuing global coordination commands, the cloud platform sets constraints for the autonomous optimization at the lower level, ensuring that while fog nodes and edge nodes are autonomously optimizing their scheduling, their optimization decisions can be guided to the most favorable direction for the whole system, thus maximizing the overall optimization scheduling benefits of the system.

[0038] In one embodiment, optionally, edge nodes can generate a first scheduling strategy for day-ahead planning based on local source-load forecasting results using constrained optimization and multi-objective genetic algorithms, following a hierarchical control and complementary advantages approach; and / or, generate a first scheduling strategy for intraday adjustment using a distributed robust model predictive control algorithm based on local source-load forecasting results; and / or, generate a first scheduling strategy for real-time control using near-end policy optimization and multi-agent reinforcement learning algorithms based on local source-load forecasting results. Day-ahead planning coordinates multiple objectives such as economic efficiency and environmental friendliness on a macro-timescale, providing a benchmark for subsequent scheduling. Intraday adjustment, based on updated ultra-short-term forecasts, corrects deviations and refines the day-ahead plan to address foreseeable fluctuations. Real-time control, acting as the "execution operation," responds to unforeseen instantaneous disturbances at the second / minute level, ensuring real-time grid balance and security.

[0039] In one embodiment, the edge node may optionally compare the local source load prediction results with the actual operating data of the end-side devices under its jurisdiction periodically to obtain the prediction deviation; if the prediction deviation is greater than or equal to a preset threshold, it may trigger incremental learning or retraining of the local prediction model to optimize and update the local prediction model, thereby improving the prediction accuracy of the local prediction model.

[0040] In one embodiment, optionally, the fog node can also perform privacy protection processing on the model parameters of the regional optimization prediction model, and upload the privacy-protected model parameters to the cloud platform for model aggregation to generate a global optimization model. The privacy protection processing method includes at least one of encryption, masking, or differential privacy. This privacy protection processing ensures the security of data during the collaborative optimization scheduling process.

[0041] Optionally, fog nodes can also interact laterally with neighboring fog nodes based on a federated learning framework. Specifically, this includes: a fog node sending its own regional optimization prediction model parameters, after privacy protection processing, to neighboring fog nodes; simultaneously, receiving similarly processed model parameters from neighboring fog nodes; subsequently, based on the received neighboring node model parameters, the fog node uses a federated averaging algorithm for local aggregation, generating a more accurate collaborative prediction model covering its own region and neighboring regions. During normal system operation, this collaborative prediction model can optimize the power exchange plan between the calculation and neighboring regions, achieving local power balance and mutual assistance, thereby reducing dependence on the cloud platform and improving response speed. When a communication interruption with the cloud platform is detected, the global optimization model is used as the basis for generating a second scheduling strategy, thus achieving regional autonomy and collaboration in failure scenarios. Here, neighboring fog nodes refer to fog nodes that have communication connections with the current fog node and are geographically close. Due to their geographical proximity, the resource complementarity between them is stronger. Interaction with neighboring fog nodes can quickly balance local power deficits, resolving regional problems at the fog layer and reducing the load pressure on the cloud platform.

[0042] In one embodiment, the cloud platform may optionally perform global optimization through the following process:

[0043] like Figure 4 As shown, the cloud platform can divide the fog nodes under its jurisdiction into one or more self-balancing units (such as...) based on the power grid topology, resource distribution, and operational objectives. Figure 4 The cloud platform decomposes the global optimization problem into subproblems corresponding to their respective balancing units, and uses the alternating direction multiplier method or the distributed Lagrange relaxation method to coordinate the solution of each subproblem in order to achieve global optimization. Each self-balancing unit contains one or more fog nodes.

[0044] Optionally, the cloud platform can set unit-level optimization targets (e.g., internal power balance and external ancillary service capability indicators) by issuing global coordination instructions to each self-balancing unit, thereby guiding and constraining the fog nodes within the self-balancing unit to perform collaborative optimization. In this way, it ensures that each self-balancing unit, as a whole, can efficiently achieve power matching within the unit and collaboratively respond to the grid's regulation needs and ancillary service requests.

[0045] Optionally, the cloud platform can also model each equilibrium unit as an agent based on a multi-agent reinforcement learning algorithm, and dynamically optimize the coordination and control strategy for each equilibrium unit through game learning among the agents.

[0046] For example, a cloud platform can optimize the coordination and control strategy of its respective balance units through the following process:

[0047] First, the cloud platform models the optimization scheduling problem of the entire virtual power plant as a cooperative multi-agent reinforcement learning problem. Each self-balancing unit is defined as an agent. These agents share a common goal: to minimize the overall operating cost and maximize the benefits of the virtual power plant while ensuring the safe and stable operation of the entire network.

[0048] State space: The state observation information of each agent (self-balancing unit) includes: net power inside the unit, status of key equipment inside the unit (such as the overall state of charge of energy storage), power exchanged between the unit and the main grid, the latest global coordination instructions received, market electricity price, weather forecast and other external environmental information.

[0049] Action Space: The action of each agent is to issue unit-level coordination and control commands to the fog nodes it contains. For example, setting the planned power exchange between the self-balancing unit and the main network in the next time period; setting the capacity of auxiliary services (such as frequency regulation backup) that the self-balancing unit needs to provide; adjusting the weight of the collaborative optimization objectives of its internal resources (such as temporarily emphasizing economy or balancing).

[0050] Reward Function: The reward function guides the agent's behavior towards the global optimum. Its core can consist of two parts: Individual Reward: Based on the performance of each self-balancing unit, such as the accuracy of instruction execution and internal operating costs. Team Reward: All agents share a reward based on the global operating state of the virtual power plant, such as penalties for grid-wide power balance deviations, total operating costs, and total carbon emissions. Team rewards force each agent to consider its impact on the global system and other agents, rather than solely pursuing its own interests.

[0051] Furthermore, the cloud platform acts as a central trainer, running multi-agent reinforcement learning algorithms. During the training phase, each agent (self-balancing unit) uploads its state, actions, and rewards to the cloud platform. The cloud platform maintains a policy network for each agent and a commentator network that can evaluate the value of actions from a global perspective. The commentator network can assess the merits of an agent's actions in the global state, thereby guiding its policy updates. After training is complete, each agent's policy network is deployed to the cloud platform. During runtime, the cloud platform, based on the current global state, calls the policy networks of each agent in parallel to generate a complete set of coordinated control policies, which are then distributed to their respective balancing units for execution. Through this process, the agents in each balancing unit, under the coordination of the cloud platform, continuously engage in games with the environment and other agents. For example, if a unit discovers that its own photovoltaic output is excessive while neighboring units have high loads, it will learn that "offering to sell electricity to that unit" is an action that yields a high reward. Through millions of such game-based learning iterations, the cloud platform can dynamically and intelligently coordinate the optimized scheduling strategies of each balancing unit.

[0052] In this embodiment, by introducing multi-agent reinforcement learning, the optimal scheduling strategy can be learned through trial and error in uncertain environments, which can better cope with the random fluctuations of renewable energy and load. Through the "team reward" mechanism, each agent can achieve global optimization scheduling collaboration while pursuing its own reward.

[0053] Furthermore, the cloud platform can periodically evaluate the operational performance of each self-balancing unit. This performance includes at least one of the following: power balance within the self-balancing unit, responsiveness to higher-level scheduling commands, and resource adjustment potential. If the operational performance of each self-balancing unit meets preset reconfiguration conditions, the platform reconfigures the self-balancing units of the fog nodes under its jurisdiction. (See also...) Figure 4 , Figure 4 Region 2 of the dynamic self-balancing unit can be the reconstructed self-balancing unit. After reconstruction, the cloud platform generates and issues new global coordination instructions and global optimization models for the new self-balancing unit to guide the fog nodes contained in the new self-balancing unit to perform regional-level collaborative optimization calculations.

[0054] Specifically, power balance refers to the ability of a self-balancing unit to balance its internal power within a statistical period. The ability to respond to upper-level dispatch instructions can be comprehensively measured by indicators such as instruction completion rate and response delay. Resource adjustment potential refers to the total capacity of dispatchable resources (such as energy storage and interruptible loads) within the self-balancing unit, which can be quantified by calculating the maximum upward adjustment potential and the maximum downward adjustment potential.

[0055] When the balancing capacity of a self-balancing unit remains low, its response capability is insufficient, or its regulation potential cannot meet the needs of balancing or response commands, the cloud platform can reconfigure the self-balancing unit. The cloud platform can combine the evaluation results of the operational efficiency of each balancing unit to determine the target unit and reconfiguration objectives (such as splitting, merging, or adjusting boundaries). With the goal of improving overall efficiency and using grid topology connectivity as a basic constraint, the platform treats all fog nodes in the entire network as allocable elements and uses clustering or community discovery algorithms to find the optimal self-balancing unit partitioning scheme. The optimization objective function of the algorithm can be to maximize the internal resource complementarity of the self-balancing unit, minimize the electrical distance between fog nodes within the self-balancing unit, and balance the regulation potential and load scale of each balancing unit. After obtaining the reconfigured new self-balancing unit, the cloud platform can combine the historical data of the new self-balancing unit to quickly train or fine-tune a global optimization model specific to that new self-balancing unit through federated learning, and then distribute it to each fog node within the new self-balancing unit, issuing new global coordination commands to the new self-balancing unit. After receiving the new global optimization model and the new global coordination instructions, the new self-balancing unit takes them as new constraints and optimization objectives, performs regional-level collaborative optimization calculations within the new boundary, and generates a new second scheduling strategy, thereby ensuring that the reconstructed new self-balancing unit can be put into operation quickly and efficiently.

[0056] Through the aforementioned dynamic reconfiguration mechanism, fog nodes that are geographically and electrically close and have complementary resource characteristics can be proactively identified and integrated. For example, merging fog node A, which is mainly photovoltaic (PV) in the jurisdiction, with fog node B, which is mainly load (HV), into a new self-balancing unit allows for local consumption of PV power generation within this unit. This reduces long-distance power transmission losses and fluctuations in power exchange with the upper-level grid, thereby directly reducing operating costs and improving the local consumption rate of new energy. Furthermore, in the face of long-term changes in grid structure, load growth, and resource distribution, regular evaluation and reconfiguration ensure that the internal architecture of the virtual power plant always matches the external environment and operational objectives. This enables the system to overcome the problem of static partitioning schemes gradually becoming ineffective over time, thus guaranteeing the long-term economic efficiency, reliability, and high efficiency of the virtual power plant.

[0057] In one embodiment, see Figure 1 Furthermore, an optimized scheduling system for a virtual power plant based on a cloud-fog-edge architecture is provided. This system can run on the optimized scheduling method for a virtual power plant based on a cloud-fog-edge architecture provided in any of the above embodiments. The system may include a cloud platform, fog nodes, and edge nodes that communicate and interact sequentially.

[0058] Specifically, edge nodes are used to receive the operating status information of the end-side equipment under their jurisdiction; using a local prediction model, they perform source-load prediction based on the operating status information to obtain the power output prediction curve and load demand prediction curve; based on the local source-load prediction results, combined with electricity price, carbon emissions and grid operation constraints, they construct and solve the local scheduling model, generate the first scheduling strategy and send it to the end-side equipment under their jurisdiction for execution; and upload the model parameters of the local prediction model and the local source-load prediction results to the fog node to which they belong.

[0059] Fog nodes are used to receive model parameters and local source-load prediction results uploaded by edge nodes within their jurisdiction; based on the received model parameters, they perform aggregation training using a federated averaging algorithm to generate a regional optimized prediction model; they receive a global optimized model from the cloud platform; they update the regional optimized prediction model using the global optimized model; they distribute the updated regional optimized prediction model to each edge node under their jurisdiction so that each edge node updates its local prediction model; they receive global coordination instructions from the cloud platform, which include boundary conditions generated based on the global optimized model to constrain the operation of the region; based on the received local source-load prediction results of each edge node and the global coordination instructions, they perform regional-level collaborative optimization calculations to generate a second scheduling strategy for coordinating each edge node; and they distribute the second scheduling strategy to each edge node under their jurisdiction to guide each edge node in generating a first scheduling strategy.

[0060] The cloud platform is used to receive model parameters from the regional optimization prediction model from each fog node; based on the received model parameters, it performs global federated aggregation and training to generate a global optimization model; based on the global optimization model, it generates global coordination instructions to be distributed to each fog node; and it distributes the global optimization model and global coordination instructions to each fog node.

[0061] In one embodiment, the edge node is optionally further configured to generate a first scheduling strategy for the day-ahead plan based on the local source load prediction results through constraint optimization and a multi-objective genetic algorithm; generate a first scheduling strategy for intraday adjustment based on the local source load prediction results through a distributed robust model predictive control algorithm; and generate a first scheduling strategy for real-time control based on the local source load prediction results through near-end policy optimization and a multi-agent reinforcement learning algorithm.

[0062] In one embodiment, the edge node is optionally configured to periodically compare the local source load prediction results with the actual operating data of the end-side devices under its jurisdiction to obtain the prediction deviation; if the prediction deviation is greater than or equal to a preset threshold, it triggers incremental learning or retraining of the local prediction model to optimize and update the local prediction model.

[0063] In one embodiment, optionally, the fog node is also used to perform privacy protection processing on the model parameters of the regional optimization prediction model, the privacy protection processing including at least one of encryption, masking or differential privacy; upload the privacy-protected model parameters to the cloud platform for the cloud platform to perform model aggregation to generate a global optimization model; and / or upload the privacy-protected model parameters to neighboring fog nodes for collaborative optimization calculations between regions.

[0064] In one embodiment, optionally, the cloud platform is further configured to divide the fog nodes under its jurisdiction into one or more self-balancing units based on the power grid topology, resource distribution and operational objectives, and set unit-level optimization objectives and constraints for each self-balancing unit; wherein each self-balancing unit contains one or more fog nodes; the global optimization problem is decomposed into sub-problems corresponding to their respective balancing units, and the alternating direction multiplier method or the distributed Lagrange relaxation method is used to coordinate the solution of each sub-problem in order to achieve global optimization.

[0065] In one embodiment, optionally, the cloud platform is also used to model each balance unit as an agent based on a multi-agent reinforcement learning algorithm, and dynamically optimize the coordination control strategy for each balance unit through game learning among the agents.

[0066] In one embodiment, optionally, the cloud platform is also used to periodically evaluate the operational performance of each self-balancing unit, including at least one of power balance within the self-balancing unit, responsiveness to upper-level scheduling instructions, and resource adjustment potential; when the operational performance of each self-balancing unit meets the preset reconfiguration conditions, the cloud platform reconfigures the fog nodes under its jurisdiction; after reconfiguration, the cloud platform generates and issues new global coordination instructions and global optimization models for the new self-balancing unit to guide the fog nodes contained in the new self-balancing unit to perform regional-level collaborative optimization calculations.

[0067] Figure 5 A schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 5 As shown, the computer device includes a processor 50, a memory 51, an input device 52, and an output device 53; the number of processors 50 in the computer device can be one or more. Figure 5 Taking a processor 50 as an example; the processor 50, memory 51, input device 52, and output device 53 in this computer device can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0068] The memory 51, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules (e.g., edge nodes, fog nodes, and cloud platforms) corresponding to the optimized scheduling method of the virtual power plant based on the cloud-fog-edge architecture in the embodiments of this application. The processor 50 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 51, thereby implementing the optimized scheduling method of the virtual power plant based on the cloud-fog-edge architecture provided in any of the above embodiments.

[0069] The memory 51 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function. The data storage area may store data involved in the optimized scheduling process of the virtual power plant based on the cloud-fog-edge architecture provided in the above embodiments. Furthermore, the memory 51 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 51 may further include memory remotely located relative to the processor 50, which can be connected to devices / terminals / servers via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0070] Input device 52 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the computer device. Output device 53 may include display devices such as a display screen.

[0071] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the optimized scheduling method for a virtual power plant based on a cloud-fog edge architecture provided in any of the above embodiments.

[0072] In one embodiment, a computer program product is also provided, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the optimized scheduling method for a virtual power plant based on a cloud-fog edge architecture provided in any of the above embodiments.

[0073] The optimized scheduling system, computer equipment, computer-readable storage medium, and computer program product based on cloud-fog edge architecture of the virtual power plant provided in the above embodiments can execute the optimized scheduling method of the virtual power plant based on cloud-fog edge architecture provided in any embodiment of this application, and have the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in the above embodiments can be found in the optimized scheduling method of the virtual power plant based on cloud-fog edge architecture provided in any embodiment of this application.

[0074] Based on the above description of the implementation methods, those skilled in the art can clearly understand that this application can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0075] It is worth noting that the units and modules included in the above embodiments are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of this application.

[0076] Note that the above description is merely a preferred embodiment and the technical principles employed in this application. Those skilled in the art will understand that this application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments. Many other equivalent embodiments may be included without departing from the concept of this application, and the scope of this application is determined by the scope of the appended claims.

Claims

1. An optimized scheduling method for virtual power plants based on a cloud-fog edge architecture, characterized in that, The cloud-fog-edge architecture comprises a cloud platform, a fog node and an edge node which are sequentially communicated; the method comprises: The edge node receives the running state information of the terminal-side equipment under jurisdiction; uses a local prediction model to perform source-load prediction based on the running state information to obtain a power output prediction curve and a load demand prediction curve; based on the local source-load prediction result, combined with the electricity price, carbon emissions and power grid operation constraints, a local scheduling model is constructed and solved to generate a first scheduling strategy and is delivered to the terminal-side equipment under jurisdiction for execution; the model parameters of the local prediction model and the local source-load prediction result are uploaded to the fog node to which the edge node belongs; The fog node receives the model parameters and the local source-load prediction result uploaded by each edge node in the jurisdiction; based on the received model parameters, a federal average algorithm is used for aggregated training to generate a regional optimization prediction model; a global optimization model is received from the cloud platform; the global optimization model is used to update the regional optimization prediction model; the updated regional optimization prediction model is delivered to each edge node under jurisdiction to enable each edge node to update its local prediction model; a global coordination instruction is received from the cloud platform, the global coordination instruction includes boundary conditions generated based on the global optimization model for constraining the operation of the region; based on the received local source-load prediction result of each edge node and the global coordination instruction, regional-level collaborative optimization calculation is performed to generate a second scheduling strategy for coordinating each edge node; the second scheduling strategy is delivered to each edge node under jurisdiction to guide each edge node to generate the first scheduling strategy; The cloud platform receives the model parameters of the regional optimization prediction model from each fog node; based on the received model parameters, global federal aggregation and training are performed to generate a global optimization model; based on the global optimization model, a global coordination instruction is generated and delivered to each fog node; the global optimization model and the global coordination instruction are delivered to each fog node.

2. The method of claim 1, wherein, The process of generating the first scheduling strategy by the edge node comprises: Based on the local source-load prediction result, a first scheduling strategy of day-ahead planning is generated by constraint optimization and a multi-objective genetic algorithm; And / or, based on the local source-load prediction result, a first scheduling strategy of intra-day adjustment is generated by a distributed robust model predictive control algorithm; And / or, based on the local source-load prediction result, a first scheduling strategy of real-time control is generated by a proximal policy optimization and multi-agent reinforcement learning algorithm.

3. The method of claim 1, wherein, The method further comprises: The edge node periodically compares the local source-load prediction result with the actual operation data of the terminal-side equipment under jurisdiction to obtain a prediction deviation; In the case that the prediction deviation is greater than or equal to a preset threshold, the edge node triggers incremental learning or retraining of the local prediction model to optimize and update the local prediction model.

4. The method of claim 1, wherein, The method further comprises: The fog node performs privacy protection processing on the model parameters of the regional optimization prediction model, the privacy protection processing comprising at least one of encryption, masking or differential privacy; The fog node uploads the model parameters after privacy protection processing to the cloud platform for the cloud platform to generate a global optimization model through model aggregation; and / or, the fog node uploads the model parameters after privacy protection processing to a neighboring fog node for collaborative optimization calculation between regions.

5. The method of claim 1, wherein, The method further comprises: The cloud platform divides the fog nodes under its jurisdiction into one or more self-balancing units based on the power grid topology, resource distribution, and operation target, and sets an optimization target and constraint condition at the unit level for each self-balancing unit; wherein each self-balancing unit contains one or more fog nodes; The cloud platform decomposes the global optimization problem into sub-problems corresponding to each self-balancing unit, and uses an alternating direction multiplier method or a distributed Lagrange relaxation method to coordinate and solve each sub-problem to achieve global optimization.

6. The method of claim 5, wherein, The method further comprises: The cloud platform models each self-balancing unit as an agent based on a multi-agent reinforcement learning algorithm, and dynamically optimizes the coordination control strategy for each self-balancing unit through game learning between agents.

7. The method of claim 5, wherein, The method further comprises: The cloud platform periodically evaluates the operation efficiency of each self-balancing unit, which includes at least one of power balance degree, response ability to superior dispatching instructions, and resource regulation potential within the self-balancing unit; The cloud platform reconfigures the self-balancing units of the fog nodes under its jurisdiction when the operation efficiency of each self-balancing unit meets a preset reconfiguration condition; After reconfiguration, the cloud platform generates and issues new global coordination instructions and a global optimization model for the new self-balancing unit to guide each fog node included in the new self-balancing unit to perform regional-level collaborative optimization calculation.

8. An optimal scheduling system of a virtual power plant based on a cloud-fog edge architecture, characterized in that, An optimization scheduling method of a virtual power plant based on a cloud-fog-edge architecture, the system comprising a cloud platform, a fog node, and an edge node that communicate with each other in sequence; The edge node is configured to receive operation state information of the terminal-side devices under its jurisdiction, use a local prediction model to perform source-load prediction based on the operation state information to obtain a power source output prediction curve and a load demand prediction curve, construct and solve a local scheduling model based on the local source-load prediction result, combined with electricity price, carbon emissions, and power grid operation constraints, generate a first scheduling strategy and issue it to the terminal-side devices under its jurisdiction for execution, and upload the model parameters of the local prediction model and the local source-load prediction result to the fog node to which the edge node belongs. The fog node is configured to receive model parameters and local source load prediction results uploaded by each edge node within the jurisdiction, aggregate and train based on the received model parameters using a federated average algorithm to generate a regional optimization prediction model, receive a global optimization model from a cloud platform, update the regional optimization prediction model using the global optimization model, distribute the updated regional optimization prediction model to each edge node within the jurisdiction to enable each edge node to update its local prediction model, receive a global coordination instruction from the cloud platform, the global coordination instruction including boundary conditions generated based on the global optimization model for constraining the operation of the region, perform regional-level collaborative optimization calculation based on the received local source load prediction results of each edge node and the global coordination instruction to generate a second scheduling strategy for coordinating each edge node, and distribute the second scheduling strategy to each edge node within the jurisdiction to guide each edge node to generate the first scheduling strategy. The cloud platform is configured to receive model parameters of the regional optimization prediction model from each fog node, perform global federated aggregation and training based on the received model parameters to generate a global optimization model, generate a global coordination instruction based on the global optimization model and distribute the global coordination instruction to each fog node, and distribute the global optimization model and the global coordination instruction to each fog node. The computer program, when executed by a processor, implements the steps of the method of any one of claims 1 to 7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the method of any one of claims 1 to 7.

10. A computer program product having stored thereon a computer program, characterized in that, ​

Images