Mass virtual power plant flexible collaboration method and system fusing mechanism knowledge

Abstract

The invention discloses a massive virtual power plant flexible collaboration method and system fusing mechanism knowledge, and the method comprises the steps: processing global information through Encoder and Decoder, and integrating all Agent data; each virtual power plant Agent inputs local and global integration information, the local information comprises adjustable capacity and cost, and the global information is an adjustment value. The adjustment value calculation is fused with next-day load data predicted by the front Encoder and Decoder, the value of each time period is determined based on a load center value, an adjustment threshold value is set to guide VPP actions, value functions of different time periods are defined, a capacity cost function is constructed in combination with a power market bidding result, a total value function is decomposed to each VPP according to the capacity cost function, the capacity value coefficient and the adjustment value of each VPP are determined, and the capacity value coefficient and the adjustment value of each VPP are calculated. And finally, learning a VPP adjustment strategy by using a full-connection neural network, and training an Agent by taking the minimum total peak-valley difference and total cost of the system as targets to achieve flexible collaboration of the virtual power plant. According to the method, the rewards are distributed and adjusted in real time, and centralized Critic is replaced to solve the calculation problem.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence technology and collaborative optimization of virtual power plants, and in particular to a flexible collaboration method and system for massive virtual power plants that integrates mechanism knowledge. Background Art

[0002] With the large-scale integration of renewable energy and the rapid growth of electricity demand, new power systems face severe challenges in energy balance and operational stability. Traditional power systems rely primarily on centralized generation resources to meet load demand, but the high volatility and uncertainty of renewable energy make this model difficult to adapt to the needs of modern power grids. At the same time, load-side resources (such as industrial, commercial, and residential electrical equipment) have great regulatory potential. By aggregating these decentralized load resources into virtual power plants (VPPs), they can effectively participate in the optimization and regulation of the power grid and mitigate the impact of renewable energy volatility.

[0003] As a flexible form of load aggregation, virtual power plants (VPPs) can achieve grid optimization goals such as peak shaving and valley filling and frequency regulation by adjusting load demand in real time. However, the regulation capacity of a single VPP is typically small, making it difficult to significantly impact the system-level load curve. Therefore, a large number of VPPs are required to work together to achieve global optimization. However, the coordinated optimization of massive VPPs faces the following technical challenges:

[0004] Reward allocation problem: In traditional reinforcement learning methods, reward signals are typically distributed based on the final global goal (such as reducing the daily peak-to-valley difference). However, this reward allocation method has a time lag and cannot reflect the contribution of each virtual power plant's adjustment actions in different time periods to the global goal in real time. For example, the traditional temporal difference (TD) method allocates the final reward to the previous time step through a discount factor. However, this method cannot accurately match the relationship between the adjustment action and the actual contribution at each time step, resulting in poor policy optimization results.

[0005] Computational complexity: The collaborative optimization of massive virtual power plants requires processing large amounts of real-time data and complex computational tasks. Existing methods typically use a centralized critic network to calculate global rewards. However, as the number of virtual power plants increases, the computational complexity increases exponentially, making it difficult to meet real-time requirements. Summary of the Invention

[0006] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0007] This paper proposes a flexible collaboration method for massive virtual power plants that integrates mechanism knowledge. By constructing a time-domain value splitting architecture based on mechanism decomposition, it realizes the real-time distribution of adjustment rewards. At the same time, it designs a flexible multi-agent collaboration architecture, breaking through the limitations of traditional methods in computational complexity and communication bottlenecks, and providing a new solution for the efficient collaborative optimization of massive virtual power plants.

[0008] Another object of the present invention is to propose a flexible collaborative system for massive virtual power plants that integrates mechanism knowledge.

[0009] To achieve the above objectives, the present invention proposes a flexible collaboration method for massive virtual power plants that integrates mechanism knowledge, including:

[0010] Build a massive virtual power plant collaborative model; including a global information processing module, a mechanism calculation module and multiple virtual power plant agents;

[0011] Input the multimodal global information into the global information processing module to calculate and obtain the load forecast result;

[0012] The mechanism calculation module is used to integrate the local information of each virtual power plant and the load forecast results of the global information processing module to calculate the global regulation value of the agent;

[0013] A fully connected neural network model is constructed for each virtual power plant agent, and training data is input into the fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the regulation strategy and achieve flexible collaboration of the virtual power plant; wherein the training data includes load forecast results, its own adjustable capacity, unit regulation cost and global regulation value.

[0014] The flexible coordination method for massive virtual power plants integrating mechanism knowledge according to the embodiment of the present invention may also have the following additional technical features:

[0015] In one embodiment of the present invention, the multimodal global information is input into the global information processing module for calculation to obtain the load forecast result, including:

[0016] Acquire multimodal global information, including electricity, meteorological, and social information;

[0017] The global information processing module includes an Encoder model and a Decoder model;

[0018] The global information processing module is used to fuse multimodal global information, predict the load curve of the next day, output the maximum load and the minimum load, and calculate the central value of the daily load.

[0019] In one embodiment of the present invention, the method further includes:

[0020] Use IoT devices to collect load data and regulation costs of virtual power plants in real time to obtain real-time data;

[0021] Input the real-time collected data into the trained fully connected neural network model to output the actual adjustment result of load regulation;

[0022] The actual regulation results are compared with the load forecast results through the feedback mechanism, and the results with deviations greater than the preset threshold are obtained to optimize the network model parameters.

[0023] In one embodiment of the present invention, the mechanism calculation module integrates the local information of each virtual power plant and the load forecast results of the global information processing module to calculate the global regulation value of the agent, including:

[0024] Set correction parameters for adjustment thresholds;

[0025] According to the actual load of each period and the central value of the daily load, the value function of reducing the peak-to-valley difference of the load is calculated based on the correction parameters;

[0026] Based on the value function, the adjustable capacity and unit regulation cost of the i-th virtual power plant VPP in period t on the next day are obtained from the virtual power plant market to calculate the regulation value of the virtual power plant.

[0027] In one embodiment of the present invention, inputting training data into a fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the adjustment strategy includes:

[0028] Assign an agent to each virtual power plant and initialize the neural network parameters of all agents;

[0029] Generate adjustment actions based on the initialized neural network according to the current state and global information;

[0030] The total load adjustment is calculated based on the adjustment actions of all agents, and the immediate reward of each agent is calculated according to the defined training target. The global reward is distributed to each agent and the local reward of each time step through the time domain value splitting model;

[0031] The neural network parameters of each agent are updated based on the local reward to optimize the regulation strategy.

[0032] To achieve the above objectives, the present invention further proposes a flexible collaborative system for massive virtual power plants that integrates mechanism knowledge, including:

[0033] Collaborative model building module, used to build massive virtual power plant collaborative models; including global information processing module, mechanism calculation module and multiple virtual power plant agents;

[0034] The load forecasting module is used to input the multimodal global information into the global information processing module to calculate and obtain the load forecast result;

[0035] The regulation value calculation module is used to calculate the global regulation value of the Agent by integrating the local information of each virtual power plant and the load forecast results of the global information processing module using the mechanism calculation module;

[0036] The regulation strategy optimization module is used to build a fully connected neural network model for each virtual power plant agent, and input training data into the fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the regulation strategy and achieve flexible coordination of the virtual power plant; wherein, the training data includes load forecast results, its own adjustable capacity, unit regulation cost and global regulation value.

[0037] The flexible coordination method and system for massive virtual power plants that integrate mechanism knowledge in the embodiments of the present invention distribute and adjust rewards in real time, replace the centralized critic to solve computational problems, break through the limitations of traditional multi-agent reinforcement learning, efficiently coordinate massive virtual power plants, and ensure the stable operation of the new power system.

[0038] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0040] Figure 1 is a flow chart of a flexible collaboration method for massive virtual power plants integrating mechanism knowledge according to an embodiment of the present invention;

[0041] Figure 2 is an architectural diagram of a flexible collaboration method for massive virtual power plants integrating mechanism knowledge according to an embodiment of the present invention;

[0042] Figure 3 is a flow chart for calculating adjustment value according to an embodiment of the present invention;

[0043] Figure 4 4 is a structural diagram of a flexible collaborative system for massive virtual power plants integrating mechanism knowledge according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0046] The following describes a flexible collaboration method and system for massive virtual power plants integrating mechanism knowledge proposed in accordance with an embodiment of the present invention with reference to the accompanying drawings.

[0047] Figure 1 is a flow chart of a flexible collaborative method for massive virtual power plants integrating mechanism knowledge according to an embodiment of the present invention. Figure 1 As shown, including:

[0048] S1, builds a massive virtual power plant collaborative model; including a global information processing module, a mechanism calculation module and multiple virtual power plant agents;

[0049] S2, inputting the multimodal global information into the global information processing module to calculate and obtain the load forecast result;

[0050] S3, using the mechanism calculation module to integrate the local information of each virtual power plant and the load forecast results of the global information processing module to calculate the global regulation value of the agent;

[0051] S4, constructing a fully connected neural network model for each virtual power plant agent, and inputting training data into the fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the regulation strategy and realize flexible coordination of the virtual power plant; wherein, the training data includes load forecast results, its own adjustable capacity, unit regulation cost and global regulation value.

[0052] Specifically, a flexible collaborative architecture for massive virtual power plants that can be flexibly expanded is designed, such as Figure 2 As shown in Figure 3, the overall architecture consists of two parts, where the Encoder and Decoder are responsible for processing global information, on the one hand giving the load forecast curve, and on the other hand integrating the information of each Agent.

[0053] The embodiment of the present invention designs a collaborative model for a large number of virtual power plants that can be flexibly expanded to participate in peak shaving and valley filling, and adopts a 1+N multi-agent collaborative architecture.

[0054] Among them, the global information processing module (Encoder and Decoder) is responsible for processing global information, including load forecast curve and regulation value calculation of virtual power plant.

[0055] Specifically, the global information processing module integrates multimodal global information such as electricity, meteorology, and society, predicts the load curve for the next day, outputs the maximum load and minimum load, and calculates the central value of the daily load.

[0056] Among them, the mechanism calculation module integrates the local information of each virtual power plant and the load value predicted by the global information processing module, calculates the global regulation value, and distributes the result to each agent.

[0057] Among them, there are multiple virtual power plant agents, and each virtual power plant is assigned an agent. The agents are all neural network decision-making modules in deep reinforcement learning. They input their own local information (such as adjustable capacity, unit adjustment cost) and globally integrated information, and output the adjustment action of the virtual power plant.

[0058] Agent input format: The input of each agent includes local information observed by the virtual power plant (such as adjustable capacity, unit regulation cost) and global information (load value predicted by the global information processing module, regulation value calculated by the mechanism calculation module, etc.), ensuring that the dimension of the input data does not change with the number of virtual power plants.

[0059] Specifically, each virtual power plant agent is isomorphic. In order to achieve flexible collaboration of virtual power plants, the input of each virtual power plant agent is defined as its own local information S l and the globally integrated information S g In the business of virtual power plants participating in peak shaving and valley filling, local information mainly includes adjustable capacity, unit regulation cost, etc., and global information mainly includes the regulation value of the virtual power plant.

[0060] In one embodiment of the present invention, a time domain value splitting architecture based on mechanism decomposition is constructed. The process of calculating using the mechanism calculation module is as follows: Figure 3 shown.

[0061] Understandably, in traditional reinforcement learning methods, reward signals are distributed based on the ultimate global goal (such as reducing the daily peak-to-valley difference). However, this approach has a time lag and cannot reflect the contribution of each virtual power plant's regulatory actions to the global goal in real time. To this end, this paper proposes a time-domain value splitting architecture based on mechanism decomposition, which decomposes the total reward at the end of the day into each time step.

[0062] Specifically, the calculation of the regulation value needs to comprehensively consider the contribution of regulation to the reduction of peak-to-valley difference and the regulation cost. First, based on the front-end Encoder and Decoder for the next day's maximum load and minimum load P f Make a forecast and get the central value of the next day's load

[0063] In each time period, VPP adjustment makes the load demand change towards the daily load center value, which can reduce the peak-to-valley difference. Therefore, the value of the total peak-to-valley difference reduction is It can be transformed into the value of each time step:

[0064]

[0065] in, and P l are the maximum and minimum values ​​of the original daily load curve, and P a are the maximum and minimum values ​​of the daily load curve after VPP adjustment, P t is the load power at time t before VPP participates in regulation.

[0066] Considering the uncertainty of load forecast results and to prevent the VPP from operating too frequently, an adjustment threshold ε is set. When the difference between the load value in period t and the central value is greater than the adjustment threshold ε, the VPP adjustment is valuable, guiding the VPP to operate. The value of reducing the total peak-to-valley difference is rewritten as;

[0067]

[0068] In addition, due to the load P in each period t Therefore, the value of VPP regulation in different periods of time is different. Therefore, the value function of reducing the load peak-valley difference is defined as: V t =-γ|P t -P mid |;

[0069] VPP needs to pay a cost to participate in peak shaving and valley filling regulation. Through bidding in the electricity market, it can obtain the adjustable capacity C of the i-th VPP in the t period of the next day. i,t and unit adjustment cost U i,t For a VPP cluster, the optimization and adjustment goals are:

[0070]

[0071] The bidding results of all VPPs can form the capacity cost function: C t =f(U), which is a monotonically decreasing function. The higher the unit cost, the more adjustable capacity.

[0072] The present invention obtains the total load P through the global information processing module t and load median P mid The overall adjustment value is defined as the peak-to-valley difference of the current period |P t -Pmid |, decomposing the overall value function to each VPP requires knowing the position of the VPP cost in the entire capacity cost function. The capacity value coefficient of the i-th VPP is:

[0073]

[0074] Therefore, the regulation value of the i-th VPP is the product of the overall regulation value and the capacity value:

[0075]

[0076] Furthermore, each virtual power plant agent learns the optimal regulation strategy through a fully connected neural network model, resulting in a trained fully connected neural network model. Its input training data includes the total load for the current period, the predicted median load, its own adjustable capacity, unit regulation cost, and regulation value. The total load and median load for the current period are derived from the large-scale model's load forecast, its own adjustable capacity is sensed by IoT devices, and the unit regulation cost is obtained from the electricity market.

[0077] Specifically, the network structure adopts a multi-layer fully connected neural network. At the adjustment time t, for the i-th virtual power plant, its input layer is 5 nodes (total load P t , load median P mid , adjustable capacity C i,t , unit adjustment cost U i,t , Adjustment value v i,t ), the output layer is 1 node (adjusting action ΔP i,t ).

[0078] Training goal: Optimize the adjustment strategy of each agent with the goal of minimizing the overall peak-to-valley difference and total cost of the system:

[0079]

[0080] Among them, μ is the balance coefficient of cost and peak shaving and valley filling. When μ is larger, the strategy is more inclined to reduce the peak-valley difference, otherwise the strategy is more inclined to reduce the adjustment cost.

[0081] The training process is:

[0082] Assign an agent to each virtual power plant and initialize the neural network parameters of all agents.

[0083] Each agent generates adjustment actions based on the current state and global information.

[0084] Integrate the adjustment actions of all agents, calculate the total load adjustment amount, and according to the defined training target J i(θ) calculates the immediate reward of each agent and distributes the global reward to the local reward of each agent at each time step through the temporal value splitting architecture.

[0085] Based on the local rewards, the neural network parameters of each agent are updated to optimize the adjustment strategy.

[0086] The embodiment of the present invention constructs a fully connected neural network to learn the VPP regulation strategy, whose input includes the total load P in the current period t , its own adjustable capacity C i,t and unit adjustment cost U i,t , adjust the value v i,t , and then train the agent with the system's overall peak-to-valley difference and the minimum total cost as a template, and finally realize the flexible coordination of the virtual power plant.

[0087] Furthermore, after obtaining the trained fully connected neural network model, the trained model is applied to the online scenario to optimize the regulation strategy of the virtual power plant in real time.

[0088] Specifically, real-time data is collected from the virtual power plant's load and regulation costs using Internet of Things (IoT) devices. This real-time data is fed into the agent's trained fully connected neural network model, which rapidly outputs the actual load regulation results to achieve dynamic response, ensuring stable system operation in the face of sudden load fluctuations. A feedback mechanism compares the actual regulation results with the load forecast results, collecting results with significant deviations. Further training is then conducted periodically based on this collected data to optimize the model parameters.

[0089] The beneficial effects of the present invention are:

[0090] (1) A 1+N multi-agent collaborative architecture is proposed, and a unified input format for each agent is designed. This preserves global information while ensuring that the dimension of the input data does not change with the number of participating virtual power plants, thus enabling flexible collaboration among multiple virtual power plants.

[0091] (2) A fusion mechanism for distributing regulation rewards is constructed to achieve temporal decoupling of regulation effects, thereby replacing centralized critics, solving the problem of calculating the advantage function for the collaboration of massive virtual power plants, and achieving efficient training in a critic-free environment. Based on this patent invention, a breakthrough is made in the problem that the learning effect of traditional multi-agent reinforcement learning decreases as the number of collaborative agents increases, and it can achieve rapid and flexible collaboration of massive virtual power plants.

[0092] The flexible collaborative method for massive virtual power plants that integrates mechanism knowledge according to an embodiment of the present invention breaks through the limitations of traditional multi-agent reinforcement learning methods in computational complexity and communication bottlenecks by designing a flexible multi-agent architecture and a time-domain value splitting architecture based on mechanism decomposition, thereby achieving efficient collaborative optimization of massive virtual power plants. By learning the adjustment strategy through a fully connected neural network and combining transfer learning and distributed training technology, the training efficiency and generalization ability of the model are further improved, providing an innovative solution for the stable operation of new power systems. By designing a flexible multi-agent architecture and a time-domain value splitting architecture based on mechanism decomposition, efficient collaborative optimization of virtual power plants is achieved.

[0093] In order to implement the above embodiment, Figure 4 As shown, this embodiment also provides a massive virtual power plant flexible coordination system 10 integrating mechanism knowledge, including:

[0094] The collaborative model building module 100 is used to build a massive virtual power plant collaborative model; it includes a global information processing module, a mechanism calculation module and multiple virtual power plant agents;

[0095] The load forecasting module 200 is used to input the multimodal global information into the global information processing module for calculation to obtain the load forecast result;

[0096] The regulation value calculation module 300 is used to calculate the global regulation value of the Agent by integrating the local information of each virtual power plant and the load forecast results of the global information processing module using the mechanism calculation module;

[0097] The regulation strategy optimization module 400 is used to construct a fully connected neural network model for each virtual power plant agent, and input training data into the fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the regulation strategy and achieve flexible coordination of the virtual power plant; wherein the training data includes load forecast results, its own adjustable capacity, unit regulation cost and global regulation value.

[0098] Furthermore, the load forecasting module 200 is further configured to:

[0099] Acquire multimodal global information, including electricity, meteorological, and social information;

[0100] The global information processing module includes an Encoder model and a Decoder model;

[0101] The global information processing module is used to fuse multimodal global information, predict the load curve of the next day, output the maximum load and the minimum load, and calculate the central value of the daily load.

[0102] Furthermore, it also includes an optimization module for:

[0103] Use IoT devices to collect load data and regulation costs of virtual power plants in real time to obtain real-time data;

[0104] Input the real-time collected data into the trained fully connected neural network model to output the actual adjustment result of load regulation;

[0105] The actual regulation results are compared with the load forecast results through the feedback mechanism, and the results with deviations greater than the preset threshold are obtained to optimize the network model parameters.

[0106] Furthermore, the adjustment value calculation module 300 is further configured to:

[0107] Set correction parameters for adjustment thresholds;

[0108] According to the actual load of each period and the central value of the daily load, the value function of reducing the peak-to-valley difference of the load is calculated based on the correction parameters;

[0109] Based on the value function, the adjustable capacity and unit regulation cost of the i-th virtual power plant VPP in period t on the next day are obtained from the virtual power plant market to calculate the regulation value of the virtual power plant.

[0110] Furthermore, the adjustment strategy optimization module 400 is further configured to:

[0111] Assign an agent to each virtual power plant and initialize the neural network parameters of all agents;

[0112] Generate adjustment actions based on the initialized neural network according to the current state and global information;

[0113] The total load adjustment is calculated based on the adjustment actions of all agents, and the immediate reward of each agent is calculated according to the defined training target. The global reward is distributed to each agent and the local reward of each time step through the time domain value splitting model;

[0114] The neural network parameters of each agent are updated based on the local reward to optimize the regulation strategy.

[0115] The flexible collaborative method for massive virtual power plants that integrates mechanism knowledge according to an embodiment of the present invention breaks through the limitations of traditional multi-agent reinforcement learning methods in computational complexity and communication bottlenecks by designing a flexible multi-agent architecture and a time-domain value splitting architecture based on mechanism decomposition, thereby achieving efficient collaborative optimization of massive virtual power plants. By learning the adjustment strategy through a fully connected neural network and combining transfer learning and distributed training technology, the training efficiency and generalization ability of the model are further improved, providing an innovative solution for the stable operation of new power systems. By designing a flexible multi-agent architecture and a time-domain value splitting architecture based on mechanism decomposition, efficient collaborative optimization of virtual power plants is achieved.

[0116] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0117] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A flexible collaborative method for massive virtual power plants integrating mechanism knowledge, characterized by: include: Build a massive virtual power plant collaborative model; It includes global information processing module, mechanism calculation module and multiple virtual power plant agents; Input the multimodal global information into the global information processing module to calculate and obtain the load forecast result; The mechanism calculation module is used to integrate the local information of each virtual power plant and the load forecast results of the global information processing module to calculate the global regulation value of the agent; A fully connected neural network model is constructed for each virtual power plant agent, and training data is input into the fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the regulation strategy and achieve flexible collaboration of the virtual power plant; wherein the training data includes load forecast results, its own adjustable capacity, unit regulation cost and global regulation value.

2. The method according to claim 1, characterized in that The multimodal global information is input into the global information processing module to calculate the load forecast results, including: Acquire multimodal global information, including electricity, meteorological, and social information; The global information processing module includes an Encoder model and a Decoder model; The global information processing module is used to fuse multimodal global information, predict the load curve of the next day, output the maximum load and the minimum load, and calculate the central value of the daily load.

3. The method according to claim 1, characterized in that The method further comprises: Use IoT devices to collect load data and regulation costs of virtual power plants in real time to obtain real-time data; Input the real-time collected data into the trained fully connected neural network model to output the actual adjustment result of load regulation; The actual regulation results are compared with the load forecast results through the feedback mechanism, and the results with deviations greater than the preset threshold are obtained to optimize the network model parameters.

4. The method according to claim 1, wherein The mechanism calculation module integrates the local information of each virtual power plant and the load forecast results of the global information processing module to calculate the global regulation value of the agent, including: Set correction parameters for adjustment thresholds; According to the actual load of each period and the central value of the daily load, the value function of reducing the peak-to-valley difference of the load is calculated based on the correction parameters; Based on the value function, the adjustable capacity and unit regulation cost of the i-th virtual power plant VPP in period t on the next day are obtained from the virtual power plant market to calculate the regulation value of the virtual power plant.

5. The method according to claim 1, wherein The training data is input into the fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the adjustment strategy, including: Assign an agent to each virtual power plant and initialize the neural network parameters of all agents; Generate adjustment actions based on the initialized neural network according to the current state and global information; The total load adjustment is calculated based on the adjustment actions of all agents, and the immediate reward of each agent is calculated according to the defined training target. The global reward is distributed to each agent and the local reward of each time step through the time domain value splitting model; The neural network parameters of each agent are updated based on the local reward to optimize the regulation strategy.

6. A flexible collaborative system for massive virtual power plants integrating mechanism knowledge, characterized by: include: Collaborative model building module, used to build massive virtual power plant collaborative models; It includes global information processing module, mechanism calculation module and multiple virtual power plant agents; The load forecasting module is used to input the multimodal global information into the global information processing module to calculate and obtain the load forecast result; The regulation value calculation module is used to calculate the global regulation value of the Agent by integrating the local information of each virtual power plant and the load forecast results of the global information processing module using the mechanism calculation module; The regulation strategy optimization module is used to build a fully connected neural network model for each virtual power plant agent, and input training data into the fully connected neural network model for training to obtain a trained fully connected neural network model to optimize the regulation strategy and achieve flexible coordination of the virtual power plant; wherein, the training data includes load forecast results, its own adjustable capacity, unit regulation cost and global regulation value.

7. The system according to claim 6, characterized in that The load forecasting module is also used to: Acquire multimodal global information, including electricity, meteorological, and social information; The global information processing module includes an Encoder model and a Decoder model; The global information processing module is used to fuse multimodal global information, predict the load curve of the next day, output the maximum load and the minimum load, and calculate the central value of the daily load.

8. The system according to claim 6, wherein: Also includes: Optimization modules for: Use IoT devices to collect load data and regulation costs of virtual power plants in real time to obtain real-time data; Input the real-time collected data into the trained fully connected neural network model to output the actual adjustment result of load regulation; The actual regulation results are compared with the load forecast results through the feedback mechanism, and the results with deviations greater than the preset threshold are obtained to optimize the network model parameters.

9. The system according to claim 6, wherein: The adjustment value calculation module is also used to: Set correction parameters for adjustment thresholds; According to the actual load of each period and the central value of the daily load, the value function of reducing the peak-to-valley difference of the load is calculated based on the correction parameters; Based on the value function, the adjustable capacity and unit regulation cost of the i-th virtual power plant VPP in period t on the next day are obtained from the virtual power plant market to calculate the regulation value of the virtual power plant.

10. The system according to claim 6, wherein: The regulation strategy optimization module is also used to: Assign an agent to each virtual power plant and initialize the neural network parameters of all agents; Generate adjustment actions based on the initialized neural network according to the current state and global information; The total load adjustment is calculated based on the adjustment actions of all agents, and the immediate reward of each agent is calculated according to the defined training target. The global reward is distributed to each agent and the local reward of each time step through the time domain value splitting model; The neural network parameters of each agent are updated based on the local reward to optimize the regulation strategy.

Images