Adjustable load capacity analysis, regulation and control method for virtual power plant
By integrating multi-source data and optimizing dynamic control strategies, the problems of data silos and prediction biases in virtual power plants have been solved, enabling accurate assessment of load capacity and improved control effectiveness, thus ensuring the stability and efficient operation of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 翁俊
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-01
AI Technical Summary
Traditional virtual power plants lack effective data sharing and integration mechanisms, resulting in low data quality, making it difficult to fully and accurately reflect the complex characteristics of the load, affecting the control effect. Furthermore, relying on experience-based judgment or simple models leads to large deviations between prediction results and actual results, affecting grid stability and control success rate.
By collecting and processing multi-source data, data cleaning and fusion techniques are used to integrate the data, construct a multi-dimensional adjustable load capacity assessment index system, calculate weights using an improved particle algorithm and entropy weight method, establish a dynamic hierarchical collaborative control strategy, and use an LSTM-attention mechanism to build a load response prediction model to optimize the control strategy in real time and perform closed-loop correction.
It has achieved efficient data integration and quality improvement, provided accurate load capacity assessment, improved the reliability and success rate of regulation, and ensured the stability of the power grid and power quality.
Smart Images
Figure CN121965637A_ABST
Abstract
Description
A method for analyzing and controlling the adjustable load capacity of a virtual power plant Technical Field
[0001] This invention relates to the field of power system operation and control technology, and in particular to a method for analyzing and controlling the adjustable load capacity of a virtual power plant. Background Technology
[0002] Distributed power sources are greatly affected by natural conditions (such as sunlight and wind), and their output power is intermittent and fluctuating. Although energy storage systems can smooth out power fluctuations to some extent, different types of energy storage have different charging and discharging characteristics, lifespan, and costs, which increases the complexity of grid operation.
[0003] Virtual power plants have emerged as an innovative power system management model. By leveraging digital technology and advanced regulation strategies, they effectively aggregate dispersed and fragmented adjustable resources such as distributed power sources, energy storage devices, and flexible loads, thereby achieving efficient interaction between power generation, grid, load, and storage.
[0004] In traditional virtual power plant (VPS) operation systems, various adjustable load data are scattered across different entities and systems. The lack of effective sharing and integration mechanisms among these data creates data silos, making it difficult for VPS to obtain comprehensive and unified data support for integrated analysis and decision-making. This leads to biased and limited decision-making. Furthermore, due to differences in the accuracy of data acquisition equipment, interference during transmission, and environmental factors, the quality of the acquired data is low, severely impacting the stable operation and control effectiveness of the VPS. Secondly, traditional VPS operation systems classify adjustable loads based on only one or a few factors, failing to comprehensively and accurately reflect the complex characteristics of the load and making it difficult to meet the needs of refined control in VPS, resulting in poor control performance. Additionally, traditional VPS operation systems rely on experience-based judgment or simple linear prediction models, which cannot accurately predict complex and variable load responses, leading to significant deviations between prediction results and actual performance. This not only affects grid stability and power quality but also reduces the success rate and quality of VPS control. Summary of the Invention
[0005] The purpose of this invention is to provide a method for analyzing and controlling the adjustable load capacity of a virtual power plant, thereby solving the technical problems existing in the prior art.
[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:
[0007] A method for analyzing and controlling the adjustable load capacity of a virtual power plant includes the following steps: S1: Multi-source data acquisition and processing. Electrical and non-electrical parameters of various adjustable loads, as well as grid dispatch instructions, are collected through a novel virtual power plant management platform, load aggregators, and direct-participation user terminals. Data cleaning algorithms are used to remove outliers and noise data, and data fusion technology is employed to integrate data from different sources. S2: Adjustable load classification. Based on the load's response time threshold, adjustment rigidity coefficient, economic cost sensitivity, environmental impact factors, and invited control priority, adjustable loads are divided into multiple levels. S3: Construction of a multi-dimensional adjustable load capacity assessment index system, including basic capacity indicators, response quality indicators, economic indicators, stability indicators, and collaborative adaptation indicators. S4: Weighted calculation. Weights are calculated for the multi-dimensional adjustable load capacity assessment index system based on an improved particle algorithm. Combined with the entropy weight method and dynamic demand revision weights for cascade control scenarios, the weights of various adjustable loads under different conditions are obtained. S5: Establish a dynamic hierarchical coordinated control strategy. Based on the type of power grid dispatch demand, establish a dynamic hierarchical coordinated control strategy, and optimize the response priority, combination ratio, and control stage switching logic of different load levels in real time through reinforcement learning algorithms. The types of power grid dispatch demand include emergency frequency regulation, peak shaving, valley filling, and optimized dispatch. S6: Predictive model construction. Based on the real-time operation data of the virtual power plant, construct a load response prediction model, pre-evaluate the execution effect of each cascade control command, and make a judgment based on the pre-evaluation results. S7: Command execution. At different stages, issue control commands to the load aggregator according to the scheduled time, real-time, regional self-balancing, and load control system cascades, directly participate in the user terminal, collect load response data of each cascade in real time, adjust the control parameters through the closed-loop correction mechanism of command feedback and real-time monitoring, and analyze the cause of the deviation when the monitored load response data deviates from the expected control target, and adjust the control parameters accordingly.
[0008] Furthermore, the multiple levels in step S2 are divided into Level 1 fast adjustable load, Level 2 time-adjustable flexible adjustable load, and Level 3 time-adjustable rigid adjustable load.
[0009] Furthermore, the first-level rapidly adjustable load is divided into energy storage load and charging pile load; the energy storage load has an adjustment rigidity coefficient ≤ 0.2, moderate economic cost sensitivity, and an environmental impact factor ≥ 0.8; the charging pile load has an adjustment rigidity coefficient of 0.5, high economic cost sensitivity, and an environmental impact factor ≥ 0.6; the second-level time-based flexible adjustable load is divided into commercial building air conditioning load and industrial process load; the commercial building air conditioning load has an adjustment rigidity coefficient of 0.45, moderate economic cost sensitivity, and an environmental impact factor ≥ 0.4; the industrial process load has an adjustment rigidity coefficient of 0.6, low economic cost sensitivity, and an environmental impact factor ≥ 0.3; the third-level time-based rigid adjustable load is divided into industrial production plan load and district heating load; the industrial production plan load has an adjustment rigidity coefficient of 0.9, low economic cost sensitivity, and an environmental impact factor ≥ 0.25; the district heating load has an adjustment rigidity coefficient of 0.65, moderate economic cost sensitivity, and an environmental impact factor ≥ 0.4.
[0010] Furthermore, the improved particle algorithm in step S4 is improved by introducing an adaptive inertia weight factor, dynamically adjusting the search capability based on the number of iterations, adding a chaotic perturbation factor, and adopting an elite retention strategy, while also combining the entropy weight method to calculate the objective weight of the index.
[0011] Furthermore, in step S5, the dynamic hierarchical coordinated control strategy prioritizes the first-level fast adjustable load when there is an emergency frequency regulation demand, with a response ratio greater than 75%. If this ratio is insufficient, regional self-balancing assistance is triggered, and finally, the load control system is activated to rigidly cut off the load. During peak shaving, the first-level fast adjustable load and the second-level time-controlled flexible adjustable load respond in coordination, with the coordination ratio dynamically adjusted according to the real-time load rate. During valley filling, the third-level time-controlled rigid load responds first, with a response ratio greater than 70%. During optimized scheduling, the coordinated combination of the three-level loads is achieved based on the comprehensive capability score, and the adjustment amplitude, response timing, and tiered switching nodes of each load are optimized through the DQN algorithm.
[0012] Furthermore, the prediction model in step S6 adopts a fusion model of LSTM-attention mechanism and cascade scenario embedding. The input features of the model are historical response data, real-time grid frequency, voltage deviation, load factor, ambient temperature and control command parameters. If the deviation between the prediction response success rate and the required power of the command is ≤4%, the pre-evaluation is passed and the corresponding control command is executed; otherwise, return to step S5 to re-optimize the cascade strategy and load combination.
[0013] Compared with the prior art, the present invention has the following beneficial effects:
[0014] (i) This invention collects adjustable load and power grid dispatch command data from multiple sources, solving the problems of data silos and low data quality, and improving the reliability of decision-making; Based on multi-dimensional characteristics, this invention scientifically divides adjustable load into three levels, providing strong support for precise control; and constructs a multi-dimensional evaluation index system to comprehensively evaluate load capacity and achieve optimal resource allocation.
[0015] (ii) The present invention adopts an improved particle algorithm that introduces an adaptive inertia weight factor, dynamically adjusts the search capability based on the number of iterations, adds a chaotic disturbance factor, adopts an elite retention strategy, and combines the entropy weight method to calculate the objective weight of the index. This algorithm can more accurately determine the weight of the multidimensional adjustable load capacity assessment index system, and establish a dynamic hierarchical collaborative control strategy for different power grid dispatching needs and optimize it in real time, so as to flexibly respond to various situations and improve the control effect.
[0016] (III) This invention uses an LSTM-attention mechanism and a cascade scenario embedding fusion model to construct a load response prediction model, which pre-evaluates the execution effect of each cascade control command, can accurately predict the load response, ensure that the execution effect of the control command meets expectations, and improve the success rate and quality of control. Attached Figure Description
[0017] Figure 1 is a flowchart of the present invention;
[0018] Figure 2 is a flowchart of the dynamic hierarchical collaborative control strategy of the present invention. Detailed Implementation
[0019] To make the content of this invention easier to understand, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Identical components are represented by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.
[0020] As shown in Figure 1, this embodiment provides a method for analyzing and controlling the adjustable load capacity of a virtual power plant, including the following steps:
[0021] S1: Multi-source data acquisition and processing. Through a new virtual power plant management platform, load aggregators, and direct participating user terminals, electrical and non-electrical parameters of various adjustable loads, as well as grid dispatch instructions, are collected. The new virtual power plant management platform monitors and records a large amount of key data related to adjustable loads in real time. The load aggregator contains information such as the electricity consumption patterns and load change patterns of different users, as well as data on users' willingness and ability to participate in grid dispatch. The direct participating user terminals are the direct carriers of adjustable loads. These data can accurately reflect the real-time status and adjustable range of various adjustable loads in actual operation. The above data is processed using data cleaning algorithms to remove outliers and noise data, and data fusion technology is used to integrate data from different sources.
[0022] S2: Adjustable load classification. Based on the load response time threshold, adjustment rigidity coefficient, economic cost sensitivity, environmental impact factor, and invited control priority, adjustable loads are divided into multiple levels. Specifically, they are divided into Level 1 fast adjustable loads, Level 2 time-based flexible adjustable loads, and Level 3 time-based rigid adjustable loads. This helps to allocate various load resources more rationally according to different power grid dispatching needs in practical applications.
[0023] The first-level fast adjustable load is divided into energy storage load and charging pile load. The energy storage load has an adjustment rigidity coefficient ≤0.2, moderate economic cost sensitivity, and an environmental impact factor ≥0.8. When the power grid experiences instantaneous power fluctuations, the energy storage load can respond quickly, absorbing or releasing electrical energy to effectively stabilize the power grid frequency. The energy storage load can adjust flexibly in response to power grid dispatch instructions, and has a large elasticity in changing its output power. The charging pile load has an adjustment rigidity coefficient of 0.5, high economic cost sensitivity, and an environmental impact factor ≥0.6. During electric vehicle charging, when the power grid needs to balance the load, the charging pile can flexibly adjust the charging power according to dispatch instructions. The charging pile load can respond to power grid jump commands in a short time.
[0024] The secondary time-adjustable load is divided into commercial building air conditioning load and industrial process load. The commercial building air conditioning load has a rigidity coefficient of 0.45, moderate economic cost sensitivity, and an environmental impact factor ≥0.4. During peak grid periods, commercial buildings can adjust the air conditioning temperature setpoint or the number of operating units according to dispatch instructions. Both economic cost and environmental impact are within the preset range, while effectively regulating the grid load. The industrial process load has a rigidity coefficient of 0.6, low economic cost sensitivity, and an environmental impact factor ≥0.3. Industrial enterprises can respond to grid dispatch instructions by adjusting the operating parameters of their production processes. Although the rigidity is high, the low economic cost sensitivity and relatively small environmental impact allow it to cooperate with the grid to a certain extent in load regulation.
[0025] The three-level time-controlled rigid adjustable load is divided into industrial production plan load and district heating load. The industrial production plan load has an adjustment rigidity coefficient of 0.9, low economic cost sensitivity, and an environmental impact factor ≥0.25. Large manufacturing enterprises usually operate according to a predetermined production plan. If they need to respond to grid dispatch for load adjustment, they need to make significant adjustments to the entire production process. However, due to its low economic cost sensitivity and controllable environmental impact, it can still play a certain role in grid load adjustment when the time is agreed in advance. The district heating load has an adjustment rigidity coefficient of 0.65, medium economic cost sensitivity, and an environmental impact factor ≥0.4. During grid off-peak hours, the district heating system can appropriately increase the operating power of heating equipment to store heat energy, while reducing the power during grid peak hours. Although the adjustment rigidity is relatively high, it can be kept within the preset range in terms of economic cost and environmental impact.
[0026] S3: Construct a multi-dimensional adjustable load capacity assessment index system, including basic capacity indicators, response quality indicators, economic indicators, stability indicators, and coordination and adaptation indicators. The basic capacity indicators include adjustable capacity and regulation rate, which reflect the load's performance in basic regulation. The response quality indicators include response time and regulation accuracy, which measure the responsiveness of adjustable loads to grid dispatch commands. The economic indicators include regulation cost and benefit coefficient, which consider the cost-benefit relationship of adjustable loads participating in grid regulation. The stability indicators include continuous regulation duration and fluctuation coefficient, which assess the stability and reliability of adjustable loads during regulation, ensuring their continuous and reliable response to grid dispatch commands over a long period. The environmental friendliness indicators include carbon emission reduction and energy consumption reduction rate, which assess the coordination and cooperation capabilities between different types of adjustable loads and between adjustable loads and other components of the grid, to achieve efficient operation of the virtual power plant as a whole.
[0027] S4: Weighted calculation: Based on the improved particle algorithm, the weights of the multidimensional adjustable load capacity assessment index system are calculated. Combined with the entropy weight method and the dynamic demand revision weights of the cascade control scenario, the comprehensive capacity scores of various adjustable loads under different cascade scenarios are obtained.
[0028] The improved particle algorithm is achieved by introducing an adaptive inertia weight factor to dynamically adjust the search capability based on the number of iterations, incorporating a chaotic perturbation factor, and employing an elite retention strategy, while also combining the entropy weight method to calculate the objective weights of the indicators; the inertia weight factor... ,in, This represents the current iteration number. The maximum number of iterations, In the early stages of iteration, high weights facilitate global search and prevent the omission of optimal weight combinations. In the later stages of iteration, smaller weights are focused on local fine-tuning to improve weight accuracy. Chaos exhibits characteristics such as randomness, ergodicity, and sensitivity to initial conditions. Introducing a chaotic perturbation factor into the particle algorithm effectively prevents particles from getting trapped in local optima during the search process. When a particle gets trapped in a local optimum, the chaotic perturbation factor randomly perturbs its position, causing it to jump out of the local optimum region and search again in the solution space, thereby increasing the probability of finding the global optimum. During each iteration, elite solutions are recorded and retained for the next generation, ensuring that elite solutions are not lost due to changes in subsequent iterations, helping the algorithm converge to the global optimum more quickly. The entropy weight method determines weights based on the dispersion of indicator data. The greater the data dispersion, the more information the indicator contains, and the higher its weight. This is existing technology and will not be elaborated upon here. The objective weights calculated by the entropy weight method fully reflect the actual importance of each indicator in the evaluation process, avoiding interference from human factors in weight determination.
[0029] S5: Establish a dynamic hierarchical coordinated control strategy. Based on the type of power grid dispatch demand, establish a dynamic hierarchical coordinated control strategy, and use reinforcement learning algorithms to optimize the response priority, combination ratio, and control stage switching logic of different load levels in real time to adapt to the dynamic changes in the power grid operation status. The types of power grid dispatch demand include emergency frequency regulation, peak shaving, valley filling, and optimized dispatch.
[0030] When the dynamic hierarchical coordinated control strategy is in the event of an emergency frequency regulation demand, the first-level fast adjustable load has the highest priority and a response ratio of more than 75%. When the grid frequency suddenly drops, the energy storage load can quickly release the stored electrical energy to supplement the grid power deficit and restore the grid frequency to the normal range as soon as possible. If this is insufficient, the regional self-balancing auxiliary is triggered, and finally the load control system is activated to rigidly cut off the load. The load control system will forcibly cut off some non-critical loads according to preset rules to quickly reduce the power demand of the grid and restore the grid frequency to the safe range.
[0031] During peak shaving, the primary fast-adjustable load and the secondary time-adjustable flexible load respond in concert, with the coordination ratio dynamically adjusted based on the real-time load rate. The real-time load rate reflects the current load tension of the power grid. By monitoring and analyzing it in real time, the system can intelligently determine the optimal coordination ratio between the primary fast-adjustable load and the secondary time-adjustable flexible load. When the real-time load rate is high and the power grid load pressure is high, the response ratio of the primary fast-adjustable load is appropriately increased to quickly alleviate the load pressure. Conversely, when the load rate decreases and the power grid pressure is relatively reduced, the participation ratio of the secondary time-adjustable flexible load is increased to achieve the peak shaving target more economically and efficiently.
[0032] During off-peak hours, the three-level rigid loads respond first, with a response rate greater than 70%. Industrial enterprises can increase the operating time or power of production equipment during off-peak hours, and district heating systems can store heat energy in advance.
[0033] During optimized scheduling, the coordinated combination of three-level loads is achieved based on the comprehensive capability score. The adjustment amplitude, response timing and cascade switching nodes of each load are optimized through the DQN algorithm. The DQN algorithm optimization ensures that the switching process between different loads is smooth and efficient, further improving the overall operating efficiency of the virtual power plant.
[0034] S6: Predictive Model Construction. Based on real-time operation data of the virtual power plant, a load response prediction model is constructed to pre-evaluate the execution effect of each cascade control command and make a judgment based on the pre-evaluation results. The prediction model adopts a fusion model of LSTM-attention mechanism and cascade scenario embedding. The input features of the model are historical response data, real-time grid frequency, voltage deviation, load factor, ambient temperature, and control command parameters. During use, the model performs deep analysis and learning based on the input multi-source data through a complex neural network structure to predict the execution effect of each cascade control command. If the deviation between the predicted response success rate and the required power of the command is ≤4%, the pre-evaluation is passed, and the corresponding control command is executed to ensure that the grid control can be carried out as expected and achieve stable and efficient operation. Otherwise, return to step S5 to re-optimize the cascade strategy and load combination. The above-mentioned feedback optimization mechanism based on the prediction model realizes the closed-loop management of the virtual power plant control process and effectively improves the accuracy and reliability of control.
[0035] S7: Command Execution. At different stages, in a tiered sequence of scheduled, real-time, regional self-balancing, and load control systems, control commands are issued to load aggregators, user control terminals, large users, and load control terminals, enabling relevant entities to rationally arrange production or operation plans to meet the grid's control needs. Control parameters are adjusted through a closed-loop correction mechanism of command feedback and real-time monitoring. When the monitored load response data deviates from the expected control target, the cause of the deviation is analyzed, and control parameters are precisely adjusted accordingly. The closed-loop correction mechanism forms a dynamic feedback loop, enabling the virtual power plant's control process to be continuously optimized and improved, thereby continuously enhancing the stability and reliability of grid operation.
[0036] For routine control plans, the general principles and sequence of tracking and control are as follows: first, call the scheduled resources for control according to the prepared strategy, and at the same time, dynamically monitor whether the load meets the requirements; if it does not meet the requirements, call the real-time adjustable resources, and then continue to monitor dynamically. If it is found that the control requirements are still not met and the real-time adjustable resources are exhausted, then activate the forced disconnection function of the power load management terminal to disconnect the load.
[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for analyzing and controlling the adjustable load capacity of a virtual power plant, characterized in that: Includes the following steps: S1: Multi-source data acquisition and processing. Through a new virtual power plant management platform, load aggregators, and direct participating user terminals, electrical and non-electrical parameters of various adjustable loads and grid dispatch instructions are collected. Data cleaning algorithms are used to remove outliers and noise data, and data fusion technology is used to integrate data from different sources. S2: Adjustable load classification, based on load response time threshold, adjustment rigidity coefficient, economic cost sensitivity, environmental impact factors and invited regulation priority, adjustable loads are divided into multiple levels; S3: Construct a multi-dimensional adjustable load capacity assessment index system, including basic capacity index, response quality index, economic index, stability index, and collaborative adaptation index. S4: Weighted calculation: Based on the improved particle algorithm, the weights of the multidimensional adjustable load capacity assessment index system are calculated. Combined with the entropy weight method and the dynamic demand revision weights of the cascade control scenario, the comprehensive capacity scores of various adjustable loads under different cascade scenarios are obtained. S5: Establish a dynamic hierarchical coordinated control strategy. Based on the type of power grid dispatch demand, establish a dynamic hierarchical coordinated control strategy, and use reinforcement learning algorithms to optimize the response priority, combination ratio, and control stage switching logic of different load levels in real time. The types of power grid dispatch demand include emergency frequency regulation, peak shaving, valley filling, and optimized dispatch. S6: Predictive model construction: Based on real-time operation data of the virtual power plant, construct a load response prediction model, pre-evaluate the execution effect of each level of control command, and make a judgment based on the pre-evaluation results; S7: Command execution. At different stages, control commands are issued to load aggregators according to the scheduled time, real-time, regional self-balancing and load control system. It directly participates in the user terminal, collects load response data of each level in real time, and adjusts the control parameters through a closed-loop correction mechanism of command feedback and real-time monitoring. When the monitored load response data deviates from the expected control target, the cause of the deviation is analyzed and the control parameters are adjusted accordingly.
2. The adjustable load capacity analysis and control method for virtual power plants according to claim 1, characterized in that: The multiple levels in step S2 are divided into Level 1 fast adjustable load, Level 2 time-controlled flexible adjustable load, and Level 3 time-controlled rigid adjustable load.
3. The adjustable load capacity analysis and control method for virtual power plants according to claim 2, characterized in that: The first-level rapidly adjustable load is divided into energy storage load and charging pile load; the energy storage load has an adjustment rigidity coefficient ≤ 0.2, moderate economic cost sensitivity, and an environmental impact factor ≥ 0.8; the charging pile load has an adjustment rigidity coefficient of 0.5, high economic cost sensitivity, and an environmental impact factor ≥ 0.
6. The second-level time-based flexible adjustable load is divided into commercial building air conditioning load and industrial process load; the commercial building air conditioning load has an adjustment rigidity coefficient of 0.45, moderate economic cost sensitivity, and an environmental impact factor ≥ 0.4; the industrial process load has an adjustment rigidity coefficient of 0.6, low economic cost sensitivity, and an environmental impact factor ≥ 0.
3. The third-level time-based rigid adjustable load is divided into industrial production plan load and district heating load; the industrial production plan load has an adjustment rigidity coefficient of 0.9, low economic cost sensitivity, and an environmental impact factor ≥ 0.25; the district heating load has an adjustment rigidity coefficient of 0.65, moderate economic cost sensitivity, and an environmental impact factor ≥ 0.
4.
4. The adjustable load capacity analysis and control method for virtual power plants according to claim 1, characterized in that: The improvement of the particle algorithm in step S4 is achieved by introducing an adaptive inertia weight factor, dynamically adjusting the search capability based on the number of iterations, adding a chaotic perturbation factor, adopting an elite retention strategy, and combining the entropy weight method to calculate the objective weight of the index.
5. The adjustable load capacity analysis and control method for virtual power plants according to claim 2, characterized in that: In step S5, the dynamic hierarchical coordinated control strategy prioritizes the first-level fast adjustable loads during emergency frequency regulation, with a response rate greater than 75%. If this is insufficient, regional self-balancing assistance is triggered, and finally, the load control system is activated to rigidly cut off the loads. During peak shaving, the first-level fast adjustable loads and the second-level time-controlled flexible adjustable loads respond in coordination, with the coordination ratio dynamically adjusted according to the real-time load rate. During valley filling, the third-level time-controlled rigid loads respond first, with a response rate greater than 70%. During optimized scheduling, the coordinated combination of the three-level loads is achieved based on the comprehensive capability score, and the adjustment amplitude, response timing, and tiered switching nodes of each load are optimized through the DQN algorithm.
6. The adjustable load capacity analysis and control method for virtual power plants according to claim 1, characterized in that: The prediction model in step S6 adopts a fusion model of LSTM-attention mechanism and cascade scenario embedding. The input features of the model are historical response data, real-time grid frequency, voltage deviation, load factor, ambient temperature and control command parameters. If the deviation between the prediction response success rate and the required power of the command is ≤4%, the pre-evaluation is passed and the corresponding control command is executed; otherwise, return to step S5 to re-optimize the cascade strategy and load combination.