Distributed photovoltaic cluster voltage active support control method suitable for high-proportion permeation

Abstract

The invention belongs to the technical field of photovoltaic voltage control of a power distribution network, and discloses a distributed photovoltaic cluster voltage active support control method adaptive to high proportional penetration, which is based on a two-dimensional dynamic aggregation algorithm, takes an electrical distance, a voltage influence weight and a load response coefficient as core indexes, and obtains a distributed photovoltaic cluster voltage active support control method. Dynamic division and real-time updating of a photovoltaic cluster are realized, and dynamic correlation characteristics of voltage and load under high-proportion photovoltaic permeation can be accurately matched; meanwhile, a four-level layered cooperative control framework is constructed, a virtual leader mechanism and digital twinborn verification are integrated, and the problem of adjustment conflicts easily caused by independent control of a single unit is avoided through hierarchical linkage of global scheduling, cluster coordination and unit execution; meanwhile, a multi-dimensional scene library covering a traditional scene and an extreme weather scene is constructed, a high-probability and high-influence key scene is reserved, a collaborative injection mechanism of virtual inertia, damping and energy storage is combined, and precise voltage regulation under different working conditions is achieved through multi-mode smooth switching.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic voltage control of distribution network, and particularly relates to a distributed photovoltaic cluster voltage active support control method suitable for high penetration. BACKGROUND

[0002] The global energy transformation process continues to accelerate. Distributed photovoltaics, with its outstanding advantages of being clean and renewable, has a growing penetration rate in distribution networks and is gradually moving towards high penetration scenarios. However, after high-penetration distributed photovoltaics is connected, the core challenge of voltage regulation in distribution networks becomes increasingly prominent, and existing technologies are difficult to effectively address. The main technical problems are as follows: The existing cluster coordinated regulation mechanism in the voltage control technology of distribution network is deficient. The existing technology mostly adopts single photovoltaic unit independent control or fixed grouping aggregation control mode, and has not established a multi-dimensional adaptive cluster division and coordinated regulation system combining the dynamic changes of distribution network topology, photovoltaic output characteristics and user load response capability. When high-penetration photovoltaics is connected, the voltage regulation actions of each photovoltaic unit lack global coordination, and conflicts such as over-regulation or under-regulation are prone to occur. Moreover, it is unable to flexibly allocate and regulate the control resources according to the dynamic changes of the voltage distribution of the distribution network, resulting in low voltage support efficiency. The existing voltage control technology of distribution network has weak source-load uncertainty adaptation capability. The existing control strategy is mostly based on static design under fixed operating conditions, and does not fully consider the intermittency of photovoltaic output, the randomness of sudden changes in light, the destructiveness of extreme weather and the uncertainty of load fluctuations. It lacks a dynamic adaptive control idea. Under complex operating conditions such as extreme weather or load impact, the static control strategy is difficult to quickly respond to voltage fluctuations, which may cause voltage out-of-limit and even trigger photovoltaic inverter cascading trip-off, further exacerbating the risk of voltage instability in the power grid. SUMMARY

[0003] The purpose of the present application is to provide a distributed photovoltaic cluster voltage active support control method suitable for high penetration, to solve the problems raised in the background art.

[0004] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: a distributed photovoltaic cluster voltage active support control method suitable for high penetration, comprising a collection stage, a source-load adaptation generation stage, a hierarchical architecture construction stage, a strategy solving and verification stage, a coordinated injection control stage, an autonomous and coordinated switching stage, and an evaluation and optimization stage. Preferably, the collection stage collects multi-dimensional real-time data through edge computing nodes, including voltage, current and power of each node of the power distribution network, output of the distributed photovoltaic unit, inverter state, light intensity, environmental temperature, and response capability and operating state data of the user-side controllable load, low-delay transmission and local preprocessing of the data are realized by using 5G and edge computing fusion technology, the preprocessing process includes data cleaning, format standardization and outlier removal, and a distributed trusted data sharing platform is established by integrating blockchain technology, data transmission encryption is realized by using an asymmetric encryption algorithm, and cross-cluster node identity authentication is completed based on a consensus mechanism; A two-dimensional dynamic aggregation algorithm based on graph theory community discovery and voltage sensitivity weighting is adopted, the first dimension takes the electrical distance between the photovoltaic unit and the key node and the voltage influence weight as the core index, the second dimension takes the user controllable load response coefficient as the optimization index, clustering division is realized by using an improved modularity function, the correlation strength matrix between nodes is calculated, the modularity threshold is set to 0.35 (after 100 groups of simulation calibration of different topological scenes, the average voltage influence correlation degree of the clustering result under this threshold is 0.85, which is significantly higher than that under other thresholds), and optimal clustering is realized, photovoltaic units with similar geographical location, similar output characteristics, high voltage influence correlation degree and strong user load response synergy are aggregated into a virtual cluster, and the cluster update period is dynamically adjusted according to the voltage fluctuation frequency, the update interval is shortened when the fluctuation is severe, and the dynamic correlation characteristics of voltage and load under high penetration are accurately matched; A cluster-level state perception model is established, an improved Kalman filter algorithm is used to perform noise reduction processing on the collected data, the overall output capacity of the cluster, the voltage operating state of the power distribution network, the user load response potential and the potential risk of exceeding the limit are evaluated in real time, and an evaluation report containing state parameters is generated.

[0005] Preferably, the source-load adaptation generation stage is based on the multi-dimensional state evaluation data of the collection stage, adopts an Attention-LSTM prediction model integrating reinforcement learning, describes the source-load uncertainty as a continuous Markov process, takes the decision time, photovoltaic output level and load demand state as state components, and takes the prediction error correction as action component, and the model parameters are adaptively optimized through the reward mechanism of reinforcement learning, the batch gradient descent method is used to optimize the loss function in the model training process, and short-term prediction of the distributed photovoltaic cluster output and the power distribution network load is realized; A multi-dimensional uncertainty scenario generation system is constructed, including traditional light change, load impact scenarios, and extreme weather scenario modules. The extreme weather scenario is constructed by correlation analysis of historical meteorological data and power grid operation data. Multiple source-load scenarios are generated using a composite method of Latin hypercube sampling and Monte Carlo simulation. First, uniform coverage of the parameter space is achieved through Latin hypercube sampling, and then the number of scenarios is expanded using Monte Carlo simulation. By improving the scenario reduction algorithm (combining the probability of scenario occurrence and the degree of voltage impact), high-probability and high-impact key scenarios are retained, and a scenario library covering different operating conditions is constructed. A full-process feedback and correction mechanism is established, and a dynamic evaluation index for scenario credibility is introduced. This index is calculated by comprehensively considering the probability of scenario occurrence, the degree of voltage impact, and the data matching degree. The scenario library is continuously corrected by combining real-time operation data, eliminating scenarios with credibility below 0.6 and adding new scenarios. At the same time, the prediction error is fed back to the prediction model for parameter iteration and optimization, and the prediction results and scenario library are output.

[0006] Preferably, the layered architecture construction phase combines the state assessment results of the acquisition phase with the prediction results and scenario library of the source-load adaptation generation phase to construct a four-level layered collaborative control architecture consisting of a distribution network scheduling layer, a cluster coordination layer, a photovoltaic unit control layer, and a digital twin mirror layer. The distribution network scheduling layer, based on the global voltage constraints and operational economic objectives of the distribution network, issues a total voltage support command to each photovoltaic cluster through global power balance calculation and voltage distribution simulation. The command includes core parameters such as the total reactive power regulation and the upper limit of active power output, and simultaneously receives operational status feedback information from each cluster. The cluster coordination layer integrates a virtual leader and follower mechanism. The virtual leader node receives instructions from the distribution network dispatch layer and decomposes tasks. Follower nodes synchronously execute adjustment tasks and report their status. After receiving instructions from the distribution network dispatch layer, the virtual leader node combines the status awareness information of the cluster and the scenario prediction results to allocate voltage adjustment tasks to each photovoltaic unit and associated controllable load through a distributed model predictive control algorithm. The photovoltaic unit control layer adopts a plug-and-play modular design with built-in standardized communication protocols and control interfaces. Based on the allocation instructions from the cluster coordination layer, it implements specific control strategies through inverters and reports the execution status in real time. The digital twin mirror layer constructs a virtual mirror of the distribution network and the photovoltaic cluster. It is dynamically updated based on real-time collected physical layer data, simulates the execution effect of different control strategies, and provides decision pre-verification for the distribution network dispatching layer and the cluster coordination layer. The control parameters corrected through virtual-real interaction are fed back to the instruction generation stage of the distribution network dispatching layer and the task allocation stage of the cluster coordination layer, respectively, to achieve hierarchical parameter optimization.

[0007] Virtual image construction logic: The digital twin virtual image consists of a physical entity mapping module, a data synchronization module, a simulation calculation module, and a result evaluation module. ① Physical entity mapping module: Maps the distribution network topology at a 1:1 scale, including line parameters, node locations, and photovoltaic unit / energy storage device model parameters. The topology data is synchronized in real time through edge computing nodes, with a synchronization period of 100ms. The synchronization period is consistent with the simulation step size to ensure that each simulation calculation is based on the latest physical layer data, and the virtual-physical linkage error is controlled within 0.5%. ② Data synchronization module: Establishes a mapping relationship between physical quantities and virtual quantities. The mapping rule is: Virtual voltage Uv = ,in These are real-time values ​​collected at the physical layer. The calibration coefficient is determined by comparing historical data. ③ Simulation Calculation Module: Steady-state simulation is performed using the nodal voltage method, with a simulation step size of 100ms. The core calculation formula is as follows: , The node admittance matrix of the virtual image. (for virtual injected current vector). ④ Results Evaluation Module: Set simulation error threshold When the error exceeds the threshold, the physical layer data is automatically resynchronized and the mirror parameters are corrected.

[0008] Simulation verification process: ① Input the control policy to be verified, such as the task allocation scheme of the cluster coordination layer; ②Based on the currently synchronized physical data, the voltage changes and power distribution results after the strategy is executed are simulated through the simulation calculation module; ③ Compare the simulation results with the preset target, such as voltage deviation ≤ ±2%, and output the verification pass rate. Pass rate = number of scenarios that meet the target / total number of simulated scenarios; ④ If the pass rate is ≥95%, the strategy is directly issued and executed; if 60%≤pass rate<95%, the control parameters are corrected and the simulation is repeated; if the pass rate<60%, the task allocation scheme is reformulated at the cluster coordination layer.

[0009] Preferably, the strategy solution verification stage takes minimizing the voltage deviation of key nodes in the distribution network, maximizing the output of distributed photovoltaic clusters, minimizing inverter regulation losses, and minimizing user load interruption costs as multi-objective optimization objectives. An optimization model is established that considers multiple constraints such as photovoltaic inverter capacity constraints, voltage regulation rate constraints, low voltage ride-through constraints, and user controllable load response capability constraints. The user controllable load response capability constraints are set with different adjustment upper limits according to the load type. An improved non-dominated sorting genetic algorithm is used to solve the optimization model. By introducing an adaptive crossover and mutation operator and reinforcement learning-guided search direction optimization, the algorithm's performance is improved. The crossover probability is dynamically adjusted between 0.6 and 0.9 according to the population convergence degree, while the mutation probability is fixed at 0.01. At the same time, combined with reinforcement learning-guided search direction optimization, a reward function is used to select high-quality search paths. Based on the scenario library generated during the source-load adaptation stage, the robustness of the optimization strategy is verified in multiple scenarios. The verification indicators include voltage deviation qualification rate, photovoltaic power output utilization rate, and equipment loss rate. The robust optimal control strategy that can meet the voltage support requirements under different uncertainty scenarios is selected. A dynamic adjustment mechanism for target weights is established. According to the voltage operation status of the distribution network (normal operation, critical over-limit, fault recovery, extreme weather) and voltage safety risk level, the weights of each target are adaptively adjusted to prioritize the dynamic balance between voltage stability and operational economy, and output the optimal control strategy.

[0010] Preferably, the collaborative injection control stage establishes a collaborative injection mechanism of virtual inertia, damping, and energy storage based on the optimal control strategy obtained in the strategy solution verification stage, thereby enhancing the execution effect of the control strategy. Based on the voltage fluctuation frequency, amplitude, and fluctuation trend of the distribution network, an adaptive virtual inertia control algorithm is designed to simulate the inertia characteristics of a synchronous generator by adjusting the virtual impedance output of the inverter. A dynamic adjustment model for the damping coefficient is established, in which the damping coefficient is positively correlated with the voltage change rate and is linked to the virtual inertia value for optimization, thereby avoiding voltage oscillation. Integrated energy storage equipment is controlled collaboratively. The energy storage equipment adopts a lithium battery energy storage system, which can quickly respond to replenish the power gap when the voltage fluctuates significantly or the photovoltaic output drops sharply. The design incorporates a multi-mode smooth switching mechanism. The switching trigger condition is set based on a threshold value set by the parameters. During normal operation, PQ control is used to ensure stable output. When voltage fluctuates, the system switches to virtual inertia and damping control mode. Under extreme conditions, the system switches to a virtual inertia, damping, and energy storage coordinated control mode to achieve voltage support under different operating conditions and output control execution status data.

[0011] Preferably, the autonomous collaborative switching phase establishes a distributed autonomous and cluster collaborative bidirectional switching mechanism with communication delay compensation and adaptive switching threshold; when communication is normal, a layered collaborative control mode formed in the layered architecture construction phase is adopted; an initial switching threshold for communication delay is set, and when a communication interruption or delay exceeding the threshold is detected, the system automatically switches to the distributed autonomous control mode. Each photovoltaic unit autonomously executes a voltage regulation strategy based on local measurement information, a preset voltage support threshold, and the scenario prediction results from the source-load adaptation generation phase, while simultaneously achieving short-term collaboration of units within the cluster through local edge nodes. An integrated handover criterion adaptive optimization algorithm dynamically adjusts the handover threshold based on multi-dimensional indicators including communication quality, voltage operation status, and scenario reliability; a communication delay compensation module is designed to predict scheduling instructions in advance through a prediction model when the communication delay has not reached the handover threshold. After communication is restored, a smooth transition strategy is designed. The transition path is optimized based on the simulation results of the digital twin mirror layer. The adjustment parameters of each unit are gradually adjusted to the coordinated control state, and the switching process data and control state data are output.

[0012] Preferably, the evaluation, tuning and optimization stage integrates the operation data and execution status of the entire process, and establishes a multi-dimensional control effect evaluation index system that includes collaborative control efficiency, uncertainty adaptation accuracy, voltage regulation accuracy, equipment loss and user satisfaction. The control effect is quantitatively evaluated by collecting control strategy execution data in real time through edge computing nodes. An improved reinforcement learning-based control parameter self-tuning algorithm is adopted, with the comprehensive optimality of multi-dimensional evaluation indicators as the reward function, to automatically optimize and adjust key parameters at each stage: the allocation coefficient of the cluster coordination layer, the virtual inertia and damping coefficient of the collaborative injection control stage, the prediction model parameters of the source load adaptation generation stage, and the switching threshold of the autonomous collaborative switching stage; the optimized parameters are fed back to the corresponding stages to achieve adaptive updating of parameters throughout the process. Establish an iterative optimization mechanism for control strategies, continuously optimize control models and algorithms by combining operational data and scenario verification results, integrate user feedback correction modules, collect controllable load operation experience data through user-side interactive terminals, adjust control strategies based on user operation experience of controllable loads, and the optimized parameters and strategies will be sent back to the aforementioned stages.

[0013] The process for quantifying user satisfaction is as follows: Questionnaire Survey Dimensions: Includes 3 core indicators, each rated on a 5-point scale, where 1 point = very dissatisfied and 5 points = very satisfied. ①Impact of regulation (weight 0.4): This includes whether the load operation is interrupted and whether the operating parameters fluctuate abnormally; ② Response timeliness (weight 0.3): This includes whether the response delay after the load receives the adjustment command is within an acceptable range; ③ Ease of operation and maintenance (weight 0.3): This includes whether manual intervention is required and whether control strategies increase the workload of operation and maintenance; Quantification formula: User satisfaction ,in For the survey sample size, The first Scores for moderating influence, response timeliness, and ease of operation and maintenance for each sample. ; Data update cycle: User satisfaction adopts a combination of real-time feedback and quarterly survey model. Feedback on abnormal operation of controllable load is collected in real time through user-side interactive terminals, and temporary satisfaction scores are generated weekly. A comprehensive calibration is carried out every quarter through questionnaires, which are synchronized to the evaluation indicator system to participate in parameter self-tuning. Real-time temporary scores ensure that parameter optimization responds to sudden user experience issues in a timely manner, and quarterly surveys ensure data accuracy, achieving a dynamic balance between technical indicators and user experience.

[0014] The beneficial effects of this invention are as follows: 1. This invention is based on a two-dimensional dynamic aggregation algorithm, which uses electrical distance, voltage influence weight, and load response coefficient as core indicators to achieve dynamic partitioning and real-time updating of photovoltaic clusters. It can accurately match the dynamic correlation characteristics of voltage and load under high photovoltaic penetration. At the same time, it constructs a four-level hierarchical collaborative control architecture, which integrates a virtual leadership mechanism and digital twin verification. Through hierarchical linkage of global scheduling, cluster coordination, and unit execution, it avoids the regulation conflict problem that is prone to occur when a single unit controls independently. It enables voltage regulation actions to form global coordination, reduces over-regulation or under-regulation, and improves voltage support efficiency and regulation accuracy.

[0015] 2. This invention utilizes an Attention-LSTM prediction model that integrates reinforcement learning to transform source-load uncertainty into a continuous Markov process. Combined with a full-process feedback correction mechanism, the prediction error is controlled within 5%. Simultaneously, a multi-dimensional scenario library covering traditional and extreme weather scenarios is constructed, retaining key scenarios with high probability and high impact. Furthermore, by combining a collaborative injection mechanism of virtual inertia, damping, and energy storage, precise voltage regulation under different operating conditions is achieved through smooth switching of multiple modes. This enhances the ability to resist disturbances caused by uncertainties such as intermittent photovoltaic output and extreme weather, avoids voltage overruns and inverter disconnection, and ensures the operational stability of the distribution network under complex operating conditions.

[0016] 3. This invention achieves bidirectional switching between distributed autonomy and cluster collaboration. When communication is normal, hierarchical collaborative control is adopted. Once communication is interrupted or the delay exceeds the limit, it automatically switches to distributed autonomy mode. Each photovoltaic unit autonomously adjusts its voltage based on local measurement information and scenario prediction results. After communication is restored, it returns to collaborative control through a smooth transition strategy to ensure the continuity of voltage regulation. At the same time, through a multi-dimensional evaluation index system and an improved reinforcement learning parameter self-tuning algorithm, it achieves adaptive updating of parameters throughout the process. Combined with user feedback, it continuously optimizes the control strategy, which not only reduces equipment loss and load interruption costs, but also improves the system's adaptability to high-proportion photovoltaic penetration scenarios, enhancing its feasibility and practicality in real-world applications. Attached Figure Description

[0017] Fig. 1 This is an overall flowchart of the method of the present invention; Fig. 2 This is a flowchart of the cluster dynamic partitioning and state evaluation process of the present invention; Fig. 3 This is a flowchart illustrating the multi-mode collaborative voltage regulation execution process of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figs. 1 to 3 As shown, this embodiment of the invention provides a voltage active support control method for distributed photovoltaic clusters adapted to high penetration rates, including a data acquisition stage, a source-load adaptation generation stage, a hierarchical architecture construction stage, a strategy solution verification stage, a collaborative injection control stage, an autonomous collaborative switching stage, and an evaluation, tuning, and optimization stage. The specific implementation of each stage is as follows: The data acquisition phase involves collecting multi-dimensional real-time data through edge computing nodes, including voltage, current, and power of each node in the power distribution network; output, inverter status, irradiance, and ambient temperature of distributed photovoltaic units; and response capabilities and operating status data of user-side controllable loads (such as industrial and commercial air conditioners and energy storage devices). The system utilizes 5G and edge computing fusion technology to achieve low-latency data transmission and local preprocessing, with data transmission latency controlled within 20ms. The preprocessing process includes data cleaning, format standardization, and outlier removal. Simultaneously, blockchain technology is integrated to establish a distributed trusted data sharing platform. Asymmetric encryption algorithms are used to encrypt data transmission, and a consensus mechanism is used to complete cross-cluster node identity authentication, ensuring the security and consistency of cross-cluster data interaction.

[0020] A two-dimensional dynamic aggregation algorithm based on graph theory community discovery and voltage sensitivity weighting is adopted. The first dimension uses the electrical distance between the photovoltaic unit and key nodes and the voltage influence weight as core indicators, while the second dimension uses the user's controllable load response coefficient as the optimization indicator. Clustering is achieved through an improved modularity function. By calculating the correlation strength matrix between nodes, the modularity threshold is set to 0.35 to achieve optimal clustering. Photovoltaic units with similar geographical locations, similar output characteristics, high voltage influence correlation, and strong user load response coordination are aggregated into a virtual cluster. The cluster update cycle is dynamically adjusted according to the voltage fluctuation frequency, shortened to 5 minutes / time when the fluctuation is severe, realizing dynamic partitioning and real-time updating of the cluster, and accurately matching the dynamic correlation characteristics of voltage and load under high penetration.

[0021] Core indicator calculation formula: ① Electrical distance The nodal impedance matrix is ​​used to calculate and reflect the electrical connection between photovoltaic unit i and key node j. The formula is as follows: ; in: Here is the self-impedance of photovoltaic unit i, in Ω; Here is the self-impedance of critical node j, in Ω; Let be the mutual impedance between photovoltaic unit i and critical node j, in Ω; obtained by solving the node impedance matrix through distribution network topology parameters.

[0022] ② Voltage influence weight Based on voltage sensitivity analysis, the formula is: ; in: The reactive power output of photovoltaic unit i is sensitive to the voltage of critical node j, in units of pu / Mvar; n is the total number of photovoltaic units. To perform the maximum value operation, .

[0023] ③ User-controllable load response coefficient The formula for quantifying the coordinated response capability of the load to voltage regulation is: ; in This represents the maximum adjustable power of the kth type of controllable load, in kW. Its rated power is expressed in kW; For response delay coefficient, Response delay Set the value to 1.0; if the delay is greater than 3 seconds, set the value to 0.8.

[0024] Two-dimensional aggregated comprehensive index: The aggregation priority of photovoltaic units is calculated using a weighted summation method. The formula is: ; in: For the weighting coefficients, satisfying ; Maximum electrical distance within the distribution network, in units of ; The larger the value, the better the compatibility of the photovoltaic unit with the target cluster.

[0025] A cluster-level state awareness model is established, and an improved Kalman filter algorithm (introducing a weight decay strategy to optimize sample evaluation and assigning a decay coefficient of 0.8 to old data) is used to reduce noise in the collected data. The model is used to evaluate the overall output capacity of the cluster, the voltage operation status of the distribution network, the user load response potential and potential over-limit risks in real time, and generate an evaluation report containing multiple state parameters.

[0026] The core iterative formula of the improved Kalman filter algorithm is as follows: State prediction equation: ; in: The predicted state value at time k includes state parameters such as photovoltaic output and node voltage. The state transition matrix is ​​set as a diagonal matrix based on the topology characteristics of the distribution network, with diagonal elements = 0.98; This is the filtered estimate at time k-1; The input matrix has a value of 0.02. This is the control input at time k-1, such as the inverter adjustment signal.

[0027] Covariance prediction equation: ; in: Predict the covariance matrix at time k; The covariance matrix at time k-1; This is the process noise matrix, with values ​​= diag([0.001, 0.001]), corresponding to the noise intensity of photovoltaic output and voltage.

[0028] Weight decay application: For historical data at time k-3 and earlier, a decay coefficient is introduced during covariance update. =0.8, that is This reduces the impact of outdated data on the current filtering results.

[0029] Filtering update equation: ; in For Kalman gain, ; Here is the observation value at time k, and represents the raw data collected. This is the observation matrix, with a value of 1.0. The observation noise matrix has a value of 0.002.

[0030] The source-load adaptation generation stage, based on the multi-dimensional state assessment data from the acquisition stage, employs an Attention-LSTM prediction model that integrates reinforcement learning. It describes the source-load uncertainty as a continuous Markov process, using decision time, photovoltaic output level, and load demand state as state components, and prediction error correction as action components. The model parameters are adaptively optimized through a reinforcement learning reward mechanism. During model training, batch gradient descent is used to optimize the loss function, with 1000 iterations to ensure convergence. This enables short-term prediction of distributed photovoltaic cluster output and distribution network load; the prediction duration is 15 minutes to 2 hours, and the prediction error is controlled within 5%, significantly improving the model's adaptability to random fluctuations.

[0031] Model structure details: The Attention-LSTM prediction model comprises an input layer, an Attention layer, an LSTM layer, a fully connected layer, and a reinforcement learning optimization module. The input layer is 12-dimensional, corresponding to six categories of historical time-series data collected: photovoltaic output, light intensity, ambient temperature, distribution network node voltage, current, and power. For each category, the first two time-series values ​​are used. The Attention layer employs the Bahdanau attention mechanism, calculating similarity weights between the input time-series data and the hidden layer states. The weight calculation formula is: ; in For attention vectors, The key temporal features are enhanced by a trainable weight matrix. The LSTM layer has 3 hidden layers, each with 64 neurons. The activation function is ReLU, and the dropout probability is set to 0.2 to prevent overfitting. The fully connected layer has a 2-dimensional output, corresponding to the predicted output of the photovoltaic cluster and the predicted load of the distribution network.

[0032] The fusion mechanism of reinforcement learning and Attention-LSTM: The reinforcement learning module adopts the DQN framework. Its inputs are three types of state parameters: the prediction error of the Attention-LSTM (the absolute deviation between the predicted and actual values ​​at the current time), the voltage fluctuation rate, and the scene credibility. The outputs are two types of action parameters: the weight adjustment of the LSTM hidden layer and the similarity threshold adjustment of the Attention layer. The reward function is defined as: ; in To predict output deviation, For voltage deviation, For the credibility of the scene, The weighting coefficient is used to stop parameter optimization when the reward value fluctuates less than 0.01 for 50 consecutive iterations.

[0033] Training data and steps: The training data uses historical data from the past year, divided into training, validation, and test sets in a 7:2:1 ratio, with a data sampling interval of 5 minutes. The training steps include: ① Initialize the weights of each layer of the Attention-LSTM using a Xavier normal distribution; ② Input the training set data and perform forward propagation to obtain the predicted value; ③ Calculate the loss function, using a hybrid loss function: 0.6MSE + 0.4MAE, where MSE is the mean squared error and MAE is the mean absolute error; ④ Update the parameters of the LSTM and Attention layers through backpropagation using batch gradient descent (batch size of 32, initial learning rate of 0.001, decaying to 0.9 every 200 iterations); ⑤ Introduce a reinforcement learning module to output weight corrections based on prediction errors and voltage states, and iteratively optimize the model; ⑥ Use the validation set to monitor overfitting. When the validation set loss increases for 100 consecutive iterations, stop training and save the optimal model.

[0034] A multi-dimensional uncertainty scenario generation system was constructed, including traditional light change, load impact scenarios, and special scenario modules for extreme weather (rainstorms, hail, strong gusts). The extreme weather scenario was constructed by correlation analysis of historical meteorological data and power grid operation data. Multiple source-load scenarios were generated by a composite method of Latin hypercube sampling and Monte Carlo simulation. First, Latin hypercube sampling was used to achieve uniform coverage of the parameter space. Then, Monte Carlo simulation was used to expand the number of scenarios to 1000. By improving the scenario reduction algorithm (combining the probability of scenario occurrence and the degree of voltage impact), 30-50 key scenarios were retained. Simulation verification showed that this number can cover more than 95% of high-probability and high-impact operating conditions, while keeping the robustness verification time within 10 minutes, balancing the comprehensiveness and efficiency of verification, and constructing a scenario library covering different operating conditions. A full-process feedback and correction mechanism is established, and a dynamic evaluation index for scenario credibility is introduced. This index is calculated by comprehensively considering the probability of scenario occurrence, the degree of voltage impact, and the data matching degree. The scenario library is continuously corrected by combining real-time operation data, eliminating scenarios with credibility below 0.6 and adding new scenarios. At the same time, the prediction error is fed back to the prediction model for parameter iteration and optimization, and the prediction results and scenario library are output.

[0035] The formula for calculating the scene credibility S(s) (where s is a single scene): ; in: The probability of scenario s occurring (in %) is obtained through historical data statistics (such as extreme rainstorm scenarios). After normalization ; The degree of voltage influence in scenario s; ; The data matching degree for scenario s; ; As weighting coefficients, scenarios with significant voltage impact and high data matching are prioritized for retention; when If the scenario is deemed low-belief, it will be removed.

[0036] In the layered architecture construction phase, the state assessment results from the acquisition phase are combined with the prediction results and scenario library from the source-load adaptation generation phase to construct a four-level layered collaborative control architecture consisting of a distribution network scheduling layer, a cluster coordination layer, a photovoltaic unit control layer, and a digital twin mirror layer. The distribution network scheduling layer, based on the global voltage constraints of the distribution network (voltage amplitude deviation ≤ ±2%) and the economic operation target, issues a total voltage support command to each photovoltaic cluster through global power balance calculation and voltage distribution simulation. The command includes core parameters such as total reactive power regulation and active power output limit. At the same time, it receives operating status feedback information from each cluster to achieve dynamic correction of global control.

[0037] The cluster coordination layer integrates a virtual leader and follower mechanism. The virtual leader node receives instructions from the distribution network dispatch layer and decomposes tasks, while the follower nodes synchronously execute adjustment tasks and provide status feedback. After receiving instructions from the distribution network dispatch layer, the virtual leader node combines the cluster's status awareness information and scenario prediction results to allocate voltage regulation tasks to each photovoltaic unit and associated controllable load through a distributed model predictive control (DMPC) algorithm. The algorithm's prediction time domain is set to 15 minutes, and the control time domain is set to 5 minutes to ensure that each unit in the cluster works collaboratively and avoids regulation conflicts. The photovoltaic unit control layer adopts a plug-and-play modular design with built-in standardized communication protocols and control interfaces. Based on the allocation instructions from the cluster coordination layer, it implements specific control strategies through inverters, with execution accuracy controlled within ±1%, and provides real-time feedback on execution status. It supports flexible expansion of photovoltaic and energy storage, and completes equipment debugging and protocol integration before leaving the factory, reducing on-site deployment costs and operation and maintenance difficulties. DMPC objective function (minimize regulation deviation and energy consumption): ; in: =15, which is the prediction time domain, in minutes, and m is the number of photovoltaic units in the cluster; The reactive power regulation reference value for the i-th photovoltaic unit at time t is given in Mvar and is obtained from the decomposition of the general command of the distribution network dispatching layer. Let Mvar be the actual reactive power output of the i-th photovoltaic unit at time t; The inverter regulation loss for the i-th photovoltaic unit is given in kW. ; As a weighting coefficient, priority is given to ensuring adjustment accuracy.

[0038] Constraints: ①Reactive power output constraint: ;

[0039] in , , Rated active power of the photovoltaic unit, in kW; ② Adjustment rate constraint: ,in The unit is Mvar / minute; ③ Load coordination constraints: Where I is the number of controllable loads, Let represent the regulating power of the Kth type load at time t, in kW.

[0040] The digital twin mirror layer constructs a virtual mirror of the distribution network and photovoltaic clusters. It dynamically updates based on real-time collected physical layer data, simulates the execution effect of different control strategies, and provides decision pre-verification for the distribution network dispatching layer and the cluster coordination layer. The control parameters corrected through virtual-real interaction, such as the voltage fluctuation exceeding the standard in the simulation, are fed back to the instruction generation stage of the distribution network dispatching layer and the task allocation stage of the cluster coordination layer, respectively, to achieve hierarchical parameter optimization.

[0041] In the strategy solution and verification stage, the optimization objectives are to minimize the voltage deviation of key nodes in the distribution network, maximize the output of distributed photovoltaic clusters, minimize the inverter regulation loss, and minimize the user load interruption cost. An optimization model is established that considers multiple constraints such as photovoltaic inverter capacity constraints, voltage regulation rate constraints, low voltage ride-through constraints, and user controllable load response capability constraints. The user controllable load response capability constraints are set with different adjustment upper limits according to the load type. An improved non-dominated sorting genetic algorithm (NSGA-Ⅲ) is used to solve the optimization model. By introducing an adaptive crossover and mutation operator and reinforcement learning-guided search direction optimization, the algorithm's performance is improved. The crossover probability is dynamically adjusted between 0.6 and 0.9 according to the population convergence degree, while the mutation probability is fixed at 0.01. At the same time, combined with reinforcement learning-guided search direction optimization, a reward function is used to select high-quality search paths, thereby improving the algorithm's convergence speed and solution accuracy, ensuring that a single solution is completed within 30 seconds.

[0042] Core modules for algorithm improvement: ① Adaptive crossover mutation operator: The crossover operator uses simulated binary crossover (SBX), with crossover probability... The formula for calculation is: ;

[0043] in MaxGen represents the maximum number of iterations (set to 200). The convergence of the population is judged by the variance of the crowding of non-dominated solutions. When the variance is less than 0.05, The mutation operator is fixed at 0.9; the mutation probability is polynomial mutation. =0.01, the variable asynchronous length is dynamically adjusted according to the variable boundary: ;

[0044] in To optimize the upper and lower bounds of variables.

[0045] ② Reinforcement learning guidance mechanism: The PPO algorithm is used to guide the search direction. The state space of reinforcement learning consists of the number of non-dominated solutions in the current population, the variance of the objective function value, and the voltage deviation qualification rate; the action space consists of the crossover operator type selection (SBX / single-point crossover) and the variable asynchronous length adjustment coefficient (0.8-1.2). The reward function is defined as: ,in The proportion of non-dominated solutions. The variance of the objective function value. The action with the highest reward value, representing the voltage deviation compliance rate, is used as the search strategy for the next iteration.

[0046] Algorithm and power distribution network scenario adaptation design: ① Chromosome encoding: Real number encoding is used, and the chromosome length is [missing information]. Where n is the number of photovoltaic units, set to a maximum of 30 per cluster; m is the number of controllable loads, set to a maximum of 20 per cluster; the gene values ​​correspond to the reactive power output ratio (0-1) of the photovoltaic inverter and the adjustment amount of the controllable load (-0.2pu~0.2pu).

[0047] ② Fitness function: The fitness calculation formula considering multiple constraints is as follows: ;

[0048] in For voltage deviation, Contribute to photovoltaic power To maximize the output of photovoltaic power, For inverter losses, For load interruption costs, The weighting is dynamic and adjusted according to the voltage operating status. Constraints are embedded through a penalty term: when the voltage regulation rate constraint is violated... When the inverter capacity constraint is violated, the fitness value is multiplied by 0.5; when the inverter capacity constraint is violated, the fitness value is multiplied by 0.3.

[0049] Solve for key parameters: The population size was set to 100, the maximum number of iterations was 200, the crowding threshold for non-dominated sorting was set to 0.1, the sampling interval for scenario robustness verification was 1 minute, and the time limit for a single solution was controlled within 30 seconds through parallel computing (using 4 threads).

[0050] Based on the scenario library generated during the source-load adaptation stage, the robustness of the optimization strategy is verified in multiple scenarios. Verification indicators include voltage deviation compliance rate, photovoltaic power output utilization rate, and equipment loss rate. The robust optimal control strategy that can meet voltage support requirements under different uncertainty scenarios is selected. A dynamic adjustment mechanism for target weights is established. This mechanism adaptively adjusts the weights of each target according to the voltage operation status of the distribution network (normal operation, critical limit exceedance, fault recovery, extreme weather) and voltage safety risk level. During normal operation, the focus is on economic targets (40% weighting), during critical limit exceedance and fault recovery, the focus is on voltage stability targets (60% weighting), and during extreme weather, all targets are considered in a balanced manner. Prioritizing the dynamic balance between voltage stability and operational economy, the optimal control strategy is output.

[0051] The collaborative injection control phase establishes a collaborative injection mechanism of virtual inertia, damping, and energy storage based on the optimal control strategy obtained from the strategy solution verification phase. This enhances the execution effect of the control strategy, improves the cluster's collaborative voltage support capability, source-load uncertainty anti-disturbance performance, and extreme operating condition survivability. Based on the distribution network voltage fluctuation frequency, amplitude, and fluctuation trend, an adaptive virtual inertia control algorithm is designed. By adjusting the virtual impedance output of the inverter, the inertia characteristics of the synchronous generator are simulated. The inertia value range is dynamically adjusted to 0.5-2.0H according to the voltage fluctuation amplitude, effectively suppressing rapid voltage fluctuations.

[0052] Adaptive virtual inertia calculation formula: ; in: The virtual inertia at time t is expressed in watts (H). ; for Voltage deviation at critical nodes at any given time, in units of pu. =0.02pu, which is the maximum allowable voltage deviation; This is a virtual inertia attenuation coefficient. Through simulation calibration, it is ensured that the greater the voltage deviation, the greater the inertia, and the stronger the effect of suppressing fluctuations.

[0053] Dynamic adjustment model for damping coefficient: ; And it satisfies the linkage constraint: The constant is expressed in H·pu·s / kV; where: is the damping coefficient at time t, with units of pu·s / kV; The reference damping coefficient; Let be the rate of change of voltage at time t, in units of pu / s; The damping factor is the amplification factor; the greater the voltage change rate, the greater the damping factor, thus preventing oscillation. Linkage constraints ensure coordinated optimization of inertia and damping, avoiding a decrease in voltage stability caused by adjusting a single parameter; Through multi-scenario simulation calibration, under typical operating conditions with voltage fluctuation amplitude of 0.01pu-0.05pu and change rate of 0.005-0.02pu / s, this parameter combination can shorten the voltage oscillation decay time by 30% and avoid the conflict between inertia and damping.

[0054] A dynamic adjustment model for the damping coefficient is established, and the damping coefficient is optimized in real time according to the voltage change rate. The damping coefficient is positively correlated with the voltage change rate and is linked to the virtual inertia value for optimization, so as to avoid voltage oscillation caused by excessive inertia injection. Integrated energy storage devices are controlled collaboratively. The energy storage devices adopt lithium battery energy storage systems with a response time of no more than 10ms. When there are large voltage fluctuations or a sudden drop in photovoltaic output, they can quickly respond to replenish the power gap.

[0055] The design incorporates a multi-mode smooth switching mechanism. The switching trigger condition is set based on thresholds for parameters such as voltage deviation and fluctuation rate. During normal operation, PQ control is used to ensure stable output. When voltage fluctuates, the system switches to virtual inertia and damping control mode. Under extreme conditions, the system switches to a virtual inertia, damping, and energy storage coordinated control mode to achieve precise voltage support under different operating conditions and output control execution status data.

[0056] The quantization trigger threshold for multi-mode switching is set as follows: Normal operating condition (PQ control mode): meets voltage deviation requirements. And voltage deviation ,in , For real-time voltage, Rated voltage, unit: pu; Fluctuating operating condition (virtual inertia and damping control mode): satisfies or ; Extreme operating conditions (inertia, damping, and energy storage coordinated control mode): meet the requirements or Or a sharp drop in photovoltaic output Rated output for photovoltaic power; Switching hysteresis logic: In order to avoid frequent switching, the mode switching must meet the trigger condition for a duration of more than 200ms. During the switching process, the control parameters are smoothly transitioned by linear interpolation, and the transition time is 50ms.

[0057] The autonomous and collaborative switching phase establishes a distributed autonomous and cluster collaborative bidirectional switching mechanism with communication delay compensation and adaptive switching threshold, thereby improving the reliability and anti-interference capability of the control method. When communication is normal, a hierarchical collaborative control mode formed during the hierarchical architecture construction phase is adopted to ensure the continuous execution of global optimization control. An initial switching threshold for communication delay is set. When a communication interruption or delay exceeding the threshold is detected, the system automatically switches to a distributed autonomous control mode. Each photovoltaic unit autonomously executes a voltage regulation strategy based on local measurement information, a preset voltage support threshold, and scenario prediction results from the source-load adaptation generation phase. At the same time, short-term collaboration among units within the cluster is achieved through local edge nodes to ensure local voltage stability.

[0058] An integrated adaptive optimization algorithm for switching criteria is used to dynamically adjust the switching threshold based on multiple dimensions such as communication quality, voltage operation status, and scenario reliability, thereby avoiding control instability caused by frequent switching. A communication delay compensation module is designed to predict scheduling instructions in advance through a prediction model when the communication delay has not reached the switching threshold, thereby compensating for the response lag caused by the delay.

[0059] Switching threshold Calculation formula (based on weighted fusion of multi-dimensional indicators): ; in: This is the initial switching threshold; Let t be the quantization value of communication quality. , Packet loss rate threshold = 0.05, latency fluctuation threshold = 20ms; This represents the quantized value of the voltage operating state at time t. ; Let t be the credibility of the scene. This refers to the scenario credibility index during the source-load adaptation phase. ; As a weighting factor, voltage stability is prioritized, followed by communication quality.

[0060] Additional information on switching logic: When communication delay When, autonomous mode is triggered; when When the collaborative mode is triggered, a lag threshold is used to avoid frequent switching. The real-time communication delay at time t is expressed in milliseconds (ms).

[0061] After communication is restored, a smooth transition strategy is designed. The transition path is optimized based on the simulation results of the digital twin mirror layer. The adjustment parameters of each unit are gradually adjusted to the cooperative control state to achieve seamless switching from autonomous control to cooperative control, ensuring the continuity and stability of voltage regulation, and outputting switching process data and control status data.

[0062] The evaluation, tuning and optimization stage integrates the operation data and execution status of the entire process, and establishes a multi-dimensional control effect evaluation index system that includes collaborative control efficiency, uncertainty adaptation accuracy, voltage regulation accuracy, equipment loss and user satisfaction. The control effect is quantitatively evaluated by collecting control strategy execution data in real time through edge computing nodes. An improved reinforcement learning-based control parameter self-tuning algorithm is adopted, with the comprehensive optimality of multi-dimensional evaluation indicators as the reward function, to automatically optimize and adjust key parameters at each stage: the allocation coefficient of the cluster coordination layer, the virtual inertia and damping coefficient of the collaborative injection control stage, the prediction model parameters of the source load adaptation generation stage, and the switching threshold of the autonomous collaborative switching stage; the optimized parameters are fed back to the corresponding stages to achieve adaptive updating of parameters throughout the process. Establish an iterative optimization mechanism for control strategies, continuously optimize control models and algorithms by combining operational data and scenario verification results, integrate user feedback correction modules, collect controllable load operation experience data through user-side interactive terminals, adjust control strategies based on user operation experience of controllable loads, and back down the optimized parameters and strategies to the aforementioned stages to improve the adaptability, long-term operational stability and user acceptance of control methods for high-proportion photovoltaic penetration scenarios, and ensure the feasibility and practicality of control schemes in practical applications.

[0063] Reinforcement learning model framework: The model employs the DDPG algorithm framework and includes an Actor network and a Critic network. ①Actor Network: The input layer has 8 dimensions, corresponding to collaborative control efficiency, voltage regulation accuracy, equipment loss, communication delay, voltage fluctuation amplitude, scenario credibility, user satisfaction, and photovoltaic power utilization rate; the hidden layer has 2 layers, each with 48 neurons, and the activation function is LeakyReLU; the output layer has 6 dimensions, corresponding to cluster allocation coefficient, virtual inertia coefficient, initial value of damping coefficient, prediction model learning rate, switching threshold baseline value, and switching threshold adjustment step size; the output value is mapped to the effective range of parameters through the Sigmoid function, for example, the virtual inertia coefficient is mapped to 0.5-2.0H.

[0064] ②Critic network: The input layer has 14 dimensions, including 8-dimensional state parameters and 6-dimensional action parameters; the hidden layer structure is the same as the Actor network, the output layer is single-valued, and the activation function is a linear function.

[0065] State space, action space, and reward function: ①State space definition: All state parameters are normalized to the interval [0, 1], where: Coordinated regulation efficiency = Actual regulation response time / Theoretical minimum response time; Voltage regulation accuracy = 1 - |actual voltage deviation| / upper limit of allowable deviation; Equipment loss = Actual loss / Rated loss; User satisfaction is converted into a 0-1 quantitative value through questionnaires.

[0066] ② Action space definition: The adjustment step size of each output parameter is 1% of its effective range. For example, if the effective range of the cluster allocation coefficient is 0.1-0.9, the adjustment step size is 0.008.

[0067] ③ Reward Function: The comprehensive reward function is: ; In the formula: (Voltage regulation accuracy bonus) = Voltage regulation accuracy; (Collaborative efficiency reward) = 1 - Collaborative regulation efficiency; (Loss Reward) = 1 - Equipment Loss; (User satisfaction reward) = User satisfaction; (Scene Adaptation Reward) = Scene Credibility; (Stability bonus) = 1 - Voltage fluctuation amplitude; when When the value is ≥0.85, the parameters are considered optimal, and the optimization round is stopped. 3. Training steps and iteration mechanism: ① Initialize the weights of the Actor and Critic networks using the He normal distribution, set the experience replay pool capacity to 10000, and the target network update coefficient to 0.005.

[0068] ② Collect current running status data, normalize it, input it into the Actor network, and output the initial action (parameter adjustment value).

[0069] ③ Adjust the execution parameters, collect the adjusted evaluation index data, and calculate the reward value.

[0070] ④ Store the state, action, reward, and next state in the experience replay pool, randomly sample batch data to train the Critic network, and minimize the mean squared error loss.

[0071] ⑤ The Actor network is updated using the policy gradient method to maximize the value evaluation of the Critic network.

[0072] ⑥ Update the target network every 10 iterations, and verify the parameter optimization effect every 50 iterations. If the effect is verified for 3 consecutive iterations... If the fluctuation is less than 0.02, the optimal parameters are output and fed back to the corresponding stage; if the requirement is not met after 500 iterations, the current optimal value is output.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for active voltage support control of distributed photovoltaic clusters adapted to high penetration rates, characterized in that, It includes the acquisition phase, the source payload adaptation and generation phase, the layered architecture construction phase, the strategy solving and verification phase, the collaborative injection and control phase, the autonomous collaborative switching phase, and the evaluation, tuning and optimization phase. Data Acquisition Phase: Multi-dimensional real-time data is collected and preprocessed through edge computing nodes. A two-dimensional dynamic aggregation algorithm is used to dynamically partition the cluster and update the cluster in real time. The collected data is denoised and evaluated in real time to generate an evaluation report containing state parameters. Source and load adaptation generation stage: Reinforcement learning fusion prediction model is used to predict source and load. A full-condition scenario library is generated through a combination of sampling and simulation methods. Low-confidence scenarios are eliminated, and the prediction error is fed back to the prediction model. Layered architecture construction phase: Establish a four-level layered collaborative control architecture, clarify the functional positioning and interaction logic of each level, integrate virtual leadership collaboration mechanism and modular design, realize virtual-real linkage verification through digital twin mirror, and output layered control commands; Strategy solution verification stage: Establish a multi-objective optimization model, use an improved genetic algorithm to solve the optimization model, perform multi-dimensional robustness verification in combination with the scenario library, and establish a dynamic weight adjustment mechanism to select the optimal control strategy; Collaborative injection control phase: Based on the optimal strategy, a collaborative injection mechanism for virtual inertia, damping and energy storage is established. The control mode is divided according to the voltage operating condition. Through smooth switching of multiple modes, precise voltage regulation under different operating conditions is achieved, and the control execution status and collaborative response effect are synchronously fed back. Autonomous and coordinated handover phase: Establish communication delay compensation and adaptive handover threshold mechanism to ensure the continuity and stability of voltage regulation, and synchronously upload handover process data and control status data; Evaluation, tuning, and optimization phase: Establish a multi-dimensional control effect evaluation index system, optimize and adjust key parameters in hierarchical collaborative control through control parameter self-tuning algorithms, and iteratively optimize control strategies by combining full-process data.

2. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 1, characterized in that, The data acquisition phase uses multi-dimensional real-time data collected by edge computing nodes, including voltage, current, and power of each node in the distribution network, output of distributed photovoltaic units, inverter status, irradiance, ambient temperature, and response capability and operating status data of user-side controllable loads. Low-latency transmission and local preprocessing technologies are used to process the data, and trusted data sharing technology is integrated to achieve data transmission encryption and cross-cluster node identity authentication. The dual-dimensional dynamic aggregation algorithm uses the electrical distance between the photovoltaic unit and the key node and the voltage influence weight as the first-dimensional core indicators, and the user controllable load response coefficient as the second-dimensional optimization indicator. Through a clustering algorithm, photovoltaic units with similar geographical locations, similar output characteristics, high voltage influence correlation and strong user load response coordination are aggregated into a virtual cluster. The cluster update cycle is dynamically adjusted according to the voltage fluctuation frequency. Establish a cluster-level state awareness model, perform noise reduction processing on the collected data, and evaluate the overall output capacity of the cluster, the voltage operation status of the distribution network, the user load response potential and potential over-limit risks in real time, and generate an evaluation report.

3. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 2, characterized in that, The source-load adaptation generation stage is based on the multi-dimensional state evaluation data from the acquisition stage. It adopts a time-series prediction model that integrates reinforcement learning to describe the source-load uncertainty as a continuous Markov process. The model parameters are adaptively optimized through the reward mechanism of reinforcement learning to achieve short-term prediction of the output of distributed photovoltaic clusters and the load of distribution networks. A multi-dimensional uncertainty scenario generation system is constructed, which includes traditional scenarios and extreme weather-specific scenarios. Multiple sets of source-load scenarios are generated using a composite sampling simulation method. By improving the scenario reduction algorithm, high-probability and high-impact key scenarios are retained, and a scenario library covering different operating conditions is constructed. A full-process feedback and correction mechanism is established, introducing dynamic evaluation indicators for scenario credibility. The scenario library is continuously corrected by combining real-time operation data, eliminating low-credibility scenarios and adding new scenarios. At the same time, the prediction error is fed back to the prediction model for parameter iteration and optimization.

4. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 3, characterized in that, The layered architecture construction phase combines the state assessment results from the acquisition phase with the prediction results and scenario library from the source-load adaptation generation phase to construct a four-level layered collaborative control architecture consisting of a distribution network scheduling layer, a cluster coordination layer, a photovoltaic unit control layer, and a digital twin mirror layer. Based on the global voltage constraints and operational economic objectives of the distribution network, the distribution network dispatch layer issues a general voltage support command to each photovoltaic cluster and receives operational status feedback information. The cluster coordination layer integrates a virtual leader and follower mechanism. After receiving the dispatch command, it combines local status information and scenario prediction results to allocate voltage regulation tasks through a distributed optimization algorithm. The photovoltaic unit control layer adopts a plug-and-play modular design, which executes control strategies and provides feedback on status through the inverter based on allocation commands; The digital twin mirror layer constructs a virtual mirror of the distribution network and the photovoltaic cluster, simulating the execution effect of different control strategies, and providing decision pre-verification for the distribution network dispatching layer and the cluster coordination layer. The control parameters corrected through virtual-real interaction are fed back to the instruction generation stage of the distribution network dispatching layer and the task allocation stage of the cluster coordination layer, respectively.

5. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 4, characterized in that, The strategy solution and verification phase takes minimizing the voltage deviation of key nodes in the distribution network, maximizing the output of distributed photovoltaic clusters, minimizing inverter regulation losses, and minimizing user load interruption costs as the multi-objective optimization objectives. An optimization model is established that considers photovoltaic inverter capacity constraints, voltage regulation rate constraints, low voltage ride-through constraints, and user controllable load response capability constraints. An improved non-dominated sorting genetic algorithm is used to solve the optimization model, and the solution performance is improved by optimizing the algorithm operators. Based on the scenario library, the robustness of the optimization strategy is verified in multiple scenarios, and the optimal robust control strategy is selected. Establish a dynamic adjustment mechanism for target weights, and adaptively adjust the weights of each target based on the voltage operation status and voltage safety risk level of the distribution network.

6. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 5, characterized in that, The collaborative injection control stage is based on the optimal control strategy, establishes a collaborative injection mechanism of virtual inertia, damping and energy storage, designs an adaptive virtual inertia control algorithm, establishes a dynamic adjustment model of damping coefficient, the damping coefficient is positively correlated with the voltage change rate, and forms a linkage optimization with the virtual inertia value. Integrated energy storage devices are coordinated and controlled to supplement power gaps when voltage fluctuates significantly or photovoltaic output drops sharply. The design incorporates a multi-mode smooth switching mechanism. The switching trigger condition is set based on a threshold value set by the parameters. During normal operation, PQ control is used. When voltage fluctuates, the mechanism switches to virtual inertia and damping control mode. Under extreme conditions, the mechanism switches to a virtual inertia, damping, and energy storage coordinated control mode to achieve voltage support under different operating conditions.

7. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 6, characterized in that, The autonomous and collaborative switching phase establishes a distributed autonomous and cluster collaborative bidirectional switching mechanism with communication delay compensation and adaptive switching threshold; when communication is normal, a hierarchical collaborative control mode is adopted; when communication is interrupted or the delay exceeds the threshold, it automatically switches to the distributed autonomous control mode, and each photovoltaic unit autonomously executes the voltage regulation strategy based on local measurement information, preset voltage support threshold and scenario prediction results from the source-load adaptation generation phase. An integrated adaptive optimization algorithm for handover criteria is used to dynamically adjust the handover threshold based on multi-dimensional indicators. Design a communication delay compensation module to predict scheduling instructions in advance when the communication delay has not reached the handover threshold; After communication is restored, a smooth transition strategy is adopted. The transition path is optimized based on the simulation results of the digital twin mirror layer, and the adjustment parameters of each unit are gradually adjusted to the coordinated control state.

8. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 7, characterized in that, The evaluation, tuning, and optimization phase integrates the operational data and execution status of the entire process, establishes a multi-dimensional control effect evaluation index system that includes collaborative control efficiency, uncertainty adaptation accuracy, voltage regulation accuracy, equipment loss, and user satisfaction, and quantitatively evaluates the control effect. An improved reinforcement learning-based control parameter self-tuning algorithm is adopted, which uses the comprehensive optimality of multi-dimensional evaluation indicators as the reward function to automatically optimize and adjust key parameters at each stage. The optimized parameters are fed back to the corresponding stages to achieve adaptive parameter updates throughout the entire process. Establish an iterative optimization mechanism for control strategies, continuously optimize control models and algorithms by combining operational data and scenario verification results, and integrate user feedback correction modules to adjust control strategies.

9. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 8, characterized in that, The acquisition phase employs an improved Kalman filter algorithm to reduce noise in the acquired data and assess potential risk of exceeding limits in real time.

10. The active voltage support control method for distributed photovoltaic clusters adapted to high penetration rates according to claim 9, characterized in that, The source-load adaptation generation stage retains key scenes by improving the scene reduction algorithm and combining the probability of scene occurrence with the degree of voltage influence.

Images