Photovoltaic distribution network voltage flexible adjustment method and system

Abstract

This invention belongs to the field of distribution voltage regulation technology and discloses a flexible voltage regulation method and system for photovoltaic distribution networks. It can construct a three-dimensional spatiotemporal sensitivity tensor of spatial nodes, time-series operating conditions, and regulation resources, integrating multi-dimensional operating parameters such as line impedance and equipment delay. Based on a weighted adaptive spectral clustering algorithm, it dynamically divides the voltage control area of ​​the distribution network. The system performs real-time voltage regulation in zones through edge computing devices, dynamically allocates equipment regulation amounts according to a five-level flexible regulation priority, collects real-time operating data within zones and between zone boundaries, performs neighborhood collaborative fine-tuning, and configures voltage offset transfer suppression constraint logic. This makes voltage regulation more targeted and stabilizes the voltage operating status of each node. By optimizing a long short-term memory neural network, it accurately predicts the time-series data of photovoltaic output and load power in each zone, builds a voltage risk assessment model, classifies risk levels, and matches corresponding pre-control logic to achieve graded early warning and pre-control.

Description

Technical Field

[0001] This invention belongs to the field of power distribution voltage regulation technology, specifically a method and system for flexible regulation of photovoltaic power distribution network voltage. Background Technology

[0002] Traditional distribution networks are designed with a single radiation source-load power supply mode, resulting in unidirectional power flow. Voltage regulation mainly relies on traditional equipment such as on-load tap-changing transformers (OLTCs) and switching capacitor banks. With the integration of a high proportion of distributed photovoltaic (PV) power, the power flow in the distribution network exhibits significant bidirectional characteristics. During peak PV generation at midday, a large amount of surplus power is fed back into the grid, causing voltage rises at the line ends, sometimes exceeding the upper limit. Conversely, during the evening when PV power generation is lower, peak loads may lead to lower voltage. Voltage fluctuations far exceed the regulation capabilities of traditional control methods. PV output exhibits strong randomness and volatility; cloud cover can cause drastic power changes within seconds. Existing PV distribution network voltage regulation methods are mainly divided into three categories: The first category is local control based on photovoltaic inverters, including Q(U) droop control and active power limiting strategies. Q(U) droop control has a fast response speed, but it will occupy inverter capacity and may increase line losses; active power limiting strategies are simple to control and take effect quickly, but they will reduce photovoltaic utilization and cause power generation loss.

[0003] The second type is centralized optimization control, which determines the control commands for each adjustable device by solving a global optimization problem. However, centralized control places excessive demands on system communication and data processing capabilities, making it difficult to apply widely in low-voltage distribution networks and prone to single-point failures.

[0004] The third type is distributed cooperative control, where each photovoltaic unit only exchanges information with its neighboring units and reaches a cooperative consensus through iteration. However, existing distributed control methods suffer from problems such as slow convergence speed, low adjustment accuracy, and difficulty in achieving global optimality.

[0005] Most existing methods do not fully consider the differences in time scale and regulation cost of different regulation resources, lack early prediction and graded response mechanisms for voltage over-limit risks, and fail to effectively integrate flexible regulation resources such as user-side energy storage and controllable loads, resulting in low regulation efficiency and poor economic performance. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for flexible regulation of photovoltaic distribution network voltage to solve one or more problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for flexible regulation of photovoltaic distribution grid voltage, comprising the following specific steps: Preferably, in the dynamic partitioning construction stage, the system collects real-time operating data of voltage, current, and power flow of each node in the distribution network, and simultaneously collects real-time operating status and capacity constraint data of distributed photovoltaic, user-side energy storage, flexible controllable load, on-load tap changer (OLTC), and capacitor bank. Based on the collected operating data, a three-dimensional spatiotemporal sensitivity tensor of spatial nodes, time-series operating conditions, and regulating resources is constructed. The perturbation method is used to calculate the voltage response sensitivity of each node at different times and under the action of different regulating resources online. The calculation process simultaneously incorporates multi-dimensional parameters such as line impedance, topology correlation degree, and equipment response delay to characterize the spatiotemporal operating characteristics of the distribution network. Based on the aforementioned three-dimensional spatiotemporal sensitivity tensor, the system completes the voltage control partitioning of the distribution network through a weighted adaptive spectral clustering algorithm. The clustering calculation is performed based on the spatiotemporal sensitivity similarity of nodes, voltage correlation strength, and regional power flow interaction characteristics. The voltages of different partitions exhibit weak correlation characteristics. During the algorithm operation, the clustering weights can be dynamically adjusted according to the intensity of photovoltaic fluctuations in the power grid and the peak and valley operation status of the load. The voltage control partitions are dynamically iterated and updated according to a fixed period.

[0008] Preferably, in the graded risk warning stage, the system uses meteorological time series data and historical power grid operation big data to optimize the long short-term memory neural network to predict the time series change curves of photovoltaic output and load power of each voltage zone. It also uses the real-time updated three-dimensional spatiotemporal sensitivity tensor to iteratively deduce the voltage distribution situation of each zone at each moment. At the same time, it builds a voltage risk assessment model based on four dimensions: voltage over-limit amplitude, over-limit duration, fluctuation rate, and regional influence range, and quantitatively calculates the voltage over-limit risk index of each zone. The system classifies the grid voltage operation risk into four levels: no risk, low risk, medium risk, and high risk, based on the quantified voltage over-limit risk index. Each risk level corresponds to a zone pre-control logic. Under the no risk level, the grid equipment maintains a steady-state normal operation. Under the low risk level, each zone performs a slight predictive fine-tuning operation and reserves grid regulation margin. Under the medium risk level, the system initiates the day-ahead plan correction process and completes the pre-scheduling configuration of zone regulation resources. Under the high risk level, flexible regulation resources enter a standby state and the frequent operation of voltage regulating equipment is restricted. The system also establishes a human-machine collaborative early warning system, which pushes corresponding early warning information and pre-control execution suggestions to the dispatching backend and user-side terminal equipment according to different voltage risk levels.

[0009] Preferably, in the day-ahead global pre-control stage, the system constructs a multi-objective optimization model with the three-dimensional optimization objectives of optimal operating cost, minimum voltage deviation, and maximum photovoltaic absorption under the day-ahead scheduling time scale. The multi-objective optimization model takes into account the cost characteristics and time scale characteristics of the flexible and rigid regulation resources of the distribution network. The total operating cost of the system includes the cost of photovoltaic curtailment penalty, the cost of energy storage charging and discharging loss, the cost of OLTC tap changer wear, the cost of capacitor bank switching, and the cost of user controllable load regulation compensation. The constraints of the multi-objective optimization model include node voltage constraints, line power constraints, and equipment capacity constraints. It also includes zonal correlation constraints, equipment operation frequency timing constraints, and risk reserve capacity constraints. The multi-objective optimization model combines the day-ahead voltage risk prediction results to configure a flexible adjustment capacity reserve mechanism for medium and high-risk zones. The system solves the multi-objective optimization model through an adaptive weighted improved particle swarm algorithm and iteratively outputs the next day's time-period OLTC tap position, capacitor bank switching status, energy storage charging and discharging timing plan, controllable load adjustment scheme, and photovoltaic inverter baseline operating parameters. Recently, the global pre-control strategy has been linked with the results of dynamic voltage zoning and hierarchical risk warning. The control strategy is configured according to the operating conditions of each zone, realizing the combination of overall scheduling and zone pre-control.

[0010] Preferably, in the intraday rolling correction phase, the system sets a 1-hour optimization time window and a 15-minute rolling iteration cycle, continuously collects the latest photovoltaic output and load power prediction data during the intraday regulation process, and synchronously updates the three-dimensional spatiotemporal sensitivity tensor parameters and partition association constraint parameters to realize the dynamic refresh of the optimization model input parameters; The system repeatedly solves the multi-objective optimization model with the optimization objectives of minimizing the operating cost of each zone and minimizing the voltage deviation of each zone in the next hour, and completes the dynamic correction of the intraday control plan. During the correction process, the control parameters and operation plan of fast response resources are adjusted first, and the number of daily operations of slow mechanical equipment is limited. Intraday rolling correction and day-ahead global pre-control form a time-sequential progressive operation logic. Through the operation mode of global overall planning, regional dynamic correction, and time-sharing coordination of fast and slow resources, a long-term scale control closed loop for the distribution network is constructed.

[0011] Preferably, in the real-time control phase of the partition, each voltage control partition is configured with an independent edge computing node for real-time voltage control calculation and strategy execution within the partition. The edge computing node is configured with a five-level flexible adjustment priority based on real-time spatiotemporal sensitivity data. The five-level flexible adjustment priority includes reactive power priority, active power supplementation, energy storage as the main force, load backup, and mechanical equipment steady-state protection. The edge computing node monitors the voltage and power flow operation status of each node in the partition in real time, predicts the trend of voltage exceeding the limit, and combines the response speed, adjustment cost, and spatiotemporal sensitivity coefficient of the adjustment resources to complete the dynamic allocation of the real-time adjustment amount of the control equipment. Each partition's edge computing nodes interact in real time with boundary node power flow and voltage operation data, configure inter-partition voltage offset transfer suppression constraint logic, maintain stable voltage operation through collaborative fine-tuning of neighboring nodes, and realize distributed voltage regulation of the distribution network through the edge computing architecture.

[0012] Preferably, in the parameter adaptation stage, the system adopts a reinforcement learning operation mechanism, and completes the dynamic adaptive optimization of control parameters through a dual closed-loop iterative optimization logic consisting of an outer loop adaptive parameter tuning closed loop and an inner loop iterative optimization closed loop for control effect. The outer loop adaptive parameter tuning closed loop collects real-time operating data such as grid voltage deviation, adjustment delay, line loss, photovoltaic absorption rate, and equipment operating cost, and completes online iterative adjustment of core control parameters through a deep reinforcement learning algorithm. The control parameters include QU droop control slope, active power limiting threshold, energy storage charging and discharging priority, controllable load adjustment coefficient, partition clustering weight, and rolling optimization step size. All parameters are dynamically updated according to the spatiotemporal operating conditions of the power grid. The inner loop control effect iteratively optimizes the steady-state accuracy, dynamic response speed, and operational economy of the closed-loop real-time monitoring voltage control. The control effect data is used as the input parameter of the reinforcement learning reward function to iteratively optimize the control strategy and parameter configuration. The system builds a parameter safety constraint threshold library to limit the operating range of the control parameters.

[0013] Preferably, in the fault self-healing and voltage stabilization stage, when a distribution network fault causes abnormal voltage fluctuations, the system starts a partitioned reconfiguration type rapid self-healing operation mechanism. First, it completes the fault location, fault type identification and fault impact range calculation, isolates the fault area, and blocks the fault spread. At the same time, it recalculates and constructs a three-dimensional spatiotemporal sensitivity tensor based on the power grid topology after the fault, updates the voltage control partition, and matches the power grid operation architecture under the fault state. The system implements a tiered emergency control strategy for non-faulty operating areas, smoothing out abnormal voltage fluctuations according to the priority order of photovoltaic inverter reactive power rapid response, energy storage active power support, controllable load flexible adjustment, and precise disconnection of a small number of non-critical loads. The fault control process is synchronously linked with the reinforcement learning operation mechanism, calling the control parameters adapted to the fault condition. After the grid fault is restored, the system performs a gradient-based gradual rollback of control parameters and a zone reset operation, switching to the conventional flexible control operation mode.

[0014] This invention also provides a photovoltaic distribution network voltage flexible regulation system, based on the above method, comprising the following modules: The dynamic partitioning construction module is used to collect electrical operation parameters and equipment status information of the distribution network, integrate multi-dimensional parameters such as line impedance and equipment delay, and construct a three-dimensional spatiotemporal sensitivity tensor. Based on the weighted adaptive spectral clustering algorithm, it dynamically divides the voltage control region, and can adjust the clustering parameters according to photovoltaic fluctuations and load conditions, and periodically update and output the partitioning results. The graded risk early warning module retrieves dynamic zoning and spatiotemporal sensitivity tensor data, integrates historical meteorological and power grid data, and predicts zoning photovoltaic output and load time series data by optimizing long short-term memory neural networks; it builds a voltage risk assessment model, quantifies the risk index and divides it into four risk levels, matches the corresponding pre-control logic, and pushes early warning and disposal suggestions in a graded manner. The global pre-control module, based on partition risk data, constructs a multi-objective optimization model with the goals of optimizing operation and maintenance costs, minimizing voltage deviation, and maximizing photovoltaic absorption. It solves the model using an adaptive weighted improved particle swarm algorithm and outputs the equipment operation status and resource pre-control plan for each time period of the next day. The intraday rolling correction module sets a 1-hour optimization window and a 15-minute iteration cycle to update photovoltaic and load forecasts and spatiotemporal parameters in real time, and iteratively optimize the model; it prioritizes adjusting parameters of fast-response equipment, constrains the frequency of action of slow-speed mechanical equipment, and dynamically corrects the intraday control plan. The partition real-time control module, based on the partition edge computing device, dynamically allocates the device adjustment amount according to the five-level adjustment priority, collects partition operation data and inter-partition boundary conditions in real time, performs neighborhood collaborative fine-tuning, and configures voltage offset suppression constraints. The parameter adaptive module, based on a dual-closed-loop iterative optimization architecture, collects control operation data, iteratively optimizes various core control parameters online through reinforcement learning algorithms, and dynamically updates parameter configuration strategies by combining a parameter safety constraint threshold library. The fault self-healing voltage regulation module monitors abnormal voltage fault signals in the power grid in real time, completes fault location, identification and range calculation, isolates the fault area, reconstructs the power grid topology and voltage partition, calls the appropriate parameters to perform emergency voltage regulation, and resets the partition and control parameters step by step after the fault is restored.

[0015] The beneficial effects of this invention are as follows: 1. This invention constructs a three-dimensional spatiotemporal sensitivity tensor of spatial nodes, temporal operating conditions, and regulating resources, integrates multi-dimensional operating parameters such as line impedance and equipment delay, and realizes dynamic division of distribution network voltage control areas based on weighted adaptive spectral clustering algorithm to adapt to the real-time operating conditions of the power grid; the system performs real-time voltage regulation of zones through edge computing devices, dynamically allocates equipment regulation amounts according to the five-level flexible regulation priority, collects operating data within zones and between zone boundaries in real time, performs neighborhood collaborative fine-tuning and configures voltage offset transfer suppression constraint logic; it can realize distributed voltage regulation of distribution networks, making voltage regulation more targeted, stabilizing the voltage operating status of each node, and improving the real-time performance, accuracy and zone coordination of voltage regulation.

[0016] 2. This invention accurately predicts the time-series data of photovoltaic output and load power in each region by optimizing the long short-term memory neural network, builds a voltage risk assessment model and classifies risk levels, and matches corresponding pre-control logic to realize graded early warning and pre-control. The system constructs a multi-time-scale optimization system with day-ahead global pre-control and intraday rolling correction. It solves the multi-objective optimization model with the goals of optimal operating cost, minimum voltage deviation and maximum photovoltaic absorption, dynamically corrects the equipment control plan, and coordinates the time-sharing and collaborative operation of flexible and rigid regulation resources.

[0017] 3. This invention adopts a dual-closed-loop iterative optimization architecture, using a deep reinforcement learning algorithm to iteratively adjust the core control parameters online, and combining this with a parameter safety constraint threshold library to limit the parameter operating range, thereby achieving adaptive optimization of control parameters under all operating conditions. The system has fault self-healing voltage regulation capability, and can complete fault location, area isolation and voltage partition reconstruction, call the appropriate control parameters to perform emergency voltage regulation operations, and reset the partition and control parameters step by step after the fault is eliminated. It can optimize the control strategy and parameter configuration, and improve the power grid fault response and self-healing capability. Attached Figure Description

[0018] Figure 1 This is a flowchart of the photovoltaic distribution network voltage flexible adjustment process of the present invention; Figure 2 This is a flowchart of the real-time zone control process of the present invention. Detailed Implementation

[0019] 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.

[0020] like Figures 1 to 2 As shown, this embodiment of the invention provides a method for flexible regulation of photovoltaic distribution network voltage, including the following specific steps: In this embodiment of the invention, during the dynamic partitioning construction phase, the system collects real-time operating data of voltage, current, and power flow of each node in the distribution network, and simultaneously collects real-time operating status and capacity constraint data of distributed photovoltaic, user-side energy storage, flexible controllable load, on-load tap changer (OLTC), and capacitor bank. Based on the collected operating data, a three-dimensional spatiotemporal sensitivity tensor of spatial nodes, time-series operating conditions, and regulating resources is constructed. The perturbation method is used to calculate the voltage response sensitivity of each node at different times and under the action of different regulating resources online. The calculation process simultaneously incorporates multi-dimensional parameters such as line impedance, topology correlation degree, and equipment response delay to characterize the spatiotemporal operating characteristics of the distribution network. The three-dimensional spatiotemporal sensitivity tensor is composed of spatial node dimension, temporal condition dimension, and regulating resource dimension. The spatial node dimension covers all busbars and key load nodes of the distribution network. The temporal condition dimension divides the entire day's operating segments into 15-minute time granularities. The regulating resource dimension includes all adjustable equipment such as distributed photovoltaic, energy storage, flexible loads, voltage regulating transformers, and capacitor banks. The perturbation method involves applying a constant small-amplitude perturbation signal of 5% of the rated output to a single regulating resource while keeping the operating status of other equipment unchanged. The voltage change amplitude and response delay of all network nodes are collected simultaneously. By collecting three consecutive perturbation samples and eliminating abnormal values ​​such as electromagnetic interference and data jumps, the average value is taken to obtain a stable and reliable voltage response sensitivity value. The calculation process automatically integrates line impedance parameters, topology node correlation degree, and equipment response delay parameters to complete the real-time correction of the sensitivity value.

[0021] The three-dimensional spatiotemporal sensitivity tensor adopts a synchronization mechanism that combines real-time updates and periodic refreshes. Equipment adjustment resources and node voltage data are collected and updated every second, and time-series operating condition data are refreshed as a whole every 15 minutes. After the tensor data is updated, it is immediately synchronized to all modules such as risk warning, global pre-control, and real-time regulation.

[0022] Based on the aforementioned three-dimensional spatiotemporal sensitivity tensor, the system completes the voltage control partitioning of the distribution network through a weighted adaptive spectral clustering algorithm. The clustering calculation is performed according to the spatiotemporal sensitivity similarity of nodes, voltage correlation strength, and regional power flow interaction characteristics, ensuring that the voltage of nodes within a partition has strong correlation characteristics, while the voltage of different partitions exhibits weak correlation characteristics. During the algorithm operation, the clustering weights can be dynamically adjusted according to the intensity of photovoltaic fluctuations in the power grid and the peak and valley operation status of the load. The voltage control partitions are dynamically iterated and updated according to a fixed period.

[0023] The weighted adaptive spectral clustering algorithm takes a three-dimensional spatiotemporal sensitivity tensor as input and constructs a similarity matrix based on node voltage correlation strength and power flow interaction degree. The weighted adaptive rule is as follows: when the photovoltaic fluctuation intensity is ≥30%, the temporal sensitivity weight is increased, and when the load peak-valley difference is ≥50%, the spatial sensitivity weight is increased. K-means is used to determine the number of partition clusters and outputs the weak correlation voltage partitioning results. The partitioning results are directly used as the spatial basis data for risk warning, global pre-control, and real-time regulation.

[0024] The dynamic adjustment rule for clustering weights is as follows: when the photovoltaic fluctuation intensity is in the range of 30%-50%, the weight of the time-series operating condition dimension is increased by 20%, and the weight of the adjustment resource dimension is decreased by 10%; when the photovoltaic fluctuation intensity exceeds 50%, the weight of the time-series operating condition dimension is increased by 40%, and the weight of the spatial node dimension is adjusted accordingly; when the load peak-valley difference is in the range of 50%-80%, the weight of the spatial node dimension is increased by 15%; when the load peak-valley difference exceeds 80%, the weight of the spatial node dimension is increased by 30%. All weight adjustments are executed in a closed loop within the range of 0-1.

[0025] The spatiotemporal sensitivity similarity of the nodes is determined by calculating the closeness of the voltage response values ​​between different nodes. The voltage correlation strength is comprehensively measured by the power flow transmission coefficient between nodes and the electrical distance. The regional power flow interaction characteristics are determined by the partition power exchange. A spectral clustering similarity matrix is ​​constructed based on the above indicators, and the matrix element values ​​are positively correlated with the node correlation degree. The number of clusters is determined by the adaptive contour coefficient method. The partition contour coefficients under different cluster numbers are calculated through iteration, and the cluster number corresponding to the maximum value of the coefficient is selected as the final number of voltage partitions.

[0026] The fixed update cycle is set to 15 minutes by default. Every 15 minutes, the system automatically re-collects the network operation data, refreshes the three-dimensional spatiotemporal sensitivity tensor, and re-executes the spectral clustering partition calculation. At the same time, a dynamic trigger update mechanism is set. When the photovoltaic output fluctuation exceeds 30%, the load peak-valley difference exceeds 50%, or the grid topology changes, the periodic update process is interrupted and real-time partition update is forcibly executed.

[0027] In this embodiment of the invention, during the graded risk warning stage, the system, based on meteorological time series data and historical power grid operation big data, optimizes the long short-term memory neural network to predict the time series change curves of photovoltaic output and load power of each voltage zone in the next 24 hours. Combined with the real-time updated three-dimensional spatiotemporal sensitivity tensor, it iteratively deduces the voltage distribution situation of each zone at each moment. At the same time, a voltage risk assessment model is built based on four dimensions: voltage over-limit amplitude, over-limit duration, fluctuation rate, and regional influence range, and the voltage over-limit risk index of each zone is quantitatively calculated. The optimized long short-term memory neural network adopts a three-layer structure of input layer-hidden layer-output layer: the input layer receives four types of feature vectors: meteorological time series data, power grid historical operation data, partition number, and timestamp; the hidden layer is equipped with 64 / 128 dual-dimensional LSTM neurons, and Dropout=0.2 is used to prevent overfitting; the output layer outputs the time series prediction values ​​of photovoltaic output and load power for the corresponding partition in the next 24 hours.

[0028] The specific training steps are as follows: using the historical power grid data of the past 3 years as the training set, the training set and validation set are divided in an 8:2 ratio; the learning rate is set to 0.001, the number of iterations is 200 rounds, and the batch_size is 32; the model training is completed using the mean square error as the loss function; the predicted output value is directly connected to the voltage risk assessment model as the input data for voltage situation inference.

[0029] The voltage risk index is calculated using a multi-dimensional weighted fusion method, which integrates four indicators—voltage over-limit amplitude, over-limit duration, fluctuation rate, and regional impact range—according to a preset ratio. The higher the value of each indicator, the greater the corresponding risk index. The risk index is a continuous value, and it is directly matched to the four-level risk level through numerical range mapping.

[0030] Multi-dimensional weighted fusion formula for voltage risk index: ; This represents the risk index of voltage exceeding limits in a distribution zone, used to quantitatively determine the voltage operation risk level of a distribution network zone. This represents the weighting factor for the voltage over-limit magnitude, used to determine the weight ratio of this indicator in risk calculation; This represents the weighting factor for the duration of voltage over-limit, used to determine the weighting ratio of this indicator in risk calculation; This represents the voltage fluctuation rate weighting coefficient, used to determine the weighting ratio of this indicator in risk calculation; This represents the weighting coefficient for the impact range of voltage anomalies, used to determine the weight ratio of this indicator in risk calculation; S represents the quantized value of the voltage over-limit amplitude, which is used to reflect the specific degree to which the node voltage deviates from the rated range; T represents the quantized value of voltage over-limit duration, which reflects the duration of the voltage over-limit state; V represents the voltage fluctuation rate quantization value, which is used to reflect the rate of change of node voltage in real time; N represents the quantified value of the impact range of the voltage anomaly, which is used to reflect the number of power grid nodes affected by the voltage anomaly.

[0031] The system classifies the grid voltage operation risk into four levels: no risk, low risk, medium risk, and high risk, based on the quantified voltage over-limit risk index. Each risk level corresponds to a zone pre-control logic. Under the no risk level, the grid equipment maintains a steady-state normal operation. Under the low risk level, each zone performs a slight predictive fine-tuning operation and reserves grid regulation margin. Under the medium risk level, the system initiates the day-ahead plan correction process and completes the pre-scheduling configuration of zone regulation resources. Under the high risk level, flexible regulation resources enter a standby state and the frequent operation of voltage regulating equipment is restricted. Each risk level is determined based on four indicators: voltage exceedance magnitude, duration of exceedance, voltage fluctuation rate, and affected node range. The risk-free level is defined as: node voltage remains within the acceptable range throughout, with no tendency to exceed limits and a fluctuation rate less than 1% / min of the rated voltage; the low-risk level is: voltage close to the acceptable limit but not exceeding it, duration not exceeding 5 minutes, affecting only a single node; the medium-risk level is: short-term, small voltage exceedance, exceedance magnitude ≤ 5%, duration 5-15 minutes, affecting no more than 3 related nodes; the high-risk level is: long-term, large voltage exceedance, exceedance magnitude > 5%, duration exceeding 15 minutes.

[0032] The system also establishes a human-machine collaborative early warning system, which pushes corresponding early warning information and pre-control execution suggestions to the dispatching backend and user-side terminal equipment according to different voltage risk levels.

[0033] The human-machine collaborative early warning system adopts a tiered push mechanism: no information is pushed if there is no risk level; for low-risk levels, text prompts are pushed to the dispatch backend, but not to user terminals; for medium-risk levels, audible and visual warnings and detailed pre-control plans are pushed to the dispatch backend, and reminders are pushed to user terminals; for high-risk levels, emergency audible and visual warnings and mandatory handling instructions are pushed to the dispatch backend, and hazard avoidance warnings are pushed to user terminals; all warning information comes with a response time limit, and the dispatcher must complete the confirmation or handling operation within the specified time limit.

[0034] In this embodiment of the invention, during the day-ahead global pre-control phase, the system constructs a multi-objective optimization model with the three-dimensional optimization objectives of optimal operating cost, minimum voltage deviation, and maximum photovoltaic absorption under the day-ahead scheduling time scale. The multi-objective optimization model takes into account the cost characteristics and time scale characteristics of the flexible and rigid regulation resources of the distribution network. The total operating cost of the system includes five major components: photovoltaic curtailment penalty cost, energy storage charging and discharging loss cost, OLTC tap changer wear cost, capacitor bank switching cost, and user controllable load regulation compensation cost. The multi-objective optimization model adopts an adaptive weight allocation logic. Under normal operating conditions, the weights of the three objectives—operating cost, voltage deviation, and photovoltaic (PV) consumption—are allocated in equal proportions. In medium- and high-risk zones, the weight of the voltage deviation objective is automatically increased to ensure voltage safety. During periods of high PV generation, the weight of the PV consumption objective is automatically increased to reduce curtailment. During peak load periods, the weight of the operating cost objective is automatically increased to control grid regulation losses.

[0035] The constraints of the multi-objective optimization model include node voltage constraints, line power constraints, and equipment capacity constraints. It also includes zonal correlation constraints, equipment operation frequency timing constraints, and risk reserve capacity constraints. The multi-objective optimization model combines the day-ahead voltage risk prediction results to configure a flexible adjustment capacity reserve mechanism for medium and high-risk zones. The system solves the multi-objective optimization model through an adaptive weighted improved particle swarm algorithm and iteratively outputs the next day's time-period OLTC tap position, capacitor bank switching status, energy storage charging and discharging timing plan, controllable load adjustment scheme, and photovoltaic inverter baseline operating parameters. The adaptive weighted improved particle swarm optimization algorithm takes as input the cost function, constraints, and risk reserve capacity parameters of the multi-objective optimization model; the particle dimension corresponds to five types of control parameters: OLTC tap changer, capacitor bank, energy storage, controllable load, and photovoltaic inverter; the inertia weight is adaptively adjusted through the inertia weight formula of the adaptive weighted improved particle swarm optimization algorithm, with learning factors c1=c2=2; the iteration termination condition is that the fitness value changes by less than 1% for 50 consecutive rounds, and the output is the equipment operation parameters and resource control scheme for each time period of the next day.

[0036] The formula for the inertia weight in the adaptive weight-based particle swarm optimization algorithm is as follows: ; The inertia weight represents the particle swarm optimization algorithm, which is used to balance the algorithm's global search capability and local search capability. The value is adaptively adjusted during the iteration process. This indicates the current iteration number of the algorithm, used to identify the real-time progress stage of algorithm optimization; This represents the maximum number of iterations preset by the algorithm, which is the upper limit condition for terminating the algorithm's iterative calculation.

[0037] The flexible regulation capacity reservation configuration rules are as follows: for medium-risk zones, 20% of the total regulation capacity of the zone is reserved for emergency regulation capacity, which is not included in the day-ahead regular dispatch plan; for high-risk zones, 40% of the total regulation capacity of the zone is reserved for emergency regulation capacity. The reserved capacity is only activated when the voltage risk escalates or the limit is actually exceeded. The equipment type of the reserved capacity is preferentially selected from fast-response resources such as photovoltaic inverters and energy storage systems.

[0038] After the constraints of the multi-objective optimization model are triggered, the constraint correction logic is automatically executed. When the node voltage constraint exceeds the limit, the voltage deviation target weight is automatically increased and the equipment output is adjusted. When the equipment operation frequency constraint exceeds the limit, the equipment is automatically locked and switched to a similar alternative regulation resource. When the risk reserve capacity constraint is insufficient, the regulation output of non-critical equipment is automatically reduced to ensure sufficient reserve capacity. All constraint corrections are performed under the premise of meeting the safe operation of the power grid.

[0039] Recently, the global pre-control strategy has been linked with the results of dynamic voltage zoning and hierarchical risk warning. The control strategy is configured according to the operating conditions of each zone, realizing the combination of overall scheduling and zone pre-control.

[0040] In this embodiment of the invention, during the intraday rolling correction phase, the system sets a 1-hour optimization time window and a 15-minute rolling iteration cycle. During the intraday regulation process, it continuously collects the latest photovoltaic output and load power prediction data, and synchronously updates the three-dimensional spatiotemporal sensitivity tensor parameters and partition association constraint parameters to achieve dynamic refresh of the optimization model input parameters. The system repeatedly solves the multi-objective optimization model with the optimization objectives of minimizing the operating cost and voltage deviation of each zone in the next hour, and completes the dynamic correction of the intraday control plan. During the correction process, the control parameters and operation plans of fast-response resources such as energy storage, controllable loads, and photovoltaic inverters are adjusted first, while the number of daily operations of slow-speed mechanical equipment such as OLTC and capacitor banks is limited. The fast response resources refer to adjustable devices with a response time of ≤1 second, specifically including distributed photovoltaic inverters, user-side energy storage systems, and flexible controllable loads, which have millisecond to second-level response capabilities and can be used to quickly smooth voltage fluctuations. The slow mechanical devices refer to mechanical voltage regulating devices with a response time of ≥10 seconds, specifically including on-load tap-changing transformers and capacitor banks. These devices are subject to mechanical wear and lifespan loss, and the system limits their total number of actions per day to no more than 50, with an interval of no less than 3 minutes between actions. During the daily rolling correction process, only small steady-state adjustments are made, and dynamic correction is prioritized through fast response resources.

[0041] The optimization model solution for the intraday rolling correction phase is set with dual termination conditions. The iteration will stop and the correction results will be output when either condition is met: First, after 30 consecutive iterations, the operating cost and voltage deviation indicators of each partition do not change significantly and the numerical fluctuation is less than the set stable threshold; Second, the calculation time of a single iteration reaches 10 seconds to ensure that the correction command is issued and executed in a timely manner within a 15-minute iteration cycle.

[0042] Intraday rolling correction and day-ahead global pre-control form a time-sequential progressive operation logic. Through the operation mode of global overall planning, regional dynamic correction, and time-sharing coordination of fast and slow resources, a long-term scale control closed loop for the distribution network is constructed.

[0043] In this embodiment of the invention, during the real-time control phase of the partition, each voltage control partition is configured with an independent edge computing node for real-time voltage control calculation and strategy execution within the partition. The edge computing node is configured with a five-level flexible adjustment priority based on real-time spatiotemporal sensitivity data. The five-level flexible adjustment priority includes reactive power priority, active power supplementation, energy storage as the main force, load backup, and mechanical equipment steady-state protection. The edge computing nodes are deployed on a single-zone, single-node basis, with only one edge computing node set up in each voltage control zone. The node is deployed at the electrical node in the center of the zone. The edge computing node is connected to the local control equipment via wired communication and to other zone edge computing nodes and the dispatch master station via wireless encrypted communication. The data exchange cycle between nodes is 1 second.

[0044] The five-level flexible regulation priority is executed in order of response speed and regulation cost. Among them, reactive power priority corresponds to the rapid regulation of reactive power of distributed photovoltaic inverters, with a response time of milliseconds, which does not affect active power output and has the lowest regulation cost; active power supplement corresponds to the flexible regulation of active power output of photovoltaic inverters, which is activated when reactive power regulation is insufficient; energy storage main force corresponds to the rapid regulation of charging and discharging power of user-side energy storage systems, which serves as the core support resource for zoned voltage regulation; load backup corresponds to the flexible and controllable load segmented switching regulation, which is activated when the regulation capacity of the above resources is insufficient; and mechanical equipment steady-state protection corresponds to the on-load tap changer tap adjustment and capacitor bank switching regulation, which serves as a steady-state voltage protection means and only operates when rapid regulation resources cannot meet the demand, thereby reducing mechanical wear of equipment.

[0045] The edge computing node monitors the voltage and power flow operation status of each node in the partition in real time, predicts the trend of voltage exceeding the limit, and combines the response speed, adjustment cost, and spatiotemporal sensitivity coefficient of the adjustment resources to complete the dynamic allocation of the real-time adjustment amount of the control equipment. Each partition's edge computing nodes interact in real time with boundary node power flow and voltage operation data, configure inter-partition voltage offset transfer suppression constraint logic, maintain stable voltage operation through collaborative fine-tuning of neighboring nodes, and realize distributed voltage regulation of the distribution network through the edge computing architecture.

[0046] The neighborhood collaborative fine-tuning involves real-time interaction of voltage, power flow, and regulation output data of boundary nodes between edge computing nodes of each partition. With the core objective of balanced and consistent boundary node voltage, the output values ​​of adjustable devices within the partition are synchronously and slightly adjusted according to the boundary voltage deviation ratio. The adjustment step size does not exceed 3% of the rated output, and the interval between single adjustments is not less than 5 seconds. The voltage difference between adjacent partition boundaries is gradually reduced until the boundary voltage deviation is controlled within 1% of the rated voltage, thus avoiding problems such as voltage abrupt changes and power flow reversal at the partition boundaries.

[0047] Neighborhood Cooperative Fine-tuning Boundary Voltage Deviation Constraint Formula: ; This represents the voltage deviation value between adjacent partition boundary nodes, used to reflect the degree of difference in boundary voltage between partitions; It represents the rated voltage value of a distribution network node and is a standard reference value for the voltage operation of the network node.

[0048] The voltage offset transfer suppression constraint logic collects the voltage values ​​and power flow data of the associated nodes at the boundaries of the sub-divisions in real time, calculates the voltage deviation value of the boundary nodes of adjacent sub-divisions, and automatically triggers the constraint mechanism when the deviation value exceeds 2% of the rated voltage, limiting the adjustment range of the voltage regulating equipment in this sub-division and adjacent sub-divisions, prohibiting large-scale voltage regulation operations on one side of the sub-division, avoiding the voltage offset problem from being transmitted and spread to adjacent sub-divisions due to excessive adjustment on one side, and simultaneously locking the execution rhythm of the sub-division adjustment command.

[0049] In this embodiment of the invention, during the parameter adaptation phase, the system adopts a reinforcement learning operation mechanism. It completes the dynamic adaptive optimization of control parameters through a dual-loop iterative optimization logic consisting of an outer loop adaptive parameter tuning closed loop and an inner loop iterative optimization closed loop for control effect. The outer loop adaptive parameter tuning closed loop collects real-time operating data such as grid voltage deviation, adjustment delay, line loss, photovoltaic absorption rate, and equipment operating costs. It then uses a deep reinforcement learning algorithm to complete the online iterative adjustment of core control parameters. The deep reinforcement learning adopts a dual closed-loop architecture: the outer loop state space includes the entire network voltage deviation, regulation delay, line loss, photovoltaic absorption rate, and operating cost; the action space includes regulation parameters such as QU droop slope, active power limiting threshold, and energy storage priority; the inner loop uses regulation steady-state accuracy, response speed, and economy as reward functions; the training step size is 5 minutes, and a DQN network iterative update strategy is adopted, with parameter update results synchronously sent to each regulation module.

[0050] Deep reinforcement learning reward function formula: ; R represents the deep reinforcement learning reward function value, which is used to quantitatively evaluate the overall performance of the voltage regulation strategy. It represents the voltage deviation coefficient, which is used to measure the steady-state control accuracy of the system after voltage regulation; The photovoltaic grid integration rate is the ratio of the actual photovoltaic power absorbed by the grid to the theoretical maximum photovoltaic output. This represents the total cost of system regulation, including all regulation expenditures such as equipment wear and tear costs and load regulation compensation costs.

[0051] The dual-loop iterative optimization adopts a periodic differentiated execution mechanism. The outer loop adaptive parameter tuning closed loop performs parameter iterative adjustment every 5 minutes to adapt to the slow changes in power grid conditions. The inner loop control effect iterative optimization closed loop performs control effect evaluation and reward value calculation every 10 seconds to quickly provide feedback on the control effect and correct the direction of parameter adjustment.

[0052] The control parameters include QU droop control slope, active power limiting threshold, energy storage charging and discharging priority, controllable load adjustment coefficient, partition clustering weight, and rolling optimization step size. All parameters are dynamically updated according to the spatiotemporal operating conditions of the power grid. The inner loop control effect iteratively optimizes the steady-state accuracy, dynamic response speed, and operational economy of the closed-loop real-time monitoring voltage control. The control effect data is used as the input parameter of the reinforcement learning reward function to iteratively optimize the control strategy and parameter configuration. The system builds a parameter safety constraint threshold library to limit the operating range of the control parameters.

[0053] The parameter safety constraint threshold library contains five core constraint values: node voltage safety upper and lower limits, equipment daily operation frequency upper limit, regulation output amplitude maximum and minimum threshold, parameter single adjustment step size threshold, and voltage regulation response delay threshold. Before performing online iterative adjustments, all control parameters must be compared and verified with the corresponding parameters in the threshold library. Only parameter adjustment commands within the safe range can be issued and executed. Adjustment commands exceeding the threshold range are automatically intercepted and corrected to the safe range to avoid safety issues such as equipment overload, frequent operation, and voltage exceeding limits caused by abnormal parameter adjustments.

[0054] The reward function is composed of three core indicators: voltage regulation quality, operating economy, and photovoltaic absorption level. The more stable the voltage regulation, the lower the operating cost, and the higher the photovoltaic absorption, the larger the value of the reward function. The deep reinforcement learning algorithm aims to obtain the maximum reward value and continuously iterates and optimizes the regulation parameters.

[0055] In this embodiment of the invention, during the fault self-healing and voltage stabilization stage, when a distribution network fault causes abnormal voltage fluctuations, the system initiates a partitioned reconfiguration-type rapid self-healing operation mechanism. First, it completes the fault location, fault type identification, and fault impact range calculation, isolates the fault area, and blocks the fault propagation. At the same time, it recalculates and constructs a three-dimensional spatiotemporal sensitivity tensor based on the power grid topology after the fault, updates the voltage control partition, and matches the power grid operation architecture under the fault state. The fault location is achieved by comparing the voltage and current change characteristics of each node in the partition with the fault waveform library, which quickly locates the specific node and line section where the fault occurred; the fault type is determined by the voltage drop amplitude and current surge magnitude to identify common distribution network faults such as short circuit, open line, and grounding; the scope of influence is centered on the fault node, and based on the topological correlation and power flow direction, all related nodes and line sections where the voltage fluctuates abnormally are automatically delineated.

[0056] The system implements a tiered emergency control strategy for non-faulty operating areas, smoothing out abnormal voltage fluctuations according to the priority order of rapid reactive power response of photovoltaic inverters, active power support of energy storage, flexible adjustment of controllable loads, and precise disconnection of a small number of non-critical loads. The fault control process is synchronously linked with the reinforcement learning operation mechanism, calling the control parameters adapted to the fault conditions. After the grid fault is restored, the system performs gradient parameter rollback and partition reset operations, switching to the conventional flexible control operation mode.

[0057] The gradient reset is executed smoothly in three steps: First, the emergency voltage regulation parameters under fault conditions are gradually reduced to the normal operating parameter range at a rate of 20% decrease every 3 minutes, while simultaneously monitoring node voltage and line power flow to ensure a smooth transition of operating parameters; Second, after the voltage stabilizes for 30 minutes, the voltage partition reconstructed after the fault is gradually restored to the normal partition state before the fault, while simultaneously updating the three-dimensional spatiotemporal sensitivity tensor data; Third, all control equipment parameters and system operating strategies are fully reset, switching to the normal flexible control operation mode. The entire reset process involves real-time monitoring of the power grid status, and if any abnormality occurs, the reset is paused and temporary voltage stabilization measures are initiated.

[0058] The rules for limiting the execution range and step size of emergency voltage regulation are as follows: the maximum range of reactive power regulation of photovoltaic inverters shall not exceed 50% of the rated reactive power capacity in a single operation, and the adjustment step size shall not exceed 10% of the rated capacity; the maximum range of active power regulation of energy storage shall not exceed 60% of the rated active power capacity in a single operation, and the adjustment step size shall not exceed 15% of the rated capacity; the switching ratio of controllable load regulation shall not exceed 10% of the total load of the zone in a single operation, and all emergency voltage regulation operations shall be executed in steps.

[0059] This invention also provides a photovoltaic distribution network voltage flexible regulation system, based on the above method, comprising the following modules: The dynamic partitioning construction module is used to collect electrical operation parameters and equipment status information of the distribution network, integrate multi-dimensional parameters such as line impedance and equipment delay, and construct a three-dimensional spatiotemporal sensitivity tensor. Based on the weighted adaptive spectral clustering algorithm, it dynamically divides the voltage control region, and can adjust the clustering parameters according to photovoltaic fluctuations and load conditions, and periodically update and output the partitioning results. The graded risk early warning module retrieves dynamic zoning and spatiotemporal sensitivity tensor data, integrates historical meteorological and power grid data, and predicts zoning photovoltaic output and load time series data by optimizing long short-term memory neural networks; it builds a voltage risk assessment model, quantifies the risk index and divides it into four risk levels, matches the corresponding pre-control logic, and pushes early warning and disposal suggestions in a graded manner. The global pre-control module, based on partition risk data, constructs a multi-objective optimization model with the goals of optimizing operation and maintenance costs, minimizing voltage deviation, and maximizing photovoltaic absorption. It solves the model using an adaptive weighted improved particle swarm algorithm and outputs the equipment operation status and resource pre-control plan for each time period of the next day. The intraday rolling correction module sets a 1-hour optimization window and a 15-minute iteration cycle to update photovoltaic and load forecasts and spatiotemporal parameters in real time, and iteratively optimize the model; it prioritizes adjusting parameters of fast-response equipment, constrains the frequency of action of slow-speed mechanical equipment, and dynamically corrects the intraday control plan. The partition real-time control module, based on the partition edge computing device, dynamically allocates the device adjustment amount according to the five-level adjustment priority, collects partition operation data and inter-partition boundary conditions in real time, performs neighborhood collaborative fine-tuning, and configures voltage offset suppression constraints. The parameter adaptive module, based on a dual-closed-loop iterative optimization architecture, collects control operation data, iteratively optimizes various core control parameters online through reinforcement learning algorithms, and dynamically updates parameter configuration strategies by combining a parameter safety constraint threshold library. The fault self-healing voltage regulation module monitors abnormal voltage fault signals in the power grid in real time, completes fault location, identification and range calculation, isolates the fault area, reconstructs the power grid topology and voltage partition, calls the appropriate parameters to perform emergency voltage regulation, and resets the partition and control parameters step by step after the fault is restored.

[0060] 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.

[0061] 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 flexible voltage regulation of photovoltaic distribution networks, characterized in that, The specific steps include the following: In the dynamic partitioning construction phase, electrical operation parameters and equipment status information of the distribution network are collected, a three-dimensional spatiotemporal sensitivity tensor is constructed, voltage control regions are divided based on the weighted adaptive spectral clustering algorithm, clustering parameters are adjusted according to the real-time operating conditions of the power grid, and the partitioning results are updated at a fixed period. During the graded risk warning stage, dynamic voltage zoning and spatiotemporal sensitivity tensor data are retrieved, and combined with meteorological and historical operation data, the photovoltaic output and load power time series data of each zone are predicted. The voltage risk index is calculated and the risk level is divided. The corresponding pre-control logic is matched and the warning information and handling suggestions are pushed in a graded manner. During the current global pre-control phase, regional voltage risk and operation data are retrieved, and a multi-objective optimization model incorporating control cost parameters and equipment operation constraints is constructed. The model is then solved using optimization algorithms to output the operating status of power grid equipment and resource control and allocation schemes. During the intraday rolling correction phase, based on the established day-ahead control plan, a fixed rolling optimization window and iteration cycle are set, photovoltaic and load forecast data and spatiotemporal parameters are updated in real time, the optimization model is iteratively solved to correct the intraday equipment control plan, and the operating parameters of the control equipment are adjusted according to categories. During the real-time control phase of the partition, the operating data within the partition is collected in real time based on the edge computing device. The device adjustment amount is allocated according to the preset flexible adjustment priority. The operating data of the partition boundary is collected to perform neighborhood collaborative fine-tuning and set voltage offset transfer suppression constraints. During the parameter adaptation phase, real-time control operation data and operating parameters are collected, and the core control parameters are adjusted online based on a dual-closed-loop iterative optimization architecture. A parameter safety constraint threshold library is configured to limit the operating range of control parameters, and the parameter configuration strategy is dynamically updated. During the fault self-healing and voltage stabilization phase, the system monitors voltage anomalies caused by grid faults, performs fault location, type identification, and impact range calculation, and isolates the fault area. Based on the post-fault grid topology, it reconstructs the three-dimensional spatiotemporal sensitivity tensor and voltage partitions, calls the appropriate control parameters to perform emergency voltage regulation, and resets the partitions and control parameters step by step after the fault is eliminated.

2. The photovoltaic distribution network voltage flexible regulation method according to claim 1, characterized in that, During the dynamic partitioning construction phase, the collected operational data includes voltage, current, and power flow data of each node in the distribution network, as well as the operating status and capacity constraint data of distributed photovoltaic, user-side energy storage, flexible and controllable loads, on-load tap-changing transformers, and capacitor banks. The constructed three-dimensional spatiotemporal sensitivity tensor includes three dimensions: spatial nodes, temporal operating conditions, and regulation resources. The perturbation method is used to calculate the voltage response sensitivity of each node at different times and under the action of different regulation resources online. The calculation process incorporates line impedance, topology correlation degree, and equipment response delay parameters. The weighted adaptive spectral clustering algorithm performs partitioning based on the spatiotemporal sensitivity similarity of nodes, voltage correlation strength, and regional power flow interaction characteristics. The voltages of different partitions exhibit weak correlation characteristics. The algorithm dynamically adjusts the clustering weights based on the intensity of photovoltaic fluctuations in the power grid and the peak-valley state of the load.

3. The photovoltaic distribution network voltage flexible regulation method according to claim 2, characterized in that, In the graded risk warning stage, the photovoltaic output and load power time-series change curves are predicted by optimizing the long short-term memory neural network in different regions, and the voltage distribution situation of each region at each moment is iteratively deduced by combining the real-time updated three-dimensional spatiotemporal sensitivity tensor. The voltage risk index is quantitatively calculated based on four dimensions: voltage over-limit amplitude, over-limit duration, fluctuation rate, and regional impact range. According to the voltage over-limit risk index, the operational risk is divided into four levels, and each risk level corresponds to a zoned pre-control logic. A human-machine linkage early warning system is built to push early warning information and pre-control execution suggestions to the dispatching backend and user-side terminals in a hierarchical manner.

4. The photovoltaic distribution network voltage flexible adjustment method according to claim 3, characterized in that, In the aforementioned global pre-control phase, the multi-objective optimization model takes optimal operating cost, minimum voltage deviation, and maximum photovoltaic absorption as three-dimensional optimization objectives, taking into account the cost characteristics and time scale characteristics of flexible and rigid regulation resources. The total system operating cost includes photovoltaic curtailment penalty cost, energy storage charging and discharging loss cost, on-load tap changer tap wear cost, capacitor bank switching cost, and user controllable load adjustment compensation cost. The model constraints include node voltage constraints, line power constraints, and equipment capacity constraints. It also includes zonal correlation constraints, equipment operation frequency timing constraints, and risk reserve capacity constraints. Combined with the day-ahead voltage risk prediction results, a flexible regulation capacity reserve mechanism is configured for medium and high risk zones. By improving the particle swarm optimization algorithm with adaptive weights, the model is solved, and the operating parameters and resource control schemes of various control devices for each time period of the next day are output.

5. The photovoltaic distribution network voltage flexible regulation method according to claim 4, characterized in that, During the intraday rolling correction phase, the latest photovoltaic output and load power prediction data are continuously collected, and the three-dimensional spatiotemporal sensitivity tensor parameters and zonal correlation constraint parameters are updated synchronously. The multi-objective optimization model is repeatedly solved with the goal of minimizing the operating cost of each zone and minimizing the voltage deviation of each zone at a fixed period, so as to complete the dynamic correction of the intraday control plan. During the correction process, the control parameters and operation plan of fast response resources are adjusted first, and the number of daily operations of slow mechanical equipment is limited.

6. The photovoltaic distribution network voltage flexible regulation method according to claim 5, characterized in that, During the real-time control phase of the partition, each voltage control partition is configured with an independent edge computing node to undertake the real-time voltage control calculation and strategy execution within the partition; a five-level flexible adjustment priority is configured based on real-time spatiotemporal sensitivity data, and the dynamic allocation of the real-time adjustment amount of the control equipment is completed by combining the response speed of the adjustment resources, the adjustment cost, and the spatiotemporal sensitivity coefficient; Each partition's edge computing nodes exchange boundary node power flow and voltage operation data in real time, and maintain stable voltage operation through collaborative fine-tuning by neighboring nodes.

7. The photovoltaic distribution network voltage flexible regulation method according to claim 6, characterized in that, The parameter adaptation phase employs a reinforcement learning mechanism, using a dual-loop logic consisting of an outer loop for adaptive parameter tuning and an inner loop for iterative optimization of control effects to achieve dynamic adaptive optimization of control parameters. The outer loop collects real-time data on grid voltage deviation, regulation delay, line loss, photovoltaic absorption rate, and equipment operating costs, and uses a deep reinforcement learning algorithm to perform online iterative adjustments of core control parameters. The inner closed-loop system monitors the steady-state accuracy, dynamic response speed, and operational economy of the voltage regulation in real time, and uses the regulation effect data as input parameters for the reinforcement learning reward function to iteratively optimize the control strategy and parameter configuration.

8. The photovoltaic distribution network voltage flexible regulation method according to claim 7, characterized in that, During the fault self-healing and voltage stabilization phase, when a distribution network fault causes abnormal voltage fluctuations, a partitioned reconfiguration-type rapid self-healing operation mechanism is activated to isolate the fault area and prevent the fault from spreading. Based on the post-fault grid topology, a three-dimensional spatiotemporal sensitivity tensor is recalculated and constructed to update the voltage control partition and match the grid operation architecture under the fault state. For non-faulty operating areas, a graded emergency control strategy is implemented to suppress abnormal voltage fluctuations according to a preset priority order. The fault control process is synchronized with the reinforcement learning operation mechanism to call up the control parameters adapted to the fault conditions. After the power grid fault is restored, a gradient-based gradual rollback of control parameters and a zone reset operation are performed to switch to the conventional flexible control operation mode.

9. A photovoltaic distribution network voltage flexible regulation system, based on the method described in any one of claims 1-8, comprising the following modules: The dynamic partitioning construction module is used to collect electrical operating parameters and equipment status information of the power distribution network, construct a three-dimensional spatiotemporal sensitivity tensor, dynamically divide the voltage control region based on the weighted adaptive spectral clustering algorithm, adjust the clustering parameters according to photovoltaic fluctuations and load conditions, and update the partitioning results periodically. The graded risk early warning module is used to retrieve dynamic zoning and spatiotemporal sensitivity tensor data, integrate meteorological and historical operation data to predict the photovoltaic output and load time series data of the zoning, calculate the voltage risk index and classify the risk level, match the corresponding pre-control logic and push early warning and disposal suggestions in a graded manner. The global pre-control module is used to build a multi-objective optimization model based on partition risk data, incorporating control cost parameters and equipment operation constraints. The model is solved by optimization algorithms and outputs the equipment operation status and resource pre-control plan for each time period of the next day. The intraday rolling correction module is used to set a fixed rolling optimization window and iteration cycle, update photovoltaic and load forecasts and spatiotemporal parameters in real time, iterate and optimize the model, prioritize the adjustment of parameters of fast-response equipment and constrain the frequency of action of slow-speed mechanical equipment, and dynamically correct the intraday control plan. The partition real-time control module is used to dynamically allocate the device adjustment amount according to the preset flexible adjustment priority based on the partition edge computing device, collect partition operation data and inter-partition boundary conditions, perform neighborhood collaborative fine-tuning and configure voltage offset suppression constraints. The parameter adaptive module is used to collect control operation data based on a dual-closed-loop iterative optimization architecture, and to iteratively optimize the core control parameters online through reinforcement learning algorithms. It is also used in conjunction with a parameter safety constraint threshold library to dynamically update the parameter configuration strategy. The fault self-healing voltage regulation module is used to monitor abnormal voltage fault signals in the power grid in real time, complete fault location, identification and range calculation and isolate the fault area, reconstruct the power grid topology and voltage partition, call the adaptive parameters to perform emergency voltage regulation, and reset the partition and control parameters step by step after the fault is restored.

10. The photovoltaic distribution network voltage flexible regulation system according to claim 9, characterized in that, The system also includes a data interaction bus module, which is used to realize real-time data transmission and command interaction between various modules; the output results of the dynamic partition construction module are synchronously transmitted to the hierarchical risk warning module, the day-ahead global pre-control module, the intraday rolling correction module and the partition real-time control module as the basic data for the operation of each module; The parameter update results of the parameter adaptive module are sent to all control execution modules in real time, so as to realize the synchronous adjustment of the control parameters of the whole system. The fault self-healing voltage regulator module has the highest priority when a fault occurs, and can interrupt the normal control process and take over the control of the system.

Images