A reactive power regulation method and system for a distributed photovoltaic cluster

Abstract

This application relates to the field of distributed photovoltaic (PV) cluster control technology, specifically to a reactive power control method and system for distributed PV clusters. The method includes: collecting online monitoring data of the distributed PV cluster and constructing cluster partitioning indices; partitioning the cluster with the goal of maximizing the sum of the cluster partitioning indices; moving nodes with the cluster's reactive power self-sufficiency as a constraint; calculating the operating state deviation based on the deviation between the measured and predicted voltages of nodes in the cluster to correct the sensitivity; adjusting the voltage of each node in the cluster based on the corrected sensitivity matrix to implement voltage safety constraints; constructing a secondary planning optimization objective by minimizing the reactive power adjustment amount of each node; and outputting reactive power control commands for each node. This application solves the problems of cluster partitioning mismatch, control response lag, and insufficient reliability of control results in distributed PV scenarios.

Description

Technical Field

[0001] This application relates to the field of distributed photovoltaic cluster control technology, specifically to a reactive power control method and system for distributed photovoltaic clusters. Background Technology

[0002] With the integration of distributed photovoltaic power, the output of the power grid is subject to uncontrollable weather conditions and is widely distributed in space. This results in highly volatile and uncertain operating conditions of the power distribution network, significantly increasing safety risks such as voltage overruns, backflow of power, and heavy equipment overload. This places higher demands on large-scale and clustered control technologies.

[0003] Existing technologies for regulating distributed photovoltaic (PV) clusters generally assume a "relatively stable operating condition" as the physical premise for cluster partitioning and regulation decisions. However, this assumption conflicts with the time-varying physical characteristics of actual operating conditions. Firstly, the determination of cluster boundaries in existing cluster partitioning methods often relies on static indicators such as day-ahead forecasts, electrical distance, or voltage sensitivity, neglecting changes in node output trends and real-time regulation capabilities. Secondly, regulation strategies are typically generated based on a fixed cluster structure and a pre-set operating model. When the actual operating state deviates from the model assumptions, the state variables, controlled objects, and constraints in the control model all change. Existing centralized optimization methods struggle to respond to these changes promptly, while purely data-driven methods, lacking physical constraints in the decision-making process, are prone to producing control results that do not conform to actual operating patterns when the operating state changes. Based on the above analysis, under the assumption of stable operating conditions, existing technologies struggle to achieve synchronous dynamic adaptation between cluster partitioning and regulation strategies, resulting in problems such as "cluster partitioning mismatch, delayed regulation response, and insufficient reliability of control results" in distributed PV scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this application is to provide a reactive power control method and system for distributed photovoltaic clusters, the specific technical solution of which is as follows:

[0005] This application provides a reactive power control method for distributed photovoltaic clusters, comprising the following steps:

[0006] Step 1: Collect online monitoring data of the distributed photovoltaic cluster;

[0007] Step 2: Based on the changes in the operating parameters of each node in the distributed photovoltaic cluster, an improved Louvain algorithm is used to dynamically partition the cluster. Specifically, cluster partitioning indices are constructed based on the matching of reactive power supply and demand within the cluster, reactive power-voltage sensitivity, the consistency of the output change trends of each node within the cluster, and the mutual influence intensity of reactive power regulation between nodes. The cluster partitioning is performed with the goal of maximizing the sum of the cluster partitioning indices, and node movement is constrained by the cluster's reactive power self-sufficiency capability.

[0008] Step 3: Dynamically correct the sensitivity matrix of the partitioned cluster. Specifically, calculate the operating state deviation based on the deviation between the measured voltage and the predicted voltage of the nodes in the cluster to correct the sensitivity. Adjust the voltage of each node in the cluster based on the corrected sensitivity matrix to perform voltage safety constraints. Then, construct a target by minimizing the reactive power adjustment of each node for secondary planning optimization and output the reactive power control command of each node.

[0009] Step 4: Perform dynamic control according to the reactive power control command.

[0010] Preferably, the specific method for acquiring the online monitoring data in step one is as follows: importing feeder topology connection relationships, switch status, and line impedance parameters through the distribution automation system; collecting three-phase voltage, current, active power, and reactive power at the photovoltaic grid connection point through the intelligent fusion terminal; collecting DC side voltage and current, AC side active power, reactive power, apparent power, grid connection frequency, internal temperature, and available reactive power capacity of each photovoltaic inverter through the edge gateway; collecting total horizontal irradiance and ambient temperature through radiation meters and temperature sensors; acquiring active and reactive power data from smart meters at each distribution area's access points through the electricity consumption information acquisition system; and uploading the above data to the cloud after pre-verification by the edge gateway.

[0011] Preferably, the matching of reactive power supply and demand within the cluster in step two includes: obtaining the current available reactive power capacity of each node in cluster K and the reactive power demand of cluster K, wherein the current available reactive power capacity is the minimum of the current dynamic reactive power capacity upper limit of each node and the rated reactive power upper limit of the inverter, and the current dynamic reactive power capacity upper limit is obtained by the square root of the square difference between the apparent power of the inverter and the current active power output; and then calculating the reactive power balance metric of cluster K. This is used to characterize the matching of reactive power supply and demand within the cluster: Where ε is the division-by-zero protection constant, and ∑ represents the summation over all nodes in cluster K. It is a cluster The Middle The current available reactive power capacity of each node. It is a cluster The sum of the absolute values ​​of the reactive power changes estimated for all nodes within the system.

[0012] Preferably, the consistency characteristics of the power output change trends of each node in the cluster in step two include: taking a sliding window of a preset duration as a measurement window, using the correlation coefficient between the active power sequences of any two nodes in the same cluster within the same measurement window as the power output dynamic similarity measure of the two nodes; and using the mean of the power output dynamic similarity measures of all node pairs in the cluster as the power output dynamic similarity measure of the cluster, which is used to characterize the consistency characteristics of the power output change trends of each node in the cluster.

[0013] Preferably, the mutual influence intensity of reactive power regulation between nodes in step two includes: calculating the electrical distance between nodes based on the equivalent mutual impedance and node self-impedance in the equivalent impedance matrix, and clustering. Middle node and electrical distance between And calculate the electrical coupling strength between nodes, nodes and electrical coupling strength between Furthermore, the electrical distance metric of the cluster is calculated to characterize the mutual influence intensity of reactive power regulation among nodes within the cluster. Electrical distance measurement ;

[0014] in, , , The nodes in the equivalent impedance matrix and Equivalent mutual impedance between nodes Self-impedance, nodes Self-impedance, It is a preset attenuation coefficient. These are the nodes The sum of the electrical coupling strengths of all node pairs, It is the sum of the electrical coupling strength of all node pairs.

[0015] Preferably, the acquisition of cluster partitioning indicators in step two includes:

[0016] ;

[0017] in For the reactive power balance metric of cluster K, Let be the normalized mean of reactive power-voltage sensitivity for all nodes in cluster K. For the dynamic similarity measure of output of cluster K, For the electrical distance metric of cluster K; They are respectively The weight of the indicator;

[0018] The calculation method for each weight is as follows: where a and b are Four indicator items index, The weight of indicator b, These are the information entropy of the a-th and b-th indicators corresponding to the candidate partitioning schemes, respectively. The candidate partitioning schemes are multiple partitioning schemes obtained by performing Louvain partitioning on data at different time nodes within a sliding time window.

[0019] Preferably, the constraints of the improved Louvain algorithm in step two include: for cluster K, the constraint is... When a node moves, if both the original cluster and the target cluster satisfy the above constraints after the move, or if the move gain is greater than zero, the node move is executed; wherein, the move gain is the difference between the sum of the cluster partitioning indices of the original cluster and the target cluster after the node moves and the sum of the cluster partitioning indices of the two clusters before the move.

[0020] Preferably, the specific method for dynamically correcting the sensitivity matrix of the partitioned cluster in step three is as follows: calculating the operating state deviation of cluster K. In the formula, It is a cluster operational status deviation, It is a cluster The number of nodes in It is a node The measured voltage, It is a node Voltage predicted by a digital twin model It is the voltage safety threshold;

[0021] The ratio of the operational status deviation of each cluster to the maximum operational status deviation of all clusters is used as the sensitivity mismatch risk coefficient of each cluster. Corrected sensitivity matrix ,in It is the original cluster. The sensitivity matrix, It is a cluster Sensitivity mismatch risk coefficient It is the maximum correction factor and satisfies .

[0022] Preferably, the objective function of the quadratic programming optimization in step three is: ,in It is the objective function. It is a cluster Middle node reactive power adjustment amount, It is a cluster The Middle The current available reactive power capacity of each node;

[0023] The constraints of the quadratic programming optimization include: voltage safety constraints. ,in It is a vector The first in Each element represents a node. The adjusted voltage, of which, , Let K be the vector composed of the adjusted voltages of each node in cluster K. It is a cluster The vector formed by the measured voltages of each node. For clusters A vector composed of the reactive power adjustment amounts of each node in the process; , These are the preset minimum and maximum safe voltage values; power circle constraint. , It is a node The reactive power output before the adjustment This is the lower limit of reactive power.

[0024] This application also provides a reactive power control system for a distributed photovoltaic cluster, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the reactive power control method for a distributed photovoltaic cluster described above.

[0025] As can be seen from the above, the reactive power control method and system for distributed photovoltaic clusters provided in this application have at least the following beneficial effects:

[0026] Overall, this application analyzes the changes in operating parameters of each distributed photovoltaic (PV) node and utilizes the Louvain algorithm for dynamic cluster partitioning. This helps eliminate the impact of fluctuations in node output caused by changes in operating conditions, reduces cluster partitioning boundary mismatch, and improves the matching degree between cluster partitioning and changes in operating conditions. By dynamically correcting the cluster sensitivity matrix after partitioning, dynamic control of physical constraints in the secondary planning process is achieved. The adjusted sensitivity matrix dynamically corrects the reactive power margin allocated by the optimizer, ensuring that voltage safety constraints are met even under dynamic operating condition changes, and avoiding control deviations and capacity overrun risks caused by sensitivity mismatch. This application can improve the ability of cluster partitioning and control decisions to track changes in operating conditions in distributed PV scenarios, ensuring the timeliness and safety of voltage overrun mitigation.

[0027] Specifically, this application achieves dynamic adaptation between cluster partitioning and changes in operating conditions. Existing technologies, when partitioning clusters, often rely on static indicators such as day-ahead forecasts, electrical distance, or voltage sensitivity to determine cluster boundaries, neglecting the dynamic changes in node output trends and real-time adjustment capabilities, leading to a mismatch between cluster partitioning results and operating conditions. This application constructs cluster partitioning indicators by introducing a comprehensive set of four dimensions: reactive power balance metric, output dynamic similarity metric, electrical distance metric, and reactive power-voltage sensitivity metric. An improved Louvain algorithm is used for dynamic cluster partitioning, enabling the cluster boundaries to be updated in real-time according to changes in the operating conditions of each node. Specifically, the reactive power balance metric quantifies the cluster's self-sufficiency by calculating the matching degree between available reactive power capacity and reactive power demand within the cluster; the output dynamic similarity metric characterizes the consistency of output trends among nodes due to shared meteorological conditions through the correlation coefficient of active power sequences between nodes; and the electrical distance metric characterizes the mutual influence intensity of reactive power regulation among nodes within the cluster through negative logarithmic mapping of standardized mutual impedance and weighted modularity calculation. The four indicators comprehensively characterize the cluster division criteria from four dimensions: reactive power supply balance, output coordination, electrical coupling tightness, and voltage response sensitivity. This results in strong complementarity of reactive power resources, high consistency of control strategies, and good voltage response coordination within the divided clusters. It effectively eliminates the impact of node output fluctuations caused by changes in operating conditions on cluster division and reduces the mismatch rate of cluster division boundaries.

[0028] This application proposes an adaptive weighting method for indicators based on information entropy. Existing cluster partitioning methods mostly use fixed weights for multi-indicator weighting, which cannot adapt to the characteristics of distributed photovoltaic scenarios where the importance of each indicator changes with the operating state. This application determines the weights by calculating the information entropy of each indicator corresponding to the candidate partitioning scheme. When the change of an indicator within a time window is more disordered and contains more information, its contribution to the current partitioning should be lower, and its corresponding weight should be smaller; conversely, indicators with more stable changes and higher distinguishability should be assigned greater weights. This adaptive weighting method based on information entropy enables cluster partitioning to dynamically adjust the weight ratio between indicators according to the effectiveness of each indicator in reflecting changes in operating conditions within the time window, avoiding the arbitrariness of manually setting weights and improving the adaptability of cluster partitioning to different operating scenarios. At the same time, the calculation of information entropy is based on candidate partitioning schemes for multiple time sections within a sliding time window, ensuring that the weight update is synchronized with changes in operating conditions, further guaranteeing the matching degree between cluster partitioning and real-time operating status.

[0029] This application establishes a dynamic sensitivity matrix correction mechanism based on operational state deviation. Existing control strategies are typically generated based on fixed cluster structures and preset operating models. When the actual operating state deviates from the model assumptions, the state variables, controlled objects, and constraints in the control model all change, and existing centralized optimization methods struggle to respond to these changes in a timely manner. This application predicts the voltage of each node using a digital twin model, takes the deviation between the measured voltage and the predicted voltage as the operational state deviation, quantifies the reliability of the current sensitivity matrix's representation of the actual physical relationship, and determines a sensitivity mismatch risk coefficient based on the operational state deviation to correct the sensitivity matrix. This corrected sensitivity matrix decreases appropriately as the operational state deviation increases, requiring the optimizer to allocate a larger reactive power margin to meet voltage safety constraints, thereby avoiding control deviations and capacity overrun risks caused by sensitivity mismatch. This correction mechanism allows the control strategy to be tuned in real time according to the operating conditions, and the strategy update is constrained by the physical feasible domain, avoiding control results that do not conform to actual operating laws produced by purely data-driven methods due to the lack of physical constraints in the decision-making process.

[0030] This application organically links dynamic cluster partitioning, dynamic sensitivity correction, quadratic programming optimization, and the issuance and execution of control commands to form a closed-loop cluster control technology system. First, the cluster boundary is dynamically updated through an improved Louvain algorithm, ensuring the cluster structure adapts to real-time operating conditions. Then, dynamic correction of the sensitivity matrix ensures the control model accurately represents the current physical relationships. Next, quadratic programming optimization solves for the optimal reactive power adjustment under physical feasibility constraints. Finally, control commands are issued to each inverter for execution through edge gateways and intelligent fusion terminals, and closed-loop feedback is achieved through continuous monitoring of voltage recovery. These four interconnected and progressively advancing steps collectively ensure the ability of cluster partitioning and control decisions to track changes in operating conditions, the timeliness and security of voltage exceedance mitigation, and the reliability of control results in distributed photovoltaic scenarios. Attached Figure Description

[0031] Figure 1 A flowchart illustrating the steps of a reactive power control method for distributed photovoltaic clusters provided in this application;

[0032] Figure 2 A schematic diagram of the dynamic cluster partitioning process provided in this application;

[0033] Figure 3 A schematic diagram of the secondary planning and optimization process provided for this application. Detailed Implementation

[0034] The following description, in conjunction with the accompanying drawings, details a specific scheme for a reactive power control method and system for distributed photovoltaic clusters provided in this application.

[0035] Please see Figure 1 The document illustrates a flowchart of a reactive power control method for a distributed photovoltaic cluster according to an embodiment of this application, including the following steps:

[0036] Step 1: Collect online monitoring data from the distributed photovoltaic cluster.

[0037] Distribution network topology and line parameters: Through the interface between the main station of the distribution automation system and the GIS platform, the feeder topology connection relationship, switch status and line impedance R / X parameters are imported according to the IEC61968 standard CIM model and updated in real time when the topology changes.

[0038] Photovoltaic grid-connected point data: Smart integrated terminals are installed at distributed photovoltaic grid-connected points and on the low-voltage side of the distribution transformer in the area to collect three-phase voltage, current, active power, and reactive power. Among them, voltage and current are used for rapid condition monitoring, with a sampling period of 1 second, and power data is collected at a period of 5 seconds.

[0039] Inverter operating data: The edge gateway connects to each photovoltaic inverter via RS485 / ModbusTCP / CAN bus to collect DC side voltage and current, AC side active power, reactive power, apparent power, grid connection frequency, internal temperature and available reactive power capacity, with the sampling period set to 1 second.

[0040] Meteorological data: A primary total radiation meter and temperature sensor are installed inside the photovoltaic power station to collect the total horizontal irradiance and ambient temperature, with a sampling period of 1 minute.

[0041] Load data: Active and reactive power data of smart meters at each transformer substation are obtained through the electricity information collection system (interface is the marketing business application system API, and the protocol follows DL / T645-2007), with a collection cycle of 15 minutes.

[0042] The above data is pre-verified by the edge gateway (removing bad data and filling in missing values ​​in accordance with the data quality requirements of Q / GDW10250-2016) and then sent to the cloud for updating the topology model, measurement section and meteorological boundary conditions of the digital twin.

[0043] Step 2: Based on the changes in the operating parameters of each node in the distributed photovoltaic cluster, an improved Louvain algorithm is used to dynamically partition the cluster. Specifically, cluster partitioning indices are constructed based on the matching of reactive power supply and demand within the cluster, reactive power-voltage sensitivity, the consistency of the output change trends of each node within the cluster, and the mutual influence intensity of reactive power regulation between nodes. The cluster partitioning is performed with the goal of maximizing the sum of the cluster partitioning indices, and node movement is constrained by the cluster's reactive power self-sufficiency capability.

[0044] In distributed photovoltaic (PV) scenarios, the proportion of distributed PV is increasing. However, due to the strong randomness and dynamic nature of distributed PV, it has a significant impact on the stable operation of the distribution network, thus placing higher demands on the large-scale and clustered control of distributed PV. Existing distributed PV cluster control technologies typically operate under the assumption of "relatively stable operating conditions," but this assumption affects the accuracy of subsequent cluster partitioning and control strategies, leading to problems such as cluster partitioning mismatch, delayed control response, and insufficient reliability of control results. In fact, with the dynamic changes in distributed PV, the cluster boundaries during partitioning must be dynamically updated according to operating conditions, and the control strategies must also be adjusted in real time according to the operating conditions, with strategy updates constrained by the physical feasible domain. Therefore, a more suitable dynamic cluster control method for distributed PV scenarios is needed.

[0045] The specific process of dynamic cluster partitioning based on the improved Louvain algorithm with four-dimensional comprehensive index is as follows:

[0046] In the process of distributed photovoltaic cluster regulation, it is necessary to first divide the distributed photovoltaic cluster. The physical meaning of cluster division is mainly to group the nodes that have the most closely related voltages, consistent output change trends, and complementary reactive power regulation capabilities into a group, thereby ensuring the balance of resources within the group. The essence of regulation is to change the injected power of the nodes through reactive power means, and then adjust the voltage amplitude change through network impedance transmission.

[0047] First, the sensitivity of node voltage to reactive power injection is not a fixed constant, but rather changes dynamically with the drift of node operating conditions. For example, during periods of high photovoltaic power generation, node voltage is generally higher, and the voltage amplitude change caused by the same unit of reactive power injection is significantly different from that during periods of light load. Furthermore, the coupling strength between nodes fluctuates with changes in line impedance and operating conditions. These factors all affect the sensitivity to reactive power injection. Therefore, when dividing the cluster, it is necessary to first quantify the reactive power balance within each cluster. The formula for obtaining the reactive power balance metric for each cluster is as follows:

[0048] ;

[0049] in It is a cluster Reactive power balance measurement It is a cluster The Middle The current available reactive power capacity of each node, i.e., the node Current dynamic reactive power capacity limit and node The minimum value in the inverter's rated reactive power limit, where the current dynamic reactive power limit can be obtained by the square root of the difference between the inverter's apparent power and the square of the current active power output; Represents a node Belongs to cluster The summation is performed on all nodes in the cluster. It is a cluster The sum of the absolute values ​​of the estimated reactive power changes required by all nodes within the cluster is used to characterize the cluster. The reactive power demand within the system (obtained from power flow calculations; to eliminate the current voltage over-limit state, the estimated reactive power change required at each node) That is, the sum of the absolute values ​​of the estimated reactive power changes mentioned above. It is the division-by-zero protection constant, taking the minimum value. Units and Consistent.

[0050] Inverter available reactive power capacity The calculation reflects the physical coupling characteristic that "the greater the active power output, the smaller the residual reactive power," and introduces... Make It can reflect changes in current operating conditions; and The calculation measures the degree of matching between reactive power supply and demand within the cluster. A higher degree of matching indicates a stronger "self-sufficiency" capability of the cluster and a smaller need for cross-cluster resource allocation, corresponding to... The closer it gets to 1.

[0051] Secondly, because distributed photovoltaic (PV) systems are significantly affected by weather fluctuations, geographically proximate PV nodes tend to exhibit consistent power output trends due to shared weather conditions. This is especially true under the influence of localized cloud formations, where nearby nodes experience almost synchronous power output changes, while more distant nodes may remain unaffected as the clouds have not yet reached them. This consistency in power output trends determines whether nodes within the cluster can respond collaboratively under the same control strategy. For example, if the power output trends of nodes within the cluster diverge, some nodes may increase their output and need to absorb reactive power, while others may decrease their output and need to generate reactive power. A unified strategy cannot accommodate all of these, leading to contradictions in the control strategies within the cluster.

[0052] In this embodiment, a sliding window of 15 minutes is used as the measurement window. Preferably, in this embodiment, the Pearson correlation coefficient between the active power sequences of any two nodes in the same cluster within the same measurement window is used as a measure of their dynamic output similarity. When the output trends of the two nodes are highly consistent, the same control strategy (such as uniformly adjusting the droop coefficient) can be applied to both nodes simultaneously without the need for differentiated strategies. When the output trends of the two nodes have diverged, a uniform strategy will lead to over-adjustment and under-adjustment of different nodes simultaneously, resulting in a deterioration in the control effect. It should be noted that the window length is 15 minutes because typical cloud cover events usually last 5 to 20 minutes, and a 15-minute window can ensure that a complete cloud cover cycle is basically covered. At the same time, when calculating the Pearson correlation coefficient, a very small division-to-zero protection parameter needs to be introduced in the square root calculation of the denominator to avoid calculation errors when the node output is constant (such as at night or on cloudy days with no fluctuations).

[0053] However, relying solely on the consistency of output change trends is insufficient to accurately reflect the electrical coupling characteristics within the network. Since different nodes in the network are connected via network feeders, even nodes with consistent output change trends will have relatively weak mutual influence on reactive power regulation if they are connected to different feeders, resulting in significant electrical distances. This makes effective intra-cluster coordination impossible. Therefore, it is necessary to measure the electrical distance between nodes and the electrical distance of the entire cluster. Specifically, the formula for measuring the electrical distance of each cluster is as follows:

[0054] ;in, ; ;

[0055] In the formula, It is a cluster Electrical distance measurement It is a cluster Middle node and Electrical distance between them , , The nodes in the equivalent impedance matrix and Equivalent mutual impedance between nodes Self-impedance, nodes The self-impedance, and the equivalent impedance matrix can be obtained by inverting the admittance matrix. It is a node and The electrical coupling strength between them It is a preset attenuation coefficient, taken from all node pairs. of the median. These are the nodes The sum of the electrical coupling strengths of all node pairs, It is the sum of the electrical coupling strengths of all node pairs. Here, i and j represent node i and node j in cluster K.

[0056] Understandable, In calculation It is a standardized mutual impedance, a process known to those skilled in the art and will not be elaborated here. It is used to measure the strength of mutual coupling relative to self-coupling. The comparison value is taken as the negative logarithm, which maps the value of mutual impedance in the interval (0,1] to [0,+∞), enhancing the distinguishability of electrical distance. The closer the electrical distance, the smaller the value, and the easier it is to merge during clustering.

[0057] Then, combining the classic weighted modularity calculation form, the first term of the summation measures the actual intra-group coupling strength of nodes within the cluster, and the second term measures the stochastic expected coupling strength, ultimately... This implies that the electrical coupling within the cluster is stronger than the stochastic expectation, and A larger value indicates tighter coupling within the cluster. It should be noted that the electrical distance metric for each cluster is calculated by dividing the inner layer by... The calculation is based on the needs of classical weighted modularity calculation, and its purpose is to make the dimensions of the second term the same as those of the first term. Maintain consistency; outer layer divided by The reason for this is to normalize the data and avoid the impact of too many or too few nodes in the cluster on the quantification results of electrical distance, thus ensuring comparability across clusters.

[0058] Therefore, when partitioning a cluster, it is necessary to consider the fluctuations in physical quantities such as sensitivity and output trend that the partitioning is based on, caused by the drift of node operating states. In this embodiment, the specific formula for obtaining the cluster partitioning index is as follows:

[0059] ;

[0060] in, It is a cluster Cluster partitioning metrics For the reactive power balance metric of cluster K, Let be the normalized mean reactive power-voltage sensitivity of all nodes in cluster K, where is the cluster's mean value. The diagonal elements in the reactive power-voltage sensitivity matrix are also the reactive power-voltage sensitivity of each node. The normalization method is maximum value normalization, which is the maximum value among all diagonal elements of the reactive power-voltage sensitivity matrix. The calculation of the reactive power-voltage sensitivity matrix is ​​a well-known technique in power flow analysis, and will not be elaborated further. Let be the mean of the dynamic similarity measure of the output of cluster K. For the electrical distance metric of cluster K; They are respectively The weight of the indicator.

[0061] Furthermore, the calculation method for each weight is as follows: ,in These are the information entropy values ​​of the a-th and b-th indicators corresponding to the candidate partitioning schemes, where a and b are... The four index items are ∑, which represents the summation of the four index items. The candidate partitioning scheme is a plurality of partitioning schemes obtained by performing Louvain partitioning on node running data at different times within the sliding time window. Specifically, in this embodiment, the candidate partitioning scheme is defined as follows: within the sliding time window, based on the current time... Based on this, Louvain partitioning is performed once for each minute of node operation data within the cluster over the past 15 minutes (a total of 15 time segments), resulting in 15 candidate partitioning schemes. The node operation data includes the node's voltage amplitude, current amplitude, active power, and reactive power. It should be noted that in this embodiment, when the system is in the cold start phase or when fewer than 15 time segments of data are acquired within the sliding time window, a preset... Weights of the four indicators The values ​​are equal, and are all preset constants (set to 0.25 in this embodiment) to perform the initial partitioning. In actual application scenarios, the implementer can also set them themselves.

[0062] It should be noted that, to ensure the computability of information entropy, the first cross-section corresponding to the 15 cross-sections within the acquired time window is used. The data sequence of an indicator is first subjected to frequency statistics or normalization to obtain the probability distribution of the indicator in each state interval, ensuring that the sum of the probabilities of the sequence is 1. Based on this, the information entropy of the corresponding item is calculated using the transformed probability distribution data. It should be noted that the information entropy is calculated using a logarithm to the base 2. Therefore, when normalizing the information entropy, it is necessary to divide the information entropy by a logarithm to the base 2 of 15 (where 15 is the number of sections in the sliding time window). The specific calculation process of information entropy is a well-known technique to those skilled in the art, and this embodiment does not impose any special restrictions on it, so it will not be elaborated here.

[0063] Finally, when using the Louvain algorithm for dynamic partitioning, the following was used: Construct an objective function that enables all clusters obtained from this partitioning to... The sum reaches its maximum, and a moving gain is constructed. ,in This refers to the first iteration in a round of the Louvain algorithm. The mobility gain of the nth node, the A node is moved from its original cluster to another target cluster. This refers to the sum of the cluster partitioning metrics of the original cluster and the target cluster after the move. This refers to the sum of the cluster partitioning metrics of the two clusters before the move. When a node is moved, and all nodes have undergone the above judgment once, a complete iteration round is completed. It should be noted that when... Move to multiple target clusters When both are greater than 0, take the one that satisfies all constraints and The largest cluster will be the final moving target cluster in this round.

[0064] In the Louvain algorithm iteration process, too few clusters result in a loss of synergistic and complementary benefits, while too many clusters increase the edge computing burden and communication latency. Therefore, the recommended range for the number of clusters is [3, 20]. In this embodiment, it is set to 10, and each cluster has... (The cluster's self-sufficiency is a hard constraint, ensuring feasible control within the cluster after partitioning). When a node moves, both the original and target clusters after the move are checked to ensure they meet the above constraints. If they do, the node move is executed; otherwise, the move is rejected. The iterative convergence control of the Louvain algorithm is as follows:

[0065] First requirement: The maximum number of iterations is 50. If this number is exceeded, the process will be forcibly terminated and the current optimal partition will be selected.

[0066] Second requirement: If 5 consecutive rounds If all values ​​are less than the minimum threshold, then the convergence is determined. This refers to the sum of the movement gains of all nodes in one iteration. The minimum threshold can be set as needed, and it represents the optimal result from the Louvain partitioning over the past 15 minutes. The mean is used as the threshold. It should be noted that in this case, for the Louvain partition in the first 15 minutes, the optimal partition needs to be obtained after reaching the maximum number of iterations (that is, the first requirement).

[0067] Specifically, in this embodiment, the dynamic cluster partitioning process is illustrated as follows: Figure 2 As shown.

[0068] Step 3: Dynamically correct the sensitivity matrix of the partitioned cluster. Specifically, calculate the operating state deviation based on the deviation between the measured voltage and the predicted voltage of the nodes in the cluster to correct the sensitivity. Adjust the voltage of each node in the cluster based on the corrected sensitivity matrix to perform voltage safety constraints. Then, construct a target by minimizing the reactive power adjustment of each node for secondary planning optimization and output the reactive power control command of each node.

[0069] The above steps achieve dynamic cluster partitioning based on changes in the operating conditions of each node. Further, it is necessary to determine the reactive power regulation of each node's inverter within each cluster to achieve the control of the distributed photovoltaic cluster. Specifically, based on the cluster partitioning, each cluster corresponds to a reactive power-voltage sensitivity matrix. This matrix reflects the linear mapping relationship between reactive power and voltage within the cluster and serves as the physical basis for subsequent control decisions. However, this matrix is ​​the result of linearizing the power flow equations under the current operating conditions, and its value changes with the drift of the operating conditions. Therefore, it is necessary to adaptively correct this matrix and make control decisions based on the corrected matrix to ensure a match between the control and the actual operating conditions.

[0070] First, the deviation of the cluster's operating status needs to be quantified. The formula for obtaining the operating status deviation of each cluster is as follows:

[0071] ;

[0072] In the formula, It is a cluster operational status deviation, It is a cluster The number of nodes in It is a node The measured voltage, The voltage is predicted using a digital twin model. This is the voltage safety threshold, which, according to the national standard GB / T12325, is taken as 0.07 p.u. of the nominal voltage. It is 0.07 times the nominal voltage.

[0073] Since the changes in photovoltaic power generation exhibit continuous and gradual characteristics, they can usually be continuously predicted over a certain period of time using a digital twin model. When the deviation in operating conditions increases, it indicates that the current sensitivity matrix can no longer accurately reflect the actual physical relationship, meaning that there are significant changes in operating conditions. In this case, the sensitivity matrix needs to be corrected.

[0074] It should be noted in the above formula that the prediction implementation of the digital twin model can be achieved by inputting the current cluster's corresponding distribution network topology connection relationship, switch status, line impedance, and transformer parameters into the digital twin distribution network model. In this embodiment, the digital twin distribution network model adopts the DistFlow linear model, ignoring the second-order term of line loss, and using the active power and reactive power of the photovoltaic grid connection point and load node as the power injection amount of the corresponding node. The upstream grid node is set as the balancing node, and a node-branch model of the current operating section is constructed. Then, based on the node-branch model, the AC power flow forward-backward substitution method is used to calculate the power flow. First, the injected current of each node is calculated according to the node power injection amount, and the branch current is substituted back according to the direction of the distribution network from the end to the power source side. Then, the node voltage is forwarded according to the direction of the power source side to the end. The current substitution and voltage forwarding are cyclically executed until the maximum node voltage change between two adjacent iterations is less than 10. -6 Once the maximum number of iterations is reached, the voltage amplitude of each node is output as the result of the power flow calculation. DistFlow model and AC power flow forward-backward substitution method are commonly used techniques in power flow analysis, and will not be elaborated further.

[0075] Then, the sensitivity matrix is ​​dynamically corrected using the operating state deviation, and quadratic programming optimization is performed based on the corrected sensitivity matrix to predict the regulated voltage. The objective function in the quadratic programming optimization process is:

[0076] ;

[0077] in, It is the objective function. It is a cluster Middle node reactive power adjustment amount, It is a cluster The Middle The current available reactive power capacity of each node, when A value of 0 indicates that the node is powerless and cannot be adjusted, and it is removed from the optimization variables. The power circle constraint is then processed synchronously in the subsequent process.

[0078] The physical feasibility constraints are as follows:

[0079] (1) First, the sensitivity matrix of the partitioned cluster is dynamically corrected. The sensitivity matrix is ​​adjusted according to the deviation of the cluster's operating status. The specific method for obtaining the corrected sensitivity matrix is ​​as follows:

[0080] ;

[0081] in, This is the sensitivity matrix after cluster K correction. It is a cluster The sensitivity mismatch risk coefficient, i.e., cluster The ratio of the operational status deviation of the current cluster to the maximum operational status deviation of all clusters. It should be noted that, to avoid errors in division by zero calculations, a preset minimum constant (in this embodiment, this value is [value]) can be added to the denominator to prevent the denominator from being zero when obtaining this ratio. ), It is the original cluster. The sensitivity matrix, It is the maximum correction factor. The settings need to ensure Preferably, in this embodiment The value is 0.2, which is a statistical estimate of the sensitivity variation under different operating conditions based on the IEEE 33-node standard example. The maximum sensitivity offset is 20%, which is sufficient to cover drift under normal operating conditions, while avoiding over-correction that could cause the optimization results to deviate from the feasible region.

[0082] Understandably, in distributed photovoltaic (PV) operation scenarios, the reverse injection of numerous distributed power sources alters the original power flow distribution of the distribution network, causing reverse power flow in some branches. This leads to a deviation between the reactive power-voltage response relationship obtained under the traditional operating model and the actual operating state. Especially when the degree of reverse power flow increases and there is a significant deviation between the model-predicted voltage and the measured voltage, it indicates a high probability of changes in operating conditions, reducing the reliability of the reactive power-voltage sensitivity obtained from the original operating point in representing the current operating condition. Therefore, by reducing the weight of the model prediction deviation under the current operating condition in the optimization model, the optimizer determines the reactive power adjustment amount according to a more conservative voltage response relationship. To meet voltage safety constraints, the optimizer must allocate a larger reactive power margin, thereby reducing the risk of voltage exceedances failing to be eliminated due to sensitivity mismatch.

[0083] Furthermore, based on the corrected sensitivity matrix of each cluster, and combined with the measured voltage and reactive power regulation of each node in the cluster, the voltage of each node in the cluster is adjusted. The specific adjustment formula is as follows:

[0084] In the formula, Let K be the vector composed of the adjusted voltages of each node in cluster K. It is a cluster The vector formed by the measured voltages of each node. For clusters The vector is composed of the reactive power adjustments of each node in the cluster. Therefore, voltage safety constraints are constructed based on the adjusted voltages of each node in the cluster, as follows:

[0085] For the i-th node in cluster K, voltage safety constraints are: ,in, It is a vector The first in Each element represents a node. Adjusted voltage, , These are the preset minimum and maximum safe voltage values, which are 85% and 110% of the nominal voltage, respectively. According to GB / T29319-2024, when the grid connection point voltage is between 85% and 110% of the nominal voltage, the photovoltaic power generation system should be able to operate normally and continuously.

[0086] (2) Power circle constraint: , It is a node The reactive power output before the adjustment The lower limit of reactive power is determined according to the reactive power output boundary specified in GB / T29319-2024.

[0087] Using the above objective function and physical constraints, the cluster at the current time... The sensitivity matrix, measured voltage, available reactive power capacity of each node, reactive power output, and operating state deviation are used as inputs. The OSQP solver (convex quadratic programming solver) is invoked to output the optimal reactive power adjustment. The OSQP solver tolerance is set to... The maximum number of iterations is set to 200. Quadratic programming is a well-known technique, so its details will not be elaborated further.

[0088] If the quadratic programming terminates due to the lack of feasible solutions under the current constraints, the reactive power control instructions of each node will be set to the maximum output boundary of the currently available reactive power capacity by default and issued for execution. This will serve as a transitional instruction for the subsequent voltage recovery observation period and the triggering of the reserve active power reduction strategy.

[0089] Specifically, the flowchart of the secondary programming optimization process is as follows: Figure 3 As shown.

[0090] Step 4: Perform dynamic control according to the reactive power control command.

[0091] Furthermore, the reactive power output of node i is adjusted according to the reactive power control command output in step three. and will As a reactive power control command, the reactive power control command and the cluster structure are transformed into actual control actions, and a closed-loop feedback is completed. Preferably, in this embodiment, the specific implementation process is as follows:

[0092] Reactive power control command As input, the edge gateway encapsulates it into adjustment commands, with the format adaptively selected according to the specifications of each inverter. These commands are then sent to the smart converged terminal in the distribution area via a protocol conversion module. The terminal then forwards them to the respective inverter controllers via RS485 / CAN. The inverter controllers receive... Then, the reactive power output is adjusted to the target value, achieving autonomous management of voltage over-limit situations without waiting for upper-level instructions. The intelligent fusion terminal continuously collects the voltage at the grid connection point, and when the voltage recovers to the target value... After 5 minutes, a "control completed" event is reported. If the voltage exceedance duration exceeds the system's communication and computing end-to-end evaluation delay (set to 15 seconds in this embodiment, but can be set by the implementer in actual applications), and reactive power regulation still cannot suppress the exceedance, the Active Power Reduction (APC) backup strategy is activated. After regulation, voltage / power measurements are reported to the cloud via the MQTT protocol, and the cloud updates the digital twin model parameters accordingly. This completes the dynamic control of the distributed photovoltaic cluster.

[0093] Based on the same inventive concept as the above method, this application embodiment also provides a reactive power control system for distributed photovoltaic clusters, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above-described reactive power control methods for distributed photovoltaic clusters.

[0094] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the protection scope of this application.

Claims

1. A reactive power control method for distributed photovoltaic clusters, characterized in that, Includes the following steps: Step 1: Collect online monitoring data of the distributed photovoltaic cluster; Step 2: Based on the changes in the operating parameters of each node in the distributed photovoltaic cluster, an improved Louvain algorithm is used to dynamically partition the cluster. Specifically, cluster partitioning indices are constructed based on the matching of reactive power supply and demand within the cluster, reactive power-voltage sensitivity, the consistency of the output change trends of each node within the cluster, and the mutual influence intensity of reactive power regulation between nodes. The cluster partitioning is performed with the goal of maximizing the sum of the cluster partitioning indices, and node movement is constrained by the cluster's reactive power self-sufficiency capability. Step 3: Dynamically correct the sensitivity matrix of the partitioned cluster. Specifically, calculate the operating state deviation based on the deviation between the measured voltage and the predicted voltage of the nodes in the cluster to correct the sensitivity. Adjust the voltage of each node in the cluster based on the corrected sensitivity matrix to perform voltage safety constraints. Then, construct a target by minimizing the reactive power adjustment of each node for secondary planning optimization and output the reactive power control command of each node. Step 4: Perform dynamic control according to the reactive power control command.

2. The reactive power control method for distributed photovoltaic clusters as described in claim 1, characterized in that, The specific method for acquiring online monitoring data in step one is as follows: Importing feeder topology connections, switch status, and line impedance parameters through the distribution automation system; collecting three-phase voltage, current, active power, and reactive power at the photovoltaic grid connection point through the intelligent fusion terminal; collecting DC-side voltage and current, AC-side active power, reactive power, apparent power, grid connection frequency, internal temperature, and available reactive power capacity of each photovoltaic inverter through the edge gateway; collecting total horizontal irradiance and ambient temperature through radiation meters and temperature sensors; acquiring active and reactive power data from smart meters at each distribution area's access points through the electricity consumption information acquisition system; and uploading the above data to the cloud after pre-verification by the edge gateway.

3. The reactive power control method for distributed photovoltaic clusters as described in claim 1, characterized in that, The matching of reactive power supply and demand within the cluster in step two includes: obtaining the current available reactive power capacity of each node in cluster K and the reactive power demand of cluster K. The current available reactive power capacity is the minimum of the current dynamic reactive power capacity upper limit of each node and the rated reactive power upper limit of the inverter. The current dynamic reactive power capacity upper limit is obtained by the square root of the square difference between the apparent power of the inverter and the current active power output. Then, the reactive power balance metric of cluster K is calculated. This is used to characterize the matching of reactive power supply and demand within a cluster: Where ε is the division-by-zero protection constant, and ∑ represents the summation over all nodes in cluster K. It is a cluster The Middle The current available reactive power capacity of each node. It is a cluster The sum of the absolute values ​​of the reactive power changes estimated for all nodes within the system.

4. The reactive power control method for distributed photovoltaic clusters as described in claim 3, characterized in that, The consistency characteristics of the power output change trend of each node in the cluster in step two include: taking a sliding window of a preset duration as a measurement window, using the correlation coefficient between the active power sequences of any two nodes in the same cluster within the same measurement window as the power output dynamic similarity measure of the two nodes; and using the mean of the power output dynamic similarity measures of all node pairs in the cluster as the power output dynamic similarity measure of the cluster, which is used to characterize the consistency characteristics of the power output change trend of each node in the cluster.

5. The reactive power control method for distributed photovoltaic clusters as described in claim 4, characterized in that, The mutual influence intensity of reactive power regulation between nodes in step two includes: calculating the electrical distance between nodes based on the equivalent mutual impedance and node self-impedance in the equivalent impedance matrix, and clustering. Middle node and electrical distance between And calculate the electrical coupling strength between nodes, nodes and electrical coupling strength between Furthermore, the electrical distance metric of the cluster is calculated to characterize the mutual influence intensity of reactive power regulation among nodes within the cluster. Electrical distance measurement ; in, , , The nodes in the equivalent impedance matrix and Equivalent mutual impedance between nodes Self-impedance, nodes Self-impedance, It is a preset attenuation coefficient. These are the nodes The sum of the electrical coupling strengths of all node pairs, It is the sum of the electrical coupling strength of all node pairs.

6. The reactive power control method for distributed photovoltaic clusters as described in claim 5, characterized in that, The acquisition of cluster partitioning metrics in step two includes: ； in For the reactive power balance metric of cluster K, Let be the normalized mean of reactive power-voltage sensitivity for all nodes in cluster K. For the dynamic similarity measure of output of cluster K, For the electrical distance metric of cluster K; They are respectively The weight of the indicator; The calculation method for each weight is as follows: where a and b are Four indicator items index, The weight of indicator b, These are the information entropy of the a-th and b-th indicators corresponding to the candidate partitioning schemes, respectively. The candidate partitioning schemes are multiple partitioning schemes obtained by performing Louvain partitioning on data at different time nodes within a sliding time window.

7. The reactive power control method for distributed photovoltaic clusters as described in claim 3, characterized in that, The constraints for improving the Louvain algorithm in step two include: for cluster K, the constraint is... When a node moves, if both the original cluster and the target cluster satisfy the above constraints after the move, or if the move gain is greater than zero, the node move is executed; wherein, the move gain is the difference between the sum of the cluster partitioning indices of the original cluster and the target cluster after the node moves and the sum of the cluster partitioning indices of the two clusters before the move.

8. The reactive power control method for distributed photovoltaic clusters as described in claim 1, characterized in that, The specific method for dynamically correcting the sensitivity matrix of the partitioned cluster in step three is as follows: calculate the operating state deviation of cluster K. In the formula, It is a cluster operational status deviation, It is a cluster The number of nodes in It is a node The measured voltage, It is a node Voltage predicted by a digital twin model It is the voltage safety threshold; The ratio of the operational status deviation of each cluster to the maximum operational status deviation of all clusters is used as the sensitivity mismatch risk coefficient of each cluster. Corrected sensitivity matrix ,in It is the original cluster. The sensitivity matrix, It is a cluster Sensitivity mismatch risk coefficient It is the maximum correction factor and satisfies .

9. The reactive power control method for distributed photovoltaic clusters as described in claim 1, characterized in that, The objective function for quadratic programming optimization in step three is: ,in It is the objective function. It is a cluster Middle node reactive power adjustment amount, It is a cluster The Middle The current available reactive power capacity of each node; The constraints of the quadratic programming optimization include: voltage safety constraints. ,in It is a vector The first in Each element represents a node. The adjusted voltage, of which, , Let K be the vector composed of the adjusted voltages of each node in cluster K. It is a cluster The vector formed by the measured voltages of each node. For cluster A vector composed of the reactive power adjustment amounts of each node in the process; , These are the preset minimum and maximum safe voltage values; power circle constraint. , It is a node The reactive power output before the adjustment This is the lower limit of reactive power.

10. A reactive power control system for distributed photovoltaic clusters, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the reactive power control method for distributed photovoltaic clusters as described in any one of claims 1-9.

Images