Active power distribution network distributed photovoltaic inverter cluster management and control method and system

Abstract

The invention discloses an active power distribution network distributed photovoltaic inverter cluster management and control method and system, and relates to the field of inverter cluster management and control, and the method comprises the steps: constructing cluster edge nodes, and converging the real-time state variables of all inverters in a cluster; carrying out cooperative calculation on the converged real-time state variables of the inverter by adopting a consistency algorithm; configuring a first data channel between the edge node and the central centralized control system and a second data channel between each inverter and the central centralized control system; establishing a data updating mechanism based on event triggering; constructing a cluster management and control model of each edge node, and outputting a power allocation weight vector of each edge node; and based on the LSTM neural network model, outputting stability labels and abnormal mark points of each edge node cluster management and control model in a future period of time. A collaborative management and control framework of edge autonomous optimization and central cluster management and control is constructed, and advanced early warning and accurate intervention of system operation risks are realized while real-time and stable operation of the power distribution network is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of inverter cluster management and control, specifically to a method and system for managing and controlling distributed photovoltaic inverter clusters in an active power distribution network. Background Technology

[0002] Distributed photovoltaic (PV) power, as a clean and flexible energy form, has seen its penetration rate in distribution networks continue to rise. However, the massive number of distributed PV inverters has transformed from traditional "loads" into "sources" with randomness and intermittency, bringing fundamental changes and challenges to the operation of active distribution networks. Traditional distribution networks are top-down, passive radial networks with relatively controllable power flow and voltage. However, the high proportion of PV integration transforms the distribution network into a complex active network with bidirectional power flow, leading to a series of new technical problems: First, when sunlight is abundant, PV power generation far exceeds local load demand, causing reverse power transmission in the distribution network lines, which is highly susceptible to... These issues can cause overloads in lines and transformers, as well as local voltage exceedances, threatening equipment safety and power quality. Secondly, drastic fluctuations in photovoltaic output (such as those caused by drifting clouds) can trigger rapid fluctuations in grid frequency and voltage, posing a serious threat to power quality. Furthermore, traditional distribution network dispatching and control models primarily target a few centralized and controllable power sources, making it difficult to handle thousands of dispersed and uncontrollable distributed photovoltaic units. To address these challenges, the concept of "cluster management" has emerged. This involves using communication and control technologies to organize dispersed photovoltaic inverters within a certain area into a coordinated group (cluster), treating it as a virtual, controllable whole. As the interface between the photovoltaic system and the grid, the inverter possesses the ability to rapidly adjust active and reactive power, making it an ideal execution unit for achieving refined control.

[0003] Existing technologies simply overlay distributed control with central monitoring, with the central system merely acting as a data dashboard or performing simple logical judgments, failing to deeply explore the value of the data. Secondly, the centralized management and control architecture relies on the central processor for unified decision-making, requiring the transmission of raw data from all inverters, resulting in excessive load on the communication links, and the decision latency is difficult to meet the millisecond-level control requirements. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a method and system for managing distributed photovoltaic inverter clusters in active power distribution networks. This technical solution solves the problems mentioned in the background, where the central system merely acts as a data dashboard or performs simple logical judgments, failing to deeply explore the value of the data. Furthermore, the centralized management architecture relies on a central processor for unified decision-making, requiring the transmission of raw data from all inverters, resulting in excessive load on the communication links and decision delays that cannot meet millisecond-level control requirements.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An active power distribution network distributed photovoltaic inverter cluster management method, comprising: According to the spatial distribution characteristics of the distributed photovoltaic inverter, the cluster edge node is constructed, and the real-time state variables of all inverters in the cluster are converged through the edge node; In all edge nodes, a consistency algorithm is used to cooperatively calculate the converged inverter real-time state variables, and the inverter state variables are adjusted according to the cooperative calculation results; Establish a central control system, configure a first data channel between the edge node and the central control system and a second data channel between each inverter and the central control system; Set the data update frequency of the first data channel to be higher than the data update frequency of the second data channel, and establish an event-triggered data update mechanism; Based on a deep learning model, a cluster management model of each edge node is constructed, and the power distribution weight vector of each edge node is output combined with the first data channel data; Based on an LSTM neural network model, combined with the second data channel data, the stability label and abnormal marker point of each edge node cluster management model in a future period of time are output.

[0006] Preferably, the consistency algorithm is used to cooperatively calculate the converged inverter real-time state variables in all edge nodes, and the inverter state variables are adjusted according to the cooperative calculation results, which specifically includes: Based on the edge computing node device, the real-time state variables of all inverters in each edge node cluster are collected, including but not limited to: AC side output voltage, output current, output active power, output reactive power, grid frequency, and DC side input voltage and input current; According to the communication relationship between the inverters in the cluster, its adjacency matrix is constructed; According to the state variables of each inverter, at least one state variable is selected as a consistency cooperative control variable; According to the adjacency matrix, the consistency deviation of each inverter state variable and all neighbor node state variables is calculated; According to the consistency deviation, the corresponding active or reactive power adjustment instruction is generated by using the feedback control law; According to the adjustment instruction, the power output set value of each inverter is updated, and the inverter is driven to execute, so that the inverter cluster operating state reaches consistency and remains stable.

[0007] Preferably, the data update frequency of the first data channel is set to be higher than the data update frequency of the second data channel, and the event-triggered data update mechanism is established, which specifically includes: According to the division of functions of the edge nodes and the central control system, different basic update periods are set for the first data channel and the second data channel; The update period of the first data channel matches the control period of the internal consistency algorithm of the edge node, and the update period of the second data channel matches the global optimization period of the central control system and the demand of device-level data analysis; A set of system-level trigger events is defined, including at least: steady-state deviation events, communication interruption events and transient fault events; The steady-state deviation events are determined by the edge node according to the cluster consistency control effect, the communication interruption events are determined when the system detects the loss of communication heartbeat with any edge node, and the transient fault events are determined by the inverter according to the local electrical quantity protection threshold; An event priority mapping strategy and a data update mode matching the trigger emergency level are established to assign priority values to the events and trigger the emergency update mode; The event priority mapping strategy specifically includes: The transient fault events and the communication interruption events causing the connectivity destruction of the adjacency matrix are defined as the highest priority; The steady-state deviation events are defined as high priority; The communication interruption events occurring only in a small number of end nodes marked by humans and not affecting the connectivity of the adjacency matrix are defined as medium priority; The data update mode matching the trigger emergency level specifically includes: For the highest priority events, the emergency update mode of the first and second data channels is activated simultaneously; For high-priority events, map to the first data channel and trigger its batch emergency update mode with data compression; For medium-priority events, map to the second data channel and trigger its emergency update mode with shortened period.

[0008] Preferably, the deep learning model-based construction of the cluster control model of each edge node, combined with the first data channel data, outputs the power allocation weight vector of each edge node specifically includes: Based on the deep learning model, a cluster control model of each edge node is constructed, and the input features are the aggregated data of each edge node, the cluster state evaluation results and the system-level demand in the first data channel data; The output of the cluster control model of each edge node is a power allocation weight vector for each edge node; The model takes the minimization of the predicted comprehensive operation cost function value and the target comprehensive operation cost function value as the goal, and constructs the loss function of the cluster control model of each edge node based on the mean square error formula; The comprehensive operation cost function value refers to a weighted sum value of power quality deviation items, line load safety items, power balance items and regulation cost items; Based on the loss function of each edge node cluster management and control model, the learning parameters of each edge node cluster management and control model are iteratively optimized by using a back propagation algorithm; According to the trained edge node cluster management and control model, a preliminary power distribution weight vector is output by inputting the preprocessed input features; Based on the preliminary power distribution weight vector and traditional power flow calculation, the power flow distribution of each key line is obtained to determine whether the line power flow exceeds the safety limit; If yes, a new and safe power distribution weight vector is solved based on the weight vector output by the model as a high-quality initial solution under the premise of meeting all line safety constraints, and a regulation instruction is issued accordingly, and if no, the weight vector is used to generate a regulation instruction and is issued to each edge node.

[0009] Preferably, the LSTM neural network model is combined with the second data channel data to output the stability label and abnormal marker point of each edge node cluster management and control model in the future period of time, which specifically includes: All inverters are digitally marked, and the digital marks are mapped to the operating variables of the inverters; Through the second data channel, the historical monitoring data of the original state variables of each inverter is obtained and preprocessed; Based on the preprocessed historical monitoring data, the stability label of each edge node cluster management and control model is marked by artificial marking to form a training sample set, which is divided into a training set and a validation set; The stability label includes stable, general and unstable. A multi-task learning model with an LSTM neural network as the core is constructed and trained, and the input of the model is the preprocessed state variable time series data of all inverters in a fixed length time window; The model has two output branches, including a first output branch as a classifier and a second output branch as an anomaly detector; The classifier outputs the stability label and probability of each edge node cluster management and control model in the future period of time through a Softmax activation function; The anomaly detector outputs an anomaly score and an abnormal marker point for each inverter at the last time point in the sequence through a Sigmoid activation function; According to the model output and the true label in the training sample, the cross-entropy loss value of the classification task and the binary cross-entropy loss value of the anomaly detection task are calculated respectively, and a weighted sum is obtained to obtain the comprehensive loss value of the current batch; The multi-task LSTM model is trained using the training set to minimize a comprehensive loss function, and is iteratively optimized by an Adam optimizer. The anomaly detector outputs an anomaly score and an anomaly marker point for each inverter at the last time point in the sequence through a Sigmoid activation function, and specifically comprises: The global time series data obtained by splicing all inverter state variables is processed by a main LSTM network to output a global context vector. A Transformer attention mechanism layer is introduced, which takes the global context vector as a query vector and the state variables of all inverters at the last time point as key vectors to calculate an attention weight distribution. The attention weight distribution directly reflects the degree of correlation between the state of each inverter and the global anomaly under the current system state. The calculated attention weight is directly used as the anomaly score of each inverter, and if the attention weight of a certain inverter exceeds a preset threshold, it is determined to be an anomaly marker point, wherein the preset threshold is selected by the F1 score maximization criterion of the validation set.

[0010] Further, the present scheme proposes a distributed photovoltaic inverter cluster management and control system for active power distribution networks, which is used to implement the distributed photovoltaic inverter cluster management and control method for active power distribution networks as described above, comprising: The regional balance module is configured to construct cluster edge nodes according to the spatial distribution characteristics of the distributed photovoltaic inverters, and to aggregate real-time state variables of all inverters in the cluster through the edge nodes; a consistency algorithm is used inside all edge nodes to collaboratively calculate the aggregated inverter real-time state variables, and the inverter state variables are adjusted according to the collaborative calculation results. The cluster management and control module is configured to establish a central control system, configure a first data channel between the edge nodes and the central control system and a second data channel between each inverter and the central control system, set the data update frequency of the first data channel to be higher than that of the second data channel, establish an event-triggered data update mechanism, construct a cluster management and control model for each edge node based on a deep learning model, output a power distribution weight vector for each edge node in combination with the first data channel data, and output a stability label and an anomaly marker point for each edge node cluster management and control model within a future period of time based on an LSTM neural network model in combination with the second data channel data.

[0011] Preferably, the regional balance module comprises: The edge node unit is configured to construct cluster edge nodes according to the spatial distribution characteristics of the distributed photovoltaic inverters, and to aggregate real-time state variables of all inverters in the cluster through the edge nodes. A regional balancing unit is configured to cooperatively calculate the converged inverter real-time state variables by using a consensus algorithm within all edge nodes, and adjust the inverter state variables according to the result of the cooperative calculation.

[0012] Preferably, the cluster management module comprises: A channel configuration unit is configured to establish a central control system, and configure a first data channel between the edge nodes and the central control system and a second data channel between the inverters and the central control system. An update priority unit is configured to set the data update frequency of the first data channel to be higher than the data update frequency of the second data channel, and establish an event-triggered data update mechanism. A power distribution unit is configured to construct a cluster management model of each edge node based on a deep learning model, combine the first data channel data, and output a power distribution weight vector of each edge node. An evaluation and positioning unit is configured to combine the second data channel data based on an LSTM neural network model, and output a stability label and an abnormal marker point of the cluster management model of each edge node within a future period of time.

[0013] Compared with the prior art, the present application has the following advantages: The present application provides a kind of active power distribution network distributed photovoltaic inverter cluster management method and system, the real-time running state of all inverters in cluster is collected and cooperatively calculated by edge node, the global running situation and system level power demand are analyzed in combination with central control system, constructs double-layer data channel to carry high-frequency control mirror and depth diagnosis data respectively, then generates power distribution weight vector based on deep learning model, and the stability of system and equipment anomaly are judged in advance using LSTM neural network model, this scheme realizes the deep cooperation of distributed rapid autonomy and centralized global optimization, effectively improves the intelligent decision-making level and active defense capability of power distribution network in response to power fluctuation and equipment failure, and then significantly enhances the operation flexibility and energy consumption efficiency of power distribution network under high proportion photovoltaic access. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 It is a kind of active power distribution network distributed photovoltaic inverter cluster management method flow chart of the present application; Figure 2 It is a kind of active power distribution network distributed photovoltaic inverter cluster management method flow chart of the present application; Figure 3For the construction of each edge node cluster management and control model of the application, combine the first data channel data, output the power distribution weight vector of each edge node flow chart; Figure 4 For the construction of each edge node cluster management and control model of the application, combine the first data channel data, output the power distribution weight vector of each edge node flow chart; DETAILED DESCRIPTION

[0015] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be made by those skilled in the art.

[0016] REFERENCE Figure 1 As shown in the figure, a kind of active power distribution network distributed photovoltaic inverter cluster management and control method, comprising: According to the spatial distribution characteristics of distributed photovoltaic inverter, the cluster edge node is constructed, and the real-time state variables of all inverters in the cluster are converged through the edge node; Consistency algorithm is used inside all edge nodes to cooperatively calculate the converged inverter real-time state variables, and adjust the inverter state variables according to the cooperative calculation results; Establish a central control system, configure the first data channel between the edge node and the central control system and the second data channel between each inverter and the central control system; Set the data update frequency of the first data channel to be higher than the data update frequency of the second data channel, and establish an event-triggered data update mechanism; Based on deep learning model, construct each edge node cluster management and control model, combine the first data channel data, output the power distribution weight vector of each edge node; Based on LSTM neural network model, combine the second data channel data, output the stability label and abnormal mark point of each edge node cluster management and control model in a period of time in the future.

[0017] It can be explained that the scheme builds a collaborative management architecture of edge autonomous optimization and central cluster management and control, realizes the rapid local collaborative control of distributed photovoltaic inverters based on a consistency algorithm by the edge node, at the same time, the new photovoltaic node only needs to access the edge node to participate in the cluster collaboration, which will not increase the calculation burden of the central controller, forming the first defense barrier; then through the first data channel to synchronize the edge cluster state to the central control system at high frequency, support the edge node cluster management and control model to generate optimization strategy in real time, and rely on the original data of the inverter transmitted through the second data channel to drive the LSTM model for stability prediction and abnormal diagnosis; the update priority unit realizes the collaborative control with the subsequent model through the differential strategy: the high-frequency update of the first data channel ensures that the edge node cluster management and control model responds to system changes in time, the regular low-frequency update of the second data channel meets the analysis needs of the LSTM neural network model, and the emergency update mode triggered by the event can provide a rapid response channel for the highest priority event, and can also actively trigger related models to intervene in decision-making, finally forming an organic whole from intelligent perception to precise execution, significantly improving the system operation efficiency and reliability.

[0018] The cluster edge node is constructed according to the spatial distribution characteristics of the distributed photovoltaic inverter, and the real-time state variables of all inverters in the cluster are converged by the edge node, which specifically includes: Obtain the installation position information of each photovoltaic inverter in the power distribution network topology; Divide the cluster boundary with the power distribution transformer power supply range as the basic unit; Deploy edge computing node devices inside each cluster boundary; Establish communication connection between the edge node and all photovoltaic inverters in the cluster, and collect the real-time state variables of all inverters in the geographical cluster.

[0019] It can be explained that the installation position information of each photovoltaic inverter in the power distribution network topology is obtained, and the cluster boundary is divided with the power distribution transformer power supply range as the basic unit, which is a division method based on the actual power grid structure and power supply relationship, which can fully consider the actual situation of power transmission and distribution. Deploy edge computing node devices inside each cluster boundary, and establish communication connection between the edge node and all photovoltaic inverters in the cluster, so that data collection and instruction transmission are more efficient.

[0020] Referring to Figure 2 It is shown that the real-time state variables of the converged inverters are collaboratively calculated by using a consistency algorithm inside all edge nodes, and the inverter state variables are adjusted according to the collaborative calculation results, which specifically includes: Based on the edge computing node device, real-time state variables of all inverters in each edge node cluster are collected, including but not limited to: AC side output voltage, output current, output active power, output reactive power, grid frequency, and DC side input voltage and input current; According to the communication relationship between the inverters in the cluster, an adjacency matrix is constructed; According to the state variables of each inverter, at least one state variable is selected as the variable for consistent cooperative control; According to the adjacency matrix, the consistency deviation of each inverter state variable and all neighbor node state variables is calculated; According to the consistency deviation, the corresponding active or reactive power regulation instruction is generated by using the feedback control law; According to the regulation instruction, the power output set value of each inverter is updated, and the inverter is driven to execute, so that the operation state of the inverter cluster is consistent and stable.

[0021] This solution achieves a paradigm shift in the power distribution network from centralized dispatch to distributed autonomy by constructing a distributed cooperative control architecture based on a consensus algorithm. Traditional centralized control methods require all data to be transmitted to a central processor for calculation, which has inherent limitations such as high communication bandwidth requirements, high risk of single-point failures, and slow response speed. The core of this solution lies in constructing a continuously operating local cooperative control loop within each edge node. This loop operates normally and aims to proactively maintain grid stability and prevent abnormal states. Specifically, the system accurately describes the information interaction relationship between inverter nodes through an adjacency matrix, enabling each inverter to continuously exchange state information (such as voltage) with neighboring nodes. By continuously calculating the deviation between its own state and the average state of its neighbors, and using a feedback control law to map it into power adjustment commands, each inverter can autonomously and proactively adjust its power. For example, when an inverter senses that its voltage is trending higher than the average of its neighbors, it will autonomously increase reactive power absorption before the voltage actually exceeds the limit, suppressing the voltage rise trend in its infancy. This mechanism simulates the collective intelligence in nature. Each node follows simple local interaction rules, enabling the entire system to exhibit global stability. This continuous "fine-tuning" mechanism forms the first line of defense against disturbances in the power grid, aiming to achieve "seamless regulation," meaning the system automatically recovers stability before users perceive voltage fluctuations. Only in extreme situations where this distributed collaborative mechanism cannot maintain stability will an event-based upper-level protection mechanism (such as forced tripping) be triggered, realizing a closed-loop conversion from electrical quantity deviations to control actions. This ensures the system can quickly recover stable operation after being disturbed. More importantly, this architecture has good scalability and robustness. New photovoltaic nodes only need to be connected to the local communication network to participate in cluster collaboration without increasing the computational burden on the central controller. When a node fails, its neighboring nodes automatically reorganize the communication topology to maintain normal system operation. This distributed autonomous mechanism effectively solves the technical challenge of "massive nodes and decentralized decision-making" brought about by high-proportion distributed photovoltaic access, providing core technical support for building a highly resilient and self-healing smart distribution network. The consistency deviation expression is:

[0022] In the formula, For the first Inconsistency deviation of individual inverters These are elements in the adjacency matrix. For the first Inverter state variable values, For the first Inverter state variable values, The total number of inverters in the cluster, where, It is an adjacency matrix The element in the element is a connection indicator, which explicitly specifies the communication relationship between the inverter and the inverter , and only when there is a direct communication connection between the two nodes , otherwise ; in a specific implementation case, assuming that there is a cluster containing 4 inverters (nodes 1, 2, 3, 4), and the communication connection is a chain structure (1-2, 2-3, 3-4), then its adjacency matrix A is:

[0023] Among them, 1 represents that two nodes are connected, for example, indicates that node 2 can receive information from node 3, indicates that node 3 can receive information from node 2, and 0 indicates that two nodes are not connected, for example, indicates that node 1 cannot directly receive information from node 3. The feedback control law means that each inverter continuously monitors its own state variable, calculates the deviation of its state consensus trend with its neighbors, and then follows a clear proportional rule (control law) that "if the deviation is high, absorb reactive power, and the greater the deviation, the greater the adjustment force" to generate instructions. This "negative feedback" mechanism ensures that any state variable fluctuation can be automatically and quickly counteracted, thereby giving the power distribution network a steady-state self-balancing ability similar to that of a living organism, which is the mathematical basis for achieving coordinated and stable operation of the cluster.

[0024] The central control system is established, and the first data channel between the edge node and the central control system and the second data channel between each inverter and the central control system are configured, which specifically includes: The first data channel is established based on MQTT or HTTP / HTTPS protocol, and is used to transmit aggregated data of edge nodes, cluster state evaluation results and coordination control instructions from the central control system; The second data channel is established based on ModbusTCP or IEC61850, 104 protocol, and is used to transmit original state variables, fault alarm information and power set value of the inverter.

[0025] It can be explained that the scheme realizes the optimized separation of data flow and control flow by defining a heterogeneous double-layer data channel. The core task of the first data channel is to carry the real-time state and decision of the edge autonomous system. It adopts MQTT or HTTP / HTTPS protocol suitable for remote communication, is responsible for the high-frequency transmission of the aggregation data of the edge node and the cluster state evaluation result, and has high real-time requirement and must keep up with the control rhythm of the edge consistency algorithm (usually in hundreds of milliseconds to seconds) to ensure that the central control system can judge whether the edge autonomy is effective in near real time. The second data channel is used as the "on-site control network", adopts industrial Ethernet and power system standard specification (such as IEC61850), directly interacts with the inverter controller, and is responsible for transmitting high-precision raw data for deep diagnosis, historical data analysis and performance evaluation of the distributed photovoltaic inverter cluster of the active power distribution network. These tasks have relatively loose real-time requirements.

[0026] The data update frequency of the first data channel is higher than that of the second data channel, and the event-triggered data update mechanism is established, which specifically includes: Based on the functional division of the edge node and the central control system, the first data channel and the second data channel are respectively set with different basic update periods; The update period of the first data channel matches the control period of the internal consistency algorithm of the edge node, and the update period of the second data channel matches the global optimization period of the central control system and the demand of device-level data analysis; Define a set of trigger events at the system level, which includes at least: steady-state deviation event, communication interruption event and transient fault event; The steady-state deviation event is determined by the edge node according to the cluster consistency control effect, the communication interruption event is determined when the system detects the loss of communication heartbeat with any edge node, and the transient fault event is determined by the inverter according to the local electrical quantity protection threshold; An event priority mapping strategy and a data update mode matching the emergency degree of triggering are established to assign priority to the event and update the mode in an emergency; The event priority mapping strategy specifically includes: The transient fault event and the communication interruption event causing the connectivity destruction of the adjacency matrix are defined as the highest priority; The steady-state deviation event is defined as high priority; The communication interruption event of a few end nodes that only occurs in manually marked nodes and does not affect the connectivity of the adjacency matrix is defined as medium priority; The data update mode matching the emergency degree of triggering specifically includes: For the highest priority event, the emergency update mode of the first and second data channels is activated at the same time; For high-priority events, mapped to the first data channel, trigger its batch emergency update mode with data compression; For medium-priority events, mapped to the second data channel, trigger its emergency update mode with shortened period.

[0027] It can be explained that the differentiated setting of the basic update period is a direct embodiment of the "edge fast control, central slow optimization" division principle, secondly, the data update mode matching the event priority and the triggering emergency level, through the priority matching the corresponding data transmission channel, improves the overall regulation and control ability of the system, the first data channel is the "control state mirror" channel, its primary task is to mirror the results of edge node autonomous control (i.e. the overall running state of the cluster) to the central control system in near real time, and the central system needs to judge whether the edge autonomy is effective and whether the upper strategy needs to be adjusted based on this near real-time mirror, therefore, its update speed must keep up with the control rhythm; the second data channel is the "device details and emergency" channel, its main purpose is not to let the central system micro-manage each inverter, but to provide device-level details for deep diagnosis, historical data analysis and performance evaluation, these tasks have no high real-time requirements, and act as an emergency channel, that is, when edge autonomy fails or extreme events occur, the central system needs to directly intervene in the single inverter through this channel, the basic update period of the second data channel is set according to the refresh rate required by the central control system for global optimization, to meet the timeliness requirements of its system state perception, while avoiding the storage and computing pressure caused by excessive data frequency on device-level historical data analysis, the steady-state deviation event is judged by the edge node according to the consistency control effect of the cluster, specifically, the consistency control algorithm dominated by the edge node, after several iterations, still cannot make the system return to a stable equilibrium state, that is, the mean value of the consistency deviation always exceeds the deviation threshold, wherein the deviation threshold is set based on the 3σ criterion, by statistically analyzing the historical consistency deviation data of the system in the steady state, and taking 2 to 3 times of the standard deviation, the local electrical quantity protection threshold refers to the hard protection action value of variables such as voltage and frequency set by the inverter according to its grid-connected technical standard, when the local electrical quantity (such as voltage and frequency) monitored by the inverter exceeds these boundaries, it will automatically execute protective action.

[0028] Referring to Figure 3 As shown in the figure, the construction of the cluster control model of each edge node combines the first data channel data to output the power distribution weight vector of each edge node. Based on the deep learning model, the cluster control model of each edge node is constructed, and the input features are the aggregated data of each edge node in the first data channel data, the cluster state evaluation result and the system-level demand; The output of the edge node cluster management and control model is a power distribution weight vector for each edge node; The model aims to minimize the predicted comprehensive operation cost function value and the target comprehensive operation cost function value, and constructs a loss function of the edge node cluster management and control model based on a mean square error formula; The comprehensive operation cost function value refers to a weighted sum value of power quality deviation items, line load safety items, power balance items and regulation cost items; Based on the loss function of the edge node cluster management and control model, a back propagation algorithm is used to iteratively optimize the learnable parameters of the edge node cluster management and control model; According to the trained edge node cluster management and control model, a preliminary power distribution weight vector is output by inputting the preprocessed input features; Based on the preliminary power distribution weight vector and traditional power flow calculation, the power flow distribution of each key line is obtained to determine whether the line power flow exceeds the safety limit; If yes, a new safe power distribution weight vector is solved based on the weight vector output by the model as a high-quality initial solution under the premise of meeting all line safety constraints, and a regulation instruction is issued accordingly, and if not, the weight vector is used to generate a regulation instruction and issue it to each edge node.

[0029] It can be explained that in the power system, no matter how complex the control target is (stabilize voltage, balance power, eliminate overload), it must ultimately be realized by changing the distribution of active power and reactive power in the power grid. The core of the scheme is to solve the problem of unequal and intelligent distribution of each interrelated edge node in the active distribution network when responding to the system-level power demand. Traditional optimization methods are prone to local optimal value calculation and time-consuming when dealing with the strong nonlinearity and multiple constraint characteristics of distribution networks, making it difficult to meet application requirements. In view of the fact that when there is no new distributed power supply, the power grid topology and operation mode are relatively stable, and the power distribution weight does not need to be updated frequently, the scheme adopts an intelligent paradigm of "offline training and periodic reasoning" to skillfully avoid the pressure of real-time decision-making. Specifically, through a deep learning model based on a multi-layer perception mechanism, an excellent distribution strategy is extracted from a large amount of historical optimal operation data, thereby generating a high-quality power distribution weight vector initial solution for any given system state. To ensure that the power distribution weight vector initial solution meets the physical safety of the power grid, an online power flow safety checking mechanism is introduced. The mechanism takes the preliminary weight vector as the initial point, the target function takes the minimum deviation of the predicted weight and the initial weight as the target, and the constraint condition is that the line apparent power is not greater than the rated capacity multiplied by a safety factor less than 1. Through constraint optimization algorithms such as sequential quadratic programming or interior point method, a new and safe power distribution weight vector is obtained (for example, by using the sequential quadratic programming algorithm, the Matlab fmincon function is used for solving). The aggregated data is a set of mean values of state variables, the cluster state assessment result refers to the consistency deviation value, and the system-level demand refers to the total active / reactive power amount that needs to be increased or reduced in the entire network to maintain the safe and stable operation of the distribution network. The input features need to be preprocessed before inputting the model, which at least includes: one-hot encoding for category variables to convert them into numerical vectors; for time series data with missing values, linear interpolation or forward filling method is used for data completion; for electrical quantity data with high-frequency noise, Kalman filtering or low-pass digital filter is used for smoothing and denoising; all numerical features are normalized to unify their scales to the [0, 1] interval to eliminate the influence of dimension.The power quality deviation term is obtained by querying real-time measurement data of voltage transformers deployed at power distribution transformers and key nodes, obtaining system average voltage, and calculating the square value of the standardized deviation based on the average voltage and the target rated voltage value as the value of the term. The line load safety term is obtained by querying real-time measurement data of current transformers or smart meters deployed at key feeder outlets, obtaining line current, combining the known rated carrying capacity of the line to calculate the real-time load rate, and taking the square value of the part of the load rate exceeding the preset safety threshold as the value of the term. The power balance term is obtained by receiving cluster total active power data from each edge node, performing algebraic summation in the central control system to obtain cluster net active power, and comparing it with the target power value given by the upper-level dispatching plan to calculate the square value of the deviation degree as the value of the term. The control cost term is obtained by accumulating the absolute value of the active power reduction instructions issued to each edge node and multiplying it by the unit penalty coefficient determined by the local on-grid electricity price policy, and taking the calculation result as the value of the term. The determination method of the weight coefficient is: based on the analytic hierarchy process, a judgment matrix is constructed, which is filled in by power operation and maintenance experts according to industry operation regulations and experience, by comparing the importance of the four optimization criteria of power quality, safety margin, balance accuracy and economic cost two by two, the characteristic vector corresponding to the maximum eigenvalue of the judgment matrix is solved, and the vector is normalized to finally obtain a set of weight coefficients.

[0030] Referring to Figure 4 The output of the future period of time includes: Digitally marking all inverters and mapping the digital marks to the operating variables of the inverters; Through the second data channel, the historical monitoring data of the original state variables of each inverter is obtained, and the historical monitoring data is preprocessed; Based on the preprocessed historical monitoring data, the running stability labels of each edge node cluster control model are marked by artificial marking to form a training sample set, and the training sample set is divided into a training set and a validation set; The stability labels include: stable, general and unstable; A multi-task learning model with an LSTM neural network as the core is constructed and trained, and the input of the model is the preprocessed state variable time series data of all inverters in a fixed length time window; The model has two output branches, including a first output branch as a classifier and a second output branch as an anomaly detector; The classifier outputs the stability labels and probabilities of each edge node cluster control model in the future period of time through a Softmax activation function; The anomaly detector outputs an anomaly score and an anomaly marker point for each inverter at the last time point in the sequence through a Sigmoid activation function; According to the model output and the true label in the training sample, the cross-entropy loss value of the classification task and the binary cross-entropy loss value of the anomaly detection task are calculated respectively, and the weighted sum is obtained to obtain the comprehensive loss value of the current batch; The multi-task LSTM model is trained using the training set, and the comprehensive loss function is minimized as the target, and the Adam optimizer is used for iterative optimization; The anomaly detector outputs an anomaly score and an anomaly marker point for each inverter at the last time point in the sequence through a Sigmoid activation function, and the anomaly detector outputs an anomaly score and an anomaly marker point for each inverter at the last time point in the sequence through a Sigmoid activation function. The global time series data obtained by splicing all inverter state variables is processed by a main LSTM network to output a global context vector; A Transformer attention mechanism layer is introduced, which takes the global context vector as the query vector and the state variables of all inverters at the last time point as the key vector to calculate an attention weight distribution; The attention weight distribution directly reflects the degree of correlation between the state of each inverter and the global anomaly under the current system state; The calculated attention weight is directly used as the anomaly score of each inverter, and if the attention weight of a certain inverter exceeds a predetermined threshold, it is determined to be an anomaly marker point, wherein the predetermined threshold is selected by the F1 score maximization criterion of the validation set.

[0031] It can be explained that the primary task of each edge node cluster management model is to mirror the results of edge node autonomous control (i.e. the overall running state of the cluster) to the central control system in near real time, and the central system needs to judge whether the edge autonomy is effective and whether the upper strategy needs to be adjusted based on this near real-time mirror, which has real-time performance, and the main purpose of the LSTM neural network model is to be used for deep diagnosis, historical data analysis and performance evaluation, which has low real-time performance requirement, and serves as an emergency channel to judge whether each edge node cluster management model is stable or each inverter is abnormal, that is, when edge autonomy fails or an extreme event occurs, the central system needs to directly intervene in the emergency of a single inverter based on the event priority and data update mode through this channel, therefore, based on the LSTM neural network model, through two output branches, based on the original data (and the processed original data) of each inverter, whether each edge node cluster management model is stable and whether there is inverter abnormality can be accurately analyzed, wherein, for a stable system, the state variables (such as voltage and power) of all inverters show a coordinated and smooth change mode in time sequence, and when the system slides to instability, the coordination will be destroyed, and the variables will oscillate, step out of step or diverge, the LSTM network can capture these weak and early signs of instability from the time window data input by virtue of its inherent memory ability, thereby realizing the forward-looking classification of the future stable state of the system, and the stable state of the system reflects the stability of each edge node cluster management model, and the abnormal diagnosis and marking of the inverter are due to the main LSTM network which first encodes the time sequence behavior of all inverters in the network to form a global context vector, which reflects the current overall running mode and health tone of the system, focuses on the abnormality, and through the attention mechanism, takes the global vector as a "benchmark" or "query" to review the state of each inverter at the latest time, and the core function is to calculate the "deviation" or "incoordination degree" of the current state of each inverter from the global health tone, and the attention weight calculated by the core function is essentially a quantitative value of the "deviation", and an inverter that is out of sync with the global mode and behaves abnormally will obtain a significantly higher attention weight than other inverters, therefore, the weight naturally becomes an accurate "abnormal score", thereby realizing the intelligent diagnosis of automatically locating the abnormal inverter from the group behavior.

[0032] Further, based on the same inventive concept as the above active power distribution network distributed photovoltaic inverter cluster management method, the present scheme proposes an active power distribution network distributed photovoltaic inverter cluster management system, comprising: A regional balance module is configured to construct cluster edge nodes according to spatial distribution characteristics of the distributed photovoltaic inverters, gather real-time state variables of all inverters in the cluster through the edge nodes, and perform collaborative calculation on the gathered real-time state variables of the inverters by using a consensus algorithm in all edge nodes, and adjust the state variables of the inverters according to a result of the collaborative calculation. A cluster management and control module is configured to establish a central control system, configure a first data channel between the edge nodes and the central control system and a second data channel between each inverter and the central control system, set a data update frequency of the first data channel to be higher than a data update frequency of the second data channel, establish an event-triggered data update mechanism, construct a cluster management and control model of each edge node based on a deep learning model, output a power distribution weight vector of each edge node in combination with first data channel data, and output a stability label and an abnormal marker point of the cluster management and control model of each edge node in a future period of time in combination with second data channel data based on an LSTM neural network model. The regional balance module comprises: An edge node unit is configured to construct cluster edge nodes according to spatial distribution characteristics of the distributed photovoltaic inverters, and gather real-time state variables of all inverters in the cluster through the edge nodes. A regional balance unit is configured to perform collaborative calculation on the gathered real-time state variables of the inverters by using a consensus algorithm in all edge nodes, and adjust the state variables of the inverters according to a result of the collaborative calculation.

[0033] The cluster management and control module comprises: A channel configuration unit is configured to establish a central control system, configure a first data channel between the edge nodes and the central control system and a second data channel between each inverter and the central control system. An update priority unit is configured to set a data update frequency of the first data channel to be higher than a data update frequency of the second data channel, and establish an event-triggered data update mechanism. A power distribution unit is configured to construct a cluster management and control model of each edge node based on a deep learning model, output a power distribution weight vector of each edge node in combination with first data channel data. An evaluation and positioning unit is configured to output a stability label and an abnormal marker point of the cluster management and control model of each edge node in a future period of time in combination with second data channel data based on an LSTM neural network model.

[0034] In summary, the application has the advantages that a collaborative management architecture of edge autonomous optimization and central cluster management is constructed, real-time stable operation of the power distribution network is ensured, and early warning and precise intervention of system operation risks are realized.

[0035] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection required by the present application is defined by the appended claims and their equivalents.

Claims

1. A method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network, characterized in that, include: Based on the spatial distribution characteristics of distributed photovoltaic inverters, cluster edge nodes are constructed, and real-time state variables of all inverters in the cluster are aggregated through the edge nodes. A consensus algorithm is used within all edge nodes to collaboratively calculate the aggregated real-time state variables of the inverters, and the inverter state variables are adjusted based on the collaborative calculation results. Establish a central control system, and configure the first data channel between the edge nodes and the central control system, as well as the second data channel between each inverter and the central control system. Set the data update frequency of the first data channel to be higher than that of the second data channel, and establish an event-triggered data update mechanism. Based on a deep learning model, a cluster management and control model for each edge node is constructed. Combined with data from the first data channel, the power allocation weight vector of each edge node is output. Based on the LSTM neural network model and combined with the data from the second data channel, the stability labels and anomaly markers of the cluster management model of each edge node are output for a period of time in the future.

2. The method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network according to claim 1, characterized in that, The step of constructing cluster edge nodes based on the spatial distribution characteristics of distributed photovoltaic inverters, and aggregating the real-time state variables of all inverters within the cluster through the edge nodes, specifically includes: Obtain the installation location information of each photovoltaic inverter in the power distribution network topology; Cluster boundaries are defined using the power supply range of distribution transformers as the basic unit. Deploy edge computing node devices within each cluster boundary; Establish communication connections between edge nodes and all photovoltaic inverters within the cluster, and collect real-time status variables of all inverters within the geographical cluster.

3. The method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network according to claim 2, characterized in that, The process of using a consensus algorithm within all edge nodes to collaboratively calculate the aggregated real-time state variables of the inverters, and adjusting the inverter state variables based on the collaborative calculation results, specifically includes: Based on edge computing node devices, real-time status variables of all inverters in each edge node cluster are collected. These status variables include, but are not limited to: AC side output voltage, output current, output active power, output reactive power, grid frequency, and DC side input voltage and input current. Construct its adjacency matrix based on the communication relationships between inverters within the cluster; Based on the state variables of each inverter, at least one state variable is selected as the variable for consistent collaborative control. Based on the adjacency matrix, calculate the consistency deviation between the state variables of each inverter and the state variables of all its neighboring nodes; Based on the consistency deviation, the corresponding active or reactive power adjustment command is generated using the feedback control law. The power output setting value of each inverter is updated according to the adjustment command, and the inverter is driven to execute, so that the inverter cluster can achieve a consistent and stable operating state.

4. The method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network according to claim 3, characterized in that, The establishment of a central control system, the configuration of a first data channel between the edge nodes and the central control system, and a second data channel between each inverter and the central control system specifically include: The first data channel is established based on the MQTT or HTTP / HTTPS protocol and is used to transmit aggregated data from edge nodes, cluster status evaluation results, and coordination control commands from the central control system. The second data channel is established based on ModbusTCP or IEC61850,104 protocols and is used to transmit the inverter's original state variables, fault alarm information and issued power settings.

5. The method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network according to claim 4, characterized in that, The establishment of an event-triggered data update mechanism by setting the data update frequency of the first data channel to be higher than that of the second data channel specifically includes: Based on the functional division of edge nodes and central control system, differentiated basic update cycles are set for the first data channel and the second data channel respectively; Among them, the update cycle of the first data channel matches the control cycle of the internal consistency algorithm of the edge node, and the update cycle of the second data channel matches the global optimization cycle of the central control system and the needs of device-level data analysis. Define a set of triggering events at the system level, which includes at least: steady-state offset events, communication interruption events, and transient fault events; Among them, steady-state offset events are determined by the edge nodes based on the cluster consistency control effect, communication interruption events are determined by the system when it detects the loss of communication heartbeat with any edge node, and transient fault events are determined by the inverter based on the local electrical quantity protection threshold. Establish an event priority mapping strategy and a data update pattern that matches the urgency of triggering events, and assign priority values ​​and urgency update patterns to events; The event priority mapping strategy specifically includes: Transient failure events and communication interruption events that disrupt the connectivity of the adjacency matrix are defined as having the highest priority. Define steady-state offset events as high priority; Communication interruption events that occur only at a small number of manually marked end nodes and whose failure does not affect the connectivity of the adjacency matrix are defined as medium priority; The data update patterns that match the urgency level specifically include: For the highest priority events, activate the emergency update mode of both the first and second data channels simultaneously; For high-priority events, map them to the first data channel and trigger their batch emergency update mode with data compression; For medium-priority events, they are mapped to the second data channel, triggering an emergency update mode with a shortened cycle.

6. The method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network according to claim 5, characterized in that, The process of constructing a cluster management model for each edge node based on a deep learning model, and combining the data from the first data channel to output the power allocation weight vector for each edge node, specifically includes: Based on a deep learning model, a cluster management and control model for each edge node is constructed. Its input features are the aggregated data of each edge node in the first data channel, the cluster status evaluation results, and system-level requirements. The output of the cluster management and control model for each edge node is a power allocation weight vector for each edge node; The model aims to minimize the predicted comprehensive operating cost function value and the target comprehensive operating cost function value. Based on the mean square error formula, it constructs the loss function of the cluster management model for each edge node. The comprehensive operating cost function value refers to the weighted sum of the power quality deviation item, the line load safety item, the power balance item, and the regulation cost item; Based on the loss function of each edge node cluster management model, the backpropagation algorithm is used to iteratively optimize the learnable parameters of each edge node cluster management model. Based on the trained edge node cluster management model, input the preprocessed input features and output a preliminary power allocation weight vector. Based on the preliminary power allocation weight vector and traditional power flow calculation, the power flow distribution of each critical line is obtained to determine whether the power flow of any line exceeds its safety limit. If so, the weight vector output by the model is used as a high-quality initial solution. Under the premise of satisfying all line safety constraints, a new and safe power allocation weight vector is solved, and control commands are issued accordingly. If not, the weight vector is used to generate control commands and issued to each edge node.

7. The method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network according to claim 6, characterized in that, The specific output of stability labels and anomaly markers for each edge node cluster management model over a future period, based on the LSTM neural network model and combined with data from the second data channel, includes: All inverters are digitally labeled, and these labels are mapped to the inverters' operating variables. The historical monitoring data of the original state variables of each inverter are obtained through the second data channel and preprocessed. Based on the preprocessed historical monitoring data, the stability labels of the cluster management and control model of each edge node are manually marked to form a training sample set, which is then divided into a training set and a validation set. The stability labels include: stable, normal, and unstable. Construct and train a multi-task learning model with an LSTM neural network as its core. The input of the model is the time series data of the preprocessed state variables of all inverters within a fixed-length time window. The model has two output branches: the first output branch is a classifier and the second output branch is an anomaly detector. The classifier outputs the stability label and probability of the cluster management model for each edge node over a future period of time through the Softmax activation function. The anomaly detector outputs an anomaly score and an anomaly marker for each inverter at the last time point in the sequence using the Sigmoid activation function. Based on the model output and the real labels in the training samples, calculate the cross-entropy loss value for the classification task and the binary cross-entropy loss value for the anomaly detection task respectively, and then sum them up by weight to obtain the comprehensive loss value for the current batch. The multi-task LSTM model is trained using the training set with the goal of minimizing the comprehensive loss function, and iterative optimization is performed using the Adam optimizer. The anomaly detector, through the Sigmoid activation function, outputs an anomaly score and an anomaly marker for each inverter at the last time point in the sequence, specifically including: A global time series data composed of concatenated state variables of all inverters is processed by a main LSTM network, and a global context vector is output. A Transformer attention mechanism layer is introduced, which uses the global context vector as the query vector and the state variables of all inverters at the last time point as the key vector to calculate an attention weight distribution; This attention weight distribution intuitively reflects the degree of correlation between the state of each inverter and the global anomaly under the current system state; The calculated attention weights are directly used as the anomaly scores of each inverter. If the attention weight of an inverter exceeds a preset threshold, it is determined to be an anomaly marker. The preset threshold is selected by maximizing the F1 score of the validation set.

8. A distributed photovoltaic inverter cluster management and control system for active power distribution networks, characterized in that, The method for managing and controlling a cluster of distributed photovoltaic inverters in an active power distribution network as described in any one of claims 1-7 includes: The regional balancing module is used to construct cluster edge nodes based on the spatial distribution characteristics of distributed photovoltaic inverters, and to aggregate the real-time state variables of all inverters in the cluster through the edge nodes; within all edge nodes, a consensus algorithm is used to perform collaborative calculation on the aggregated real-time state variables of the inverters, and the inverter state variables are adjusted according to the collaborative calculation results; The cluster management module is used to establish a central control system, configure a first data channel between edge nodes and the central control system, and a second data channel between each inverter and the central control system; set the data update frequency of the first data channel to be higher than that of the second data channel, and establish an event-triggered data update mechanism; construct a cluster management model for each edge node based on a deep learning model, and output the power allocation weight vector of each edge node by combining the data from the first data channel; and output the stability label and anomaly marker points of the cluster management model for each edge node over a future period by combining the data from the second data channel with the LSTM neural network model.

9. The active power distribution network distributed photovoltaic inverter cluster management and control system according to claim 8, characterized in that, The regional balancing module includes: An edge node unit is used to construct cluster edge nodes based on the spatial distribution characteristics of distributed photovoltaic inverters, and to aggregate the real-time status variables of all inverters in the cluster through the edge nodes. The regional balancing unit is used to perform collaborative calculations on the converged real-time state variables of the inverters within all edge nodes using a consensus algorithm, and to adjust the inverter state variables based on the collaborative calculation results.

10. The active power distribution network distributed photovoltaic inverter cluster management and control system according to claim 9, characterized in that, The cluster management module includes: A channel configuration unit is used to establish a central control system and configure a first data channel between edge nodes and the central control system, and a second data channel between each inverter and the central control system. An update priority unit is used to set the data update frequency of the first data channel to be higher than the data update frequency of the second data channel, and to establish an event-triggered data update mechanism. A power allocation unit is used to construct a cluster management and control model for each edge node based on a deep learning model, and output the power allocation weight vector for each edge node by combining the data from the first data channel. An evaluation and localization unit is used to output stability labels and anomaly markers for each edge node cluster management model over a future period of time, based on an LSTM neural network model and combined with data from the second data channel.

Images