Dynamic compensation predictive control method and system for low-voltage photovoltaic array

Abstract

The invention discloses a dynamic compensation prediction control method and system for a low-voltage photovoltaic array, and relates to the technical field of low-voltage photovoltaics, and the method comprises the following steps: building a disturbance response sample set, training a Light GBM prediction model with a disturbance feedback label based on the disturbance response sample set, and recognizing the self-excitation disturbance characteristics in a prediction value; based on the prediction model, multi-step rolling prediction is carried out on the current system state, uncertainty scores are output for nodes with self-excitation disturbance paths in prediction results, and disturbance correction is carried out on prediction control quantities of all the nodes; and on the basis of actual feedback after control execution, updating a disturbance response sample set, performing local incremental learning updating on the prediction model, reconstructing a regulation and control priority, and adjusting a trigger strategy and a target range, thereby realizing multi-node cooperative interference avoidance. According to the method, the disturbance response sample set is constructed, the prediction model with the disturbance feedback label is trained, and the self-excitation disturbance path caused by the control behavior can be accurately recognized.

Description

Technical Field

[0001] The present invention relates to the field of low-voltage photovoltaic technology, and more particularly to a dynamic compensation prediction control method and system for a low-voltage photovoltaic array. Background Art

[0002] In existing low-voltage distribution networks, improvements to address voltage fluctuations caused by PV grid integration have introduced machine learning-based voltage prediction methods. For example, the LightGBM model is used to provide short-term predictions of future voltage trends at a single node, driving the inverter to perform reactive power compensation control accordingly. However, these prediction methods generally employ a single-node-centric modeling strategy, ignoring the spatial coupling effects between PV nodes caused by the grid-connected structure. In particular, the voltage cascades caused by compensation control at a single node, leading to voltage disturbances at neighboring nodes or upstream and downstream branches.

[0003] When a node triggers an inverter boost compensation action due to a prediction result, this regulation action, influenced by the resistive network structure and topological paths of the distribution network, often causes a short-term increase or unstable change in the voltage of adjacent nodes, forming a second-order disturbance response. However, the existing LightGBM model does not have the ability to model a control-disturbance closed loop, and its training data does not include the characteristics of secondary disturbances caused by its own regulation behavior. As a result, in subsequent time steps: the model misidentifies this second-order disturbance as a new round of light fluctuations or load changes; the continuous prediction value errors accumulate, triggering erroneous control commands; and the control behaviors of multiple nodes interfere with each other, causing systemic oscillations. These problems are particularly prominent in low-voltage branches with a high proportion of photovoltaic access and frequent control actions. Ultimately, the LightGBM control system enters a self-disturbance misregulation cycle during dynamic operation, unable to accurately predict or stably control, seriously affecting voltage quality and operational safety.

[0004] The above-mentioned disclosed technical solutions have at least the following technical problems: existing prediction methods generally adopt a single-node-centered modeling strategy, ignoring the spatial coupling effect between photovoltaic nodes caused by the grid-connected structure, especially the voltage chain disturbances to adjacent nodes or upstream and downstream branches caused by the execution of compensation control at a certain node.

[0005] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a dynamic compensation predictive control method and system for low-voltage photovoltaic arrays. By constructing a disturbance response sample set, dynamic prediction of the voltage response of the photovoltaic node is realized based on the LightGBM model, and local incremental updates are performed in combination with the control strategy simulation feedback mechanism to achieve accurate characterization and dynamic control compensation of the disturbance response capability of the photovoltaic node under variable operating scenarios, thereby solving the problems of insufficient prediction accuracy and poor adaptability of the control strategy under complex disturbance conditions in the prior art.

[0007] To achieve the above object, the present invention provides the following technical solutions: A dynamic compensation predictive control method for low-voltage photovoltaic arrays includes the following steps: constructing and training a LightGBM prediction model with a disturbance feedback label based on a disturbance response sample set to identify the self-excited disturbance characteristics in the predicted value; based on the LightGBM prediction model, performing a multi-step rolling prediction of the current system state, and scoring the output uncertainty of the node with a self-excited disturbance path in the prediction result; based on the current prediction step disturbance uncertainty score, constructing a control inhibitor structure to perform disturbance correction on the predicted control quantity of each node; executing the corrected control strategy, and based on the actual feedback of the system state after the control is executed, updating the disturbance response sample set, and performing local incremental learning update on the LightGBM prediction model; based on the updated LightGBM prediction model, reconstructing the control priority, adjusting the trigger strategy and target range, and realizing multi-node collaborative interference avoidance.

[0008] In a preferred embodiment, the steps for constructing the disturbance response sample set are specifically as follows: obtaining first data of each distributed photovoltaic node in the target area, the first data including historical control behavior data, voltage status data, and network structure and topological path; based on the first data, combined with the physical distance between nodes and the voltage delay modeling results of the propagation path, identifying the set of perception nodes affected by the control behavior to form a candidate response node set; based on the control trigger timestamp and the voltage change inflection point time of each candidate response node, calculating the response delay offset, and eliminating nodes with an offset less than the system detection resolution threshold; combining the effective response node, the corresponding control behavior and the response delay offset into a triplet to form a "control trigger-response node-time offset" data structure; grouping the triplet according to the control behavior type and the node role to construct a disturbance response sample set in a unified format.

[0009] In a preferred embodiment, the method identifies the set of sensing nodes affected by the control behavior based on the first data in combination with the physical distance between nodes and the voltage delay modeling results of the propagation path, specifically: based on the distribution network topology information, extract the electrical paths between all photovoltaic nodes and their adjacent nodes, and calculate the physical distance between each pair of nodes; in combination with historical disturbance propagation data, estimate the propagation delay time through a preset delay model based on the physical distance between nodes, line impedance parameters and voltage disturbance propagation characteristics; obtain the occurrence time of the control behavior of the control node, and calculate the expected response time interval of each adjacent node based on the propagation delay time; for the voltage time series of each node, use the first-order difference and sliding window fitting method to detect the voltage change inflection point time; if the voltage disturbance inflection point time of the node is within the expected response time interval, then include the node in the set of sensing nodes.

[0010] In a preferred embodiment, the LightGBM prediction model with a disturbance feedback label is trained to identify the self-excited disturbance features in the predicted value, specifically: trend fitting and change point detection operations are performed on each disturbance response sequence in the disturbance response sample set to extract the main change inflection point and response delay characteristics of the voltage fluctuation; a disturbance response label is constructed based on the response amplitude, duration and spatial diffusion path, and a disturbance feedback label of the self-excited disturbance phenomenon is constructed in combination with the disturbance label type; the LightGBM model is trained with the pre-disturbance state vector as the input feature and the disturbance feedback label as the supervision target; the trained LightGBM model is applied to the voltage prediction task of the new time period or new node, and the regional response node containing the self-excited disturbance is identified based on the disturbance response score and decision path output by the model.

[0011] In a preferred embodiment, the LightGBM prediction model is used to perform a multi-step rolling prediction on the current system state, specifically: taking the current system state as the initial input, performing the first round of single-step prediction through the LightGBM prediction model, and decomposing the contribution of each feature in the prediction result through the SHAP value of the model; screening the features whose contribution exceeds the preset threshold to form a core feature subset; constructing an input update mechanism based on the first round of prediction results and the core feature subset to generate input features for the second round of prediction; repeatedly performing multi-step rolling prediction, and comparing the fluctuation amplitude and direction of the current prediction value with the previous prediction value after each round of prediction; combining the node disturbance propagation coefficient to identify the self-excited disturbance path nodes that meet the current fluctuation amplitude greater than the previous fluctuation amplitude and the consistent fluctuation direction.

[0012] In a preferred embodiment, the uncertainty score of the node output with a self-excited disturbance path in the prediction result is specifically as follows: for the identified self-excited disturbance path node, its multi-step prediction sequence is obtained; the integrated prediction method of Bayesian optimization is adopted to generate multiple prediction sub-models through different hyperparameter combinations of the LightGBM prediction model; the prediction results of each sub-model are weighted and fused using the node self-excited disturbance intensity as the weight to generate a prediction distribution; the uncertainty score is corrected according to the standard deviation, mean and node self-excited disturbance intensity of the prediction distribution to obtain a corrected uncertainty score.

[0013] In a preferred embodiment, the control inhibitor structure is constructed based on the current prediction step disturbance uncertainty score, and disturbance correction is performed on the predicted control quantity of each node, specifically: obtaining the disturbance uncertainty score value of the target node in the current prediction step; constructing a disturbance inhibitor response function based on the disturbance uncertainty score value, and calculating the dynamic suppression factor of the target node; and updating the predicted control quantity of the target node to a correction value based on the dynamic suppression factor and historical data.

[0014] In a preferred embodiment, the modified control strategy is executed, and based on the actual feedback of the system state after the control execution, the disturbance response sample set is updated, specifically: the system state time series data before and after the control action is executed is obtained, and the propagation characteristic parameters of the voltage state on the topological path are calculated; based on the time difference between the control execution moment and the voltage state change, a disturbance tracing judgment rule is constructed: if the voltage change meets the control action propagation characteristics and the time difference is less than the threshold, it is marked as a control source disturbance, otherwise it is marked as an external disturbance; a correlation index is constructed in combination with the coupling relationship between the control correction amount and the voltage deviation, and new samples are screened according to preset rules; the screened samples are added to the sample set according to the correlation weight, and high correlation samples are given high weights, and low correlation historical samples are eliminated through a sliding time window.

[0015] The LightGBM dynamic compensation predictive control system for low-voltage photovoltaic arrays includes the following modules: a disturbance feedback perception module: used to construct and train a LightGBM prediction model with a disturbance feedback label based on a disturbance response sample set, and identify the self-excited disturbance characteristics in the predicted value; an uncertainty assessment module: used to perform a multi-step rolling prediction of the current system state based on the LightGBM prediction model, and to score the node output uncertainty of the node with a self-excited disturbance path in the prediction result; a disturbance correction module: used to construct a control inhibitor structure based on the current prediction step disturbance uncertainty score, and perform disturbance correction on the predicted control quantity of each node; a disturbance feedback incremental learning module: used to execute the corrected control strategy, and based on the actual feedback of the system state after the control is executed, update the disturbance response sample set, and perform local incremental learning updates on the LightGBM prediction model; a collaborative control reconstruction module: used to reconstruct the control priority, adjust the trigger strategy and target range based on the updated LightGBM prediction model, and achieve multi-node collaborative interference avoidance.

[0016] The technical effects and advantages of the dynamic compensation predictive control method and system for low-voltage photovoltaic arrays of the present invention are as follows: 1. This invention constructs a disturbance response sample set and trains a LightGBM model with disturbance feedback labels, enabling precise identification of secondary voltage rebound (i.e., self-excited disturbance) paths triggered by control actions. Combining rolling prediction with path dependency analysis, this method captures disturbance propagation links. Through SHAP analysis and feature contribution path backtracking, it identifies high-risk factors and propagation nodes, effectively enabling intelligent prediction and traceable control of nonlinear disturbances in complex photovoltaic systems. This mechanism surpasses traditional response identification approaches based on empirical or linear models and, in practical systems, significantly reduces the risk of uncontrolled voltage fluctuations caused by misjudgments or omissions.

[0017] 2. This invention dynamically updates the LightGBM model decision tree subset based on the system feedback data after control execution, combining the disturbance-control correlation and the disturbance feedback label through real-time updating of the disturbance response sample set and the local incremental learning mechanism of the model. This allows for real-time learning and self-correction of the system disturbance evolution and control effect. In addition, through the nonlinear quantification of the disturbance uncertainty score, as well as the multi-node control coordination and priority reconstruction strategy, the control stability and safety margin in a multi-disturbance coupling environment are ensured, disturbance isolation and coordinated compensation between multiple nodes are achieved, self-excited chain reactions are effectively avoided, and the overall robustness and adaptability of the system are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the flow of the dynamic compensation predictive control method for low-voltage photovoltaic arrays of the present invention; Figure 2This is a schematic diagram of the structure of the LightGBM dynamic compensation prediction control system for low-voltage photovoltaic arrays of the present invention. DETAILED DESCRIPTION

[0019] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] Example 1, Figure 1 The present invention provides a dynamic compensation predictive control method for a low-voltage photovoltaic array, comprising the following steps: S1, build and train a LightGBM prediction model with disturbance feedback labels based on the disturbance response sample set to identify the self-excited disturbance features in the predicted value; In this embodiment, the steps of constructing the disturbance response sample set are specifically as follows: Acquire first data of each distributed photovoltaic node in the target area, the first data including historical control behavior data, voltage state data, and network structure and topology path. The historical control behavior data includes active / reactive power regulation instructions, start / stop commands, and voltage set point adjustments. The voltage state data is a time-series voltage change curve of the regulation node and its surrounding nodes. Based on the first data, combined with the physical distance between nodes and the voltage delay modeling results of the propagation path, a set of sensing nodes affected by the control behavior is identified to form a set of candidate response nodes; Based on the control trigger timestamp and the voltage change inflection point time of each candidate response node, the response delay offset of each candidate response node relative to the control behavior is calculated, and responses with offsets smaller than the system detection resolution threshold are eliminated; Combine the identified valid response node with its corresponding control behavior and response delay offset to form a "control trigger-response node-time offset" triple; The above triplets are grouped according to the control behavior type and node role to construct a disturbance response sample set in a unified format.

[0021] In this embodiment, based on the first data and in combination with the physical distance between nodes and the voltage delay modeling results of the propagation path, a set of sensing nodes affected by the control behavior is identified. The specific steps are: Based on the distribution network topology information, the electrical paths between all photovoltaic nodes and their adjacent nodes are extracted, and the physical distance between each pair of nodes is calculated. ; Combined with historical disturbance propagation data, the propagation delay time is estimated through a preset delay model based on the physical distance between nodes, line impedance parameters and voltage disturbance propagation characteristics. ; Get the control behavior occurrence time of the control node, and take each adjacent node as the object to obtain its expected response time interval based on the delay model result ;in, is the time when the control behavior of the control node occurs, is the preset disturbance response fluctuation tolerance; For each node's voltage time series, the first-order difference and sliding window fitting method are used to detect the inflection point time of significant voltage change; If the voltage disturbance inflection point time of the node is within the expected response time interval, it is included in the sensing node set.

[0022] The specific calculation formula of the propagation delay time is:

[0023]

[0024] in, is the propagation delay time, 、 、 is the propagation coefficient obtained by fitting the historical disturbance samples, is the physical distance between node i and its neighboring nodes, is the equivalent impedance of the corresponding path, is the coordinate of node i, is the coordinate of node j.

[0025] In this embodiment, a LightGBM prediction model with a disturbance feedback label is trained to identify the self-excited disturbance features in the predicted values, specifically: Perform trend fitting and change point detection on each disturbance response sequence in the disturbance response sample set, extract the main change inflection points and response delay characteristics of voltage fluctuations, and construct a disturbance response label based on the response amplitude, duration, and spatial diffusion path; A disturbance feedback label is constructed based on the disturbance response label and the disturbance label type to identify whether there is a self-excited disturbance. Self-excited disturbance is defined as a secondary voltage rebound behavior caused by regulation in the uncontrolled area, which has delay and rebound characteristics. The LightGBM model is trained with the pre-disturbance state vector as input features and the disturbance feedback label as the supervision target to learn the regional response pattern after the disturbance and the conditions for the occurrence of self-excited disturbances. The trained LightGBM model is applied to the voltage prediction task of the new period or new node. Based on the disturbance response score and decision path output by the model, the regional response nodes that may contain self-excited disturbances in the predicted value are identified.

[0026] Furthermore, based on the disturbance response scores and decision paths output by the model, regional response nodes that may contain self-excited disturbances in the predicted values ​​are identified, specifically: Input the state vector of the new period or new node into the trained LightGBM prediction model to obtain the voltage prediction value of each node at multiple future moments and the corresponding disturbance response score; Combined with the internal splitting path of the model, the characteristic decision path corresponding to the disturbance response score is traced back. Based on correlation analysis, the main characteristic variables that lead to the sudden increase in the score are extracted and combined with their thresholds to form a high-risk factor set; If a node with a high-risk factor concentration meets one of the following two conditions, it can be determined as a regional response node with potential self-excited perturbations: 1. The perturbation response score of the node is greater than the preset self-excited perturbation threshold in multiple consecutive prediction steps, and the high-risk factors activated in the model path come from non-directly regulated regions; 2. The node exhibits a high-scoring path dependency clustering phenomenon in the prediction time window, that is, its upstream and downstream nodes also have a high perturbation score trend in a similar path structure; Regional clustering is formed for the identified suspicious nodes, and combined with their propagation path relationships in the power grid topology, it is determined whether they constitute possible self-excited disturbance propagation links.

[0027] The specific calculation formula of the disturbance response score is:

[0028]

[0029]

[0030]

[0031] in, Score the disturbance response of node i at the future time t, is the absolute value of the predicted voltage deviation at node i, is the predicted voltage at node i, is the reference steady-state voltage of node i, is the path dependency of node i, is the voltage variation of node i, is the time interval during which the disturbance changes. is the neighborhood coupling degree of node i, is the number of neighbor nodes of node i, 、 are the disturbance response time series of node i and neighbor node j respectively, is the Pearson correlation coefficient between nodes i and j, is the continuous time length of the node i voltage deviation exceeding the set threshold, 、 、 、 is the weighting coefficient, which is obtained by performing correlation analysis on historical data (the higher the correlation, the larger the set value). .

[0032] In this embodiment, the specific steps of determining whether a possible self-excited disturbance propagation link is formed are as follows: For all the nodes identified as high-risk nodes, an undirected graph structure is constructed based on their physical connection relationships in the electrical topology diagram. ,in, Indicates high-risk nodes, Indicates the edges where electrical paths are connected; Perform the connected subgraph extraction operation on the graph G to obtain several local perturbation subgraphs ,Each subgraph represents a group of high-risk nodes with path connectivity; For each perturbation subgraph, the perturbation response score sequence of all nodes in the multi-step prediction time window is counted. If the judgment condition is met, the subgraph is determined to constitute a self-excited perturbation propagation link; The specific judgment conditions are as follows: Rating consistency condition: more than 60% of the nodes in the subgraph have perturbation scores higher than the set threshold (such as 0.7) in NNN consecutive prediction steps (such as N=3); Path time progressiveness condition: After the nodes are sorted by the power flow direction, their disturbance score peak time satisfies monotony and does not decrease (for example, node A reaches a high score at t1, node B reaches a high score when t2>t1, and so on).

[0033] S2, based on the LightGBM prediction model, performs a multi-step rolling prediction of the current system state and scores the output uncertainty of nodes with self-excited perturbation paths in the prediction results, specifically: In this embodiment, based on the LightGBM prediction model, a multi-step rolling prediction of the current system state is performed, specifically: The current system state (historical control behavior, voltage state, and network topology path data) is used as the initial input. The LightGBM prediction model is used to perform the first round of single-step prediction. At the same time, the contribution of each feature in the prediction result is decomposed by the SHAP value of the model. The core driving features whose influence on the prediction result exceeds the preset threshold are screened out to form a core feature subset. Based on the first-round forecast results and the core feature subset, an input update mechanism for rolling forecasts is constructed. This input update mechanism includes incorporating the first-round forecast values ​​as new lagged features into the dynamic feature set, and adjusting the time-varying weights of external variables based on the contribution of each feature in the core feature subset to generate input features for the second-round forecast. Repeat the multi-step rolling forecast. After each round of forecast, by comparing the fluctuation amplitude and direction of the current forecast value with the previous forecast value, combined with the node disturbance propagation coefficient, identify the self-excited disturbance path nodes that meet the current fluctuation amplitude greater than the previous fluctuation amplitude and the same fluctuation direction.

[0034] In this embodiment, the output uncertainty score of the node with the self-excited disturbance path in the prediction result is specifically: For the identified self-excited perturbation path nodes, a Bayesian optimization ensemble prediction method is adopted based on their multi-step prediction sequence: multiple prediction sub-models are generated by combining different hyperparameters of the LightGBM prediction model. The node self-excited perturbation intensity (the ratio of the current fluctuation amplitude to the previous fluctuation amplitude) is used as the weight to perform weighted fusion on the prediction results of each sub-model to obtain the prediction distribution. Based on the predicted distribution, the uncertainty score of the self-excited disturbance path node is calculated. The ratio of the standard deviation of the predicted distribution to the mean is used as the basic value, and then multiplied by the cumulative coefficient of the node self-excited disturbance intensity (the product of the self-excited disturbance intensity of each step) to obtain the corrected uncertainty score. This score increases nonlinearly with the increase of the self-excited disturbance intensity.

[0035] In this embodiment, the specific calculation formula of the disturbance uncertainty scoring function is:

[0036]

[0037] in, Score the perturbation uncertainty from step t+1 to step t+k, is the standard deviation of the prediction distribution, is the mean of the prediction distribution, is a small positive number used to avoid numerical overflow when the mean is close to 0. is the self-excited perturbation intensity of step i, is the preset nonlinear amplification factor, is the step size of the rolling prediction window, , , are the predicted values ​​at time t+i, t+i-1, and t+i-2 respectively.

[0038] S3, based on the current prediction step disturbance uncertainty score, builds a control inhibitor structure to perform disturbance correction on the predicted control quantity of each node, specifically: Get the disturbance uncertainty score value of the target node in the current prediction step; Based on the disturbance uncertainty score value, a disturbance suppressor response function is constructed to obtain the dynamic suppression factor of the target node; According to the dynamic inhibition factor and historical data of the target node, its predicted control quantity is updated to a corrected value.

[0039] In this embodiment, the disturbance suppressor response function is specifically:

[0040] The update of the predicted control amount is a correction value, and the specific calculation formula is:

[0041] in, is the dynamic suppression factor of target node i, is the preset global sensitivity factor, is the preset node electrical topology influencing factor, Score the perturbation uncertainty, is the corrected control quantity after disturbance suppression, is the predicted control quantity, is the steady-state reference control value of node i (obtained through historical average value).

[0042] S4, executes the revised control strategy, and based on the actual feedback of the system state after the control execution, updates the disturbance response sample set and performs local incremental learning updates on the LightGBM prediction model; In this embodiment, the modified control strategy is executed, and based on the actual feedback of the system state after the control is executed, the disturbance response sample set is updated, specifically: After executing the revised control strategy, the system's actual state data and the system voltage state time series data before and after the control action are executed are obtained. The actual state data includes the real-time monitoring value of each node and the control variable execution deviation. The sampling frequency of the time series data matches the control period, and the propagation delay and attenuation coefficient of the voltage state on the topological path are calculated. Calculate the deviation between the actual state data and the predicted value, analyze the matching degree between the voltage state difference value and the topological path propagation characteristics, and build a disturbance tracing judgment rule in combination with the timing offset. The voltage state difference value is the voltage after control minus the voltage before control, and the timing offset is the time difference between the disturbance occurrence time and the control action execution time. If the propagation law of the voltage difference value along the topological path is consistent with the influence range of the control action, and the timing offset is less than the preset delay threshold, then the disturbance is determined to be caused by the current round of control action. Otherwise, it is marked as an external disturbance, and the judgment result is used as the "disturbance feedback label". When the value of the “disturbance feedback tag” is 1, it indicates a control source disturbance, and when the value is 0, it indicates an external disturbance; Based on the coupling relationship between the deviation amount, the control correction amount and the "disturbance feedback label", a "disturbance-control" correlation quantitative index is constructed. For control source disturbances, the correlation calculation requires superimposing the topological attenuation factor of the voltage difference; Newly collected data is screened in three stages based on the "disturbance-control" correlation quantification index: high-value samples with correlations above a preset threshold are retained. These high-value samples include critical state data where significant disturbances are suppressed or amplified, and control-source disturbance samples marked as 1 are prioritized. Samples with medium correlations are subjected to feature dimensionality reduction, retaining only core features related to disturbance conduction, including voltage state timing features. Redundant samples with correlations below the threshold are eliminated. The screened samples are dynamically added to the disturbance response sample set according to the correlation weight. High-correlation samples are given higher weights, and the weight coefficient of the control source disturbance samples is multiplied by an additional 1.2 times adjustment factor. At the same time, a sliding time window mechanism is used to eliminate historical samples in the sample set whose correlation decays below the preset threshold.

[0043] The LightGBM prediction model is updated through local incremental learning. The specific steps are as follows: Perform local incremental learning on the LightGBM model: Based on the feature space distribution of the new sample, which includes voltage topology features, locate the decision tree subset in the model with the highest overlap with the feature space. This decision tree subset is determined by calculating the match between the sample features and the tree node splitting threshold. The remaining decision tree parameters are frozen, and only the located decision tree subset is incrementally trained, using a weighted cross-entropy loss function, where the weight is the product of the sample association and the "perturbation feedback label". During the incremental training process, a "control correction effect label" is added to the new sample. The label value of the "control correction effect label" is the difference between the actual state and the predicted state when no control is applied. The "disturbance feedback label" is used as the auxiliary learning target of the model. The splitting rule of the decision tree is optimized through the multi-objective loss function. The main target of the multi-objective loss function is the state prediction error, and the auxiliary targets are the correction effect deviation and the disturbance tracing accuracy, so that the model can simultaneously learn the disturbance conduction law, the interaction effect of the control strategy and the disturbance tracing logic.

[0044] S5, based on the updated LightGBM prediction model, reconstructs the control priority, adjusts the trigger strategy and target range, and implements multi-node collaborative interference avoidance. Specifically: Based on the updated LightGBM prediction model, a multi-step rolling prediction is performed on each node of the low-voltage photovoltaic array to obtain the latest self-excited disturbance characteristics and disturbance uncertainty scores of each node: When a high-confidence self-disturbance trend is detected (self-excited disturbance intensity ≥ 1.5 for 3 consecutive steps or more and no), the adaptive control path update logic is triggered; Combined with the PV array topology (such as the distribution of series and parallel branches and the connection relationship of the combiner box), the control coordination coefficient between adjacent nodes is recalculated using the inverse of the node uncertainty score as the weight (a higher coefficient indicates a more significant mutual influence between nodes) and the control priority order is reconstructed; The priority order of the regulation is specifically as follows: the priority of the adjacent nodes of the high-confidence self-disturbance node is lowered, and the priority of the non-self-disturbance node with low coordination coefficient is increased; For high-confidence self-disturbance nodes, the "pre-trigger + step-by-step adjustment" mode is adopted (trigger control one step in advance and the adjustment amount per step does not exceed 10% of the rated value), while the normal trigger threshold is maintained for ordinary nodes; Synchronously shrink the target value setting range of high-confidence self-disturbance nodes (for example, narrow the voltage stability range to ±3%), and expand the target value buffer zone between adjacent nodes to achieve interference isolation; The final control strategy generated is as follows: the control quantity correction value of each node is positively correlated with its own uncertainty score and negatively correlated with the coordination coefficient of the adjacent nodes. In addition, the total control quantity change rate of the cluster control node group does not exceed the safe regulation threshold of the photovoltaic array, realizing interference avoidance and dynamic compensation of self-excited disturbances under multi-node coordinated control. The regulation synergy coefficient is specifically:

[0045] in, is the control coordination coefficient of node i and node j, is the connection weight between node i and node j in the topology (here expressed as the inverse of the physical distance), Historical control quantity 、 Pearson correlation coefficient, 、 Score the uncertainty of the perturbation of the output of node i, j by the updated LightGBM model.

[0046] Example 2, Figure 2 The LightGBM dynamic compensation predictive control system for low-voltage photovoltaic arrays is proposed, which includes the following modules: Perturbation feedback perception module: used to build and train a LightGBM prediction model with disturbance feedback labels based on a disturbance response sample set, and identify self-excited disturbance features in the predicted values; Uncertainty assessment module: used to perform multi-step rolling prediction of the current system state based on the LightGBM prediction model, and output uncertainty scores for nodes with self-excited perturbation paths in the prediction results; Disturbance correction module: used to construct a control inhibitor structure based on the current prediction step disturbance uncertainty score and perform disturbance correction on the predicted control quantity of each node; Disturbance feedback incremental learning module: used to execute the revised control strategy and update the disturbance response sample set based on the actual feedback of the system state after control execution, and perform local incremental learning updates on the LightGBM prediction model; Collaborative control reconstruction module: used to reconstruct control priorities, adjust trigger strategies and target ranges based on the updated LightGBM prediction model, and achieve multi-node collaborative interference avoidance.

[0047] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0048] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0049] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0050] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0051] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0052] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A dynamic compensation predictive control method for low-voltage photovoltaic arrays, characterized in that: The following steps are involved: Construct and train a LightGBM prediction model with disturbance feedback labels based on a disturbance response sample set to identify self-excited disturbance features in the predicted values; Based on the LightGBM prediction model, a multi-step rolling prediction of the current system state is performed, and the output uncertainty of nodes with self-excited perturbation paths in the prediction results is scored; Based on the current prediction step disturbance uncertainty score, a control suppressor structure is constructed to perform disturbance correction on the predicted control quantity of each node; Execute the revised control strategy, and based on the actual feedback of the system state after the control execution, update the disturbance response sample set and perform local incremental learning updates on the LightGBM prediction model; Based on the updated LightGBM prediction model, the control priority is reconstructed, the trigger strategy and target range are adjusted, and multi-node collaborative interference avoidance is achieved.

2. The dynamic compensation predictive control method for low-voltage photovoltaic arrays according to claim 1, characterized in that: The steps for constructing the disturbance response sample set are specifically as follows: Acquire first data of each distributed photovoltaic node in the target area, the first data including historical control behavior data, voltage state data, and network structure and topology path; Based on the first data, combined with the physical distance between nodes and the voltage delay modeling results of the propagation path, a set of sensing nodes affected by the control behavior is identified to form a set of candidate response nodes; Based on the control trigger timestamp and the voltage change inflection point time of each candidate response node, the response delay offset is calculated, and nodes with offsets smaller than the system detection resolution threshold are eliminated; Combine the valid response node, corresponding control behavior, and response delay offset into a triplet to form a data structure of "control trigger-response node-time offset"; The triplets are grouped according to the control behavior type and node role, and a disturbance response sample set with a unified format is constructed.

3. The dynamic compensation predictive control method for low-voltage photovoltaic arrays according to claim 2, characterized in that: The method of identifying the set of sensing nodes affected by the control behavior based on the first data and the physical distance between nodes and the voltage delay modeling result of the propagation path is as follows: Based on the topological structure information of the distribution network, the electrical paths between all photovoltaic nodes and their neighboring nodes are extracted, and the physical distance between each pair of nodes is calculated; Combined with historical disturbance propagation data, the propagation delay time is estimated using a preset delay model based on the physical distance between nodes, line impedance parameters, and voltage disturbance propagation characteristics. Obtain the control behavior occurrence time of the control node and calculate the expected response time interval of each adjacent node based on the propagation delay time; For each node’s voltage time series, the first-order difference and sliding window fitting method are used to detect the voltage change inflection point time; If the voltage disturbance inflection point time of a node is within the expected response time interval, the node is included in the set of sensing nodes.

4. The dynamic compensation predictive control method for low-voltage photovoltaic arrays according to claim 3, characterized in that: The LightGBM prediction model with disturbance feedback labels is trained to identify the self-excited disturbance features in the predicted values, specifically: Perform trend fitting and change point detection operations on each disturbance response sequence in the disturbance response sample set to extract the main change inflection point and response delay characteristics of voltage fluctuation; Construct a disturbance response label based on the response amplitude, duration, and spatial diffusion path, and construct a disturbance feedback label for the self-excited disturbance phenomenon based on the disturbance label type; The LightGBM model is trained with the pre-perturbation state vector as the input feature and the perturbation feedback label as the supervision target; The trained LightGBM model is applied to the voltage prediction task of the new period or new node. Based on the disturbance response score and decision path output by the model, the regional response nodes containing self-excited disturbances are identified.

5. The dynamic compensation predictive control method for low-voltage photovoltaic arrays according to claim 4, characterized in that: The LightGBM prediction model is used to perform multi-step rolling prediction of the current system status, specifically: The current system state is used as the initial input, and the first round of single-step prediction is performed through the LightGBM prediction model. At the same time, the contribution of each feature in the prediction result is decomposed by the SHAP value of the model; Filter features whose contribution exceeds the preset threshold to form a core feature subset; Build an input update mechanism based on the first-round prediction results and the core feature subset to generate input features for the second-round prediction; Repeatedly perform multi-step rolling forecasts, and compare the fluctuation range and direction of the current forecast value with the previous forecast value after each round of forecast; Combined with the node disturbance propagation coefficient, the self-excited disturbance path nodes are identified, where the current fluctuation amplitude is greater than the previous fluctuation amplitude and the fluctuation direction is consistent.

6. The dynamic compensation predictive control method for low-voltage photovoltaic arrays according to claim 5, characterized in that: The output uncertainty score of the node with self-excited perturbation path in the prediction result is specifically: For the identified self-excited perturbation path nodes, obtain their multi-step prediction sequences; Using the Bayesian optimization integrated prediction method, multiple prediction sub-models are generated by combining different hyperparameters of the LightGBM prediction model; The prediction results of each sub-model are weighted and fused using the node self-excited disturbance intensity as the weight to generate the prediction distribution; The uncertainty score is corrected according to the standard deviation, mean and node self-excited disturbance intensity of the predicted distribution to obtain the corrected uncertainty score.

7. The dynamic compensation predictive control method for low-voltage photovoltaic arrays according to claim 6, characterized in that: Based on the current prediction step disturbance uncertainty score, a control inhibitor structure is constructed to perform disturbance correction on the predicted control quantity of each node, specifically: Get the disturbance uncertainty score value of the target node in the current prediction step; Based on the disturbance uncertainty score value, a disturbance suppressor response function is constructed to calculate the dynamic suppression factor of the target node; According to the dynamic inhibition factor and historical data, the predicted control quantity of the target node is updated to a correction value.

8. The dynamic compensation predictive control method for low-voltage photovoltaic arrays according to claim 7, characterized in that: In this embodiment, the modified control strategy is executed, and based on the actual feedback of the system state after the control is executed, the disturbance response sample set is updated, specifically: The system state time series data before and after the control action is executed is obtained, and the propagation characteristic parameters of the voltage state along the topological path are calculated. Based on the time difference between the control execution moment and the voltage state change, a disturbance source tracing judgment rule is established: if the voltage change meets the control action propagation characteristics and the time difference is less than a threshold, it is marked as a control-source disturbance; otherwise, it is marked as an external disturbance. A correlation index is constructed based on the coupling relationship between the control correction amount and the voltage deviation, and new samples are screened according to preset rules. The screened samples are added to the sample set according to the correlation weight, and high-correlation samples are given high weights. At the same time, low-correlation historical samples are eliminated through a sliding time window.

9. A system using the dynamic compensation predictive control method for a low-voltage photovoltaic array according to any one of claims 1 to 8, characterized in that: Includes the following modules: Perturbation feedback perception module: used to build and train a LightGBM prediction model with disturbance feedback labels based on a disturbance response sample set, and identify self-excited disturbance features in the predicted values; Uncertainty assessment module: used to perform multi-step rolling prediction of the current system state based on the LightGBM prediction model, and output uncertainty scores for nodes with self-excited perturbation paths in the prediction results; Disturbance correction module: used to construct a control inhibitor structure based on the current prediction step disturbance uncertainty score and perform disturbance correction on the predicted control quantity of each node; Disturbance feedback incremental learning module: used to execute the revised control strategy and update the disturbance response sample set based on the actual feedback of the system state after control execution, and perform local incremental learning updates on the LightGBM prediction model; Collaborative control reconstruction module: used to reconstruct control priorities, adjust trigger strategies and target ranges based on the updated LightGBM prediction model, and achieve multi-node collaborative interference avoidance.

Images