A transformer area voltage out-of-limit tracing method based on multi-source data

Abstract

The present application relates to power distribution network operation monitoring and data analysis technical field, disclose a kind of based on multi-source data's transformer area voltage over-limit tracing method, the transformer first end voltage and user side power data are obtained in this method;When voltage over-limit occurs, the first-order time derivative extreme point of voltage is extracted to establish physical time anchor point, reconstruct pseudo high-frequency power sequence and generate diagonal time confidence matrix;Topological drift residual index is calculated to check and update sensitivity vector;Effective node is screened and original basis vector sequence is constructed according to descending order of comprehensive influence factor;Diagonal time confidence matrix and regularization term are used to perform iterative Schmidt orthogonalization, and the orthogonal responsibility component vector of each node is extracted, and finally the responsibility index is calculated in combination with the updated sensitivity vector and the grading sequence is output.The present application solves the problem of multi-source data high-low frequency sampling mismatch and multi-node concurrent disturbance coupling, and improves the tracing accuracy under the dynamic topology working condition of distribution network.

Description

Technical Field

[0001] This invention relates to the field of power distribution network operation monitoring and data analysis technology, specifically a method for tracing voltage over-limit in transformer areas based on multi-source data. Background Technology

[0002] With the widespread integration of new loads and distributed power sources into distribution transformer areas, the power flow distribution of distribution networks is becoming increasingly complex, and voltage exceedance issues frequently occur. To develop effective power quality management strategies, it is necessary to accurately trace the responsible nodes that cause voltage exceedances. Existing voltage exceedance tracing methods typically utilize operational data collected from distribution terminals at the transformer head end and smart meters on the user side, combined with voltage and power sensitivity models to perform correlation analysis or multiple regression calculations to assess the impact of each node within the distribution transformer area on voltage exceedances.

[0003] However, existing technologies have several limitations in practical engineering applications. In actual power distribution systems, the voltage acquisition equipment at the transformer head end has a high sampling frequency, while the power acquisition frequency of the smart meters on the user side is low, resulting in a difference in time resolution. Conventional analysis methods often use simple forward assignment or linear interpolation to handle the alignment problem of high and low frequency data. This approach does not consider the confidence decay of reconstructed data over time, making it difficult to accurately reflect the instantaneous power fluctuations, thus leading to synchronization errors during data fusion calculations. Furthermore, power distribution substations undergo physical topology changes during actual operation due to line switching, fault isolation, or branch reconfiguration. Existing causal analysis methods typically rely on static sensitivity models based on historical steady-state data. When the physical topology of the substation changes, the actual electrical distance between nodes changes accordingly, rendering the original static model invalid. Continuing to use this model will cause the causal analysis results to deviate from the actual physical state. In addition, voltage exceedances within a substation are often caused by simultaneous power disturbances at multiple user nodes. Under conditions of concurrent disturbances at multiple nodes, the collected power time series of each node exhibit collinearity mathematically. Existing conventional analysis methods are unable to decouple this concurrent correlation feature, resulting in overlapping computational responsibilities among nodes and making it impossible to accurately and independently quantify the actual physical responsibility of a single node for this voltage overshoot. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for tracing the source of voltage overruns in transformer substations based on multi-source data. This method solves the technical problem that existing technologies suffer from serious deviations in the quantification and tracing results of voltage overrun responsibility due to inaccurate disturbance feature extraction under complex operating conditions.

[0005] To address the above problems, the present invention provides the following technical solution: The first aspect of this invention provides a method for tracing out voltage limits in transformer substations based on multi-source data, comprising: Acquire real-time voltage sequence data collected by the monitoring and acquisition terminal at the transformer head end, as well as power data collected by the smart meters at each user-side node; A static empirical sensitivity vector is constructed based on real-time voltage sequence data and power data under historical operating conditions. When a voltage over-limit event is detected, the extreme points of the first time derivative of the real-time voltage sequence data are extracted and established as physical time anchor points. Based on physical time anchors, the power data is reconstructed by equivalent assignment using the zero-order hold rule to generate a pseudo-high-frequency power sequence; the sampling time difference before and after reconstruction is calculated to obtain the time confidence weights, and the time confidence weights are arranged on the main diagonal to generate a diagonal time confidence matrix; Using the physical time anchor point as a reference, a time window is extracted. Based on pseudo-high-frequency power sequence, real-time voltage sequence data, and static empirical sensitivity vector, topology drift residual is calculated to perform topology state verification. The updated sensitivity vector is determined according to the verification results. Specifically: if the topology drift residual is not greater than the preset tolerance threshold, it is determined that the physical topology has not changed, and the static empirical sensitivity vector is confirmed as the updated sensitivity vector; if the topology drift residual is greater than the preset tolerance threshold, it is determined that the physical topology has changed. Historical short-band sample data that is closest to the physical time anchor point and has not experienced voltage over-limit is retrieved, regression calculation is re-executed to obtain the dynamically recalculated sensitivity vector, and the dynamically recalculated sensitivity vector is confirmed as the updated sensitivity vector. Effective nodes are selected from the pseudo-high frequency power sequence based on the power fluctuation range and static dead zone threshold of each user-side node. The comprehensive impact factor is calculated by combining the updated sensitivity vector. The effective nodes are then sorted in descending order based on the comprehensive impact factor to construct the original basis vector sequence. The diagonal time confidence matrix is ​​used as the weight matrix to define the joint weighted inner product operator. Based on the joint weighted inner product operator, the original basis vector sequence is subjected to iterative Schmitt orthogonalization with positive real number regularization terms. The orthogonal responsibility component vector of each effective node is extracted by sequentially stripping the concurrent collinearity components. Calculate the root mean square effective value of each orthogonal responsibility component vector, and multiply the root mean square effective value of each effective node by the absolute value of the coefficient corresponding to that node in the updated sensitivity vector to obtain the responsibility index of each effective node; calculate the normalized responsibility weight ratio of each effective node, and sort the effective nodes in descending order according to the value of the responsibility index; accumulate the normalized responsibility weight ratio of each effective node in descending order from top to bottom, and define the cumulative set of nodes that first reach or exceed the preset responsibility division threshold as the primary over-limit responsibility nodes, and define the remaining effective nodes as secondary over-limit responsibility nodes, thereby generating a responsibility classification sequence reflecting the primary and secondary relationship of node responsibility; output the responsibility classification sequence to the scheduling terminal, so that the scheduling terminal can perform corresponding over-limit tracing and load control according to the responsibility classification sequence.

[0006] Using the physical time anchor point of voltage mutation as a benchmark, low-frequency power data is reconstructed using the zero-order hold rule and time confidence weights to construct a unified time axis for multi-source data, suppressing time misalignment caused by heterogeneous sampling frequencies. A physical topology state verification mechanism for transformer areas is established by calculating topology drift residuals, which can adaptively update the sensitivity vector when structural changes are detected. In the attribution calculation stage, a joint weighted inner product operator including time confidence weights is defined, and a regularization term is introduced in the Schmitt orthogonalization process to remove concurrent collinearity components, thereby decoupling mutually coupled power fluctuations into independent orthogonal responsibility components. The classification sequence is divided based on the decomposition results, improving the objectivity and reliability of the over-limit tracing results.

[0007] Furthermore, a static empirical sensitivity vector is constructed, specifically including: extracting the time-aligned voltage fluctuation increment vector and the original power increment matrix within the historical steady-state time window; performing principal component analysis (PCA) on the original power increment matrix for rank reduction and denoising to generate a denoised reconstructed power increment matrix; based on the denoised reconstructed power increment matrix and the voltage fluctuation increment vector, using a multiple linear regression algorithm with ridge regression bias constants to calculate the static partial derivative values ​​of each user-side node relative to the head-end voltage; using the static partial derivative values ​​as the static empirical sensitivity coefficients of each user-side node, and concatenating the static empirical sensitivity coefficients of all users column-wise to construct a static empirical sensitivity vector of dimension N×1, where N is the total number of nodes. The combination of PCA rank reduction and ridge regression reduces the interference of measurement noise and multicollinearity in historical operating data on the regression results, ensuring the numerical stability of the static empirical sensitivity vector.

[0008] Furthermore, establishing physical time anchors specifically includes: when real-time voltage sequence data is detected to exceed a preset operating boundary, recording the timestamp of the initial sampling point exceeding the boundary as the over-limit trigger time; using the over-limit trigger time as a benchmark to extract an event analysis window, and using a finite difference algorithm to calculate the first-order time derivative sequence of the real-time voltage sequence data within the event analysis window; traversing the first-order time derivative sequence to retrieve the maximum value; if the retrieved maximum value is greater than a preset disturbance change rate threshold, then establishing the discrete sampling timestamp corresponding to the maximum value as the physical time anchor point; if the maximum value is not greater than the disturbance change rate threshold, then triggering a fault-tolerant mechanism, designating the over-limit trigger time as the physical time anchor point. By extracting the extreme points of the first-order time derivative of the voltage, the initial physical trigger action of the over-limit disturbance is directly locked at the time series level.

[0009] Furthermore, generating a pseudo-high-frequency power sequence specifically includes: defining a discrete time step with the same time step size as the real-time voltage sequence data as a high-frequency time step; for adjacent real low-frequency sampling time intervals that cross physical time anchors, using the physical time anchors as boundaries, forcibly assigning the power values ​​at the high-frequency time step as the actual measured values ​​at the start and end times of the adjacent real low-frequency sampling time intervals; for intervals that do not cross physical time anchors, using a zero-order hold for horizontal forward assignment; for intervals that do not cross physical time anchors, using a zero-order hold for horizontal forward assignment to generate a pseudo-high-frequency power sequence; calculating the absolute time difference between each reconstructed time step in the pseudo-high-frequency power sequence and the actual sampling timestamp of the assignment reference, and using a preset time decay function to calculate the time confidence weight of each reconstructed time step based on the absolute time difference; arranging all the calculated time confidence weights in chronological order to the main diagonal of the matrix to construct a diagonal time confidence matrix. The forced assignment and forward assignment operations on the data range, combined with the time confidence weight generated based on the absolute time difference, not only make up for the lack of data sampling density, but also introduce the reliability quantification constraint of the reconstructed data.

[0010] Furthermore, topology state verification is performed, specifically including: reversing the data from the physical time anchor point to extract a steady-state verification sub-window; calculating the reconstructed power increment at each high-frequency time step within the steady-state verification sub-window compared to the sub-window's initial time, combining these increments to generate a verification reconstructed power increment matrix, and calculating the topology drift residual index by combining the actual voltage increment vector and the static empirical sensitivity vector; if the topology drift residual index is not greater than a preset tolerance threshold, the physical topology is determined to be unchanged, and the nearest valid historical shortwave band sample window from the physical time anchor point that has not experienced voltage exceedance is retrieved, regression calculation is re-executed to obtain a dynamically recalculated sensitivity vector, which is then confirmed as the updated sensitivity vector. By reversing the residual calculation based on the data within the steady-state verification sub-window, changes in electrical topology characteristics caused by line switching can be perceived in real-time.

[0011] Furthermore, if the topology drift residual exceeds a preset tolerance threshold, regression calculation is re-executed based on historical valid data to obtain a dynamically recalculated sensitivity vector. Specifically, when a physical topology change judgment is triggered and the database cannot retrieve sufficient valid low-frequency communication sampling point data pairs, the common resistance value of the physical lines from the transformer head to each user-side node recorded in the system ledger is extracted. This common resistance value is used as an algebraic approximation of the active power disturbance sensitivity, combined to form the ledger-calculated sensitivity vector, which is then confirmed as the updated sensitivity vector for output. An algebraic approximation alternative based on the ledger's common resistance value is provided to ensure continuous execution of the algorithm even when communication sampling data is missing.

[0012] Furthermore, the original basis vector sequence is constructed, specifically including: calculating the difference between the maximum and minimum power values ​​of each user-side node in the pseudo-high-frequency power sequence as the power fluctuation range; comparing the power fluctuation range with a preset static dead zone threshold; if the power fluctuation range is not greater than the static dead zone threshold, the corresponding node is removed; if the power fluctuation range of all monitored nodes is not greater than the static dead zone threshold, the internal attribution operation is skipped, and all monitored nodes are directly encapsulated as zero-responsibility nodes and output to the scheduling terminal; if there are valid nodes with power fluctuation ranges greater than the static dead zone threshold, the absolute value of the coefficient corresponding to each valid node in the updated sensitivity vector is extracted, and multiplied by the corresponding power fluctuation range to calculate the comprehensive influence factor; based on the value of the comprehensive influence factor, each valid node is sorted in descending order, and the corresponding pseudo-high-frequency power column vector is extracted synchronously according to the sorting order, and the original basis vector sequence is constructed by concatenating the columns. The static dead zone threshold is used to filter out nodes without obvious power fluctuations, eliminate redundant interference data, and reduce the dimension of the calculation matrix for subsequent Schmitt orthogonalization.

[0013] Furthermore, the orthogonal responsibility component vectors of each effective node are extracted, specifically including: for any two column vectors of the same dimension in the original basis vector sequence, the inner product of the vectors is constrained by the intermediate product using the diagonal time confidence matrix, and a joint weighted inner product operator is defined; the basis vector ranked first in the original basis vector sequence is extracted as the first orthogonal responsibility component vector; for the original basis vector ranked k in the original basis vector sequence, the joint weighted inner product projection coefficients of the original basis vector in the direction of each preceding orthogonal responsibility component vector are calculated, wherein the denominator of the projection coefficient is configured as the sum of the self-weighted inner product of the corresponding preceding orthogonal responsibility component vector and a preset positive real number regularization term to avoid division by zero; the corresponding preceding projection component is obtained according to each joint weighted inner product projection coefficient, and the kth orthogonal responsibility component vector stripped of concurrent collinearity is generated based on the vector difference between the original basis vector at the kth position and the sum of all preceding projection components; by iterating the above process, the concurrent collinearity coupling features in the original basis vector sequence are stripped in turn, and mutually independent orthogonal responsibility component vectors are extracted.

[0014] Furthermore, the independent orthogonal responsibility component vectors are extracted, specifically including: calculating the joint weighted inner product of the generated k-th orthogonal responsibility component vector itself; if the calculated joint weighted inner product is less than a preset collinearity removal threshold, it is determined that the physical disturbance represented by the original basis vector of the k-th position has been covered by the preceding node, and the k-th orthogonal responsibility component vector is reset to a zero vector. By setting a collinearity removal threshold to form a hard truncation constraint, subordinate responsibility components generated by data following changes are eliminated.

[0015] Furthermore, the generation of a responsibility grading sequence specifically includes: normalizing the responsibility index of each valid node and extracting the corresponding normalized responsibility weight ratio; sorting each valid node in descending order based on the value of the responsibility index; iteratively accumulating the normalized responsibility weight ratio of each valid node according to the sorting order; when the cumulative ratio obtained by the iteration first falls within the target interval formed by the preset responsibility division threshold and the maximum normalization limit value, the cumulative node set of the current iteration is extracted and designated as the main over-limit responsibility node, and the remaining valid nodes are designated as secondary over-limit responsibility nodes, and the responsibility grading sequence is generated; the responsibility grading sequence is concatenated with the zero responsibility node status data corresponding to the removed nodes and the reset nodes, and encapsulated into a data packet in the standard industrial communication protocol format and published to the scheduling terminal.

[0016] A second aspect of this invention provides a transformer area voltage over-limit tracing system based on multi-source data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various calculation and processing steps in the transformer area voltage over-limit tracing method based on multi-source data described in the first aspect. This system possesses the program control logic and data processing capability corresponding to the aforementioned method, and its implementation principle and technical effects correspond to those of the method.

[0017] This invention provides a method for tracing voltage exceedances in transformer substations based on multi-source data. It offers the following advantages: 1. This invention establishes physical time anchors by extracting the extreme points of the first-order time derivative of voltage and constructing a diagonal time confidence matrix based on the sampling time difference. This matrix serves as an intermediate product constraint for inner product operations. This technical feature solves the problem of misaligned sampling rates between high-frequency voltage and low-frequency power data. By using a time decay function to quantify the confidence of reconstructed data, it reduces the calculation error caused by forward interpolation of low-frequency data and improves the accuracy of multi-source heterogeneous data fusion operations.

[0018] 2. This invention calculates the topology drift residual index by extracting a steady-state verification sub-window before the physical time anchor point and constructs a rollback update mechanism that includes dynamic recalculation and the retrieval of physical parameters from the ledger. This feature enables the system to quantitatively determine the validity of the original static empirical sensitivity parameters and perform timely updates when switching or reconfiguration occurs in the physical lines of the distribution area. This prevents calculation errors caused by the failure of the baseline model due to changes in the topology structure and ensures the applicability of the method under complex operating conditions of the distribution network.

[0019] 3. This invention sorts nodes in descending order based on comprehensive influence factors, and performs Schmitt orthogonalization by using a reconstructed joint weighted inner product operator and introducing a positive real number regularization term into the denominator of the iterative projection. This processing method progressively removes the collinear vector characteristics generated when multiple nodes experience concurrent power disturbances, enabling independent extraction and quantification of the physical responsibility of each individual node for exceeding limits. Simultaneously, it avoids matrix singularities and division-by-zero anomalies caused by the constantization of low-frequency power data, maintaining the stability of the attribution and source tracing operation. Attached Figure Description

[0020] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the voltage over-limit tracing method based on orthogonal projection of asynchronous data from a supervisory control and data acquisition system according to the present invention; Figure 3 This is a timing diagram of multi-source heterogeneous measurement data acquisition for the present invention; Figure 4 This is a schematic diagram illustrating the principle of pure data-driven empirical sensitivity extraction in this invention. Figure 5 This is a timing diagram for extracting over-limit feature anchor points in this invention; Figure 6 This is a schematic diagram illustrating the asynchronous data trapezoidal synchronization and confidence quantization principle of the present invention. Figure 7 This is a flowchart of the topology status verification and dynamic rollback update process of the present invention; Figure 8 This is a flowchart of the sensing dead zone filtering and invalid dimension isolation process of the present invention; Figure 9 This is a schematic diagram of the basis vector sorting principle guided by the combined electrical and wave characteristics of the present invention. Figure 10 This is a schematic diagram illustrating the decoupling principle of regularized weighted orthogonal projection and responsibility component in this invention. Figure 11 This is a flowchart of the calculation and closed-loop output of the excess liability attribution index of the present invention.

[0021] Among them, 10 is the monitoring and acquisition terminal; 20 is the smart meter; 30 is the data processing master station; 110 is the data preprocessing module; 120 is the sensitivity extraction module; 130 is the data synchronization module; 140 is the status verification module; 150 is the filtering and sorting module; 160 is the orthogonal projection module; and 170 is the responsibility output module. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 , Figure 1 This is a system architecture diagram according to an embodiment of the present invention. The present invention provides a voltage over-limit tracing system based on the orthogonal projection of asynchronous data from a monitoring and control and data acquisition system. Relying on the physical architecture of the distribution network in the transformer substation, it may include a monitoring and acquisition terminal 10, a smart meter 20, and a data processing master station 30.

[0024] The monitoring and acquisition terminal 10 is deployed at the head end of the transformer in the distribution area and collects real-time voltage sequence data.

[0025] The smart meter 20 is deployed at the user-side node and collects active power data and reactive power data.

[0026] The data processing master station 30 communicates with the monitoring and acquisition terminal 10 and the smart meter 20 to obtain measurement data.

[0027] The data processing master station 30 integrates a data preprocessing module 110, a sensitivity extraction module 120, a data synchronization module 130, a status verification module 140, a filtering and sorting module 150, an orthogonal projection module 160, and a responsibility output module 170.

[0028] See attached document Figure 2 , Figure 2 This is a flowchart of a voltage limit violation tracing method based on orthogonal projection of asynchronous data from a supervisory control and data acquisition system, according to an embodiment of the present invention. The present invention provides a voltage limit violation tracing method based on orthogonal projection of asynchronous data from a supervisory control and data acquisition system, comprising the following steps: S10, the data preprocessing module 110 receives real-time voltage sequence data from the monitoring and acquisition terminal 10 and power data from the smart meter 20; S20, Sensitivity extraction module 120 constructs a static empirical sensitivity vector based on voltage and power data within the normal operating time window; S30, when the monitoring and acquisition terminal 10 detects a voltage over-limit event, the data synchronization module 130 extracts the first derivative extreme point of the real-time voltage sequence data as a physical time anchor point. S40, the data synchronization module 130 reconstructs the power data into a pseudo-high-frequency power sequence based on the physical time anchor point and generates a diagonal time confidence matrix by combining the sampling time difference; S50, the state verification module 140 calculates the topology drift residual index within a specific time window before the physical time anchor point, and triggers the recalculation and update operation of the sensitivity vector based on the topology drift residual index. S60, the filtering and sorting module 150 removes invalid nodes in the pseudo-high frequency power sequence whose fluctuation range is less than the dead zone threshold; S70, the filtering and sorting module 150 calculates the comprehensive influence factor by combining the updated sensitivity vector and the node power range, and constructs the original basis vector sequence by arranging the pseudo high-frequency power sequence of the effective nodes in descending order based on the comprehensive influence factor; S80, the orthogonal projection module 160 reconstructs the joint weighted inner product operator using the diagonal time confidence matrix, performs iterative Schmitt orthogonalization according to the original basis vector sequence and introduces a regularization term, and extracts the orthogonal responsibility component vector of each effective node; S90, the responsibility output module 170 calculates the effective value of each orthogonal responsibility component vector, multiplies it by the corresponding effective sensitivity coefficient to obtain the responsibility index, and generates a responsibility grading sequence based on the responsibility index and outputs it to the scheduling terminal.

[0029] See attached document Figure 3 , Figure 3 This is a timing diagram of multi-source heterogeneous measurement data acquisition according to an embodiment of the present invention. In the method provided by the present invention, step S10 specifically includes the following processing steps performed by the data preprocessing module 110: S11, the data preprocessing module 110 receives the real-time voltage sequence data collected by the monitoring and acquisition terminal 10.

[0030] The monitoring and acquisition terminal 10 is deployed at the low-voltage side of the transformer in the distribution network area. Its internal analog-to-digital conversion circuit operates at the first sampling time step. The main line voltage waveform is continuously discretized and sampled. The acquired real-time voltage sequence data is represented as a high-frequency discrete-time series. ,in For discrete sampling time points. In a specific implementation, the first sampling time step is... The value ranges from 0.1 seconds to 5 seconds to accurately capture the transient change characteristics of the voltage in the distribution area.

[0031] S12, the data preprocessing module 110 receives power data uploaded by smart meters 20 deployed at various user-side nodes.

[0032] The total number of effective user-side nodes under the jurisdiction of the distribution network of the transformer area is denoted as: All monitored user-side nodes constitute a node set. For any user-side node in the set... Smart meters measure user-side nodes. The active power is denoted as With reactive power The data preprocessing module 110 extracts the active power with positive and negative signs. With reactive power To preserve the true power direction characteristics in scenarios such as photovoltaic reverse power transmission, reactive power compensation, and energy storage charging and discharging, the power is converted and combined based on the average reactive-active power sensitivity weight of the long-term operation of the distribution area to construct a one-dimensional equivalent node injection power sequence. The combined calculation formula for the one-dimensional equivalent node injection power sequence is as follows:

[0033] in, The scalar represents the equivalent node injection power after dimensionality reduction and transformation of the user-side node. The positive and negative signs of active power and reactive power are used to characterize the absorption state and feedback / injection state of the corresponding power, respectively. This is the reactive power conversion weighting coefficient calibrated based on the historical steady-state operation data of the transformer area. The smart meter 20 performs periodic data freezing and reporting using a second sampling time step. Limited by the low-voltage side power line carrier communication bandwidth and the terminal's underlying storage capacity, the second sampling time step is set to a minute-level time scale, specifically 15 minutes, 30 minutes, or 60 minutes. In actual operating conditions, if a user-side node experiences a communication outage resulting in missing power measurement data, the user-side node with the outage will be marked as invalid in the current data window and temporarily excluded from subsequent feature calculations to avoid invalid null values ​​causing interruptions in system logic operations.

[0034] S13, The data sequence obtained by the data preprocessing module 110 establishes a strong asynchronous time constraint condition for multi-source heterogeneous data.

[0035] Due to differences in hardware architecture and communication bandwidth between the monitoring terminal 10 and the smart meter 20, high-frequency voltage measurement and low-frequency power measurement satisfy a specific asynchronous proportional relationship. First sampling time step. With the second sampling time step Numerically, there is an order-of-magnitude difference. To clarify the boundary conditions for the system to initiate the cross-timescale reconstruction algorithm, the strongly asynchronous proportional constraint formula is set as follows: ; in, This is a preset strong asynchronous activation threshold. Based on the second-level or minute-level data acquisition characteristics of monitoring and acquisition terminals in actual distribution networks, and the 15-minute-level or hour-level freezing characteristics of smart meters, The optimal value is a positive integer not less than 15. When the sampling time step ratio of the underlying sensors satisfies the above constraint formula, the system determines that there is a serious time granularity mismatch in the current operating condition, thereby activating the subsequent multi-source asynchronous data fusion and pseudo-high-frequency reconstruction algorithm. The above constraint conditions precisely define the asymmetry mechanism of the communication frequency of the underlying sensors in the actual power distribution network operation monitoring and the applicable scope of its algorithm. Under this strong asynchronous time scale constraint, when the voltage at the head end of the transformer in the distribution area experiences a short-term out-of-limit fluctuation, the data preprocessing module 110 has difficulty directly obtaining the high-frequency power measurement point at the same moment from the smart meter 20. For the persistent storage architecture of heterogeneous discrete data within the main station system, those skilled in the art can use a time-series database to construct independent data tables to partition and write voltage characteristic values ​​and power characteristic values. The underlying data queue caching and storage addressing process is a well-known technology in the field and will not be described in detail here.

[0036] See attached document Figure 4 , Figure 4 This is a schematic diagram of a pure data-driven empirical sensitivity extraction principle according to an embodiment of the present invention. In the method provided by the present invention, step S20 specifically includes the following processing steps performed by the sensitivity extraction module 120: S21, the sensitivity extraction module 120 retrieves and extracts historical steady-state time windows from the historical operation database.

[0037] The historical steady-state time window is defined as the continuous operating period during which the voltage sequence does not touch the overvoltage warning threshold or the undervoltage warning threshold. The values ​​of the overvoltage warning threshold and the undervoltage warning threshold are set according to relevant national power quality standards (such as power supply voltage deviation limit standards) and the actual operating procedures of the regional power grid; in practice, they are usually based on the rated operating voltage of the distribution area. Based on this, the overvoltage warning threshold is set to 1.05. Up to 1.07 Between these values, the undervoltage warning threshold is set to 0.90. up to 0.95 The preset period is typically between 7 and 30 days to cover the complete rhythm of user electricity consumption behavior. In actual transformer operation, if the single continuous steady-state time period is insufficient for the preset period, multiple discontinuous short-term steady-state time windows need to be extracted and combined. To avoid directly splicing discontinuous time windows and disrupting the continuity of the load cycle, thus causing distortion of sensitivity features in multiple linear regression calculations, a similar attribute matching mechanism needs to be introduced before performing the splicing operation. Specifically, candidate short-term steady-state time windows are classified by weekday and holiday labels, aligned with the same time period, and outlier points are removed to ensure that the time window segments to be spliced ​​have high similarity in meteorological factors and user electricity consumption behavior patterns; finally, the high-similarity short window data after screening are normalized and smoothly spliced ​​according to time series features to construct an equivalent steady-state historical dataset that meets the overall sample size requirements and conforms to the real physical evolution law. Since there is a time scale difference between the two types of measurement data, the sensitivity extraction module 120 uses the second sampling time step of the smart meter 20. Using the time axis as a reference, the characteristic values ​​of the transformer's front-end voltage at the corresponding sampling time are extracted to establish time-aligned observation samples within the historical steady-state time window. The numerical difference between adjacent sampling times is calculated to form a sample containing... Voltage fluctuation increment vector of discrete data pairs User-side power increment vectors at each node .in:

[0038] Represents a node The equivalent power change during the d-th sampling period. Concatenate the power increment vectors of the N nodes column-wise to construct the original power increment matrix of dimension D×N. .

[0039] S22, Sensitivity extraction module 120 performs a process on the original power increment matrix. Perform principal component analysis to reduce rank and noise.

[0040] During the operation of power distribution area loads, the power fluctuations of adjacent physical users often exhibit high similarity and random unidirectional superposition characteristics. This characteristic can cause the data matrix to exhibit multicollinearity. Directly substituting a multicollinear matrix into regression calculations can easily lead to matrix inversion divergence or equation solution failure. The sensitivity extraction module 120 calculates the covariance matrix of the original power increment matrix and extracts the eigenvalues ​​and corresponding eigenvectors of the covariance matrix. The eigenvalues ​​are sorted in descending order of numerical value, and the top q principal components whose cumulative variance contribution rate reaches the set noise reduction threshold are selected, while small eigenvalue components representing high-frequency random environmental disturbances are discarded. The noise reduction threshold is typically set between 90% and 95%. The power matrix is ​​reconstructed using the eigenvectors corresponding to the retained top q principal components, generating the reconstructed power increment matrix after noise reduction. For the specific matrix decomposition calculations of principal component analysis, those skilled in the art can use the singular value decomposition method. Its feature reduction and data transformation are well-known techniques in the field and will not be elaborated here.

[0041] S23, Sensitivity extraction module 120 is based on reconstructed power increment matrix With voltage fluctuation increment vector The static partial derivatives of each node with respect to the head-end voltage are calculated using a multiple linear regression algorithm.

[0042] The actual physical topology and line impedance parameters of a distribution network often undergo minor adjustments during long-term operation due to equipment aging or load cutovers. A purely data-driven calculation method can directly extract the hidden true electrical distance relationships from the measured data. The multiple linear regression equation is constructed as follows: ; in Let be the Gaussian white noise residual vector. The least squares method combined with regularization constraints is used to estimate the parameters of this model, and the derivation formula is as follows: ; In the formula, to ensure that the tracing results can be directly mapped to specific physical users, The dimension is strictly defined as N×1, representing the static empirical sensitivity vector to be determined for N real physical user nodes; After the original power data is denoised by extracting q principal components through principal component analysis (PCA), it is remapped back to the physical space to generate a denoised reconstructed power increment matrix. Its final dimension is D×N, which is the number of effective samples within the historical steady-state time window. Let be the voltage fluctuation increment vector of dimension D×1; α be the ridge regression bias constant (α>0); N be the N×N diagonal identity matrix.

[0043] It is important to note that in this computational logic, the PCA process is used only as a filtering and denoising method and is not used to change the final dimension of the solution space. Because It is obtained by reconstructing the N-dimensional physical space from q principal components (q < N). This D×N dimensional matrix has an algebraic rank deficiency (its rank is at most q). Therefore, its autocovariance matrix product term... It is definitely an N×N dimensional singular matrix, which cannot be directly inverted. Based on this, this step involves adding positive real number regularization terms to the diagonal. This mechanism forcibly compensates for the Nq missing ranks at the mathematical level, restoring the full-rank invertibility of the matrix. This allows the algorithm to avoid getting physically meaningless solutions in the q-dimensional abstract principal component space that cannot correspond to actual users, and instead directly and stably solve for the static partial derivative values ​​of each real node with dimension N×1 in the N-dimensional physical topological space.

[0044] Through this matrix solution process, the actual electrical impact factors of each user node on the head-end voltage fluctuation under normal operating conditions, i.e., the static empirical sensitivity coefficients, can be obtained. Due to principal component dimensionality reduction operation... Exhibiting non-full-rank characteristics, direct inversion is not possible; therefore, a ridge regression bias constant with a minimal positive real number is introduced during computation. Numerical diagonal protection is provided to ensure that the matrix has a non-singular inverse. The calculation results of all nodes are aggregated to construct a static empirical sensitivity vector of dimension N×1. This extraction mechanism uses purely historical measurement data to extract the electrical distance attributes of the physical topology, which helps to overcome the technical limitations of lagging updates and inaccurate parameters in the geographic information system of power distribution networks in traditional engineering.

[0045] See attached document Figure 5 , Figure 5 This is a timing diagram for extracting over-limit feature anchor points according to an embodiment of the present invention. In the method provided by the present invention, step S30 specifically includes the following processing steps performed by the data synchronization module 130: S31, the data synchronization module 130 obtains the trigger time of the voltage over-limit event and captures the event analysis window.

[0046] Monitoring and data acquisition terminal 10 continuously monitors real-time voltage sequence data. The preset upper and lower limits of the operating voltage are pre-issued and fixed in the terminal by the distribution area dispatch and control system according to national power quality standards. For example, for a low-voltage distribution area with a nominal voltage of 220V, the upper limit threshold of the operating voltage can be configured as +7% of the nominal value, and the lower limit threshold can be configured as -10% of the nominal value. When the voltage amplitude of multiple consecutive sampling points exceeds the preset upper and lower limits of the operating voltage, the system generates an over-limit alarm event and records the timestamp of the initial sampling point that first exceeds the upper and lower limit boundaries as the over-limit trigger time. .

[0047] The specific values ​​of the preset upper and lower limits of the operating voltage are based on the nominal operating voltage of the distribution network system. And the national power quality supply voltage deviation standard setting. In specific embodiments, for low-voltage distribution substations, the upper limit of the operating voltage (overvoltage trigger threshold) is usually set to +7% of the nominal voltage deviation, i.e. Set the operating voltage lower limit (undervoltage trigger threshold) to -10% of the nominal voltage deviation, i.e. The monitoring and acquisition terminal 10 compares the real-time voltage amplitude with the absolute voltage threshold calculated above in real time to reliably determine over-limit events.

[0048] To capture the leading physical disturbance that triggers the out-of-limit event, the data synchronization module 130 uses the out-of-limit trigger time as the time of the event. Based on the baseline, the cut length is The event analysis window. The time interval of the event analysis window is defined as... .in, For the preceding observation margin, This is for post-observation margin. In the system parameter configuration, to ensure coverage of the low-frequency communication cycle of the smart meter 20 to capture complete event information, a pre-observation margin is defined. The value should not be less than the second sampling time step of the smart meter 20. Post-observation margin The value is typically a time interval of 1 to 10 seconds, containing several high-frequency sampling periods. This is to address the issue of insufficient historical data length during system startup. In such cases, the preceding endpoint of the event analysis window is automatically truncated to the system startup time.

[0049] S32, the data synchronization module 130 performs differential calculations on the high-frequency voltage sequence within the event analysis window to extract fluctuation characteristics.

[0050] In actual operation of distribution networks, voltage exceedances caused by physical events such as high-power load switching or sudden changes in output from distributed photovoltaic units often involve changes in the slope of the voltage waveform at the moment the event occurs. The data synchronization module 130 uses a finite difference algorithm to approximate continuous differentials and calculates the first-order time derivative of the discrete voltage sequence within the event analysis window, thereby transforming the voltage exceedance transient process into pulse characteristics. This is then applied to discrete time points within the analysis window. Its first time derivative value The calculation formula is as follows: ; In the formula, and Representing the first Real-time voltage sampling values ​​at adjacent times before and after each sampling point. To monitor the first sampling time step of the data acquisition terminal 10, For the first The rate of change of voltage with directional attributes at each sampling point.

[0051] It is important to note that when calculating the rate of change of voltage, absolute values ​​are strictly avoided to preserve the direction of the disturbance in the actual physical evolution.

[0052] The positive and negative signs (i.e., positive and negative slopes) correspond to different types of physical disturbances: for overvoltage exceeding the limit caused by a voltage rise, the sudden surge characteristics (such as sudden grid connection backfeeding of photovoltaic power) must be extracted based on the positive slope; for undervoltage exceeding the limit caused by a voltage drop, the sudden drop characteristics (such as the starting of a large-capacity inductive motor absorbing reactive power) must be extracted based on the negative slope. This ensures that the disturbance direction and the exceeding limit type are strictly matched when extracting the physical time anchor point subsequently.

[0053] Since absolute values ​​have been removed, the system needs to execute targeted extreme value retrieval logic based on the specific type of the out-of-limit event, which specifically includes: If the current over-limit event is an overvoltage over-limit, it indicates that the transformer area has experienced an upward voltage surge. Then, the first-order time derivative sequence is traversed to retrieve the positive maximum (i.e., the maximum positive slope). If the current over-limit event is an undervoltage over-limit, it indicates that the transformer area has experienced a downward voltage drop. Then, the negative minimum value (i.e., the maximum negative slope) is retrieved by traversing the first-order time derivative sequence. Extract the corresponding extreme point. If the absolute value of the extreme point is greater than the preset perturbation change rate threshold, then the discrete sampling timestamp corresponding to the extreme point is established as the physical time anchor point.

[0054] Using a central difference calculation method helps to reduce the interference of high-frequency pulse noise from single-point measurements on derivative calculations. By traversing the sampling points within the event analysis window, the data synchronization module 130 generates a sequence of first-order time derivatives characterizing the rate of voltage change. To prevent data index out-of-bounds errors, the first and last endpoints of the event analysis window do not participate in the aforementioned central difference calculation; alternatively, those skilled in the art can use forward or backward difference algorithms to replace these endpoints.

[0055] S33, the data synchronization module 130 extracts the extreme points of the first-order time derivative sequence and establishes the real physical time anchor point.

[0056] The data synchronization module 130 retrieves the extreme point with the largest amplitude from the generated first-order time derivative sequence and records its corresponding calculated value as... . The maximum value Compared with the preset disturbance change rate threshold Perform numerical comparisons. Threshold for the rate of change of disturbance. The setting is established based on the maximum normal voltage fluctuation slope recorded during the historical normal steady-state operation of the transformer area, and is generally set to 1.5 to 2.0 times the average of this normal voltage fluctuation slope. When Greater than the disturbance change rate threshold When the extreme point is determined to be a sudden change caused by a physical electrical event, the data synchronization module 130 extracts the extreme point. The corresponding discrete sampling timestamps are established as the real physical time anchors. If the voltage exceedance in the transformer area is a drift-type exceedance caused by a slow increase in load without a sudden change, resulting in a calculated maximum value... Not greater than the threshold of the rate of change of disturbance The data synchronization module 130 will trigger the fault tolerance mechanism, directly recording the time of exceeding the limit. Designated as physical time anchor point Establish physical time anchor points. This provides a reference time constraint for overcoming the asynchronous misalignment problem of high and low frequency sampling equipment.

[0057] See attached document Figure 6 , Figure 6 This is a schematic diagram illustrating the asynchronous data trapezoidal synchronization and confidence quantification principle according to an embodiment of the present invention. In the method provided by the present invention, step S40 specifically includes the following processing steps performed by the data synchronization module 130: S41, the data synchronization module 130 extracts the low-frequency power data of each user-side node in the event analysis window, and based on the physical time anchor point... Perform zero-order hold-for-reconstruction.

[0058] Due to the second sampling time step of smart meter 20 Much larger than the first sampling time step of the monitoring and acquisition terminal 10 The timing details of actual load surges occurring between two adjacent low-frequency power sampling periods are difficult to observe directly. To address the problem of fusing and reconstructing multi-source asynchronous data, this embodiment does not assume at the physical level that load changes for all users occur simultaneously. Instead, it uses physical time anchors... As a "macroscopic mathematical alignment benchmark boundary" for pseudo-high-frequency sequence reconstruction. Specifically, the system utilizes The step switching point of the zero-order hold is used for horizontal alignment. To eliminate the artificial synchronization artifacts and attribution biases that may arise from this forced mathematical alignment, this invention introduces a dual correction mechanism in the subsequent source tracing calculation: on the one hand, it uses the node power fluctuation range and static dead zone threshold for pre-filtering to remove nodes that are not in the zero-order hold. Stable nodes experiencing substantial abrupt changes in the vicinity are directly removed; on the other hand, a diagonal time confidence matrix based on the actual sampling time difference is constructed to assess distance. Asynchronous sampling points that are further away are penalized with algebraic weight decay. Through the synergy of the above mechanisms, the system can accurately filter out non-correlated disturbances and restore the causal relationship between the actual physical disturbances of each node and voltage over-limit events, provided that the mathematical alignment is maintained.

[0059] For any node i, set and To cross physical time anchor points Two adjacent real low-frequency sampling times, that is, satisfying The data synchronization module 130 will synchronize the interval... All power values ​​within the high-frequency time step are forcibly assigned as The actual measurement value at time At the same time, the interval All power values ​​within the high-frequency time step are forcibly assigned as The actual measurement value at time For events that do not cross physical time anchors within the event analysis window. For other low-frequency sampling intervals, the data synchronization module 130 uses a standard zero-order hold for horizontal forward assignment. Specifically, for any interval not containing... regular interval The data synchronization module 130 assigns the power values ​​corresponding to all high-frequency time steps within it to the equivalent interval start time. Measurement values Through this zero-order hold-and-reconstruction rule, the data synchronization module 130 transforms the original low-frequency discrete sequences of each node into sequences with the same time step size as the high-frequency voltage. pseudo-high frequency power sequence ,in This is the aligned high-frequency discrete-time index within the event analysis window. In a real distribution network communication environment, if a low-frequency communication is missed, causing the span of a certain interval to change... Even longer, the refactoring rule will still automatically use the most recent previous real measurement value to ensure that the computation queue does not experience a dead zone due to null values.

[0060] S42, the data synchronization module 130 calculates the absolute time difference between each reconstruction time step in the pseudo-high frequency power sequence and its assigned reference real sampling timestamp.

[0061] Pseudo-high frequency power sequence Data at non-real sampling points is generated by the zero-order hold principle, and its numerical values ​​reflect the reliability of the real physical state, which decreases with time deviation. Let... For the first For each reconstruction time step, the high-frequency physical timestamp is used. The data synchronization module 130 retrieves the corresponding real low-frequency sampling timestamp that provides the actual power assignment value for the current high-frequency time step from the historical low-frequency timestamp sequence, denoted as . Absolute time difference The calculation formula is: ; By forcibly anchoring the time base to the numerical source base, the logical fallacy of misaligned weighting due to short time intervals but the use of outdated sampling data can be avoided. By traversing all reconstructed time steps within the event analysis window, the data synchronization module 130 generates an absolute time difference sequence representing the degree of time deviation. Based on the aforementioned assignment and comparison rules, this absolute time difference value is distributed across... Within the range.

[0062] S43, the data synchronization module 130 generates a diagonal time confidence matrix based on the absolute time difference sequence.

[0063] The data synchronization module 130 introduces a time decay function to quantify the effectiveness of the reconstructed data. This is specifically designed for the reconstruction time steps in pseudo-high-frequency sequences. Its time confidence weight The calculation formula is as follows: ; In the formula, This is the preset time decay coefficient. The value is determined based on the communication cycle of the smart meter 20 and the average fluctuation frequency of the load in the distribution area, and in engineering implementation, it is usually taken in the range of 0.01 to 0.15. (When the timestamp...) The absolute time difference when it exactly corresponds to the actual low-frequency sampling time And confidence weight This indicates that the data at that point has the highest confidence level; when When far from the actual sampling point, It exhibits non-linear decay. The data synchronization module 130 extracts all confidence weights calculated from the M reconstruction time steps within the event analysis window. Arrange them in chronological order to the main diagonal of the matrix, constructing the main diagonal elements as follows: An M×M dimensional diagonal time confidence matrix in which all off-diagonal elements are 0. .

[0064] ; Diagonal time confidence matrix The generation process mathematically transforms the asynchronous delay characteristics of physical communication of sensing devices into an algebraic weight matrix of multi-dimensional spatial projection, providing clear boundary constraints for solving the reliable fusion of multi-source asynchronous data.

[0065] See attached document Figure 7 , Figure 7 This is a flowchart of topology state verification and dynamic rollback update according to an embodiment of the present invention. In the method provided by the present invention, step S50 specifically includes the following processing steps performed by the state verification module 140: S51, the status verification module 140 uses physical time anchor points Steady-state verification sub-windows are extracted from the inverse time axis as a reference.

[0066] In actual operation of distribution networks, short-term leading transient fluctuations often precede over-limit events. To avoid interference from these transient fluctuations on topology verification, the state verification module 140 sets an isolation time margin. The end time of the steady-state verification sub-window is defined as... Quarantine time margin The time span is typically set to 10 to 30 seconds. The steady-state check sub-window time span is set to... Its start time is defined as one To ensure that the verification window contains sufficient low-frequency communication sampling points to reflect the true steady-state electrical relationships, the time span... This is typically set as the second sampling time step of the smart meter 20. The time span is 2 to 4 times that of the time span, and in one specific embodiment of this application, the time span is specifically 30 to 60 minutes.

[0067] S52, the state verification module 140 calculates the topology drift residual index based on the data in the steady-state verification sub-window.

[0068] The state verification module 140 extracts the pseudo-high-frequency power sequence generated in step S40 within the steady-state verification sub-window, calculates the reconstructed power increment at each high-frequency time step within the sub-window compared to the start time of the sub-window, and combines the generated dimensions as follows: ×N reconstructed power increment matrix ,in This represents the total number of high-frequency sampling points within the steady-state verification sub-window. Simultaneously, real-time voltage sequence data from the monitoring and acquisition terminal 10 is extracted, and the actual voltage increment vector for the corresponding time step is calculated. Combined with the static empirical sensitivity vector obtained in step S20 The state verification module 140 calculates the topology drift residual index. The calculation formula is as follows: ; In the formula, The first in the actual voltage increment vector One element, To reconstruct the power increment matrix of the th The above formula, by comparing the difference between the actual observed voltage increment and the theoretically predicted voltage increment calculated based on historical experience, mathematically quantifies the root mean square deviation when predicting current voltage fluctuations using historical topological experience.

[0069] S53, the status verification module 140 compares the topology drift residual index with the preset tolerance threshold to determine the physical topology change.

[0070] Preset tolerance threshold The setting is based on 1.5 to 2.0 times the average residual value calculated under long-term historical normal operation of the transformer area. The status verification module 140 executes the condition judgment logic: if the topology drift residual index... The physical topology of the transformer area has not changed, and the original static empirical sensitivity vector remains unchanged. It remains valid; it can be directly recognized as the updated sensitivity vector. If the topology drift residual index If the physical topology of the transformer area has changed, such as line switching or branch reconstruction, it is determined that the original electrical distance attribute has become invalid, thus triggering a dynamic rollback update operation.

[0071] S54, after determining a topology change, the state verification module 140 extracts the most recent valid sample and re-executes the regression calculation to update the sensitivity vector.

[0072] After triggering the rollback update, the status verification module 140 checks the database along the physical time anchor point. Reverse backtracking, retrieve and extract distance The most recent valid historical shortwave band sample window that has not experienced voltage exceedances is used. Discrete voltage and power data pairs are extracted from this sample window. The time span of this valid historical shortwave band sample window must ensure that the number of valid low-frequency communication sampling point data pairs it contains is strictly greater than the total number of monitored nodes N. This satisfies the underlying matrix degree of freedom requirement for multiple linear regression and avoids singular covariance matrices caused by low-frequency power data being constant within the same sampling period. Subsequently, a multiple linear regression algorithm incorporating principal component analysis for dimensionality reduction and the introduction of a ridge regression penalty term is re-executed to resolve the partial derivative relationships of each node with respect to the head-end voltage, obtaining a dynamically recalculated sensitivity vector. The state verification module 140 will dynamically recalculate the sensitivity vector. Confirmed as the updated sensitivity vector This is for use in subsequent steps. In actual engineering implementation, if insufficient valid samples cannot be obtained during the backtracking retrieval process due to extreme circumstances such as widespread communication interruption, the state verification module 140 executes the underlying fallback logic, directly reading the default line impedance ledger parameters in the distribution management system of the transformer area, converting them into an initial empirical sensitivity vector, and outputting it. The specific algebraic transformation rules are as follows: The specific algebraic transformation rule is as follows: Extract the data recorded in the system ledger from the transformer head end to the corresponding user-side node. The physical circuit has a total resistance value With common reactance value According to the classic physical model of voltage drop in low-voltage distribution network lines, the line voltage deviation can be equivalent to... (in (Rated operating voltage of the transformer substation).

[0073] To obtain the static partial derivative of voltage with respect to power, the active power increment in the above voltage drop formula is calculated. and reactive power increment By taking partial derivatives, we can achieve dimensional transformation and physical equivalence from impedance parameters to sensitivity parameters: The physical sensitivity component of voltage to active power The physical sensitivity component of voltage to reactive power .

[0074] Combining the two with the reactive power conversion weighting factor Linear fusion is performed to obtain the one-dimensional static physical sensitivity scalar of the corresponding user-side node, and its fallback calculation formula is as follows: ; In the formula, This is the one-dimensional static physical sensitivity equivalent scalar value of the corresponding user-side node calculated based on the ledger parameters; This is an algebraic approximation of the active disturbance sensitivity, whose dimensions can be obtained from the derivation process. (and (Equivalent, meeting the sensitivity dimension requirements); This is an algebraic approximation of the reactive power disturbance sensitivity. The system uses the equivalent scalar value of the static physical sensitivity as the initial empirical sensitivity vector after the fallback update of this node.

[0075] See attached document Figure 8 , Figure 8 This is a flowchart of sensor dead zone filtering and invalid dimension isolation according to an embodiment of the present invention. In the method provided by the present invention, step S60 specifically includes the following processing steps performed by the filtering and sorting module 150: S61, the filtering and sorting module 150 extracts the calculation features of the fluctuation range of the pseudo high-frequency power sequence of each node.

[0076] For each user-side node under the jurisdiction of the transformer substation The filtering and sorting module 150 reads the pseudo high-frequency power sequence within the event analysis window. By traversing the discrete reconstructed data points in the sequence, the maximum and minimum power values ​​in the sequence are extracted respectively, and the difference between the two is calculated to obtain the fluctuation range of the node during the out-of-limit event. The calculation formula is as follows: ; In the formula, M represents the total number of aligned high-frequency discrete time steps within the event analysis window. This fluctuation range numerically reflects the maximum absolute disturbance amplitude of a single node within the event analysis window timescale. In practical engineering applications, to avoid interference from single-point data jumps caused by underlying communication errors in the range calculation, those skilled in the art can use the difference between the 95th percentile and the 5th percentile of the sequence to replace the absolute maximum and minimum values ​​for range calculation, thereby improving the algorithm's anti-interference capability.

[0077] S62, the filter sorting module 150 sets the specification for the static dead zone threshold of industrial sensors.

[0078] In the underlying measurement environment of the distribution network, due to the quantization truncation error of the smart meter analog-to-digital converter and the background electromagnetic white noise of low-voltage lines, the reconstructed power sequence collected and generated will still exhibit random fluctuations even if no load switching physical action occurs at the node. To shield the interference of this kind of background noise on subsequent source tracing calculations, the filtering and sorting module 150 establishes a static dead zone threshold. The threshold value is determined based on the measurement accuracy level and range baseline of the smart meter 20, or calibrated based on a fixed multiple of the historical root mean square value of noise under long-term steady-state operation of the distribution area. In a specific embodiment of this application, for conventional 220V low-voltage single-phase users, the static dead zone threshold is... The specific value range is set to 10 watts to 50 watts.

[0079] S63, the filtering and sorting module 150 compares the fluctuation range with the dead zone threshold and executes the invalid dimension isolation logic.

[0080] The filtering and sorting module 150 checks the fluctuation range of each node one by one. With static dead zone threshold Perform a condition check. If... If the node does not generate a substantial physical disturbance within the current voltage over-limit event analysis window, the minor change in its power data is considered random measurement noise within the sensing dead zone. The filtering and sorting module 150 marks this type of node as an invalid node and removes it from the global node set of the transformer area. Remove from the middle, and generate a result containing Dimensionally reduced feature subset of each effective participating node At the underlying matrix operation level, this elimination action directly reduces the column dimensions of the multidimensional data matrix, removes redundant vectors of constant level or near-white noise, and alleviates the risk of matrix multicollinearity caused by perturbationless variables.

[0081] Meanwhile, the filtering and sorting module 150 filters out the subset of dimensionality-reduced features of the effective participating nodes. The power sequences corresponding to the retained nodes are extracted synchronously from the original pseudo-high-frequency power matrix and recombined to generate a dimension-reduced effective power matrix; and the updated sensitivity vector is used to generate the effective power matrix. The sensitivity coefficients corresponding to the aforementioned retained nodes are extracted synchronously and assembled into a dimension-reduced sensitivity sub-vector, thereby ensuring that the algebraic dimension of each variable in the subsequent computation flow remains consistent.

[0082] During the above screening and comparison process, if the fluctuation range of the monitored nodes within the transformer area is not greater than the dead zone threshold, The situation, namely The filtering and sorting module 150 will trigger the dead-zone prevention logic, determining that the voltage over-limit is caused by the downward penetration of voltage fluctuations in the upstream main grid, rather than by load disturbances within the distribution area. At this time, the system generates an externalized over-limit alarm and skips the internal attribution calculations in steps S70 to S92, directly jumping to step S94 to encapsulate all nodes as zero-responsibility nodes and output them to the dispatch terminal. This avoids subsequent matrix operations from causing program abnormalities due to empty set inputs while maintaining the communication closed loop for system status monitoring.

[0083] See attached document Figure 9 , Figure 9 This is a schematic diagram of a basis vector sorting principle guided by a combination of electrical and wave characteristics according to an embodiment of the present invention. In the method provided by the present invention, step S70 specifically includes the following processing steps performed by the filtering sorting module 150: S71, the filtering and sorting module 150 calculates the combined electrical and wave impact factor by multiplying the updated sensitivity coefficient by the node fluctuation range.

[0084] In the distribution network disturbance tracing logic, a large power fluctuation amplitude at a single node usually does not independently cause voltage exceedance at the transformer substation's head end; its ultimate impact is constrained by the electrical distance between that node and the transformer substation. To measure the actual physical contribution of each node to voltage exceedance, the filtering and sorting module 150 extracts a dimensionality-reduced feature subset. Fluctuation range of each node Simultaneously extract the updated sensitivity coefficient corresponding to that node in the dimensionality-reduced sensitivity subvector. For valid nodes Its comprehensive influence factor The calculation formula is as follows: ; In the formula, the absolute value operation is used to eliminate the interference of the positive or negative sign of the sensitivity coefficient (representing the direction of active or reactive power injection / absorption) on the quantitative evaluation of the disturbance amplitude. At the physical principle level, due to... The dimension of voltage is the rate of change of power. The dimension of each factor is power, and the comprehensive influence factor obtained after multiplying them is... The physical dimensions are directly mapped to the voltage deviation amplitude, thus ensuring that the calculation results truly reflect the electrical amplification effect of the absolute scale of the physical disturbance.

[0085] S72, the filtering and sorting module 150 sorts the remaining valid nodes in descending order based on the comprehensive influence factor value.

[0086] The filtering and sorting module iterates through the dimensionality-reduced feature subsets 150 times. Obtain the comprehensive impact factor for all valid participating nodes within the scope. The set is sorted in descending order of its numerical values. For the specific execution of this descending sort, those skilled in the art can use conventional algorithms such as quicksort or mergesort; the internal operator calling rules are well-known in the field and will not be elaborated here. To address conflicts arising from multiple nodes having completely equal comprehensive influence factor values ​​in practical engineering, the sorting module 150 executes secondary anti-conflict logic, arranging the nodes sequentially according to the natural order of their numbers in the distribution management system to ensure the uniqueness of the sorting results. After sorting, the sorting module 150 generates and outputs a node index mapping sequence reflecting the priority of node disturbance contributions.

[0087] S73, the filtering and sorting module 150 establishes the pseudo high-frequency power sequence after descending order as the original basis vector sequence of the multidimensional algebraic space.

[0088] In subsequent multidimensional space algebraic calculations, multidimensional matrix projection algorithms such as Schmidt orthogonalization have strict computational order dependencies. Vectors ordered later in the sequence will have their collinear algebraic components removed during the projection process. If random sorting is performed, the feature information of key disturbance sources may be masked by the overlapping features of secondary disturbance sources. The filtering and sorting module 150 extracts the pseudo-high-frequency power sequences of the corresponding columns from the effective power matrix reconstructed in step S60 according to the node index mapping sequence generated in step S72. The node power sequence that is first in descending order according to the responsibility index is defined as the first original basis vector. And so on, the order will be... The node power sequence of the digit is defined as the digit of ... Original basis vectors In the algebraic definition, each primitive basis vector... All by The data is composed of discrete reconstructed data points, presented as an M×1 dimensional column vector structure. Through this mapping operation, the filtering and sorting module 150 constructs a column vector structure by concatenating the columns. ordered original basis vector sequence This mapping mechanism establishes the priority preservation of the dominant perturbation source from the underlying logic of algebraic projection, avoiding the computational dead zone of feature loss caused by mutual interference between dimensions during multidimensional dimensionality reduction projection.

[0089] See attached document Figure 10 , Figure 10 This is a schematic diagram illustrating the decoupling principle of regularized weighted orthogonal projection and responsibility components according to an embodiment of the present invention. In the method provided by the present invention, step S80 specifically includes the following processing steps performed by the orthogonal projection module 160: S81, the orthogonal projection module 160 uses the diagonal time confidence matrix to apply product constraints to the traditional vector inner product and reconstructs the joint weighted inner product operator.

[0090] To address the time confidence discrepancies arising from multi-source asynchronous data reconstruction, the orthogonal projection module 160 extracts the M×M dimensional diagonal time confidence matrix W generated in step S40. For any two column vectors x and y in the multidimensional algebraic space with the same M×1 dimension, the orthogonal projection module 160 sets the calculation rule for the joint weighted inner product as follows: ; In the formula, represents the transpose of column vector x. By embedding time decay weights in the underlying operations of the algebraic inner product, the numerical weight of high-frequency reconstructed data points that are farther from the true sampling point and have lower confidence levels will be reduced accordingly in the subsequent projection calculation, thereby suppressing the divergence of pseudo-data errors caused by zero-order hold-for-reconstruction within the mathematical calculation framework.

[0091] S82, the orthogonal projection module 160 performs iterative Schmitt orthogonalization based on the original basis vector sequence, extracts the orthogonal decoupled power basis vector, and introduces a positive real number regularization term to avoid the underlying algebra overflow dead zone.

[0092] The orthogonal projection module 160 extracts the ordered original basis vector sequence generated in step S70. According to the orthogonalization algorithm rules, let the first orthogonal decoupling power basis vector be... For the first original basis vectors (in The orthogonal projection module 160 calculates its vectors for each preceding orthogonal responsibility component. (in The projection along the direction is calculated and the vector difference is taken. The iterative calculation formula is as follows: ; In the formula, express and The joint weighted inner product, express The scalar value generated by dividing the two products by their joint weighted inner product is the weighted projection coefficient. This is a pre-defined positive real number regularization term. In the engineering implementation of multidimensional matrix iterative projection, as the orthogonality depth increases, more components are stripped from the later-ordered vectors, causing their magnitude to approach zero; the diagonal time confidence matrix is ​​superimposed. The numerical multiplication contraction effect of a large number of decaying decimal weights is used to calculate the denominator. It is prone to underflow below the computer's floating-point precision limit, causing a program dead zone that triggers a division-by-zero exception. This can be addressed by introducing a value range limited to 10 in the denominator. -8 Up to 10 -5 Positive real number regularization term This provides a hard lower bound for the division operator, ensuring the continuity of the entire matrix algebra solution process without affecting the normal orthogonal projection of large disturbance sources.

[0093] To further avoid the complete collinearity dead zone in algebraic projection, the orthogonal projection module 160 calculates the orthogonal responsibility component vector for the current iteration. Then, its own joint weighted inner product is calculated simultaneously. If the inner product value is less than the preset collinearity rejection threshold (which can be set to 10 in a specific embodiment), then... -6 ), then determine the first The physical perturbation represented by the original basis vectors has been completely covered by the preceding nodes and lacks its own independent perturbation contribution. At this time, the orthogonal projection module 160 will... Force a reset to a zero vector to prevent the tiny truncation noise generated by the underlying floating-point operations from being passed to subsequent steps.

[0094] S83, the orthogonal projection module 160 sequentially strips the concurrent collinear coupling characteristics and combines them with the actual target voltage fluctuation sequence to generate a single-node over-limit responsibility sequence.

[0095] During distribution network overload events, the power loads of multiple users may exhibit concurrent fluctuations, leading to algebraic collinearity among the generated original basis vector sequences. The orthogonal projection module 160 uses the aforementioned iterative subtraction operation to remove the overlapping portions of the pseudo-high-frequency power sequence with the waveforms of the leading nodes in the ordering. After a complete iterative traversal, the orthogonal projection module 160 transforms the original basis vector sequences into independent, decoupled power sequences that satisfy a joint weighted orthogonality relationship. .

[0096] It should be noted that the sequence obtained by the above orthogonalization step While it achieves algebraic independence between the power disturbance characteristics of each node, it has not yet established a causal mapping between these characteristics and actual node voltage exceedances, and therefore cannot be directly used as the final basis for liability assessment. To completely complete the tracing logic chain from "power disturbance" to "voltage exceedance," the orthogonal projection module 160 further extracts the incremental vector of the target event voltage fluctuation when the actual exceedance occurs in the transformer area. As a global target vector, calculate the target vector in each independent decoupling power basis vector. The target mapping projection is applied to the target. This generates the final sequence representing the single-node absolute voltage over-limit liability sequence. The calculation formula is as follows: ; In the above formula, the terms within the parentheses This is the target projection coefficient. Physically, this coefficient represents the true voltage fluctuation sequence after removing collinear coupling interference from other nodes. From the node Independent power fluctuations The resulting proportional gain is essentially equivalent to the node's "dynamic decoupling voltage sensitivity" (dimension 1) under the current over-limit operating conditions. or ).

[0097] Multiply the dynamic decoupling sensitivity coefficient back into the independent decoupling power sequence of that node. The system successfully maps and quantizes the pure power basis in algebraic space across the physical domain into a sequence of physical responsibilities (in terms of units) that substantially boosts or suppresses over- and under-voltage phenomena at each time step for that specific node. This target projection mechanism ensures that the final tracing result is a true restoration of the actual voltage over-limit phenomenon, rather than merely a comparison of the power waveform shape.

[0098] See attached document Figure 11 , Figure 11 This is a flowchart of the calculation and closed-loop output of the over-limit liability attribution index according to an embodiment of the present invention. In the method provided by the present invention, step S90 specifically includes the following processing steps performed by the liability output module 170: S91, the responsibility output module 170 extracts the root mean square effective value of the single-node over-limit responsibility sequence generated by the independent orthogonal responsibility component vector mapping of each node.

[0099] For the single-node over-limit liability sequence output in step S80 Each column vector Collinear components from concurrent fluctuations with other nodes have been removed, mathematically representing the unique net perturbation characteristics of this node and completing the dynamic voltage deviation mapping across the physical domain. The responsible output module 170 traverses the sequence, calculating the root mean square effective value of each vector. The calculation formula is as follows: ; In the formula, This represents the total number of high-frequency discrete time steps after alignment within the event analysis window. Independent orthogonal responsibility component vectors The first in A discrete reconstructed data point. This root mean square effective value physically quantifies the independent net fluctuation energy injected into the power grid by a single node after eliminating the effects of cross-coupling.

[0100] S92, the responsibility output module 170 multiplies the root mean square effective value by the corresponding effective sensitivity coefficient to obtain the quantitative responsibility index.

[0101] The net fluctuation energy of a node has a correlation with the actual voltage impact in a distribution network and is related to the network's physical topology. The responsibility output module 170 extracts the updated sensitivity coefficients from the reduced-dimensional sensitivity sub-vector. For each valid node, calculate the attribution index for liability for exceeding limits. The calculation formula is as follows: ; In the formula, the absolute value operation is used to eliminate the difference in the sign of the coefficients caused by the direction of active or reactive power flow, and to uniformly quantify the contribution of the absolute deviation of the disturbance amplitude to the voltage at the head end. After generating the basic responsibility index, in order to facilitate intuitive evaluation by the dispatching system, the responsibility output module 170 simultaneously calculates the normalized responsibility weight ratio of each effective node. The calculation rule is to divide the responsibility attribution index of a single node by the sum of the responsibility attribution indices of all effective nodes. In this algebraic transformation process, the responsibility output module 170 introduces the underlying division-by-zero verification logic: if the sum of the responsibility attribution indices of all nodes is calculated to be zero (or less than the lower limit of the computer's floating-point precision), it is determined that there is no effective responsible entity within the distribution area, the system directly interrupts the normalization calculation, and attributes the voltage over-limit responsibility entirely to the over-limit penetration of the upper-level main grid, thereby avoiding the abnormal dead zone of the system caused by the underflow of the calculation denominator.

[0102] S93, the responsibility output module 170 generates a responsibility rating sequence based on the responsibility index values ​​arranged in descending order.

[0103] The responsibility output module 170 aggregates the out-of-limit responsibility attribution index of all valid nodes. The nodes are sorted in descending order of their numerical values. Based on this descending order, the responsibility output module 170 uses a preset responsibility allocation threshold to classify and label the nodes. In specific engineering implementations, this responsibility allocation threshold is typically set to a value between 70% and 90% based on the Pareto principle (80 / 20 rule) for distribution network load management. The responsibility output module 170 accumulates the normalized responsibility weight ratio of each node from top to bottom, using a closed-loop accumulation logic that includes the first node that crosses the limit. Specifically, when accumulating to the k-th node, if the accumulated ratio reaches or exceeds the set responsibility allocation threshold for the first time, the system marks the set of nodes from the 1st to the k-th node in the sequence as the primary over-limit responsibility node, and marks the remaining valid nodes in the sequence as secondary over-limit responsibility nodes. Through this classification operation, the responsibility output module 170 generates a structured responsibility classification sequence, providing data for subsequent peak shaving or flexible control of the distribution network.

[0104] S94, the responsibility output module 170 packages and outputs the responsibility classification sequence and the zero responsibility node status to the scheduling terminal.

[0105] To achieve a closed-loop business control for distribution network status monitoring, the responsibility output module 170 uses the generated responsibility classification sequence as the core data payload. It simultaneously aggregates invalid nodes removed in step S60 due to failing to cross the static dead zone threshold, and nodes forcibly reset to zero vectors in step S80 due to meeting the complete collinearity condition. These two types of nodes are uniformly registered as zero-responsibility nodes in the data table. The responsibility output module 170 combines the core data payload with the zero-responsibility node status data to maintain the dimensional integrity of the global node status matrix, avoiding system status open-loop due to the filtering of some nodes. After combination, it is encapsulated into a data packet in a standard industrial communication protocol format and published. For the specific encapsulation of the communication protocol and the Ethernet transmission interaction to the dispatch control terminal, those skilled in the art can use conventional IoT protocol standards such as IEC61850 or MQTT. The internal message parsing and handshake rules are well-known technologies in the field and will not be elaborated here. After receiving the data packet, the dispatch terminal can issue corresponding load limiting instructions according to the responsibility classification sequence, completing the multi-source asynchronous data-driven voltage over-limit proactive management process.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for tracing voltage exceedances in transformer substations based on multi-source data, comprising: Acquire real-time voltage sequence data collected by the monitoring and acquisition terminal at the transformer head end, as well as power data collected by the smart meters at each user-side node; Based on the real-time voltage sequence data and power data under historical operating conditions, a static empirical sensitivity vector is constructed; When a voltage over-limit event is detected, the extreme points of the first time derivative of the real-time voltage sequence data are extracted and established as physical time anchor points. Based on the physical time anchor point, the power data is reconstructed by equivalent assignment using the zero-order hold rule to generate a pseudo-high-frequency power sequence; the sampling time difference before and after reconstruction is calculated to obtain the time confidence weight, and the time confidence weight is arranged on the main diagonal to generate a diagonal time confidence matrix; Using the physical time anchor point as a reference, a time window is extracted. Based on the pseudo-high-frequency power sequence, the real-time voltage sequence data, and the static empirical sensitivity vector, the topology drift residual is calculated to perform topology state verification. The updated sensitivity vector is determined according to the verification result. Specifically: if the topology drift residual is not greater than a preset tolerance threshold, it is determined that the physical topology has not changed, and the static empirical sensitivity vector is confirmed as the updated sensitivity vector; if the topology drift residual is greater than the preset tolerance threshold, it is determined that the physical topology has changed, and historical short-band sample data that is closest to the physical time anchor point and has not experienced voltage over-limit is retrieved. Regression calculation is re-executed to obtain a dynamically recalculated sensitivity vector, and the dynamically recalculated sensitivity vector is confirmed as the updated sensitivity vector. Effective nodes are selected from the pseudo-high frequency power sequence based on the power fluctuation range and static dead zone threshold of each user-side node. A comprehensive influence factor is calculated in combination with the updated sensitivity vector. The effective nodes are then sorted in descending order based on the comprehensive influence factor to construct the original basis vector sequence. The diagonal time confidence matrix is ​​used as the weight matrix to define a joint weighted inner product operator; based on the joint weighted inner product operator, an iterative Schmitt orthogonalization with positive real number regularization terms is performed on the original basis vector sequence, and the orthogonal responsibility component vector of each effective node is extracted by sequentially stripping the concurrent collinearity components; Calculate the root mean square effective value of each of the orthogonal responsibility component vectors, and multiply the root mean square effective value of each effective node by the absolute value of the coefficient corresponding to that node in the updated sensitivity vector to obtain the responsibility index of each effective node; calculate the normalized responsibility weight ratio of each effective node, and sort the effective nodes in descending order according to the value of the responsibility index. The normalized responsibility weight ratios of each effective node are accumulated from top to bottom in descending order. The set of cumulative nodes that first reach or exceed the preset responsibility division threshold are designated as primary over-limit responsibility nodes, and the remaining effective nodes are designated as secondary over-limit responsibility nodes. This generates a responsibility classification sequence that reflects the primary and secondary relationship of node responsibilities. The responsibility classification sequence is output to the scheduling terminal so that the scheduling terminal can perform corresponding over-limit tracing and load control based on the responsibility classification sequence.

2. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 1, characterized in that, The construction of the static empirical sensitivity vector specifically includes: Extract the time-aligned voltage fluctuation increment vector and the original power increment matrix within the historical steady-state time window; Principal component analysis is performed on the original power increment matrix to reduce its rank and denoise it, generating a denoised reconstructed power increment matrix; Based on the noise reduction and reconstruction power increment matrix and the voltage fluctuation increment vector, a multiple linear regression algorithm with ridge regression bias constant is used to calculate the static partial derivative values ​​of each user-side node relative to the head-end voltage. The static partial derivative values ​​are used as the static empirical sensitivity coefficients of each user-side node, and the static empirical sensitivity coefficients of all users are concatenated column-wise to construct the static empirical sensitivity vector with dimension N×1, where N is the total number of nodes.

3. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 1, characterized in that, The establishment of physical time anchor points specifically includes: When the real-time voltage sequence data is detected to have exceeded the preset operating boundary, the timestamp of the initial sampling point that exceeded the boundary is recorded as the over-limit trigger time; the event analysis window is extracted based on the over-limit trigger time, and the first-order time derivative sequence of the real-time voltage sequence data within the event analysis window is calculated using the finite difference algorithm; The first-order time derivative sequence is traversed to retrieve the maximum value. If the retrieved maximum value is greater than a preset perturbation rate threshold, the discrete sampling timestamp corresponding to the maximum value is established as the physical time anchor point. If the maximum value is not greater than the perturbation rate threshold, a fault tolerance mechanism is triggered, and the over-limit trigger time is designated as the physical time anchor point.

4. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 1, characterized in that, The generation of the pseudo-high-frequency power sequence specifically includes: A discrete time step with the same time step size as the real-time voltage sequence data is defined as a high-frequency time step. For adjacent real low-frequency sampling time intervals that cross the physical time anchor point, the power values ​​at the high-frequency time step are forcibly assigned the actual measured values ​​at the start and end times of the adjacent real low-frequency sampling time intervals, with the physical time anchor point as the boundary. For intervals that do not cross the physical time anchor point, a zero-order hold is used for horizontal forward assignment. For intervals that do not cross the physical time anchor point, a zero-order hold is used for horizontal forward assignment to generate the pseudo-high-frequency power sequence. Calculate the absolute time difference between each reconstruction time step and the actual sampling timestamp of the assignment benchmark in the pseudo high-frequency power sequence, and calculate the time confidence weight of each reconstruction time step based on the absolute time difference using a preset time decay function; All the calculated time confidence weights are arranged in chronological order on the main diagonal of the matrix to construct the diagonal time confidence matrix.

5. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 1, characterized in that, The topology state verification specifically includes: Using the physical time anchor point as a reference, a steady-state verification sub-window is extracted in reverse. Calculate the reconstruction power increment at each high-frequency time step within the steady-state verification sub-window compared to the start time of the sub-window, combine them to generate the verification reconstruction power increment matrix, and combine the actual voltage increment vector and the static empirical sensitivity vector to calculate the topology drift residual index. If the topology drift residual index is not greater than the preset tolerance threshold, it is determined that the physical topology has not changed. The valid historical short-band sample window that is closest to the physical time anchor point and has not experienced voltage over-limit is retrieved, and regression calculation is re-executed to obtain the dynamically recalculated sensitivity vector, which is then confirmed as the updated sensitivity vector.

6. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 5, characterized in that, If the topology drift residual is greater than the preset tolerance threshold, regression calculation is re-executed based on historical valid data to obtain a dynamically recalculated sensitivity vector, specifically as follows: When a physical topology change is triggered and the database cannot retrieve a sufficient number of valid low-frequency communication sampling point data pairs, the common resistance value of the physical lines from the transformer head to each user-side node recorded in the system ledger is extracted; the common resistance value of the physical lines is used as an algebraic approximation of the active power disturbance sensitivity, and combined to form a ledger-calculated sensitivity vector, which is then confirmed as the updated sensitivity vector for output.

7. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 1, characterized in that, The construction of the original basis vector sequence specifically includes: The difference between the maximum and minimum power values ​​of each user-side node in the pseudo-high-frequency power sequence is calculated as the power fluctuation range; The power fluctuation range is compared with the preset static dead zone threshold. If the power fluctuation range is not greater than the static dead zone threshold, the corresponding node is removed. If the power fluctuation range of all monitored nodes is not greater than the static dead zone threshold, the internal attribution operation is skipped, and all monitored nodes are directly packaged as zero-responsibility nodes and output to the scheduling terminal. If there are valid nodes whose power fluctuation range is greater than the static dead zone threshold, extract the absolute value of the coefficient corresponding to each valid node in the updated sensitivity vector, multiply it by the corresponding power fluctuation range, and calculate the comprehensive influence factor; sort each valid node in descending order according to the value of the comprehensive influence factor, extract the corresponding pseudo high-frequency power column vector in the sorting order, and concatenate the columns to construct the original basis vector sequence.

8. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 1, characterized in that, The extraction of orthogonal responsibility component vectors for each effective node specifically includes: For any two column vectors with the same dimension in the original basis vector sequence, the intermediate product constraint is applied to the vector inner product using the diagonal time confidence matrix, and the joint weighted inner product operator is defined. Extract the first basis vector in the original basis vector sequence as the first orthogonal responsibility component vector; For the original basis vector at position k in the original basis vector sequence, calculate its joint weighted inner product projection coefficients in the directions of each preceding orthogonal responsibility component vector. The denominator of the projection coefficients is configured as the sum of the weighted inner product of the corresponding preceding orthogonal responsibility component vector and a preset positive real number regularization term to avoid division by zero. Based on the joint weighted inner product projection coefficients, the corresponding preceding projection components are obtained. Based on the vector difference between the original basis vector of the kth position and the sum of all the preceding projection components, the kth orthogonal responsibility component vector stripped of the concurrent collinearity component is generated. By iterating through the above process, the concurrent collinear coupling features in the original basis vector sequence are stripped away in turn, and the mutually independent orthogonal responsibility component vectors are extracted.

9. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 8, characterized in that, The extraction of mutually independent orthogonal responsibility component vectors specifically includes: Calculate the joint weighted inner product of the generated k-th orthogonal responsibility component vector itself; If the calculated joint weighted inner product is less than the preset collinearity elimination threshold, it is determined that the physical disturbance represented by the original basis vector of the k-th position has been covered by the preceding node, and the k-th orthogonal responsibility component vector is reset to the zero vector.

10. The method for tracing voltage exceedances in transformer substations based on multi-source data according to claim 1, characterized in that, The generation of the responsibility grading sequence specifically includes: The responsibility index of each valid node is normalized, and the corresponding normalized responsibility weight ratio is extracted; and the valid nodes are sorted in descending order according to the value of the responsibility index. According to the order of arrangement, the normalized responsibility weight ratio of each effective node is iteratively accumulated. When the cumulative ratio obtained by the iteration falls into the target interval formed by the preset responsibility division threshold and the maximum normalization limit value for the first time, the cumulative node set of the current iteration is extracted and designated as the main over-limit responsibility node, and the remaining effective nodes are designated as the secondary over-limit responsibility nodes, and the responsibility classification sequence is generated by combining them. The responsibility rating sequence is concatenated with the zero-responsibility node status data corresponding to the removed nodes and the reset nodes, and then encapsulated into a data packet in the standard industrial communication protocol format and published to the scheduling terminal.

Images