A low-voltage transformer area household identification method and system

Abstract

This invention discloses a method and system for identifying cross-connection issues in low-voltage distribution areas. It acquires voltage and current time-series data of users in the low-voltage distribution area, cleans the data using a mask interpolation strategy, and adaptively extracts voltage drop and current surge features based on the median absolute deviation algorithm. It constructs a voltage trend similarity model based on local regularization kernels and a mutation co-occurrence similarity model based on event level weights, employing a dual-track strategy combining daily rigorous assessment and feature revival to generate suspected voltage correlation pairs. A density-based clustering algorithm is used to generate an initial population, and a voltage similarity-weighted centroid screening mechanism is introduced to remove loose outliers in the feature space. A pairwise cross-validation model is constructed based on electrical causality, and the population is topologically refined using relative voltage response criteria and a maximum fully connected clique extraction algorithm to determine the final cross-connection population. This invention significantly improves the accuracy and robustness of cross-connection identification.

Description

Technical Field

[0001] This invention belongs to the field of power system distribution network operation and maintenance technology, specifically relating to a method and system for identifying cross-connection between low-voltage distribution areas based on voltage spatiotemporal feature fusion and electrical causality verification. Background Technology

[0002] As the "last mile" of the power grid, the accuracy of the topological relationship of low-voltage distribution areas directly affects the implementation of services such as line loss management, fault repair, and three-phase load balancing. In actual operation, due to construction modifications, unauthorized connections, or delayed updates to records, discrepancies often arise between the physical connection locations of users and the records. In particular, some users illegally connect to other users' lines to obtain power, creating "cross-connection" relationships.

[0003] Currently, existing technologies for cross-connection identification are mainly divided into two categories: one is the hardware-based pulse injection method, which detects cross-connection by installing a signal generator on the user side. This method is accurate but costly and difficult to promote in large-scale distribution areas; the other is the voltage similarity analysis method based on data mining, which determines the correlation by calculating the Pearson correlation coefficient of the user voltage curve.

[0004] However, traditional voltage similarity-based methods face many challenges in practical applications: (1) Spatiotemporal deviation interference: There may be clock drift in the data acquisition of smart meters, which causes two originally overlapping curves to be misaligned on the time axis, resulting in huge errors in the traditional point-to-point Euclidean distance calculation; (2) Insufficient numerical sensitivity: The Pearson coefficient only focuses on the linear trend and ignores the absolute difference in voltage amplitude. In the same transformer area, the voltage trends of non-interconnected neighboring users are often highly similar due to the influence of bus voltage fluctuations, which can easily lead to false positives. (3) Lack of physical causal verification: Most existing methods only analyze the "correlation" of voltage as a single dimension, ignoring that the essence of the serial relationship is physical connection, and failing to make full use of the core electrical physical law that "a sudden change in current (cause) will inevitably lead to a sudden drop in voltage (effect)" to provide evidence; (4) Long chain effect misjudgment: Commonly used clustering algorithms (such as DBSCAN) are prone to transitive misjudgment, resulting in a high error rate in cross-household identification in practical applications.

[0005] Therefore, there is an urgent need for a low-voltage distribution area cross-connection identification method that can integrate voltage spatiotemporal characteristics, has noise resistance, and can perform self-verification using electrical causality. Summary of the Invention

[0006] Objective: To address the problems and shortcomings of existing technologies, this invention provides a method and system for identifying cross-connections in low-voltage distribution areas based on voltage spatiotemporal feature fusion and electrical causality verification. This aims to solve the problems of low accuracy and high false alarm rate in cross-connection identification caused by meter clock asynchrony, background noise interference, and the lack of physical verification mechanisms in existing technologies. This invention achieves accurate identification and topology reconstruction of cross-connections in low-voltage distribution areas by constructing a two-dimensional voltage similarity model and an electrical causality verification closed loop, without requiring additional hardware investment.

[0007] Technical Solution: A method for identifying cross-connections in low-voltage distribution areas based on voltage spatiotemporal feature fusion and electrical causality verification, comprising the following steps: Step 1: Obtain the voltage and current timing data of all users in the low-voltage distribution area to be identified, perform timing alignment and cleaning processing on the data based on interpolation tolerance, and generate a standardized voltage data matrix and current data matrix with reliability status markings; Step 2: Introduce the principle of robust statistics, adaptively extract feature sequences, model the background noise of each user's voltage and current data, construct an adaptive dynamic threshold that adapts to individual fluctuation characteristics, and extract the voltage mutation event feature sequence and current surge event feature sequence that characterize key power quality disturbances from the continuous time series data. Step 3: Construct a two-dimensional voltage similarity evaluation model. Based on the voltage time series data, calculate a trend similarity index based on a local regularization kernel that can tolerate sampling time deviation. Based on the voltage mutation event feature sequence, calculate a mutation co-occurrence similarity index based on event level weight that weakens the influence of amplitude. Then, use a dynamic assessment strategy that includes a revival mechanism to screen out voltage suspected related user pairs with highly coupled voltage features. Step 4: Construct a user association graph, use a density-based clustering algorithm to perform preliminary clustering on the voltage-related suspected user pairs, and perform centroid iterative adsorption and splitting on the feature space geometric distance based on the introduced mutation co-occurrence similarity weight to generate a geometrically compact preliminary user group; Step 5: Introduce a causal verification mechanism at the electrical and physical level. Use the current surge event feature sequence as the excitation signal and the voltage time series data as the response signal to construct a ternary hybrid criterion model that includes strong response pass-through, absolute threshold fallback, and relative proportion pass-through. Perform pairwise cross-causal inference on users within the group to verify the authenticity of the physical connection. Step 6: Construct a strong verification relationship graph based on the causal inference results, and use the maximum clique extraction algorithm to iteratively analyze the fully connected subgraph structure in the graph to ensure that any two users within the output group satisfy electrical connection consistency. The fully connected subgraph is then determined as the final physical user group.

[0008] Furthermore, the steps for acquiring voltage and current time series data of all users within the low-voltage distribution area to be identified, performing time series alignment and cleaning processing based on interpolation tolerance on the data, and generating standardized voltage and current data matrices with reliability status markers are as follows: Step S11: Set the standard sampling interval according to the acquisition frequency of the measurement equipment in the transformer area. and daily standard sampling points Based on the timestamps of the user's original data, the user's original data is mapped to a standard timeline, generating an original sequence containing missing values. Simultaneously, a binary reliability mask sequence of the same length as the original sequence is initialized in this step to accurately track the reliability of each data point. All data points are initialized to an "unreliable state" by default or based on the existence of the original data. The maintenance logic for this reliability mask sequence is as follows: Let the original time series data be: ; Corresponding reliability mask sequence definition: ; in: ; Step S12: Traverse the original sequence, identify consecutive missing data segments, and calculate the length of each missing segment; for each missing segment, record its start time. End time and the length of that segment This step aims to distinguish between "intermittent packet loss" and "systematic failure." Intermittent packet loss usually involves only 1-3 missing data points, which can be reasonably restored through interpolation; however, systematic failures (such as power outages or module damage) may result in several hours of data gaps, in which case the physical meaning of the data has been lost, and forced interpolation will only introduce huge errors.

[0009] Step S13: Introduce interpolation tolerance threshold As a boundary for data repair; when a continuous missing interval is detected in the time series data, and the length of the continuous missing interval is... Not exceeding the preset threshold When, i.e., consecutive missing lengths are detected If the missing data is considered occasional, it will be filled using a linear interpolation algorithm. The calculation formula is as follows: ; In the formula, The moment to be interpolated Data, and These are the known valid observations before and after the missing segment. After calculation, the reliability mask sequence is updated, and the data points within the filled time period are marked as "reliable states" to maintain the continuity of subsequent feature extraction. The logical basis of this operation is that within a very short time window (such as 45 minutes), the voltage and current changes of the low-voltage power grid are usually smooth, and linear interpolation can restore the real physical process with a very high confidence. Therefore, these interpolation points are fully qualified to participate in subsequent feature extraction.

[0010] Step S14: When a continuous missing interval is detected in the time series data, and the length of the continuous missing interval is... Greater than the preset threshold When a continuous missing length is detected Although linear interpolation is performed to maintain the mathematical integrity of the sequence, all interpolation points within the time period are forcibly maintained or marked as "unreliable" in the reliability mask sequence. Before performing difference calculation or similarity calculation, the reliability mask state corresponding to each data point participating in the difference calculation or similarity calculation is detected. When any data point participating in the calculation is detected as unreliable, the difference calculation or similarity calculation at the corresponding time point is blocked, or the weight of the corresponding time point in the difference calculation result or similarity calculation result is reset to zero, so as to suppress the influence of data introduced by interpolation of consecutive missing intervals on the calculation result.

[0011] Furthermore, the method introduces robust statistical principles, adaptively extracts feature sequences, models background noise for each user's voltage and current data, constructs an adaptive dynamic threshold that adapts to individual fluctuation characteristics, and discretizes and extracts voltage surge event feature sequences and current surge event feature sequences representing key power quality disturbances from continuous time-series data. A dynamic threshold segmentation technique capable of adaptively distinguishing between background fluctuations and significant disturbances is employed. The specific steps are as follows: Step S21: First-order difference sequence calculation. First, the cleaned user voltage time series data... Perform discrete first-order difference operations to generate difference sequences. At any given time difference Calculate using the following formula: This indicator Directly represents time The instantaneous voltage change relative to the previous moment. During this process, the reliability mask sequence generated in step S1 is read in real time; if the moment... or The corresponding mask is marked as "unreliable state", then at time The difference values ​​will be set to invalid values ​​(NaN) or zero to eliminate the interference of missing data on feature extraction.

[0012] Step S22: To overcome the shortcomings of traditional standard deviation (SSD) indicators, which are susceptible to extreme outliers (such as power grid failures or rare sharp drops) leading to inflated valuations and a threshold masking effect, this step utilizes the MAD algorithm (Median Absolute Deviation algorithm) to analyze the difference sequence. A robust estimation of the background noise level is performed. The specific calculation logic is as follows: First, the median of the difference sequence is calculated. This value reflects the central trend of voltage variation. Next, the absolute deviation sequence of each differential point from this central trend is calculated, and the median of this sequence is taken to obtain the original MAD value. In the above formula, Represents voltage differential sequence The median, representing the background noise level, is used to calculate the voltage differential point ith. The absolute difference from the background noise is obtained by taking the median of the absolute difference sequence. This can represent the inherent dispersion of voltage data. Finally, to ensure that the MAD value is consistent with the standard deviation in terms of dimensions for engineering applications, a normal distribution consistency scaling factor is introduced. The MAD value is calibrated. For a standard normal distribution, Values The final estimate of the standard deviation of background noise. The calculation is as follows: From this obtained It can accurately characterize the inherent dispersion of the user's voltage data and is not affected by occasional drastic fluctuations.

[0013] Step S23: Construction of the dynamic voltage surge threshold. Based on the above background noise standard deviation estimate, a dynamic voltage surge threshold that adapts to the user load characteristics is constructed. This threshold is defined as the boundary of the noise standard deviation that deviates from the central trend by a certain multiple. In the formula, This is a preset voltage significance coefficient used to adjust the algorithm's sensitivity to voltage drop events.

[0014] Step S24: Filtering and generating the event sequence. Traverse the first-order difference sequence. Check each moment one by one The effective difference value. If the condition is satisfied. This means that a negative drop exceeding the dynamic threshold has occurred, and at any given time... If the data mask is "reliable state", then the extraction time is... The voltage fluctuations and their corresponding drop values ​​constitute the characteristic sequence of voltage mutation events for this user.

[0015] Step S25: Event Hierarchy Labeling. To quantify the information value of different events in subsequent similarity calculations, this step labels them according to the magnitude of mutation. Standard deviation of background noise The extracted events are categorized based on their multiple relationships. Specifically, a higher-order multiple threshold is introduced. (For example ): like If the event is marked as a "significant abrupt change event," such events typically correspond to the start-up and shutdown of high-power loads, have extremely high signal-to-noise ratios, and possess core identification value; if If a significant event is identified, it is labeled as a "common mutation event". Through this hierarchical strategy, the subsequent two-dimensional similarity model can assign higher weights to significant events, thereby significantly improving the robustness of the algorithm in complex noisy environments.

[0016] Furthermore, the construction of the two-dimensional voltage similarity evaluation model, which calculates a trend similarity index based on a local regularization kernel that can tolerate sampling time deviations based on the voltage time series data, involves the following specific steps: Step S31: For user A and user B to be compared, first obtain their cleaned voltage sequences. and To address the clock skew problem, this invention introduces the concept of Dynamic Time Warping (DTW), which is simplified into a "local warping" strategy to reduce computational complexity. For any given time... Define a coverage area with several sampling points before and after (e.g.) Each point corresponds to Local time window (minutes) Within this window, at user A's time... Based on this, user B searched in Find the point within the range that is closest to the voltage value of user A at time t, and calculate the minimum voltage deviation between them. : In the formula, This represents the voltage value of user A at time t. This represents the voltage value of user B at time t+k. This step effectively eliminates the rigid error caused by clock asynchrony by finding the "best match" in a local range, making the evaluation results more focused on the shape consistency of the voltage curve itself.

[0017] Step S32: At the time of obtaining Minimum voltage deviation The similarity score is then converted into a normalized similarity score. This invention utilizes a Gaussian kernel function for nonlinear mapping. The Gaussian kernel function has a "soft threshold" characteristic: when the voltage deviation is small, the score is close to 1; as the deviation increases, the score smoothly decays until the deviation exceeds a certain level, at which point the score rapidly approaches 0. The specific instantaneous similarity score... The calculation formula is: in, This is a preset voltage tolerance parameter (e.g., 0.3V). This parameter determines the "bandwidth" of the Gaussian function, which in turn determines the rate at which the similarity score decays with increasing voltage deviation. By adjusting... It allows for flexible control over how "picky" the algorithm is about voltage value differences.

[0018] Step S33: Calculate the instantaneous similarity score sequence for all times throughout the day. Next, to eliminate the interference of individual extreme values, this step performs a truncated mean aggregation operation. First, for... Sort the sequence in ascending order, then remove the element with the lowest value. (For example, 10%) of the data points, only for the remaining The arithmetic mean of the data points is calculated as the final trend similarity. .

[0019] Furthermore, the construction of the two-dimensional voltage similarity evaluation model, which calculates a mutation co-occurrence similarity index based on event level weights with reduced amplitude influence based on the characteristic sequence of voltage mutation events, involves the following specific steps: Step S34: Based on the voltage mutation event feature sequence adaptively extracted in Step S2, obtain the voltage mutation event time sets for User A and User B throughout the entire time period. and Construct the union of the two sets. This union encompasses all moments when voltage disturbances occur on either side, forming the time reference for subsequent similarity calculations.

[0020] Step S35: For the union Every moment in We no longer focus on its specific voltage differential amplitude. Instead, it assigns a fixed base weight based on the event level (“significant” or “normal”) marked in step S25. This weight is based on user A and user B's... Time and Voltage differential amplitude at time The calculation yields: The specific weight allocation logic is as follows: If user A or user B at time... The mutation that occurs is marked as a "significant mutation event" (i.e., a sharp drop, usually corresponding to a strong disturbance source), and a higher weight value is assigned to that moment. (Value is 2); This reflects that significant events have a higher "veto power" and "evidence effect" in determining cross-party transactions. If neither party experiences a significant change, and only one or both parties experience a "common change event" (i.e., a small drop), then a lower weight value is assigned to that moment. (The value is 1).

[0021] Step S36: Define the matching indicator function Used for strict time determination Whether it constitutes a valid "co-occurrence" event. The judgment logic requires two conditions to be met simultaneously: Condition 1: Event coexistence. That is, at time... Must exist simultaneously in the set and middle( This indicates that both users were disturbed at that moment. Condition 2: Consistent direction. That is, the first-order differential signs of the voltages of the two users at that moment must be the same. For cross-user identification scenarios, the focus is on the same-direction voltage drop, i.e. and If one increases and the other decreases, it indicates that the disturbance sources experienced by the two users are of opposite nature (e.g., one is load startup, and the other is an increase in photovoltaic output), clearly indicating a lack of cross-user physical characteristics. When both of the above conditions are simultaneously met, The value is 1 if it is not 1, otherwise the value is 0.

[0022] Step S37: Mutation co-occurrence similarity The calculation is as follows: Based on the discrete weights and matching states defined above, the final mutation co-occurrence similarity is calculated using the weighted Jaccard exponent principle. The calculation formula is as follows: Furthermore, the construction of a two-dimensional voltage similarity evaluation model, and the screening of suspected voltage-related user pairs with highly coupled voltage features through a dynamic assessment strategy including a revival mechanism, specifically includes the following steps: (1) Main assessment process: Time series consistency test based on daily slices. In the actual operation of low-voltage distribution networks, the electricity consumption behavior of users and the operating environment of the power grid (such as voltage fluctuations and three-phase imbalance) have significant time-varying characteristics. Two non-inter-user users may show occasional high voltage similarity during certain specific periods (such as the late night load trough), but this similarity cannot be maintained over a long period. If the data of the whole month or the whole year are simply spliced ​​into a long series to calculate the total similarity, these short-term pseudo-correlation will be averaged, thus interfering with the final judgment. To this end, this invention proposes an assessment strategy of "breaking down the whole into parts". First, the full voltage time series data of users is divided into several daily slices according to natural days. For each daily slice, the two-dimensional voltage similarity evaluation model described in steps 2 to 3 is independently called to calculate the trend similarity of the day. Co-occurrence similarity with mutations Subsequently, a weighted fusion formula was used to calculate the user's overall score for the day. : In the formula, Set a trend weighting factor (e.g., 0.4). Set a daily passing grade. (For example, 0.60). Iterate through all valid data days. If any valid day is found to exist... If the similarity is not consistently high across all valid assessment days, it indicates that the user pair exhibited significant non-correlation during this period (i.e., the voltage characteristics deviated), which physically constitutes strong evidence of a "non-hard connection." Based on this, the user pair is initially classified as "Rejected." Conversely, only when the user pair maintains a consistently high similarity across all valid assessment days will it be initially classified as "Passed." This mechanism significantly improves the algorithm's accuracy, effectively eliminating a large number of false positive samples.

[0023] (2) Resurrection Process: Secondary review based on long-cycle characteristics and high-confidence events. Although the "daily veto" main assessment process can effectively filter false alarms, it also brings a side effect: it has extremely high requirements for the perfection of the data. In actual scenarios, there may be cases where the data quality is extremely poor on individual days (such as data drift caused by meter failure) or special electricity consumption behaviors (such as users not using high-power equipment all day). The "false positives" (false negatives) are due to low recall rates. For genuine users, these false positives reduce the algorithm's recall rate. To resolve this issue, this invention introduces a "feature revival mechanism." For all user pairs judged as "rejected" in the main process, they are not immediately discarded, but undergo a second, in-depth feature review. The following two key indicators for the user pair are statistically analyzed throughout the entire time period: Indicator 1: Average trend similarity throughout the entire time period. This indicator reflects the overall overlap of the voltage curves of two users over a long period. Indicator Two: Number of days with high-confidence abrupt changes. That is, to count how many days the mutation co-occurrence similarity was between two users. Exceeded an extremely high score threshold (For example, 0.8). The resurrection determination logic is: if the user satisfies both of the following equations simultaneously: (In the formula, The threshold for the resurgence trend is, for example, 0.85; If the resurrection time threshold is 3 days, it indicates that although the user's behavior may fluctuate on individual days, the overall pattern is extremely similar (due to high...). (Guaranteed), and exhibits irrefutable synchronous characteristics on multiple key days (by... (Guarantee). Accordingly, the resurrection mechanism is triggered, forcibly correcting the user's status to "passed". This mechanism cleverly utilizes the "special talent" logic—as long as the user exhibits extremely strong physical connection characteristics at certain times and performs well in general, occasional mistakes can be tolerated, thereby maximizing the recovery of real cross-user cases while ensuring accuracy.

[0024] Furthermore, the construction of the user association graph involves using a density-based clustering algorithm to perform preliminary clustering of the voltage-related suspected user pairs. Based on the geometric distance in the feature space with introduced mutation co-occurrence similarity weights, the preliminary clustering results are subjected to centroid iterative adsorption and splitting to generate a geometrically compact preliminary user group. The specific steps are as follows: Step S41: In the initial clusters generated by the density clustering algorithm, select the pair of users with the highest overall similarity that is greater than the preset core threshold as the core user pair. The average voltage curve of the core user pair is calculated as the characteristic centroid curve of the cluster. ; Step S42: Traverse any remaining candidate users within the cluster. Calculate its voltage curve With centroid curve root mean square error between , as the original physical metric characterizing geometric distance; Step S43: To address the issue of voltage amplitude differences due to varying line impedances (i.e., lower voltage for users farther from the core), a distance correction mechanism based on abrupt co-occurrence similarity is introduced; user data is obtained. Maximum mutation co-occurrence similarity with core user pairs Introducing mutation tolerance factor Calculate the effective feature distance using the following formula. : ; This formula shows that users with higher mutation synchronicity will have their effective distance in the feature space compressed more, making them easier to cluster and adsorb, even if their physical voltage deviation is large. Step S44: Calculate the effective feature distance With the preset adsorption radius threshold Compare; if Then determine the user Those belonging to the current core group will be retained; if Then determine the user For loosely connected or long-chained nodes that have been mixed in through weak connections, remove them from the current population; Step S45: Remove the adsorbed users from the set to be processed, and repeat steps S41 to S44 among the remaining users until no core user pairs that meet the conditions can be found, thereby splitting the loosely connected large clusters that may have misjudged connections into several subgroups with internal geometric compactness and consistent characteristics.

[0025] Furthermore, the system utilizes the current surge event feature sequence as the excitation signal and the voltage time series data as the response signal to construct a ternary hybrid criterion model that includes strong response pass-through, absolute threshold fallback, and relative proportion pass-through. This model is used to perform pairwise cross-causal inference on users within the group to verify the authenticity of the physical connection.

[0026] This invention introduces a matching verification mechanism for current surges and voltage drops, whose electrical mechanism is based on the equivalent circuit model of low-voltage distribution networks and the voltage division principle of line impedance.

[0027] In low-voltage distribution areas, the power supply lines objectively possess a line impedance that cannot be ignored. According to Ohm's law and circuit principles, the terminal voltage on the user side ( ) and power supply voltage ( ), line current ( The relationship between the line impedance and the line impedance can be approximated as follows: When a user starts a high-power load, it will cause a significant surge in the line load current. Due to the presence of line impedance, this current increment will generate an additional voltage drop on the power supply line. This directly leads to a sudden drop in voltage on the user side. Therefore, at the physical level, "sudden current rise" is the cause and "sudden voltage drop" is the effect. The two are strictly synchronous in time and have a definite opposite direction.

[0028] For user groups with a "connected" relationship, the essential characteristic is that they are physically connected to the same branch line or share the same point of access (PCC). Therefore, a sudden change in the load current of any user in this group will flow through the shared line impedance, causing a simultaneous voltage drop for all users on the series path. Conversely, if two users only have similar voltage curve shapes (e.g., affected by unified voltage regulation from the upstream transformer) but do not share a line impedance, then a drastic change in the current of one user will not produce an observable instantaneous response on the voltage of the other user. Based on this electrical mechanism, this invention uses "current surge matching voltage drop" as the standard for determining the authenticity of the physical connection, which can effectively eliminate false positive samples that only have statistical correlation but no physical connection.

[0029] The specific steps are as follows: Step S51: Adaptive Extraction of Excitation Signal. For the user pair to be verified (UA, UB), significant current events that can serve as verification "probes" are first captured. Based on the MAD adaptive threshold algorithm introduced in Step S2 above, the set of current surge moments Tspike of the originating user A is extracted. This step adds an additional physical lower limit constraint (e.g., the current increment must be greater than 3.0A) on top of the MAD statistical threshold. This is because small current fluctuations (such as light bulb switching) often cause voltage drops in low-voltage distribution networks that are often drowned out by background noise and do not qualify as verification sources.

[0030] Step S52: Windowed search of the response signal. For any time t in the set of current surge moments of the originating user extracted in Step S51, to overcome the asynchrony problem caused by communication transmission delay or small deviations in the meter's clock, a time window is defined covering several sampling points before and after time t (e.g., ±1 point, i.e., 15 minutes before and after). Within this window, the maximum voltage drop amplitude ΔVA,self of user A and the maximum voltage drop amplitude ΔVB,target of target user B are searched and calculated respectively.

[0031] Step S53: First, verify the validity of A's own response. Set a valid drop threshold Vmin_valid (e.g., -0.5V). Only when |ΔVA,self|>|Vmin_valid| is it considered that user A has experienced a valid voltage disturbance sufficient as a reference at that moment. For moments with insufficient drop, this algorithm chooses to skip the verification directly and not include them in the denominator of the success or failure statistics, thus ensuring the fairness of the assessment.

[0032] Step S54: Execution of the ternary hybrid criterion. For the moment that passes the validity prediction, a hybrid criterion model containing three different physical scenarios is executed. Electrical causality matching is considered successful at that moment if any of the following conditions are met: Condition 1 (Strong Response Straight Through): ΔVB,target ≤ ΔVA,self. That is, the voltage drop of target user B is greater than or equal to the voltage drop of reference user A. This situation typically occurs when user B is downstream of user A (at the end of the line), or when the line impedance of the branch containing user B is much greater than that of A. According to Ohm's law and the voltage divider principle, end users often experience more severe voltage fluctuations than beginning users.

[0033] Condition 2 (Absolute Threshold Cub): ΔVB,target <- Vabs_pass. That is, if the voltage drop of target user B exceeds the preset absolute pass voltage value (e.g., 0.35V), it is considered a valid response regardless of its ratio to A.

[0034] Condition 3 (Relative Proportion Passed): ΔVB,target < ΔVA,self × Rratio is satisfied. That is, the voltage drop of target user B reaches the preset ratio Rratio (e.g., 0.25) of the voltage drop of benchmark user A.

[0035] Step S55: Causal determination based on dynamic tolerance. Calculate the matching success rate at all valid verification moments. If the success rate is higher than the dynamic tolerance, the unidirectional causal relationship between user A and user B is determined to be valid. If the voltage similarity between the two users in the first stage is extremely high, the pass rate requirement for current verification can be appropriately relaxed to tolerate occasional data asynchrony. If the voltage similarity is just above the passing threshold, current verification must have an extremely high pass rate to prevent false positives.

[0036] Furthermore, the process involves constructing a strong verification graph based on the causal inference results, using a maximum clique extraction algorithm to iteratively analyze the fully connected subgraph structure within the graph, ensuring that any two users within the output group satisfy electrical connection consistency, and determining the analyzed fully connected subgraph as the final physical user group. The specific steps are as follows: Step S61: Construction of a strong validation graph. Based on the validation results of the previous steps, construct an undirected graph. Among them, the node set This includes all suspected cross-users identified through preliminary clustering. The edges in the graph... This represents a "strong connection" relationship that has been rigorously verified by electrical causality.

[0037] Step S62: For any two users and Check the causality verification records between them. Establish connecting edges in the graph. The conditions are: (1) Two-way verification passed: that is The sudden change in current led to voltage response ( (established), and The sudden change in current also led to voltage response ( (Established).

[0038] (2) Unable to be falsified (presumption of innocence): If one or both parties cannot be effectively verified due to missing data (such as no current data) or no significant change, the connection edge shall be temporarily retained to avoid accidentally deleting potential real cross-connection users.

[0039] Step S63: Apply the maximum clique search algorithm to the graph The algorithm seeks the fully connected subgraph with the most nodes. In graph theory, a fully connected subgraph (clique) is defined as a subset of a graph where every two nodes are directly connected by an edge. Applying this concept to the context of cross-connection identification, a fully connected subgraph means that all members within the group satisfy "pairwise electrical connectivity consistency." This mechanism completely eliminates transitive misjudgments such as "A connects to B, B connects to C, but A and C are unrelated," ensuring that each group in the final output is a tightly coupled physical entity.

[0040] Step S64: After finding the largest fully connected subgraph in the current graph, mark it as a specific, independent group of households (e.g., "household group #1"). To prevent duplicate calculations, from the graph... Remove all nodes and their associated edges contained in the subgraph.

[0041] Step S65: Repeat the search and removal steps above to continue searching for the next largest fully connected subgraph in the remaining graph. This process continues until the remaining nodes in the graph can no longer form a fully connected subgraph containing at least two nodes (i.e., only isolated nodes remain). Finally, the remaining isolated nodes in the graph are marked as "non-resident users" and removed.

[0042] A low-voltage transformer substation cross-connection identification system based on voltage spatiotemporal feature fusion and electrical causality verification includes: The data preprocessing module acquires the voltage and current time series data of all users in the low-voltage distribution area to be identified, performs time series alignment and cleaning processing based on interpolation tolerance on the data, and generates a standardized voltage data matrix and current data matrix with reliability status markings. The adaptive feature extraction module adaptively extracts feature sequences, models background noise for each user's voltage and current data, constructs an adaptive dynamic threshold that adapts to individual fluctuation characteristics, and discretizes and extracts voltage surge event feature sequences and current surge event feature sequences that characterize key power quality disturbances from continuous time series data. It adopts a dynamic threshold segmentation technique that can adaptively distinguish between background fluctuations and significant disturbances. The two-dimensional similarity calculation module constructs a two-dimensional voltage similarity evaluation model. Based on the voltage time series data, it calculates a trend similarity index based on a local regularization kernel that can tolerate sampling time deviation. Based on the characteristic sequence of voltage mutation events, it calculates a mutation co-occurrence similarity index based on event level weight that weakens the influence of amplitude. Based on the above similarity indices, it calculates the voltage two-dimensional similarity. The clustering and screening module constructs a dynamic assessment strategy for the revival mechanism to screen out voltage-related suspected user pairs with highly coupled voltage characteristics and constructs a user association graph. It uses a density-based clustering algorithm to perform preliminary clustering on the voltage-related suspected user pairs, and performs centroid iterative adsorption and splitting on the feature space geometric distance based on the introduction of mutation co-occurrence similarity weight to generate a geometrically compact preliminary user group. The causal verification and topology reconstruction module uses the current surge event feature sequence as the excitation signal and the voltage time series data as the response signal to construct a ternary hybrid criterion model that includes strong response pass-through, absolute threshold fallback, and relative proportion pass-through. It performs pairwise cross-causal inference on users within the group to verify the authenticity of the physical connection. Based on the causal inference results, a strong verification relationship graph is constructed. The maximum clique extraction algorithm is used to iteratively analyze the fully connected subgraph structure in the graph to ensure that any two users within the output group satisfy electrical connection consistency. The analyzed fully connected subgraph is determined as the final physical interconnected user group.

[0043] The implementation process and methods of the system are the same and will not be described again.

[0044] A computer device includes 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 steps of the low-voltage distribution area cross-unit identification method based on voltage spatiotemporal feature fusion and electrical causality verification as described in the first aspect.

[0045] A storage medium storing a computer program that, when executed by at least one processor, implements the steps of the low-voltage distribution area cross-unit identification method based on voltage spatiotemporal feature fusion and electrical causality verification as described in the first aspect.

[0046] The present invention has the following beneficial effects: (1) This invention proposes a method for identifying cross-connection between households in low-voltage distribution areas based on voltage spatiotemporal feature fusion and electrical causality verification. It can use existing measurement data to achieve accurate identification and topology reconstruction of complex cross-connection relationships in low-voltage distribution areas, providing efficient and reliable technical means for line loss management, household-transformer relationship verification and safe electricity use management in distribution areas. (2) This invention constructs a two-dimensional voltage evaluation model that integrates trend similarity and mutation co-occurrence similarity. It uses local regularization kernel to solve the waveform misalignment problem caused by the asynchronous clock of the meter. It uses event level weight to overcome the interference of inconsistent voltage amplitude caused by line impedance difference. Combined with the dual-track strategy of daily assessment and revival mechanism, it realizes the comprehensive and accurate capture of user voltage characteristics, which significantly improves the recognition accuracy and recall rate under complex working conditions. (3) This invention introduces a pairwise cross-verification mechanism based on electrical causality, using current surge as excitation and voltage drop as response, and establishes a three-element hybrid criterion including strong response pass-through, absolute threshold fallback and relative proportion pass-through. It constructs a rigorous closed-loop verification system from the physical level, eliminates false positives that are easily generated by relying solely on voltage similarity, and ensures that the identification results have solid physical evidence support. (4) This invention adopts an adaptive threshold algorithm based on MAD statistics and a topology reconstruction technique based on maximum clique extraction. It can automatically adapt to the load fluctuation characteristics of different transformer areas and different users. It can deal with long chain effects and loose clustering problems without frequent manual parameter adjustments. It has strong algorithm robustness and engineering universality. (5) The present invention is simple to calculate and has a clear principle. It can realize the automatic and accurate identification and topology reconstruction of the relationship between households in the low-voltage transformer area, help maintenance personnel to quickly locate abnormal connections, investigate safety hazards and manage line loss problems, and provide a powerful tool for verifying the relationship between households and transformers. It has good application prospects. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification disclosed in this invention. Figure 2 This is a schematic diagram of the user topology in an embodiment of the present invention; Figure 3 This is a schematic diagram of the voltage and current characteristics of serial users in an embodiment of the present invention; Figure 4 This is a schematic diagram of the identification results of the cross-unit identification algorithm of the present invention in the transformer area of ​​an embodiment. Detailed Implementation

[0048] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0049] Example 1: This embodiment takes a typical low-voltage distribution area within the jurisdiction of a power supply station in a certain city as an example to further illustrate the low-voltage distribution area cross-connection identification method proposed in this invention.

[0050] Figure 1 This paper describes the application of a low-voltage transformer substation cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification in low-voltage distribution networks, including but not limited to low-voltage transformer substation cross-connection identification. For example... Figure 1 As shown, this example employs a low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification, which mainly includes the following steps: Step 1: Obtain voltage and current timing data for all users within the low-voltage distribution area to be identified. In this example, a typical low-voltage distribution area within the jurisdiction of a power supply station in a certain city is used as the diagnostic target. The area's records show 75 low-voltage users, including 67 residential users (codes U-01 to U-67) and 8 non-residential users (codes C-01 to C-08). Data for these 75 low-voltage users from November 1st to November 7th, 2025 (7 days in total) is obtained. The ratio of cross-connected users to non-cross-connected users is approximately 1:5, with approximately 6 cross-connected groups. Figure 2 This demonstrates the physical topology differences between typical cross-line users (U-08, U-09, U-10, U-11) and non-cross-line users in this distribution area. Topology, transformer ledgers, line parameters, transformer parameters, and electrical distance data for the cross-line identification samples were obtained from the PMS (Equipment Asset Management) system and the marketing system; power, voltage, and current data for 96 points related to the cross-line identification samples were obtained from the D5000 system; and power, voltage, and current data for 96 points related to the users involved in the cross-line identification samples were obtained from the electricity consumption data acquisition system.

[0051] The acquired data undergoes time-series alignment and cleaning based on interpolation tolerance to generate standardized voltage and current data matrices with reliability status markers. This process includes the following steps: Step S11: Set the standard sampling interval according to the acquisition frequency of the measurement equipment in the transformer area. and daily standard sampling points Based on the timestamps of the user's original data, it is mapped to a standard timeline to generate an original sequence containing missing values. Simultaneously, a binary reliability mask sequence of the same length as the original sequence is initialized in this step to accurately track the reliability of each data point. All data points are initialized to an "unreliable state" by default or based on the existence of the original data. The maintenance logic for this reliability mask sequence is as follows: Let the original time series data be: ; Corresponding reliability mask sequence definition: ; in: ; Step S12: Traverse the original sequence, identify consecutive missing data segments, and calculate the length of each missing segment; for each missing segment, record its start time. End time and the length of that segment This step aims to distinguish between "intermittent packet loss" and "systemic failure".

[0052] Step S13: Introduce interpolation tolerance threshold As a boundary for data repair; when a continuous missing interval is detected in the time series data, and the length of the continuous missing interval is... Not exceeding the preset threshold When, i.e., consecutive missing lengths are detected If the missing data is considered occasional, it will be filled using a linear interpolation algorithm. The calculation formula is as follows: ; In the formula, The moment to be interpolated Data, and These are the known valid observations before and after the missing segment. After calculation, the reliability mask sequence is updated, and the data points within the filled time period are marked as "reliable states" to maintain the continuity of subsequent feature extraction. The logical basis of this operation is that within a very short time window (such as 45 minutes), the voltage and current changes of the low-voltage power grid are usually smooth, and linear interpolation can restore the real physical process with a very high confidence. Therefore, these interpolation points are fully qualified to participate in subsequent feature extraction.

[0053] Step S14: When a continuous missing interval is detected in the time series data, and the length of the continuous missing interval is... Preset threshold When a continuous missing length is detected Although linear interpolation is performed to maintain the mathematical integrity of the sequence, all interpolation points within the time period are forcibly maintained or marked as "unreliable" in the reliability mask sequence. Before performing difference calculation or similarity calculation, the reliability mask state corresponding to each data point participating in the difference calculation or similarity calculation is detected. When any data point participating in the calculation is detected as unreliable, the difference calculation or similarity calculation at the corresponding time point is blocked, or the weight of the corresponding time point in the difference calculation result or similarity calculation result is reset to zero, so as to suppress the influence of data introduced by interpolation of consecutive missing intervals on the calculation result.

[0054] In this example, the system exports data from the Electricity Consumption Information Acquisition System (AMI) for 76 users in transformer area TQ-A from November 1st to 7th, 2025. A standard sampling interval is set. Minutes, the system automatically generates 672 time points ( The system uses a standard timeline. For user U-05, some upload records in the original data are not at the exact hour (e.g., 10:14:58). The system automatically merges these records to the nearest standard time of 10:15:00 and initializes the corresponding timestamp. A reliability mask matrix of dimension, initially all values ​​are "unreliable". For steps S12 and S13, an interpolation tolerance threshold is set. The system scan revealed that user U-05 had two consecutive missing data points between 10:00 and 10:30 on November 2nd. The system reads the voltage as 225.4V at 09:45 and 224.8V at 10:45. Using linear interpolation, it calculates the voltage as 225.2V at 10:00 and 225.0V at 10:15, and updates these two interpolation points to "reliable state" in the mask matrix. Regarding step S14, the system detects that user U-12 experienced a power outage on November 3rd, resulting in the loss of 48 consecutive data points. Although the system performed linear padding to connect the breakpoints, it forcibly retained these 48 points as "unreliable" in the mask matrix. When subsequently calculating the similarity between U-12 and other users, this 12-hour data segment was automatically masked and not included in the valid sample statistics.

[0055] Step 2: Introducing robust statistical principles, adaptively extracting feature sequences, modeling background noise for each user's voltage and current data, constructing an adaptive dynamic threshold that adapts to individual fluctuation characteristics, and discretizing and extracting voltage surge event feature sequences and current surge event feature sequences representing key power quality disturbances from continuous time-series data. The specific steps are as follows: Step S21: First-order difference sequence calculation. First, the cleaned user voltage time series data... Perform discrete first-order difference operations to generate difference sequences. At any given time difference Calculate using the following formula: This indicator Directly represents time Compared to the instantaneous voltage change at the previous moment. During this process, the system will read the reliability mask sequence generated in step S1 in real time, if at time... or The corresponding mask is marked as "unreliable state", then at time The difference values ​​will be set to invalid values ​​(NaN) or zero to eliminate the interference of missing data on feature extraction.

[0056] In this example, the system performs first-order differential calculations on the voltage / current sequences of all 76 users. Taking user U-01 as an example, at 18:30 on November 4th, the voltage dropped from 225.7V at 18:15 to 221.5V. The calculation yields... Since the mask at this moment is reliable, the difference value is retained for subsequent statistics.

[0057] Step S22: This step uses the MAD algorithm to process the difference sequence. A robust estimation of the background noise level is performed. The specific calculation logic is as follows: First, the median of the difference sequence is calculated. This value reflects the central trend of voltage variation. Next, the absolute deviation sequence of each differential point from this central trend is calculated, and the median of this sequence is taken to obtain the original MAD value. In the above formula, Represents voltage differential sequence The median, representing the background noise level, is used to calculate the voltage differential point ith. The absolute difference from the background noise is obtained by taking the median of the absolute difference sequence. This can represent the inherent dispersion of voltage data. Finally, to ensure that the MAD value is consistent with the standard deviation in terms of dimensions for engineering applications, a normal distribution consistency scaling factor is introduced. The MAD value is calibrated. For a standard normal distribution, Values The final estimate of the standard deviation of background noise. The calculation is as follows: From this obtained It can accurately characterize the inherent dispersion of the user's voltage data and is not affected by occasional drastic fluctuations.

[0058] In this example, the MAD algorithm is used to perform statistical analysis on the difference sequence of user U-01. The median is calculated. The median of absolute deviation Multiplying by a calibration factor of 1.4826 yields an estimate of the user's background noise standard deviation. This indicates that the user's voltage fluctuations under normal circumstances are typically within... Within.

[0059] Step S23: Construction of the dynamic voltage surge threshold. Based on the estimated background noise level described above, a dynamic voltage surge threshold that adapts to user load characteristics is constructed. This threshold is defined as the boundary of the noise standard deviation that deviates from the central trend by a certain multiple. In the formula, This is a preset voltage significance coefficient used to adjust the algorithm's sensitivity to voltage drop events.

[0060] Table 1. Graded detection parameters and example calculations for adaptive voltage surge events in this embodiment.

[0061] As shown in Table 1, in this embodiment, for user U-01, the system first calculates its background noise to be approximately 0.45V. Based on the set coefficients, a normal mutation threshold of 1.35V and a significant mutation threshold of 2.7V are constructed respectively. The 4.2V drop detected at 18:30 on November 4th, because its amplitude exceeds both of these thresholds, is ultimately determined to be a 'significant mutation event' with a weight of 2, thus playing a dominant role in the subsequent similarity calculation.

[0062] Step S24: Filtering and generating the event sequence. Traverse the first-order difference sequence. Check each moment one by one The effective difference value. If the condition is satisfied. This means that a negative drop exceeding the dynamic threshold has occurred, and at any given time... If the data mask is "reliable state", then the extraction time is... The voltage fluctuations and their corresponding drop values ​​constitute the characteristic sequence of voltage mutation events for this user.

[0063] Step S25: Event Hierarchy Labeling. To quantify the information value of different events in subsequent similarity calculations, this step labels them according to the magnitude of mutation. Standard deviation of background noise The extracted events are categorized based on their multiple relationships. Specifically, a higher-order multiple threshold is introduced. :like If the event is marked as a "significant abrupt change event," such events typically correspond to the start-up and shutdown of high-power loads, have extremely high signal-to-noise ratios, and possess core identification value; if If a significant event is identified, it is labeled as a "common mutation event". Through this hierarchical strategy, the subsequent two-dimensional similarity model can assign higher weights to significant events, thereby significantly improving the robustness of the algorithm in complex noisy environments.

[0064] In this example, the system filtered out the drop (-4.2V) experienced by user U-01 at 18:30 on November 4th. Because... This moment is selected into the event sequence. Furthermore, a significant mutation fold is defined. (corresponding threshold) ).because This event was further labeled a "significant mutation event" and will be given high weight in subsequent calculations. Simultaneously, physical limits were set for the current data. The 12.5A current surge event corresponding to that moment was successfully extracted.

[0065] Step 3: Construct a two-dimensional voltage similarity evaluation model. Based on the voltage time series data, calculate a trend similarity index based on a local regularization kernel that can tolerate sampling time deviation. Based on the voltage mutation event feature sequence, calculate a mutation co-occurrence similarity index based on event level weights that weakens the influence of amplitude. Then, use a dynamic evaluation strategy including a revival mechanism to screen out suspected voltage-related user pairs with highly coupled voltage features. The specific steps are as follows: Step S31: For user A and user B to be compared, first obtain their cleaned voltage sequences. and To address the clock skew problem, this invention introduces the concept of Dynamic Time Warping (DTW), which is simplified into a "local warping" strategy to reduce computational complexity. For any given time... Define a coverage area with several sampling points before and after (e.g.) Each point corresponds to Local time window (minutes) Within this window, at user A's time... Based on this, user B searched in Calculate the minimum voltage deviation between the two points that are closest to its voltage value within the range. : In this example, suspected cross-user pairs U-08 and U-09 are selected for calculation. A local window is set. ( (minutes). At a certain moment U-08 has a voltage of 221.5V, and U-09 has a voltage of 220.8V. The direct difference is 0.7V. The system searches for U-09. The voltage at that moment was 221.4V. Since it is even smaller, 0.1V is taken as the local minimum deviation at that moment. This eliminates the effects of clock skew.

[0066] Step S32: At the time of obtaining Minimum voltage deviation Then, it is transformed into a normalized similarity score. This invention uses a Gaussian kernel function for nonlinear mapping. Specifically, the instantaneous similarity score... The calculation formula is: In this example, the voltage tolerance parameter is set. Substituting the aforementioned 0.1V deviation into the Gaussian kernel formula, the instantaneous similarity score is calculated. If a traditional hard threshold (such as a threshold of 0.5V) is used, a direct deviation of 0.7V will result in a score of 0, while this method accurately reflects the high similarity between the two.

[0067] Step S33: Calculate the instantaneous similarity score sequence for all times throughout the day. Next, to eliminate the interference of individual extreme values, this step performs a truncated mean aggregation operation. First, for... Sort the sequence in ascending order, then remove the element with the lowest value. (For example, 10%) of the data points, only for the remaining The arithmetic mean of the data points is calculated as the final trend similarity. .

[0068] In this example, the truncation ratio is set. The system sorts the similarity scores of 96 points for U-08 and U-09 throughout the day, removes the 9 lowest scores (corresponding to occasional data jumps or communication delays), and calculates the average of the remaining 87 points to obtain the daily trend similarity. This accurately characterizes the overall overlap of the voltage curves of the two.

[0069] Step S34: Based on the voltage mutation event feature sequence adaptively extracted in Step S2, obtain the voltage mutation event time sets for User A and User B throughout the entire time period. and Construct the union of the two sets. This union encompasses all moments when voltage disturbances occur on either side, forming the time reference for subsequent similarity calculations.

[0070] Step S35: For the union Every moment in We no longer focus on its specific voltage differential amplitude. Instead, it assigns a fixed base weight based on the event level (“significant” or “normal”) marked in step S25. The specific weight allocation logic is as follows: If user A or user B at time... The mutation that occurs is marked as a "significant mutation event" (i.e., a sharp drop, usually corresponding to a strong disturbance source), and a higher weight value is assigned to that moment. (Value is 2); This reflects that significant events have a higher "veto power" and "evidence effect" in determining cross-party transactions. If neither party experiences a significant change, and only one or both parties experience a "common change event" (i.e., a small drop), then a lower weight value is assigned to that moment. (The value is 1).

[0071] In this example, the union of mutations for U-08 and U-09 over the entire time period is calculated, encompassing a total of 15 time points. Three of these time points are marked as "significant mutations" by one or both parties, and the system assigns them basic weights. The remaining 12 time points are considered "ordinary mutations" and are assigned weights. .

[0072] Step S36: Define the matching indicator function Used for strict time determination Whether it constitutes a valid "co-occurrence" event. The judgment logic requires two conditions to be met simultaneously: Condition 1: Event coexistence. That is, at time... Must exist simultaneously in the set and middle( This indicates that both users were disturbed at that moment. Condition 2: Consistent direction. That is, the first-order differential signs of the voltages of the two users at that moment must be the same. For cross-user identification scenarios, the focus is on the same-direction voltage drop, i.e. and If one increases and the other decreases, it indicates that the disturbance sources experienced by the two users are of opposite nature (e.g., one is load startup, and the other is an increase in photovoltaic output), clearly indicating a lack of cross-user physical characteristics. When both of the above conditions are simultaneously met, The value is 1 if it is not 1, otherwise the value is 0.

[0073] Step S37: Mutation co-occurrence similarity The calculation is as follows: Based on the discrete weights and matching states defined above, the final mutation co-occurrence similarity is calculated using the weighted Jaccard exponent principle. The calculation formula is as follows: In this example, the system compared the 15 time points one by one. It found that at 12 of these time points (including all 3 significant abrupt change points), both users experienced a simultaneous drop in the same direction, satisfying the condition... The numerator is calculated as follows: Finally, the mutation co-occurrence similarity was calculated. The table below shows the daily calculation process of the co-occurrence similarity of voltage abrupt changes and the similarity of voltage trends between the significant cross-users U-13 and U-14 in this example, providing a strong basis for identifying their cross-user relationship.

[0074] Table 2. Typical Voltage Characteristics of Inter-user Users

[0075] Furthermore, a dynamic assessment strategy incorporating a revival mechanism is used to screen out suspected voltage-correlated user pairs with highly coupled voltage characteristics. Specific steps include: 1. Main Assessment Process: Time Series Consistency Verification Based on Daily Slices. First, the user's full voltage time series data is divided into several daily slices according to natural days. For each daily slice, the aforementioned algorithm is independently called to calculate the trend similarity for that day. Co-occurrence similarity with mutations Subsequently, a weighted fusion formula was used to calculate the user's overall score for the day. : In the formula, Set a trend weighting factor (e.g., 0.4). Set a daily passing grade. (For example, 0.60). The system will iterate through all valid data days. If it finds any valid day... If the similarity is negative, it indicates that the user pair exhibited significant non-correlation during this period (i.e., the voltage characteristics deviated), which physically constitutes strong evidence of "non-hard connection." Based on this, the system initially classifies the user pair as "Rejected." Conversely, only when the user pair maintains a stable high similarity across all valid assessment days will it be initially classified as "Passed."

[0076] 2. Resurrection Process: Secondary review based on long-term characteristics and high-confidence events. The system statistically analyzes the user's performance across the entire time period for the following two key indicators: Indicator 1: Average trend similarity across the entire time period. This indicator reflects the overall overlap of the voltage curves of two users over a long period. Indicator Two: Number of days with high-confidence abrupt changes. That is, to count how many days the mutation co-occurrence similarity was between two users. Exceeded an extremely high score threshold (For example, 0.8). The resurrection determination logic is: if the user satisfies both of the following equations simultaneously: (In the formula, The threshold for the resurgence trend is, for example, 0.85; If the resurrection time threshold is 3 days, it indicates that although the user's behavior may fluctuate on individual days, the overall pattern is extremely similar (due to high...). (guaranteed), and has synchronous characteristics on multiple key days (by...) (Guarantee). Accordingly, the system triggers the resurrection mechanism, forcibly correcting the user's status to "passed".

[0077] In this example, trend weights are set. Daily passing score U-08 and U-09 experienced data anomalies on day 5 due to station maintenance, resulting in a combined score of only 0.41 and triggering initial rejection. After entering the revival process, the system detected their average trend similarity over all time periods. ( ), and the number of days with high-resolution mutations ( ) for 5 days ( (Day). If the resurrection conditions are met, the system corrects its status to "passed," successfully recalling the real user.

[0078] Step 4: Construct a user association graph. Use a density-based clustering algorithm to perform preliminary clustering on the suspected voltage-related user pairs. Based on the geometric distance in the feature space with introduced mutation co-occurrence similarity weights, perform centroid iterative adsorption and splitting on the preliminary clustering results to generate a geometrically compact preliminary user group. The specific steps are as follows: Step S41: In the initial clusters generated by the density clustering algorithm, select the pair of users with the highest overall similarity that is greater than the preset core threshold as the core user pair. The average voltage curve of the core user pair is calculated as the characteristic centroid curve of the cluster. ; Step S42: Traverse any remaining candidate users within the cluster. Calculate its voltage curve With centroid curve root mean square error between , as the original physical metric characterizing geometric distance; Step S43: Introduce a distance correction mechanism based on mutation co-occurrence similarity; obtain user... Maximum mutation co-occurrence similarity with core user pairs Introducing mutation tolerance factor Calculate the effective feature distance using the following formula. : ; This formula shows that users with higher mutation synchronicity will have their effective distance in the feature space compressed more, making them easier to cluster and adsorb, even if their physical voltage deviation is large. Step S44: Calculate the effective feature distance With the preset adsorption radius threshold Compare; if Then determine the user Those belonging to the current core group will be retained; if Then determine the user For loosely connected or long-chained nodes that have been mixed in through weak connections, remove them from the current population; Step S45: Remove the adsorbed users from the set to be processed, and repeat steps S41 to S44 among the remaining users until no core user pairs that meet the conditions can be found, thereby splitting the loosely connected large clusters that may have misjudged connections into several subgroups with internal geometric compactness and consistent characteristics.

[0079] In this example, the system found that the mutation co-occurrence similarity between U-11 and core user U-08 was only [missing information]. (Extremely low). Set a mutation tolerance factor. Adsorption radius Calculate the effective feature distance. .because The system determined that U-11 was a loose node and removed it from the group. For user U-10, although the physical distance... However, because Calculated It was successfully adsorbed and retained.

[0080] Step 5: Using the current surge event feature sequence as the excitation signal and the voltage time series data as the response signal, construct a ternary hybrid criterion model that includes strong response pass-through, absolute threshold fallback, and relative proportion pass-through. Perform pairwise cross-causal inference on users within the group to verify the authenticity of the physical connection.

[0081] In low-voltage distribution areas, the power supply lines objectively possess a line impedance that cannot be ignored. According to Ohm's law and circuit principles, the terminal voltage on the user side ( ) and power supply voltage ( ), line current ( The relationship between the line impedance and the line impedance can be approximated as follows: When a user starts a high-power load, it will cause a significant surge in the line load current. Due to the presence of line impedance, this current increment will generate an additional voltage drop on the power supply line. This directly leads to a sudden drop in voltage on the user side. Therefore, at the physical level, "sudden current rise" is the cause and "sudden voltage drop" is the effect. The two are strictly synchronous in time and have a definite opposite direction.

[0082] For user groups with a "connected" relationship, the essential characteristic is that they are physically connected to the same branch line or share the same point of access (PCC). Therefore, a sudden change in the load current of any user in this group will flow through the shared line impedance, causing a simultaneous voltage drop for all users on the series path. Conversely, if two users only have similar voltage curve shapes (e.g., affected by unified voltage regulation from the upstream transformer) but do not share a line impedance, then a drastic change in the current of one user will not produce an observable instantaneous response on the voltage of the other user. Based on this electrical mechanism, this invention uses "current surge matching voltage drop" as the standard for determining the authenticity of the physical connection, which can effectively eliminate false positive samples that only have statistical correlation but no physical connection.

[0083] The specific steps are as follows: Step S51: Adaptive extraction of the excitation signal. For the user to be verified... First, significant current events that can serve as verification "probes" are captured. Based on the MAD adaptive threshold algorithm introduced in step S2 above, the set of current surge moments of the originating user A is extracted. This step adds an additional physical minimum constraint (e.g., the current increment must be greater than 3.0A) to the MAD statistical threshold. This is because small current fluctuations (such as light bulb switching) in low-voltage distribution networks often cause voltage drops that are drowned out by background noise and are not qualified as a verification source.

[0084] In this example, the retained group {U-08, U-09, U-10} is validated. Taking the data of users U-08 and U-09 on November 13, 2025 as an example... Figure 3 The system displays current and voltage data for U-08 and voltage data for U-09. The voltage data for both shows significant co-occurrence of abrupt changes and overall trend similarity, with extremely high voltage curve alignment. Furthermore, the current surge in U-08 correctly covers the voltage drop in U-09, thus passing verification based on electrical causality. For the user's pair (U-08, U-09), the system extracts the current surge ΔI=15.3A in U-08 at 04:45 on November 14th as the verification point.

[0085] Step S52: Windowed search of the response signal. For any time in the set... To overcome the asynchronous problems caused by communication transmission delays or small deviations in the meter's data acquisition clock, a coverage time is defined. Several sampling points before and after (e.g.) A time window (i.e., 15 minutes before and after the specified point) is used. Within this window, the maximum voltage drop for user A is searched and calculated. The maximum voltage drop of target user B .

[0086] Step S53: First, verify whether A's own response is valid. Set its own valid drop threshold. (e.g., -0.5V). Only when Only when a voltage fluctuation sufficient to serve as a reference is the algorithm considered to have occurred in user A. For moments when the voltage drop is insufficient, the algorithm skips the verification and does not include them in the denominator of the success or failure statistics, thus ensuring the fairness of the assessment.

[0087] In this example, within the time window of time t, a voltage drop of ΔVA,self=-4.2V is detected in U-08 itself. The effective voltage drop threshold Vmin_valid=-0.5V is set. Since |-4.2|>0.5, this time is considered a valid reference. Simultaneously, a voltage drop of ΔVB,target=-3.8V is detected in target user U-09.

[0088] Step S54: Execution of the ternary hybrid criterion. For the moment that passes the validity prediction, a hybrid criterion model containing three different physical scenarios is executed. Electrical causality matching is considered successful at that moment if any of the following conditions are met: Condition 1 (Strong Response Straight-Through): Satisfied This means the voltage drop at target user B is greater than or equal to the voltage drop at reference user A. This situation typically occurs when user B is downstream of user A (at the end of the line), or when the line impedance of the branch where user B is located is much greater than that of A. According to Ohm's law and the voltage divider principle, end users often experience more severe voltage fluctuations than beginning users.

[0089] Condition 2 (absolute threshold fallback): Satisfied That is, if the voltage drop of target user B exceeds the preset absolute pass voltage value (e.g., 0.35V), it is considered a valid response regardless of its ratio to that of A.

[0090] Condition 3 (Relative Proportion Passed): Satisfied This means that the voltage drop of target user B has reached a preset proportion of the voltage drop of benchmark user A. (e.g., 0.25).

[0091] In this example, a ternary criterion is applied: Condition 1: The voltage drop of the target user (-3.8V) is greater than the voltage drop of the baseline user (-4.2V) (strong response pass-through is not met); Condition 2: The absolute threshold Vabs_pass is set to 0.35V. Since -3.8V < -0.35V (i.e., voltage drop of 3.8V > 0.35V), the absolute threshold catch-all condition is met. The system determines that the electrical causal match is successful at this moment. Similarly, for user U-10 (voltage drop of -1.5V), since condition 3 is met (relative ratio pass, Rratio = 0.25, -1.5V < -4.2 × 0.25 = -1.05), the match is also determined to be successful.

[0092] Table 3. Measured data record for electrical causality verification in suspected cross-dwelling groups.

[0093] The table above records the key electrical linkage events and auxiliary verification process of a typical interconnected user group (users: {U-08, U-09, U-10}) identified in the embodiment during November 10th to 17th, 2025. The data in the table verifies the physical law that "a sudden increase in current (cause) leads to a sudden drop in voltage (effect)." Among them, The current surge amplitude of the originating user. The voltage drop amplitude (baseline) of the originating user. The voltage drop magnitude (target) for the receiving end user.

[0094] Step S55: Causal determination based on dynamic tolerance. Calculate the matching success rate at all valid verification moments. If the success rate is higher than the dynamic tolerance, the unidirectional causal relationship between user A and user B is determined to be valid. If the voltage similarity between the two users in the first stage is extremely high, the pass rate requirement for current verification can be appropriately relaxed to tolerate occasional data asynchrony. If the voltage similarity is just above the passing threshold, current verification must have an extremely high pass rate to prevent false positives.

[0095] Step 6: Construct a strong verification relationship graph based on the causal inference results, and use the maximum clique extraction algorithm to iteratively analyze the fully connected subgraph structure in the graph to ensure that any two users within the output group satisfy electrical connection consistency. The analyzed fully connected subgraph is determined as the final physical user group. The specific steps are as follows: Step S61: Construction of a strong validation graph. Based on the validation results of the previous steps, construct an undirected graph. Among them, the node set This includes all suspected cross-users identified through preliminary clustering. The edges in the graph... This represents a "strong connection" relationship that has been rigorously verified by electrical causality.

[0096] Step S62: For any two users and The system checks the causality verification records between them. Connection edges are established in the graph. The conditions are: 1. Two-way verification passed: that is... The sudden change in current led to voltage response ( (established), and The sudden change in current also led to voltage response ( (Established).

[0097] 2. Unfalsifiable (Presumption of innocence): If one or both parties cannot be effectively verified due to missing data (such as no current data) or no significant change, the connection edge is temporarily retained to avoid mistakenly deleting potential real cross-connection users.

[0098] Step S63: Apply the Maximum Clique Search Algorithm to the graph The algorithm seeks the fully connected subgraph with the most nodes. In graph theory, a fully connected subgraph (clique) is defined as a subset of a graph where every two nodes are directly connected by an edge. Applying this concept to the context of cross-connection identification, a fully connected subgraph means that all members within the group satisfy "pairwise electrical connectivity consistency." This mechanism completely eliminates transitive misjudgments such as "A connects to B, B connects to C, but A and C are unrelated," ensuring that each group in the final output is a tightly coupled physical entity.

[0099] Step S64: After finding the largest fully connected subgraph in the current graph, the system marks it as a specific, independent group of households (e.g., "household group #1"). To prevent duplicate calculations, from the graph... Remove all nodes and their associated edges contained in the subgraph.

[0100] Step S65: Repeat the search and removal steps above to continue searching for the next largest fully connected subgraph in the remaining graph. This process continues until the remaining nodes in the graph can no longer form a fully connected subgraph containing at least two nodes (i.e., only isolated nodes remain). Finally, the remaining isolated nodes in the graph are marked as "non-resident users" and removed.

[0101] In this example, based on the above verification results, the system established connecting edges (U-08, U-09) and (U-08, U-10) in the verification graph G. Further verification showed that U-09 and U-10 also passed bidirectional verification. The maximum clique search algorithm identified {U-08, U-09, U-10} as forming a fully connected subgraph with 3 nodes. The system marked this as the finally determined cross-connected household group "Group-1" and removed it from the graph. Ultimately, a total of 7 fully connected cross-connected household groups were identified in this transformer area. The accuracy rate of cross-connected household groups was 100% after on-site verification. The original voltage data of the users in this transformer area was then dimensionality-reduced using t-SNE and mapped to a two-dimensional visualization plane, as shown below. Figure 4 As shown, users who cross the river are well distinguished from non-users and other users who cross the river.

[0102] Example 2: Based on the same inventive concept as the method embodiment, this embodiment provides a low-voltage transformer substation cross-connection identification system based on voltage spatiotemporal feature fusion and electrical causality verification, including: The data preprocessing module acquires the voltage and current time series data of all users in the low-voltage distribution area to be identified, performs time series alignment and cleaning processing based on interpolation tolerance on the data, and generates a standardized voltage data matrix and current data matrix with reliability status markings. The adaptive feature extraction module adaptively extracts feature sequences, models background noise for each user's voltage and current data, constructs an adaptive dynamic threshold that adapts to individual fluctuation characteristics, and discretizes and extracts voltage surge event feature sequences and current surge event feature sequences that characterize key power quality disturbances from continuous time series data. It adopts a dynamic threshold segmentation technique that can adaptively distinguish between background fluctuations and significant disturbances. The two-dimensional similarity calculation module constructs a two-dimensional voltage similarity evaluation model. Based on the voltage time series data, it calculates a trend similarity index based on a local regularization kernel that can tolerate sampling time deviation. Based on the characteristic sequence of voltage mutation events, it calculates a mutation co-occurrence similarity index based on event level weight that weakens the influence of amplitude. Based on the above similarity indices, it calculates the voltage two-dimensional similarity. The clustering and screening module constructs a dynamic assessment strategy for the revival mechanism to screen out voltage-related suspected user pairs with highly coupled voltage characteristics and constructs a user association graph. It uses a density-based clustering algorithm to perform preliminary clustering on the voltage-related suspected user pairs, and performs centroid iterative adsorption and splitting on the feature space geometric distance based on the introduction of mutation co-occurrence similarity weight to generate a geometrically compact preliminary user group. The causal verification and topology reconstruction module uses the current surge event feature sequence as the excitation signal and the voltage time series data as the response signal to construct a ternary hybrid criterion model that includes strong response pass-through, absolute threshold fallback, and relative proportion pass-through. It performs pairwise cross-causal inference on users within the group to verify the authenticity of the physical connection. Based on the causal inference results, a strong verification relationship graph is constructed. The maximum clique extraction algorithm is used to iteratively analyze the fully connected subgraph structure in the graph to ensure that any two users within the output group satisfy electrical connection consistency. The analyzed fully connected subgraph is determined as the final physical interconnected user group.

[0103] It should be understood that the low-voltage distribution area cross-unit identification system in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0104] Example 3: This embodiment provides a computer device, including 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 steps of the low-voltage distribution area cross-unit identification method based on voltage spatiotemporal feature fusion and electrical causality verification as described in Embodiment 1 of the present invention.

[0105] Example 4: This embodiment provides a storage medium storing a computer program. When the computer program is executed by at least one processor, it implements the steps of the low-voltage distribution area cross-unit identification method based on voltage spatiotemporal feature fusion and electrical causality verification as described in Embodiment 1 of the present invention.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A method for identifying cross-connections in low-voltage distribution areas based on voltage spatiotemporal feature fusion and electrical causality verification, characterized in that, Includes the following steps: Step 1: Obtain the voltage and current timing data of all users in the low-voltage distribution area to be identified, perform timing alignment and cleaning processing on the data based on interpolation tolerance, and generate a standardized voltage data matrix and current data matrix with reliability status markings; Step 2: Introduce the principle of robust statistics, adaptively extract feature sequences, model the background noise of each user's voltage and current data, construct an adaptive dynamic threshold that adapts to individual fluctuation characteristics, and extract the voltage mutation event feature sequence and current surge event feature sequence that characterize key power quality disturbances from the continuous time series data. Step 3: Construct a two-dimensional voltage similarity evaluation model. Based on the voltage time series data, calculate a trend similarity index based on a local regularization kernel that can tolerate sampling time deviation. Based on the voltage mutation event feature sequence, calculate a mutation co-occurrence similarity index based on event level weight that weakens the influence of amplitude. Then, use a dynamic assessment strategy that includes a revival mechanism to screen out voltage suspected related user pairs with highly coupled voltage features. Step 4: Construct a user association graph, use a density-based clustering algorithm to perform preliminary clustering on the voltage-related suspected user pairs, and perform centroid iterative adsorption and splitting on the feature space geometric distance based on the introduced mutation co-occurrence similarity weight to generate a geometrically compact preliminary user group; Step 5: Introduce a causal verification mechanism at the electrical and physical level. Use the current surge event feature sequence as the excitation signal and the voltage time series data as the response signal to construct a ternary hybrid criterion model that includes strong response pass-through, absolute threshold fallback, and relative proportion pass-through. Perform pairwise cross-causal inference on users within the group to verify the authenticity of the physical connection. Step 6: Construct a strong verification relationship graph based on the causal inference results, and use the maximum clique extraction algorithm to iteratively analyze the fully connected subgraph structure in the graph to ensure that any two users within the output group satisfy electrical connection consistency. The fully connected subgraph is then determined as the final physical user group.

2. The low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification according to claim 1, characterized in that, The specific steps of step 1 are as follows: Step S11: Set the standard sampling interval and daily standard sampling points Based on the timestamps of the user's original data, the user's original data is mapped to a standard timeline to generate an original sequence containing missing values; a binary reliability mask sequence of the same length as the original sequence is initialized, with all data points defaulting to an "unreliable state" or initialized based on the existence of the original data; the maintenance logic of the reliability mask sequence is as follows: Let the original time series data be: ； Corresponding reliability mask sequence definition: ； in: ； Step S12: Traverse the original sequence, identify consecutive missing data segments, and calculate the length of each missing segment. ; Step S13: Introduce interpolation tolerance threshold As a boundary for data repair; when a continuous missing interval is detected in the time series data, and the length of the continuous missing interval is... Not exceeding the preset threshold When, i.e., consecutive missing lengths are detected If the missing data is considered occasional, it will be filled using a linear interpolation algorithm. The calculation formula is as follows: ； In the formula, The moment to be interpolated Data, and These are the known valid observations before and after the missing segment; the reliability mask sequence is then updated after calculation, and the data points within the filled time period are marked as "reliable state"; Step S14: When a continuous missing interval is detected in the time series data, and the length of the continuous missing interval is... Greater than the preset threshold When a continuous missing length is detected In the reliability mask sequence, all interpolation points within the time period are forcibly maintained or marked as "unreliable". Before performing differential calculation or similarity calculation, the reliability mask state corresponding to each data point participating in the differential calculation or similarity calculation is detected. When any data point participating in the calculation is detected as unreliable, the differential calculation or similarity calculation of the corresponding time point is blocked, or the weight of the corresponding time point in the differential calculation result or similarity calculation result is reset to zero.

3. The low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step S21: Calculate the first-order difference sequence of user voltage time series data ,in Characterized time The rate of change of voltage relative to the previous moment; Step S22: Use the MAD algorithm to process the difference sequence A robust estimation of the background noise level is performed. The specific calculation logic is as follows: First, the median of the difference sequence is calculated. Next, calculate the absolute deviation sequence of each difference point from the central trend, and take the median of the sequence to obtain the original MAD value: Finally, a normal distribution consistency factor is introduced. The MAD value is calibrated; for a standard normal distribution, Values The final estimated standard deviation of background noise The calculation is as follows: ； Step S23: Construct a dynamic voltage surge threshold that adapts to changes in user load characteristics. : In the formula, The preset voltage significance coefficient; Step S24: Traverse the first-order difference sequence to filter those that meet the conditions. The moment This constitutes a characteristic sequence of voltage mutation events; Step S25: Based on the mutation magnitude Standard deviation of background noise Based on the multiple relationship, voltage mutation events are further subdivided into "significant mutation events" and "ordinary mutation events".

4. The low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification according to claim 1, characterized in that, The specific steps of step 3 are as follows: Step S31: For user A and user B to be compared, first obtain their cleaned voltage sequences. and For any time Define a local time window that covers several sampling points before and after. Within this window, search for the local minimum voltage deviation of user B relative to user A. : Step S32: Use a Gaussian kernel function to perform a nonlinear mapping to map the local minimum voltage deviation into a nonlinear instantaneous similarity score. The mapping formula is: in, The preset voltage tolerance parameter determines the rate at which the similarity score decays as the voltage deviation increases; Step S33: Construct a full-time similarity sequence and sort it in ascending order. Perform a truncated mean aggregation operation, i.e., remove the sequence with the lowest value. Data points, only for the remaining The arithmetic mean of the data points is calculated as the final trend similarity. ; Step S34: Obtain the set of voltage surge events for user A and user B throughout the entire time period. and Construct the union of the two sets. ; Step S35: For Every moment in Assign basic weights based on the level of the mutation event that occurs at that moment. This weight is based on user A and user B's... Time and Voltage differential amplitude at time The calculation yields: ； If user A or user B is at time... Mutations marked as "significant mutation events" are assigned higher weight values. If neither party undergoes a significant mutation, but only a "common mutation event" occurs, a lower weight value will be assigned. ; Step S36: Define the matching indicator function Used to determine time Whether it is a valid co-occurrence time; if and only if user A and user B are at time [time missing]. Both exhibit voltage abrupt events, and their first-order difference signs are different. and When they are the same, The value is 1 if it is set to 1, otherwise the value is 0. Step S37: Calculate the final mutation co-occurrence similarity using the weighted Jaccard index principle. The calculation formula is as follows: 。 5. The low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification according to claim 1, characterized in that, The specific steps of constructing a two-dimensional voltage similarity evaluation model and screening out suspected voltage-related user pairs with highly coupled voltage features using a dynamic assessment strategy including a revival mechanism include: The main assessment process based on daily voltage is as follows: The user's full time-series data is divided into several daily slices based on natural days, and the trend similarity is calculated for each daily slice. Co-occurrence similarity with mutations The weighted fusion formula is used to calculate the comprehensive score for the day, and the calculation results are as follows: ； If there exists any valid data day Below the preset daily passing score If the result is positive, it indicates that the user has shown significant irrelevance at certain times, and is initially judged as "rejected"; otherwise, it is judged as "approved". Similarity score revival process: A second review is conducted on users who were judged to be in "rejected" status, and their average trend similarity over the entire period is calculated. And the similarity of mutation co-occurrence is greater than the high score threshold Number of days If the user satisfies both of the following equations: ； ； This indicates that although there are fluctuations, the overall pattern is extremely similar and has synchronous characteristics on multiple key days, thus triggering the revival mechanism and forcibly correcting the user's status to "passed".

6. The low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification according to claim 1, characterized in that, The process of constructing the user association graph involves using a density-based clustering algorithm to initially cluster the voltage-related user pairs. Then, based on the geometric distance in the feature space with introduced mutation co-occurrence similarity weights, the initial clustering results are subjected to centroid iterative adsorption and splitting to generate a geometrically compact initial cluster of users. The specific steps are as follows: Step S41: In the initial clusters generated by the density clustering algorithm, select the pair of users with the highest overall similarity that is greater than the preset core threshold as the core user pair. The average voltage curve of the core user pair is calculated as the characteristic centroid curve of the cluster. ; Step S42: Traverse any remaining candidate users within the cluster. Calculate its voltage curve With centroid curve root mean square error between , as the original physical metric characterizing geometric distance; Step S43: Introduce a distance correction mechanism based on mutation co-occurrence similarity; obtain user... Maximum mutation co-occurrence similarity with core user pairs Introducing mutation tolerance factor Calculate the effective feature distance using the following formula. : ； This formula shows that users with higher mutation synchronicity will have their effective distance in the feature space compressed more, making them easier to cluster and adsorb, even if their physical voltage deviation is large. Step S44: Calculate the effective feature distance With the preset adsorption radius threshold Compare; if Then determine the user Those belonging to the current core group will be retained; if Then determine the user For loosely connected or long-chained nodes that have been mixed in through weak connections, remove them from the current population; Step S45: Remove the adsorbed users from the set to be processed, and repeat steps S41 to S44 among the remaining users until no core user pairs that meet the conditions can be found, thereby splitting the loosely connected large clusters that may have misjudged connections into several subgroups with internal geometric compactness and consistent characteristics.

7. The low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification according to claim 1, characterized in that, Step 5 includes the following specific steps: Step S51: For the user to be verified First, the set of current surge moments for user A is extracted based on the MAD adaptive threshold. Although the same algorithm as the previous voltage surge capture is used to capture current surges, the specific parameters are different to ensure the accuracy of current surge capture. Step S52: For any time in the set Calculate the maximum voltage drop of user A within the time window around this moment. The maximum voltage drop of user B Step S53: Set your own effective drop threshold Only when Only when the absolute value of the current data is greater than the threshold is it determined that user A has experienced an effective voltage disturbance sufficient to serve as a reference at that moment; otherwise, the verification at that moment is skipped to prevent invalid comparisons caused by noise in the current data. Step S54: For valid verification moments, execute the ternary hybrid criterion. Electrical causality matching is considered successful at any of the following conditions: Condition 1: Satisfy That is, the voltage drop of target user B is greater than or equal to the voltage drop of reference user A. This usually indicates that B is located at a more distant end of the line or is more severely affected by the disturbance, which is strong evidence of cross-line interference. Condition 2: Satisfied That is, if the voltage drop of target user B exceeds the preset absolute pass voltage value, it is considered a valid response regardless of its ratio to A. Condition 3: Satisfied This means that the voltage drop of target user B has reached a preset proportion of the voltage drop of benchmark user A. ; Step S55: Calculate the matching success rate at all valid verification times. If the success rate is higher than the dynamic tolerance, then it is determined that the one-way causal relationship between user A and user B is valid. The dynamic tolerance is a variable generated by linear interpolation based on the voltage similarity score calculated in the first stage. The higher the voltage similarity score, the higher the upper limit of the allowable verification failure rate.

8. The low-voltage distribution area cross-connection identification method based on voltage spatiotemporal feature fusion and electrical causality verification according to claim 1, characterized in that, The process involves constructing a strong verification graph based on causal inference results, using a maximum clique extraction algorithm to iteratively analyze the fully connected subgraph structure within the graph, ensuring that any two users within the output group satisfy electrical connection consistency, and determining the analyzed fully connected subgraph as the final physical user group. The specific steps are as follows: Step S61: Construct an undirected verification graph , where the set of nodes This refers to all users within the initial user group; Step S62: For any two users and If the bidirectional electrical causality checks between them all pass, or if checks cannot be performed due to missing data, then in the diagram... Establish connection edges ; Step S63: Apply the maximum clique search algorithm to the graph Find the fully connected subgraph with the most nodes in the graph; Step S64: Mark the found largest fully connected subgraph as a defined cluster, and from the graph... Remove all nodes and their associated edges contained in the subgraph. Step S65: Repeat the above search and removal steps until the remaining nodes in the graph can no longer form a fully connected subgraph containing at least two nodes; Step S66: Mark the remaining isolated nodes as non-cross-household users and remove them, thereby outputting the final cross-household identification result.

9. A low-voltage distribution area cross-connection identification system based on voltage spatiotemporal feature fusion and electrical causality verification, characterized in that, include: The data preprocessing module acquires the voltage and current time series data of all users in the low-voltage distribution area to be identified, performs time series alignment and cleaning processing based on interpolation tolerance on the data, and generates a standardized voltage data matrix and current data matrix with reliability status markings. The adaptive feature extraction module adaptively extracts feature sequences, models background noise for each user's voltage and current data, constructs an adaptive dynamic threshold that adapts to individual fluctuation characteristics, and discretizes and extracts voltage surge event feature sequences and current surge event feature sequences that characterize key power quality disturbances from continuous time series data. It adopts a dynamic threshold segmentation technique that can adaptively distinguish between background fluctuations and significant disturbances. The two-dimensional similarity calculation module constructs a two-dimensional voltage similarity evaluation model. Based on the voltage time series data, it calculates a trend similarity index based on a local regularization kernel that can tolerate sampling time deviation. Based on the characteristic sequence of voltage mutation events, it calculates a mutation co-occurrence similarity index based on event level weight that weakens the influence of amplitude. Based on the above similarity indices, it calculates the voltage two-dimensional similarity. The clustering and screening module constructs a dynamic assessment strategy for the revival mechanism to screen out voltage-related suspected user pairs with highly coupled voltage characteristics and constructs a user association graph. It uses a density-based clustering algorithm to perform preliminary clustering on the voltage-related suspected user pairs, and performs centroid iterative adsorption and splitting on the feature space geometric distance based on the introduction of mutation co-occurrence similarity weight to generate a geometrically compact preliminary user group. The causal verification and topology reconstruction module uses the current surge event feature sequence as the excitation signal and the voltage time series data as the response signal to construct a ternary hybrid criterion model that includes strong response pass-through, absolute threshold fallback, and relative proportion pass-through. It performs pairwise cross-causal inference on users within the group to verify the authenticity of the physical connection. Based on the causal inference results, a strong verification relationship graph is constructed. The maximum clique extraction algorithm is used to iteratively analyze the fully connected subgraph structure in the graph to ensure that any two users within the output group satisfy electrical connection consistency. The analyzed fully connected subgraph is determined as the final physical interconnected user group.

10. A computer device, characterized in that: The computer device includes 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 steps of the low-voltage distribution area cross-unit identification method based on voltage spatiotemporal feature fusion and electrical causality verification as described in any one of claims 1-8.

Images