User digital portrait construction method based on multi-source data fusion of transformer area power intelligent terminal

Abstract

The application discloses a user digital portrait construction method based on multi-source data fusion of a transformer area power intelligent terminal, which runs in a local closed loop of the transformer area intelligent terminal and comprises the following steps: acquiring preconfigured transformer area topology and static archives and real-time collected power consumption sensing equipment measurement data, performing time sequence alignment and abnormal data cleaning on the terminal to obtain a standardized fusion data set; based on the data set, extracting indexes of static bases, time sequence dynamics and transformer area interaction dimensions on the terminal to obtain user hierarchical features; performing threshold self-adaptive calibration and risk grading calculation on the terminal to obtain multi-dimensional user digital labels; performing iterative update grading scheduling based on label change states to obtain updated portrait data and output the updated portrait data on the terminal. The application reduces the dependence on the computing power of a cloud host, realizes business offline autonomy, reduces real-time communication traffic and the risk of cross-network transmission of private data.

Description

Technical Field

[0001] This invention relates to the field of power system edge computing and data mining technology, and in particular to a method for constructing a digital profile of users based on the fusion of multi-source data from power distribution smart terminals. Background Technology

[0002] With the construction of new power systems, low-voltage distribution network areas are connecting to a large number of distributed power sources, charging piles, and various intelligent electricity consumption sensing devices. Deep fusion and high-frequency time series analysis of multi-source heterogeneous measurement data within the distribution area, and the construction of user electricity consumption behavior characteristic models, are the physical foundation for achieving local coordinated control of the distribution network, grid, load, and storage at the distribution level, precise location of line losses, and refined investigation of abnormal electricity consumption. Efficient data mining at the edge can improve the real-time electrical status perception and response control capabilities at the end of the distribution network.

[0003] Currently, the feature analysis and modeling of electricity consumption data in distribution transformer areas mainly rely on a system architecture with centralized processing at a cloud-based master station. Existing distribution transformer terminal equipment typically serves only as basic data transmission nodes, continuously uploading all collected raw electrical detail data, such as high-frequency meter load and power quality data, to the cloud-based master station. The master station's big data platform then uniformly completes multi-source data alignment, cleaning, feature extraction, and rating modeling. Under this architecture, the continuous uploading of massive amounts of high-frequency measurement data consumes significant power communication channels and results in high network interaction latency. If the master station's uplink communication network is interrupted or the cloud platform undergoes routine maintenance, the distribution transformer site will be unable to obtain or update any analysis results, leading to the stagnation of on-site anomaly investigation and local control operations. Furthermore, the method of stacking and storing all high-frequency data for retrospective analysis sometimes faces storage overflow and insufficient computing power when dealing with edge terminals with limited hardware resources.

[0004] In summary, existing technologies, when processing massive amounts of high-frequency measurement data from distribution transformer substations, rely excessively on centralized cloud computing power and communication network interaction. This results in poor service continuity of the underlying terminal systems when dealing with sudden network outages, and makes it difficult to balance long-term tracking and rapid response of multi-dimensional time-series characteristics under limited edge storage conditions. Therefore, there is an urgent need to research a power consumption characteristic analysis method that can reduce dependence on master station communication, improve the autonomous data processing capabilities of the transformer substations, and enhance the overall robustness of the system. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a digital profile of users based on the fusion of multi-source data from smart power terminals in power distribution areas, in order to solve the aforementioned problems in the prior art.

[0006] Technical solution: A method for constructing digital user profiles based on multi-source data fusion from smart power distribution terminals. This method operates in a closed loop locally on the smart distribution terminal and includes:

[0007] The system acquires pre-configured transformer area topology and static files, as well as real-time power consumption sensing device measurement data, and performs time-series alignment and abnormal data cleaning on the smart terminal to obtain a standardized fusion dataset.

[0008] Based on a standardized fusion dataset, indicators including static basic, time-series dynamic and station area interaction dimensions are extracted locally on the smart terminal to obtain user segmentation characteristics.

[0009] Based on user segmentation characteristics, threshold adaptive calibration and risk classification calculation are performed locally on the smart terminal to obtain multi-dimensional digital user tags.

[0010] Based on the changing status of multi-dimensional user digital tags, iterative updates and hierarchical scheduling are performed locally on the smart terminal to obtain updated user digital profile data.

[0011] The updated user digital profile data is output locally on the smart terminal.

[0012] Furthermore, the specific steps for extracting station interaction features from indicators including static basic, time-series dynamic, and station interaction dimensions extracted locally on the smart terminal include:

[0013] Based on the pre-configured distribution area topology, the power supply path of each user is extracted, and combined with the measured parameters in the standardized fusion dataset, the cross-influence of each user's power consumption behavior on the public power grid nodes of the distribution area is evaluated, and the cross-influence is used as the interaction feature of the distribution area.

[0014] Furthermore, the cross-impact of each user's electricity consumption behavior on the public power grid nodes in the distribution area is assessed, specifically including:

[0015] Obtain the impedance parameters and total current of each branch on the feed path, thereby determining the technical line loss of the branch. Based on the ratio of the measured current of a single user to the total current of the branch, the technical line loss of the branch is allocated along the feed path to obtain the line loss allocation contribution of a single user.

[0016] The product of the measured current of a single user and the equivalent impedance of the feed path is used as the marginal voltage drop. The ratio of the marginal voltage drop to the preset rated bus voltage is calculated to obtain the voltage disturbance contribution of a single user.

[0017] The contribution of line loss sharing and the contribution of voltage disturbance are used as the cross-influence quantities.

[0018] Furthermore, a local quantile online estimation mechanism is employed to dynamically track the characteristic distribution of user groups in the transformer area, thereby adaptively generating adaptive classification thresholds suitable for the current transformer area and determining the label level of each user, specifically including:

[0019] For each feature dimension, the newly accessed user stratification features are used as sample values, and incremental iterative calculations are performed in combination with an adaptive step size to track the quartile estimate of that feature dimension in real time.

[0020] The interquartile range is calculated based on the quartile estimates obtained from the tracking, and an adaptive grading threshold for the corresponding feature dimension is constructed by combining the preset grading coefficients with the median in the quartile estimates.

[0021] Calculate the deviation of each user's user segmentation features from the median on the corresponding feature dimension, use the ratio of the deviation to the interquartile range as the standardized score, and complete the label level classification for each user based on the standardized score and the adaptive classification threshold.

[0022] Furthermore, the deviation of each tag from the historical records in the multi-dimensional user digital tags is extracted, and joint detection of tag change significance is performed to obtain a multi-dimensional joint change significance index. This multi-dimensional joint change significance index is then compared with a preset hierarchical scheduling threshold. Based on the comparison results, differentiated update scheduling actions are executed, specifically including:

[0023] By comparing the current score of a single tag dimension in the multi-dimensional user digital tags with the record value of the previous historical snapshot, and combining the historical recursive standard deviation of that single tag dimension, the significance index of single tag change for each tag dimension is calculated.

[0024] Extract the maximum value from the single-label change significance index of each label dimension, and use the maximum value as the multi-dimensional joint change significance index;

[0025] The multidimensional joint change significance index is compared with the preset hierarchical scheduling threshold. When the multidimensional joint change significance index is higher than the upper threshold, the scheduling action of immediately updating and storing the portrait snapshot is executed; when it is between the middle threshold and the upper threshold, the scheduling action of immediately updating but not storing the snapshot is executed; when it is lower than the middle threshold, the scheduling action of skipping the update of this cycle is executed.

[0026] Furthermore, obtain the pre-configured transformer topology and static files, specifically including:

[0027] During the initialization phase of smart terminal deployment in the distribution area, the initial distribution area topology and static files are imported to the local terminal in batches. During the normal operation phase, according to the preset daily synchronization cycle, a small amount of user change information is synchronized to the local smart terminal in an incremental caching manner to maintain the weak dependence of core profile computing on the cloud main station and offline autonomy.

[0028] Beneficial effects: This invention reduces reliance on cloud-based main station computing power, enables offline business autonomy, and reduces the risk of cross-network transmission of real-time communication traffic and privacy data. Attached Figure Description

[0029] Figure 1 A flowchart illustrating a method for constructing a digital profile of users based on multi-source data fusion from smart power terminals in distribution areas is shown.

[0030] Figure 2 A flowchart illustrating the threshold adaptive calibration process is shown.

[0031] Figure 3 A flowchart for risk rating calculation is shown;

[0032] Figure 4 The flowchart shows the updated user digital profile data. Detailed Implementation

[0033] Example 1:

[0034] This embodiment provides a method for constructing a digital user profile based on multi-source data fusion from a power distribution area smart terminal. This method operates in a closed-loop manner locally on the smart terminal. Figure 1 As shown, it includes the following steps:

[0035] Step 101: Obtain the pre-configured transformer area topology and static files, as well as the real-time collected power consumption sensing device measurement data, and perform time-series alignment and abnormal data cleaning on the smart terminal to obtain a standardized fusion dataset.

[0036] Specifically, the operating environment of this embodiment is built within a power distribution smart terminal with limited computing power and storage resources. Based on this hardware, the pre-configured distribution topology and static files are loaded locally in a single step during the terminal's initialization and commissioning phase. This includes the distribution-household binding relationship, meter number, line impedance, household electrical equipment registration, and the user's basic electricity consumption category and historical payment information. Real-time collected electricity consumption sensing data is generated by the smart meters and sensing terminals deployed on the user side during operation, covering raw electrical parameters such as active power, reactive power, current, voltage, power factor, and real-time output of photovoltaic or charging piles.

[0037] After data access is completed, due to the different sampling frequencies of the multi-source data, data resampling is performed using a conventional linear interpolation method through a preset basic time window (e.g., 5 minutes) to unify the time base of the multi-frequency data. Furthermore, a sliding window recursive mean method is used to filter out jump noise points that exceed the normal range, generating a standardized fusion dataset with a unified format and consistent spatiotemporal base, which serves as the underlying data foundation for subsequent local closed-loop operations.

[0038] Step 102: Based on the standardized fusion dataset, extract indicators including static basic, time-series dynamic and station area interaction dimensions locally on the smart terminal to obtain user segmentation features.

[0039] In this embodiment, the standardized fusion dataset undergoes multi-dimensional dimensionality reduction extraction. The static basic dimension indicators are extracted from locally cached static archives, reflecting inherent attributes such as the user's rated capacity and installed equipment. The time-series dynamic dimension indicators are derived by smoothing and recursively extrapolating the user's time-of-use load, peak-valley electricity consumption ratio, and photovoltaic self-consumption rate to capture the evolution of the user's electricity consumption behavior over time.

[0040] Correspondingly, the indicators for the transformer area interaction dimension are mainly achieved by evaluating the impact of a single user on the line loss sharing and voltage disturbance contribution of the common nodes in the transformer area, reflecting the physical interaction impact of individuals in the group network. By aggregating the feature indicators of the above three levels, user hierarchical features are thus constructed. This type of hierarchical extraction logic avoids full analysis of high-dimensional raw measurement data and is adapted to the limited edge computing power of the terminal.

[0041] Step 103: Based on user segmentation characteristics, perform threshold adaptive calibration and risk rating calculation locally on the smart terminal to obtain multi-dimensional digital user tags.

[0042] Furthermore, the multi-dimensional user digital tags specifically encompass a feature tag system covering six dimensions: basic attributes, electricity consumption behavior, power quality, distributed energy, risk auditing, and regulation adaptation. During tag generation, an online quantile estimation algorithm running in the terminal's memory dynamically tracks the characteristic distribution of the user group in the local area. Without relying on cloud-based main stations, it adaptively generates evaluation threshold tiers applicable to the current electricity consumption level of the local group. Combined with evidence accumulation calculations from multi-dimensional anomaly channels, the risk dimensions are quantitatively graded, thereby completing the scoring and tiered classification of all dimension tags. This adaptive generation process solves the problem in traditional solutions where a lack of global coordination leads to extreme misjudgments with fixed thresholds in different-sized distribution areas, improving the accuracy and applicability of user profile classification.

[0043] Step 104: Based on the changes in the multi-dimensional user digital tags, perform iterative update hierarchical scheduling locally on the smart terminal to obtain updated user digital profile data.

[0044] Building upon this foundation, to reduce wear and tear on solid-state flash memory in edge devices and conserve computing power, after acquiring all tags, the system compares the current evaluation status of each tag dimension with the previous historical snapshot data to extract the significance of changes. Based on the magnitude of the multi-dimensional joint significance index, the system triggers different levels of scheduling actions, including immediately executing a disk-based snapshot, updating variables only in memory, or skipping the current update calculation cycle. Through this differentiated autonomous scheduling mechanism, redundant update tasks during periods of stable power consumption can be filtered out, and updated user digital profile data can be generated on demand, reducing the long-term operating load of the local system while ensuring the timeliness of feature capture.

[0045] Step 105: Output the updated user digital profile data locally on the smart terminal.

[0046] Furthermore, after generating the latest user digital profile data, the smart terminal in the distribution area opens its local serial communication interface and edge microservice interface to distribute the updated user digital profile data to the portable operation and maintenance tools or edge control devices in the distribution area. This local data transfer mechanism enables the distribution area to support real-time source-grid-load control, lean line loss auditing, and local early warning of abnormal power consumption, even when the communication link with the cloud master station is disconnected, by reading local tag data. This achieves full-cycle offline autonomous operation of the distribution area.

[0047] Example 2:

[0048] In one possible implementation, obtaining the pre-configured area topology and static files includes: during the initialization phase of the smart terminal deployment, importing the initial area topology and static files to the local terminal in a batch; during the normal operation phase, synchronizing a small amount of user change information to the local smart terminal in an incremental caching manner according to a preset daily synchronization cycle, so as to maintain the weak dependence of core profile computing on the cloud main station and offline autonomy.

[0049] Accordingly, a synchronization mechanism for the transformer area topology and static files was designed to address the balance between limited terminal memory and data timeliness. During the initialization phase of the transformer area's smart terminal deployment, the system loads the full physical topology connection relationships and electricity consumption profiles of all users within its jurisdiction into local solid-state storage via on-site debugging interfaces or a one-time command, thus constructing the initial operating environment. During normalized operation, the terminal only actively requests or receives a small amount of user change information from the cloud master station when the preset daily synchronization cycle arrives. This includes the addition of new meter binding relationships, topology changes caused by line upgrades, or changes in user payment and credit status. This incremental caching method cuts off the strong dependence on real-time data streams, enabling the local system to continue performing core profile calculations based on the previous day's historical cached data even when master station communication is interrupted.

[0050] According to one aspect of this application, a pre-configured transformer area topology and static profile, as well as real-time collected power consumption sensing device measurement data, are obtained. Time-series alignment and abnormal data cleaning are performed locally on a smart terminal to obtain a standardized fusion dataset. Specifically, this includes: recursively updating the data statistical distribution variables based on the power consumption sensing device measurement data to maintain its timeliness, calculating an adaptive anomaly judgment interval based on the recursive mean and recursive standard deviation, and removing jump data in the power consumption sensing device measurement data that exceeds the adaptive anomaly judgment interval to obtain cleaned measurement data.

[0051] Meanwhile, after receiving real-time power consumption sensing device measurement data, a conventional linear interpolation algorithm with a 5-minute time window is used to align the time axis of the multi-frequency data, ensuring that all electrical parameters are at a unified spatiotemporal reference. To address potential sensor glitches or communication errors in the measurement data, an adaptive anomaly detection interval based on the recursive mean and recursive standard deviation is constructed for data cleaning.

[0052] Specifically, for each measurement channel, newly acquired measurement values ​​are continuously accumulated over time to update the recursive mean, and the recursive standard deviation is updated based on the squared deviation between the current value and the recursive mean. On this basis, boundaries for anomaly detection intervals are set: the upper limit is set to the recursive mean + 4 * recursive standard deviation, and the lower limit is set to the recursive mean - 4 * recursive standard deviation. When the current measurement value falls outside this interval, the system determines the data as a jump and removes it, employing a previous value preservation method, i.e., using the normal historical value from the previous time series to replace the current abnormal missing value, generating cleaned measurement data with temporal continuity. In some optional implementations, the detection interval can be adjusted by changing the multiplier of the standard deviation to adapt to the noise distribution characteristics of different stations or different measurement channels.

[0053] As an optional implementation, a hierarchical compression mechanism is used to write the cleaned measurement data locally to the smart terminal to obtain a standardized fusion dataset. The hierarchical compression mechanism includes a short-term rolling storage layer for storing the original data, a summary retention layer for storing the feature statistics generated by the daily aggregation of the original data, and a long-term profile feature retention layer for storing the behavioral baseline data generated by the monthly compression of the feature statistics.

[0054] Optionally, to address the limited local storage space of smart terminals, a layered compression mechanism is constructed to achieve persistent data retention. For the short-term rolling storage layer of raw data, the system only saves the 5-minute high-frequency raw measurement sequence within the last 7 days, and cyclically overwrites the old data according to the time sequence, meeting the analysis conditions for short-term instantaneous load analysis and power quality traceability. For the feature statistics summary retention layer, the system extracts and saves the feature statistics summary generated daily for 90 days. This summary covers aggregated indicators such as average daily load, peak-to-valley ratio, and photovoltaic self-consumption rate, reducing the storage occupation of a single user to the level of hundreds of KB. For the long-term profile feature retention layer of behavioral baseline, the system extracts the typical daily load curve or seasonal feature fingerprint vector of the user compressed monthly and retains it in the terminal for a long time. Its space occupation is further compressed to the level of 50KB per household, which is used for cross-seasonal electricity consumption pattern comparison.

[0055] Through the above three-layer storage architecture, the system can compress and store months of historical operating status data of users in the distribution area into the available storage space of the terminal without expanding the edge hardware capacity, thus obtaining a stable and verifiable standardized fusion dataset.

[0056] Example 3:

[0057] In one possible implementation, the extraction of time-series dynamic features from indicators including static basic, time-series dynamic, and transformer area interaction dimensions is carried out locally on the smart terminal. The extraction of time-series dynamic features includes: based on a standardized fusion dataset, using a multi-time-scale exponential decay feature memory mechanism, performing feature smoothing recursion on historical electricity consumption behavior indicators to obtain time-series dynamic features.

[0058] As an optional implementation, edge computing terminals have limited memory, and storing long-term continuous historical measurement data incurs disk overhead. Utilizing a multi-timescale exponential decay feature memory mechanism, the current measurement value can be merged and updated with the smoothed recursive state of the previous moment without retaining the original data details. This type of feature smoothing recursion operation achieves weighted preservation of historical information. By assigning higher weights to near-end data and lower weights to far-end data, it extracts dynamic evolution patterns reflecting changes in electricity consumption habits and uses the results as time-series dynamic features for subsequent classification and early warning calculations.

[0059] Furthermore, short-term and long-term attenuation coefficients are set separately, and a dual-time-constant recursive calculation is performed on each feature dimension to obtain the short-term and long-term feature estimates for the corresponding dimensions. The day is divided into multiple time slots and the day type is distinguished. The historical electricity consumption behavior indicators of the corresponding time slots are updated using an exponential recursive method to construct a typical daily load fingerprint for a single user. The deviation between the short-term and long-term feature estimates is compared. When the deviation exceeds a preset fluctuation threshold, a seasonal offset index is generated to indicate seasonal behavior switching. The short-term feature estimates, long-term feature estimates, typical daily load fingerprints, and seasonal offset index are combined to obtain the time-series dynamic features.

[0060] In specific feature engineering processing, to simultaneously capture both short-term load fluctuation patterns and long-term electricity consumption evolution trends, two sets of independent time constants are used to recursively calculate the same feature dimension. The current moment's raw feature data is weighted and summed with the previous moment's short-term feature estimate to obtain the current moment's short-term feature estimate; similarly, the current moment's raw feature data is weighted and summed with the previous moment's long-term feature estimate to obtain the current moment's long-term feature estimate. In this embodiment, the short-term attenuation coefficient is set to 0.1 to capture the recent electricity consumption fluctuation trend; the long-term attenuation coefficient is set to 0.001 to track the cross-seasonal electricity consumption baseline evolution pattern. Those skilled in the art can adaptively adjust the above attenuation coefficients according to the actual data update cycle of the transformer area and the response time constant required by the service. The calculation formula for the above process is as follows:

[0061] μ s (t)=α s *x(t)+(1-α s )*μ s (t-△t);

[0062] Where, μ s (t) represents the short-term characteristic estimate at the current time, α s Let μ be the short-term decay coefficient, x(t) be the original data of the feature dimension at the current time, and μ be the short-term decay coefficient. s (t-△t) is the short-term characteristic estimate of the previous time step.

[0063] μ l (t)=α l *x(t)+(1-α l )*μ l (t-△t);

[0064] Where, μ l (t) represents the long-term characteristic estimate at the current time, α l Let μ be the long-term decay coefficient, x(t) be the original data of the feature dimension at the current moment, and μ be the long-term decay coefficient.l (t-△t) is the long-term characteristic estimate of the previous time step.

[0065] Furthermore, to preserve users' intraday time-sharing patterns and periodic differences, the entire day is divided into 288 five-minute time slots, and the day type is categorized into weekdays, rest days, and special days. For each time slot of each day type, an exponential update method is used to calculate the characteristic distribution state, constructing a typical daily load fingerprint for a single user. The specific calculation logic involves weighted fusion of the data collected in the current time slot with the state values ​​of the same day type and time slot from historical data. The calculation expression is as follows:

[0066] F d,k (t)=β*x(t)+(1-β)*F d,k (t prev );

[0067] Among them, F d,k (t) represents the historical electricity consumption behavior index corresponding to the day type and time slot, β is the fingerprint update rate, x(t) is the currently collected measurement data, and F d,k (t prev The last updated status value for the same day type and time slot is 0.

[0068] After obtaining the estimated short-term and long-term states, the absolute difference between the short-term and long-term characteristic estimates is calculated. This absolute difference is then compared to the sum of the long-term characteristic recursive standard deviation and the zero-prevention constant to generate the seasonality shift index. This index increases when a short-term characteristic undergoes a sudden change while the long-term characteristic has not yet followed suit, thus identifying seasonal transition periods such as summer cooling or winter heating. Based on the standard deviation calculation principle, the recursive standard deviation of the aforementioned long-term characteristics is obtained using conventional exponential recursive smoothing processing; the specific derivation is not detailed here. The formula for calculating the seasonality shift index is as follows:

[0069] D season =|μ s (t)-μ l (t)| / (σ l (t)+ε);

[0070] Among them, D season μ is the seasonal shift index. s (t) is the short-term characteristic estimate, μ l (t) is the long-term characteristic estimate, σ l (t) represents the recursive standard deviation of the long-term characteristic, and ε is the zero-prevention constant. In this embodiment, the preset fluctuation threshold is set to 2.0, that is, when the seasonal offset index D... seasonIf the value exceeds 2.0, the user is determined to be in a seasonal behavior transition period. Those skilled in the art can adjust this threshold based on the climatic and seasonal characteristics of the area where the transformer station is located.

[0071] In some alternative implementations, the density of time slot allocation can be dynamically adjusted in conjunction with the remaining hardware memory, for example, downgrading 288 time slots to 96 time slots to further relieve storage pressure.

[0072] Optionally, the specific steps for extracting the interaction features of the distribution area include: extracting the power supply path of each user based on the pre-configured distribution area topology, and evaluating the cross-influence of each user's electricity consumption behavior on the public power grid nodes of the distribution area by combining the measured parameters in the standardized fusion dataset, and using the cross-influence as the interaction feature of the distribution area.

[0073] Because of the topological connectivity of the physical power grid, the electricity consumption behavior of a single user can be transmitted radially to the source end of the distribution area through the feeder circuit. From the pre-configured distribution area topology ledger, all power supply branches and their connections from the low-voltage outgoing line of the transformer to the individual user's electricity meter are extracted to establish the user's feeder path. Standardized measured parameters are used to quantitatively analyze the common losses and electrical disturbances generated by the user on the feeder path, and the obtained cross-influence quantities are output to provide spatial topological cross-verification features for subsequent abnormal electricity consumption location and power quality tracing.

[0074] Optionally, assess the cross-impact of each user's electricity consumption behavior on the public power grid nodes in the distribution area, specifically including:

[0075] Obtain the impedance parameters and total current of each branch on the feeder path to determine the branch technical line loss. Then, allocate the branch technical line loss along the feeder path according to the ratio of the measured current of a single user to the total current of the branch, obtaining the line loss allocation contribution for a single user. Accumulate the impedance parameters of each branch along the feeder path to obtain the equivalent impedance of the feeder path. Use the product of the measured current of a single user and the equivalent impedance of the feeder path as the marginal voltage drop, and calculate the ratio of the marginal voltage drop to the preset rated bus voltage to obtain the voltage disturbance contribution for a single user. Combine the line loss allocation contribution and the voltage disturbance contribution as the cross-influence quantity.

[0076] When calculating the cross-effect, conventional power loss calculation methods in physics are used to obtain the branch technical line loss. Since branches within a transformer area often carry the superimposed current of multiple users, the impedance parameters of each common branch on the path are extracted, and combined with the measured square of the total branch current, the total technical line loss power of each physical branch is calculated.

[0077] Furthermore, the ratio of the measured current of a single user to the total current of the branch is calculated, and the power loss of the corresponding branch is allocated to the user based on this ratio. By traversing all associated branches on the feeder path and accumulating the allocated amounts, the line loss contribution of a single user is calculated. This method maps the macroscopic high loss situation of the distribution area to the allocation of individual terminal responsibilities.

[0078] In assessing the impact of voltage disturbances, the marginal contribution method is employed. The marginal voltage drop is calculated by multiplying the measured current of a single user by the equivalent impedance of the feeder path. This marginal voltage drop is then divided by the preset rated bus voltage on the low-voltage side of the transformer to obtain the percentage of voltage undervoltage caused by that single user load, i.e., the voltage disturbance contribution. The branch line losses and marginal voltage drops mentioned above are calculated using well-known Ohm's law and Joule's law; their mathematical equations are not listed here. In some optional implementations, for three-phase power supply areas, a calculation branch for the three-phase imbalance contribution can be added based on the current asymmetry of different phases, and this branch can be merged into the cross-influence calculation.

[0079] Example 4:

[0080] In one possible implementation, threshold adaptive calibration and risk rating calculation are performed locally on the smart terminal to obtain multi-dimensional user digital tags, such as... Figure 2 As shown, the specific steps for threshold adaptive calibration include:

[0081] Step 401: Based on user stratification features, a local quantile online estimation mechanism is used to dynamically track the feature distribution status of the user group in the transformer area, thereby adaptively generating an adaptive stratification threshold suitable for the current transformer area.

[0082] Different transformer substations exhibit varying baseline electricity consumption levels among their user groups; for example, urban commercial substations typically consume significantly more electricity than rural residential substations. Using fixed, empirically-based thresholds uniformly distributed from a cloud-based main station for user profiling can lead to a large number of users in some substations being misclassified as abnormal. To mitigate the assessment errors caused by differences in the absolute values ​​of characteristics across different regions, a local quantile online estimation mechanism is employed. This allows intelligent terminals to independently track the statistical distribution of characteristics within their current jurisdiction in an unsupervised manner. By extracting statistics that characterize data dispersion and centrality, a dynamic evaluation scale is adaptively generated, enabling the labeling thresholds to automatically adapt to the size and characteristic dispersion of the substation.

[0083] According to one aspect of this application, a local quantile online estimation mechanism is used to dynamically track the characteristic distribution of user groups in a transformer substation, thereby adaptively generating an adaptive grading threshold suitable for the current substation and determining the label level of each user, specifically including:

[0084] Step 401a: For each feature dimension, use the newly accessed user hierarchical features as sample values, and perform incremental iterative calculations based on the adaptive step size determined by the current feature data dispersion to track the quartile estimate of that feature dimension in real time.

[0085] As an optional implementation, to achieve continuous distribution state tracking in edge devices with limited memory and avoid storing all historical feature samples in memory for sorting, an iterative calculation method based on incremental update logic is adopted. Whenever a new user hierarchical feature is generated, it is input as a sample value into the recursive equation. During the update process, an adaptive step size parameter is constructed, which is positively correlated with the data dispersion of the current feature set, to adjust the update rate and prevent the estimated value from diverging when the sample fluctuates. Independent incremental state iterations are performed for the target quantiles of 25%, 50%, and 75%, respectively, to obtain the corresponding quartile estimates. The calculation formula for the above recursive update is as follows:

[0086] Q p (t)=Q p (t-1)+η*(sgn(x new -Q p (t-1))+2*p-1);

[0087] Among them, Q p Q(t) is the target quartile estimate at the current time. p (t-1) is the target quartile estimate at the previous time step, η is the adaptive step size parameter, sgn is the sign function for extracting symbols, and x new The sample values ​​are the stratified feature values ​​of newly connected users, and p is the preset probability value of the target quantile.

[0088] Step 401b: Calculate the interquartile range based on the tracked interquartile estimates, and combine the preset grading coefficients with the median in the interquartile estimates to construct an adaptive grading threshold for the corresponding feature dimension.

[0089] After obtaining the quartile estimates, the 75th quartile estimate is subtracted from the 25th quartile estimate to calculate the interquartile range. This interquartile range reflects the concentration of most user characteristics within the distribution area and has better resistance to outlier interference than the standard deviation.

[0090] Furthermore, using the 50th percentile (median) as the group baseline, a multi-level stepped threshold boundary is constructed by setting monotonically increasing grading coefficients for different levels. For example, the upper limit threshold for low-risk assessment can be set as the sum of the median and the low-level grading coefficient multiplied by the interquartile range, and the upper limit threshold for medium-risk assessment can be set as the sum of the median and the medium-level grading coefficient multiplied by the interquartile range. The relevant calculation formulas are as follows:

[0091] θ level =Q 50 +k level *IQR;

[0092] Where, θ level Q is the adaptive grading threshold for the corresponding level. 50 k is the median of the interquartile estimates. level The preset grading coefficient is used, and IQR is the interquartile range.

[0093] Step 401c: Calculate the deviation of each user's user stratification features from the median on the corresponding feature dimension, use the ratio of the deviation to the interquartile range as the standardized score, and complete the label level classification for each user based on the standardized score and the adaptive grading threshold.

[0094] To achieve the universality of the cross-regional data model, a standardized mapping of the data is achieved by calculating the ratio of the deviation to the interquartile range, thus eliminating the absolute differences in the original physical dimensions of the characteristics. A minimal zero-prevention constant is introduced during the calculation process to prevent computational overflow caused by a zero denominator.

[0095] A simplified normalized example is used for qualitative explanation. Assume that after previous iterations, the estimated median value for a certain feature dimension of the current station area is 0.5, and the interquartile range is 0.2. If the feature sample value of a new user is 0.8, its deviation from the median is calculated to be 0.3. The ratio of 0.3 to the interquartile range of 0.2 yields a standardized score of 1.5. The formula for calculating the standardized score is as follows:

[0096] Score x =(xQ 50 ) / (IQR+ε);

[0097] Among them, Score x For the standardized score corresponding to the feature dimension, x is the user's feature sample value, and Q is the score. 50 ε is the median estimate, IQR is the interquartile range, and ε is the zero protection constant.

[0098] Step 402: Evaluate the user stratification features of the corresponding dimension based on the adaptive grading threshold, determine the label level of each user, and generate multi-dimensional digital labels for users.

[0099] Accordingly, after obtaining standardized scores for all users, a hard decision on the feature level is performed by comparing the standardized scores with an adaptively tiered threshold. For example, if the standardized score is below the first-level coefficient, it is classified as normal; if it falls between the first and second-level coefficients, it is classified as concerning; and if it exceeds the third-level coefficient, it is classified as abnormal. The above calibration and classification process is repeated independently for multiple dimensions such as user electricity consumption behavior, distributed energy utilization rate, and power quality. The classification results of each dimension are then structured and encapsulated to generate localized multi-dimensional user digital tags that do not rely on externally issued parameters. In some embodiments, for physical feature indicators with clear safety red lines, the adaptive classification threshold can be subjected to OR or logical operations with fixed hard safety limit thresholds to ensure the basic operational safety of the equipment.

[0100] Example 5:

[0101] In one possible implementation, threshold adaptive calibration and risk rating calculation are performed locally on the smart terminal to obtain multi-dimensional user digital tags, such as... Figure 3 As shown, the specific steps for risk rating calculation include:

[0102] Step 501: Extract multidimensional anomaly indicators from user segmentation features; use a multidimensional evidence accumulation risk assessment model to perform time-dimensional evidence accumulation and multi-source fusion calculation on the multidimensional anomaly indicators to obtain risk audit labels in multidimensional user digital tags.

[0103] In traditional single-threshold anomaly detection mechanisms, occasional data disturbances can easily lead to false alarms. This embodiment employs a multi-dimensional evidence accumulation risk assessment model to extract independent-dimensional anomaly indicators for joint evaluation. Specifically, it extracts five categories of anomaly indicators from user-stratified characteristics, including load and power consumption deviations, abnormal line loss contributions, abnormal voltage and current phases, abrupt changes in power consumption patterns, and power outage and restoration behaviors. These anomaly indicators, distributed across different physical attributes, are transformed into a unified strength of evidence, and integral accumulation is performed over time. The confidence level of the status determination is obtained by trading time span, generating a credible risk audit label.

[0104] Step 501a: Assign multidimensional abnormal indicators to multiple preset independent evidence channels, and use a mapping function to convert the degree of abnormal deviation of each independent evidence channel into the corresponding single-channel evidence strength.

[0105] Optionally, after obtaining multidimensional anomaly indicators, they are mapped and input into corresponding independent evidence channels. Each channel operates independently, and the degree of anomaly deviation with different physical dimensions is processed into dimensionless form using a predefined mapping function. This transformation process is used to improve the dimensional differences between different parameters such as voltage deviation percentage and active power difference, unifying them into a weight space within a fixed range, serving as the basis for subsequent multi-source data fusion and overlay calculations.

[0106] Step 501a1: Configure sensitivity parameters and evidence activation thresholds for each independent evidence channel; use a nonlinear mapping function with an S-shaped curve distribution to convert the sensitivity parameters and evidence activation thresholds, so that when the abnormal deviation is lower than the evidence activation threshold, the evidence strength of a single channel approaches zero, and when it is higher than the evidence activation threshold, the evidence strength of a single channel approaches the full weight limit.

[0107] Based on this, to suppress system malfunctions caused by minor data disturbances, the Sigmoid function, which possesses nonlinear characteristics, is used as the mapping function. During the transformation process, a sensitivity parameter controls the steepness of the curve's state transition, and an evidence activation threshold defines the trigger point for abnormal states. This logic filters normal operational fluctuations through a nonlinear interval, and once a set threshold is exceeded, the strength of the output evidence rapidly increases and approaches a saturation limit. The calculation formula is as follows:

[0108] w j (t)=1 / (1+exp(-γ j *(e j (t)-τ j )));

[0109] Among them, w j (t) represents the single-channel evidence strength of the j-th independent evidence channel, exp represents the natural exponent operation, and γ j e represents the sensitivity parameter for the corresponding channel. j (t) represents the degree of abnormal deviation at the current moment, τ j This is the evidence activation threshold for the corresponding channel.

[0110] Step 501b introduces a time decay factor to perform decay processing on the historical cumulative evidence amount of the previous period, and then superimposes the decayed historical cumulative evidence amount with the weighted sum calculated based on the pre-configured channel weights of the single-channel evidence strength of each independent evidence channel in the current period to obtain the comprehensive cumulative evidence amount at the current moment.

[0111] Accordingly, to ensure that historical anomalies fade over time and prevent the continuous accumulation of risk, a recursive integral model incorporating a decay factor is constructed. For the current calculation period, the system retrieves the historical accumulated evidence from the previous period and multiplies it by a time decay factor less than 1 to complete the decay operation.

[0112] Simultaneously, the single-channel evidence strength output from all channels is multiplied by the pre-configured channel weights and summed to obtain the incremental evidence for the current period. The attenuated historical evidence is then added to the incremental evidence for the current period to update the overall cumulative evidence amount at the current moment.

[0113] Taking a simplified normalized example, assuming the historical cumulative evidence amount of the previous period is 2.0, the set time decay factor is 0.9, and the weighted sum of the evidence strength of each channel calculated in this period is 0.5, substituting into the formula, the comprehensive cumulative evidence amount at the current moment is calculated to be 2.0*0.9+0.5=2.3. This calculation process presents the physical mechanism of the gradual fading of past abnormal states over time, and its calculation expression is as follows:

[0114] E(t) = λ * E(t - Δt) + Σ(ω j *w j (t));

[0115] Where E(t) is the total accumulated evidence at the current time, λ is the time decay factor, E(t-Δt) is the historical accumulated evidence in the previous period, Σ represents the summation operation of the weighted product of all channels, and ω j Let w be the channel weight of the j-th independent evidence channel. j (t) represents the strength of evidence for a single channel.

[0116] Step 501c: Compare the total accumulated evidence with the pre-configured risk level threshold, determine the user's risk level based on the comparison results, and use it as a risk audit label.

[0117] Furthermore, after generating the total accumulated evidence, it is input into the tiered comparison logic unit. Multiple risk level thresholds are preset, for example, setting separate judgment boundaries for low, medium, and high levels. When the total accumulated evidence is below the low-level threshold, the judgment output is "normal"; when it falls between different boundary ranges, it is mapped to the corresponding "concerned," "medium-risk," or "high-risk" levels, respectively. Through the above integration and comparison mechanism, a single disturbance, lacking sufficient time accumulation, cannot trigger a risk warning, while persistent, weak anomalies, after accumulating through multiple calculation cycles, exceed the threshold and trigger the corresponding risk level judgment, thereby outputting a highly credible risk audit label.

[0118] Example 6:

[0119] In contrast, in a basic alternative implementation, the terminal performs a full-coverage update of incremental data at a fixed 5-minute cycle and performs a global recalculation of long-cycle features once every morning. Although this basic approach meets the real-time requirements of the data, the high-frequency full hard drive write operation will accelerate the hardware wear of the solid-state flash memory of the edge device and generate a lot of invalid computing overhead during periods of stable power consumption, such as at night.

[0120] Based on this, this embodiment provides a preferred implementation method, which performs iterative update hierarchical scheduling locally on the smart terminal based on the changing status of multi-dimensional user digital tags to obtain updated user digital profile data, such as... Figure 4 As shown, it includes the following steps:

[0121] Step 601: Extract the deviation of each tag from the historical records in the multi-dimensional user digital tags, perform joint detection of tag change significance, and obtain multi-dimensional joint change significance index.

[0122] As an optional implementation, to address the hardware wear and tear caused by high-frequency writing, this embodiment introduces a significance-based filtering mechanism. After obtaining the newly calculated scores for each dimension of the label in the current cycle, the system does not immediately write them all to memory. Instead, it retrieves the previous historical snapshot data stored in memory and quantitatively extracts the numerical fluctuation amplitude of each label dimension by comparing the current data with the historical data. By dividing this fluctuation amplitude by the data reflecting the historical fluctuation background, conventional data noise interference is eliminated, and a multi-dimensional joint judgment is made on the significance of changes in each label, outputting indicators to guide subsequent update actions.

[0123] Step 601a: Compare the current score of a single tag dimension in the multi-dimensional user digital tags with the record value written to the previous historical snapshot, and calculate the historical recursive standard deviation of the single tag dimension by combining the historical score of the single tag dimension in a recursive manner, and calculate the single tag change significance index for each tag dimension.

[0124] In the specific extraction process, for each tag dimension, the absolute difference between its current period's original score and the record value from the last snapshot write operation is calculated. Further, the historical recursive standard deviation maintained synchronously during the initial data processing phase for that single tag dimension is retrieved. The ratio of the absolute difference to the sum of the historical recursive standard deviation and the preset zero-prevention constant is used to obtain the significance index for that dimension. The formula for this calculation process is as follows:

[0125] S j (t)=|v j (t)-v j (t last )| / (σj_hist +ε);

[0126] Among them, S j (t) represents the significance index of single-label change in the j-th label dimension, v j (t) represents the current score for this dimension in the current calculation period, v j (t last ) represents the record value previously written to the history snapshot, σ j_hist ε is the historical recursive standard deviation of this label dimension, and ε is the zero protection constant.

[0127] Step 601b: Extract the maximum value from the single-label change significance index of each label dimension, and use the maximum value as the multi-dimensional joint change significance index.

[0128] Simultaneously, after obtaining the independent significance indicators for all label dimensions, it is necessary to extract a comprehensive indicator representing the overall state. Here, a maximum value extraction strategy is used instead of an arithmetic mean strategy to preserve boundary abrupt changes. The system iterates through the calculation results for all label dimensions and selects the largest value as the multi-dimensional joint change significance indicator. This strategy ensures that sudden changes in a single dimension, such as the input of high-power equipment or abrupt changes in photovoltaic output, are reflected in the comprehensive indicator without attenuation, avoiding the loss of abrupt change characteristics due to local changes being averaged by low values ​​of other stable dimensions. The extraction process formula is as follows:

[0129] S joint (t)=max(S1(t),S2(t),...,S n (t));

[0130] Among them, S joint (t) is the significance index of multidimensional joint change, max is the maximum value extraction function, and S1(t)~S n (t) represents the significance index of single-label changes for each independent label dimension.

[0131] Step 602: Compare the multidimensional joint change significance index with the preset hierarchical scheduling threshold, and perform differentiated update scheduling actions based on the comparison results to generate updated user digital profile data as needed.

[0132] Accordingly, after obtaining the multidimensional joint change significance index, it is input into a decision module containing three levels of judgment logic for conditional scheduling. Since this embodiment involves differentiated action branches under different data states, it is described in detail here as three execution paths.

[0133] In the first execution path, the multidimensional joint change significance index is compared with a preset hierarchical scheduling threshold. When the multidimensional joint change significance index exceeds the upper limit threshold, a scheduling action of immediately updating and storing the profile snapshot is executed. Specifically, when the multidimensional joint change significance index is detected to exceed the upper limit threshold (the preset value can be set to 2.0), it indicates that the user's electricity consumption behavior has changed abruptly. At this time, the system schedules computing resources to update the memory data and triggers a hard disk write instruction to generate and save a profile snapshot with a timestamp to ensure the traceability of the abrupt change.

[0134] In the second execution path, when the metric falls between the middle and high thresholds, an immediate update is performed without storing a snapshot. Specifically, when the metric is between the middle threshold (preset to be 0.5) and the high threshold, it indicates a moderate level of gradual change in user behavior. In this case, the system only updates the variable data in memory, ensuring that the tags read by the local control service are up-to-date, but blocking disk write operations to reduce the number of erase / write cycles on the solid-state flash memory.

[0135] In the third execution path, when the value is below the mid-threshold threshold, a scheduling action is executed to skip the update for the current cycle. Specifically, during periods of stable user electricity consumption, such as late-night off-peak hours, the multi-dimensional joint change significance index is usually below the mid-threshold threshold. At this time, the system determines that the features have not shifted, suspends the feature extraction and label calculation tasks for the current cycle, and skips all update actions. Based on statistical patterns, this scheduling action can reduce the amount of invalid computation during periods of stable electricity consumption and free up edge computing resources for smart terminals in the distribution area.

[0136] In one possible implementation, the updated user digital profile data is output locally on the smart terminal, including:

[0137] Optionally, the serial communication interface and edge microservice interface are locally opened on the smart terminal in the distribution area; in the offline state where the communication link with the main station is disconnected, the updated user digital profile data is output to the external operation and maintenance equipment locally through the serial communication interface or edge microservice interface.

[0138] For the generated profile data, a localized output channel is configured to support the closed-loop business at the edge. At the hardware and protocol level, a physical serial communication interface is configured on the smart terminals in the distribution area, specifically using industry-standard RS-485 or RS-232 interfaces for wired direct reading. Simultaneously, at the software level, an edge microservice interface based on a lightweight protocol is deployed. When the uplink communication module of the distribution area fails, causing the communication link with the cloud master station to be lost, external portable maintenance tools or edge control devices can access the latest profile data stored in the terminal's memory using the aforementioned local interface. This mechanism cuts off the strong network dependence of the data consumer on the master station, maintaining the continuous operation of on-site services.

[0139] Furthermore, after external maintenance equipment or control nodes acquire updated user digital profile data, they combine it with feature tags of different dimensions to carry out local business operations. In the source-grid-load control scenario, the local control program reads the control adaptation tags in the profile, generates a scheduling strategy based on the peak-shaving potential indicated by the tags, and sends it to the corresponding load equipment to complete the local load reduction during peak electricity consumption periods in the distribution area. In the on-site line loss investigation scenario, the maintenance equipment carried by the operators reads the line loss contribution tags of each user through the serial communication interface, and locates the high-loss power supply branches and corresponding users on-site according to the tag values. In the local abnormal electricity consumption early warning scenario, the intelligent terminal independently generates an abnormal status message based on the risk audit tags and pushes it locally to the distribution area maintenance terminal for display. The above process controls the business flow at the edge, avoiding data security issues caused by cross-network transmission of electricity consumption details.

[0140] Accordingly, the generated tag data, generated through the aforementioned method of local feature extraction and adaptive classification on the terminal, possesses reliable classification characteristics under predetermined application conditions. For structured tags such as residential electricity consumption classification, industrial and commercial electricity consumption classification, and charging behavior pattern recognition, verification was conducted based on measured data from typical transformer substations in a controlled testing environment. The controlled testing environment verification demonstrated that the classification results possess the reliability to meet the needs of on-site business applications. This indicator shows that, under conditions without master station computing power, the edge terminal autonomous architecture, through local historical data recursion and multi-dimensional evidence cross-verification, can achieve classification results close to those of centralized cloud computing, meeting the closed-loop usage requirements of on-site businesses.

[0141] Meanwhile, this invention provides a feature analysis and mining technology for end-side closed-loop autonomous operation. By constructing a localized archive incremental micro-synchronization and hierarchical feature extraction pipeline, the strong dependence of analysis and computation on the cloud network is reduced, enabling the terminal to maintain continuous operation of on-site source-network-load control and anomaly troubleshooting services through open local interfaces even when the main station's uplink communication fails.

[0142] Specifically, to address the limitation of limited storage space on edge devices, a multi-timescale exponential decay feature memory mechanism is introduced. This eliminates the need to retain high-frequency original detailed sequences for extended periods, requiring only a small amount of constant-level memory overhead to track users' long-term, medium-term, and short-term electricity consumption patterns and seasonal mutation behaviors.

[0143] Furthermore, in the face of complex background scale differences in unsupervised environments, this method relies on incremental quantile tracking to achieve adaptive dynamic calibration of the evaluation tier threshold, suppressing large-scale group misjudgments caused by fixed empirical thresholds. Simultaneously, combined with a time-decay integral model of multi-dimensional channel evidence, it effectively smooths single-channel measurement noise and suppresses false alarms in fault warnings caused by instantaneous electrical disturbances. Assisted by a joint scheduling strategy based on the significance of tag changes, it intelligently suspends and intercepts invalid data writing tasks during periods of stable power consumption, reducing the frequency of erasing and writing edge solid-state flash memory and freeing up terminal computing power, thereby improving the data processing efficiency and operational robustness of the underlying autonomous system of the distribution area under complex and harsh operating conditions.

[0144] It should be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A method for constructing a digital user profile based on multi-source data fusion from smart power terminals in distribution areas, characterized in that, This method operates in a closed loop locally on the smart terminal in the distribution area, including: The system acquires pre-configured transformer area topology and static files, as well as real-time power consumption sensing device measurement data, and performs time-series alignment and abnormal data cleaning on the smart terminal to obtain a standardized fusion dataset. Based on a standardized fusion dataset, indicators including static basic, time-series dynamic and station area interaction dimensions are extracted locally on the smart terminal to obtain user segmentation characteristics. Based on user segmentation characteristics, threshold adaptive calibration and risk classification calculation are performed locally on the smart terminal to obtain multi-dimensional digital user tags. Based on the changing status of multi-dimensional user digital tags, iterative updates and hierarchical scheduling are performed locally on the smart terminal to obtain updated user digital profile data. The updated user digital profile data is output locally on the smart terminal.

2. The method according to claim 1, characterized in that, The system acquires pre-configured transformer area topology and static profiles, as well as real-time power consumption sensing device measurement data. Time-series alignment and outlier cleansing are then performed locally on the smart terminal to obtain a standardized fusion dataset, specifically including: Maintain the statistical distribution variables of the data, calculate the adaptive anomaly judgment interval based on the recursive mean and recursive standard deviation, remove the jump data in the power sensing equipment measurement data that exceeds the adaptive anomaly judgment interval, and obtain the cleaned measurement data. A hierarchical compression mechanism is used to write the cleaned measurement data locally to the smart terminal to obtain a standardized fusion dataset. The hierarchical compression mechanism includes a short-term rolling storage layer for the original data, a summary retention layer for feature statistics, and a long-term profile feature retention layer for behavioral baselines.

3. The method according to claim 1, characterized in that, The specific steps for extracting temporal dynamic features from metrics including static basic, temporal dynamic, and station interaction dimensions extracted locally on the smart terminal include: Based on a standardized fusion dataset, a multi-time-scale exponential decay feature memory mechanism is used to perform feature smoothing and recursive operations on historical electricity consumption behavior indicators to obtain time-series dynamic features.

4. The method according to claim 3, characterized in that, A multi-time-scale exponential decay feature memory mechanism is employed to perform feature smoothing and recursive operations on historical electricity consumption behavior indicators, yielding time-series dynamic features, specifically including: By setting short-term and long-term decay coefficients respectively, and performing recursive calculations with dual time constants for each feature dimension, the short-term and long-term feature estimates for the corresponding dimensions are obtained. The day is divided into multiple time slots and the day type is distinguished. The historical electricity consumption behavior indicators of the corresponding time slots are updated using an exponential recursive method to construct a typical daily load fingerprint for a single user. By comparing the deviation between short-term and long-term feature estimates, a seasonal offset index is generated to indicate seasonal behavior switching when the deviation exceeds a preset fluctuation threshold. By combining short-term feature estimates, long-term feature estimates, typical daily load fingerprints, and seasonal offset indices, time-series dynamic features are obtained.

5. The method according to claim 1, characterized in that, The specific steps for threshold adaptive calibration, which involves performing threshold adaptive calibration and risk rating calculation locally on a smart terminal to obtain multi-dimensional user digital tags, include: Based on user stratification features, a local quantile online estimation mechanism is used to dynamically track the feature distribution of user groups in the transformer area, thereby adaptively generating adaptive stratification thresholds suitable for the current transformer area. The user stratification features of the corresponding dimensions are evaluated based on adaptive grading thresholds to determine the label level of each user, so as to generate multi-dimensional digital labels for users.

6. The method according to claim 1, characterized in that, The specific steps for risk rating calculation, which involves performing threshold adaptive calibration and risk classification locally on the smart terminal to obtain multi-dimensional user digital tags, include: Extracting multidimensional anomaly indicators from user segmentation features; A multidimensional evidence accumulation risk assessment model is used to accumulate evidence over time and perform multi-source fusion calculations on multidimensional abnormal indicators to obtain risk audit labels in multidimensional user digital tags.

7. The method according to claim 6, characterized in that, A multidimensional evidence accumulation risk assessment model is used to accumulate evidence over time and perform multi-source fusion calculations on multidimensional abnormal indicators, resulting in risk audit tags from multidimensional user digital tags, specifically including: Multidimensional abnormal indicators are assigned to multiple preset independent evidence channels, and the abnormal deviation degree of each independent evidence channel is converted into the corresponding single-channel evidence strength using a mapping function. A time decay factor is introduced to decay the historical cumulative evidence amount of the previous period, and the decayed historical cumulative evidence amount is superimposed with the weighted sum of the single-channel evidence strength of each independent evidence channel in the current period to obtain the comprehensive cumulative evidence amount at the current moment. By comparing the total amount of accumulated evidence with the pre-configured risk level thresholds, the user's risk level is determined based on the comparison results and used as a risk audit label.

8. The method according to claim 1, characterized in that, Based on the changing status of multi-dimensional user digital tags, iterative updates and hierarchical scheduling are performed locally on the smart terminal to obtain updated user digital profile data, specifically including: Extract the degree of deviation of each tag from the historical records in the multi-dimensional user digital tags, perform joint detection of tag change significance, and obtain multi-dimensional joint change significance index; The multidimensional joint change significance index is compared with the preset hierarchical scheduling threshold, and differentiated update scheduling actions are executed according to the comparison results to generate updated user digital profile data on demand.

9. The method according to claim 7, characterized in that, The mapping function is used to convert the degree of abnormal deviation of each independent evidence channel into the corresponding single-channel evidence strength, specifically including: Configure sensitivity parameters and evidence activation thresholds for each independent evidence channel; A nonlinear mapping function with an S-shaped curve distribution is used to transform the sensitivity parameter and the evidence activation threshold. This makes the single-channel evidence strength approach zero when the degree of abnormal deviation is lower than the evidence activation threshold, and approach the full weight limit when it is higher than the evidence activation threshold.

10. The method according to claim 1, characterized in that, The updated user digital profile data is output locally on the smart terminal, specifically including: Locally open serial communication interfaces and edge microservice interfaces on smart terminals in the distribution area; In an offline state where the communication link with the main station is disconnected, the updated user digital profile data is output to external operation and maintenance equipment locally through a serial communication interface or an edge microservice interface.

Images