A method, system, device and medium for early warning of sudden power surges in low-voltage users
By constructing multi-dimensional user profiles and a two-layer machine learning model, and combining random forest, gradient boosting decision tree, and logistic regression, we have achieved accurate identification and cause analysis of sudden power surges in low-voltage users. This solves the problems of delayed early warning and high false alarm rate in existing technologies, and improves the initiative and accuracy of power supply services.
Patent Information
- Application Number
- CN202611150898.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for monitoring low-voltage user electricity consumption rely on manual periodic meter readings or single threshold monitoring, resulting in delayed early warning responses, high false alarm rates, and an inability to accurately identify the causes of anomalies, making it difficult to support proactive services and precise prevention and control.
Construct multi-dimensional user profiles, use random forest models to identify sudden increases in battery power, combine gradient boosting decision trees to analyze the causes, use logistic regression models for probability calibration, generate scientifically graded early warnings, and establish a closed-loop feedback mechanism from early warning to response.
It enables accurate identification and cause analysis of sudden power surges in low-voltage users, improves the accuracy of early warnings and the initiative of response, significantly reduces false alarm rate and response lag, and supports the implementation of differentiated service strategies.
Smart Images

Figure CN123118299A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power big data analysis and power supply service technology, and in particular to a method, system, device and medium for early warning of sudden power consumption increases for low-voltage users. Background Technology
[0002] The current jurisdiction has a large number of low-voltage users with significantly different electricity consumption scenarios. A sudden surge in user electricity consumption often conceals multiple risks, including abnormal power usage, equipment malfunctions, or electricity theft. However, existing monitoring methods remain relatively traditional, relying mainly on manual, periodic meter readings or single-dimensional monitoring of fixed electricity consumption thresholds. This lack of comprehensive consideration of multi-dimensional information, such as users' historical electricity consumption habits and load characteristics, leads to severe delays in early warning responses and fails to accurately identify the true causes behind sudden increases, making it difficult to meet the management requirements of "early intervention and proactive service" in modern power supply service systems.
[0003] Existing technical solutions mainly fall into two categories: post-event verification and simplified threshold warning, both of which have significant shortcomings. Post-event verification requires waiting for the electricity bill to be issued before anomalies can be detected, resulting in a severely delayed processing flow and a high likelihood of user complaints. Simplified threshold warning only sets fixed standards for electricity consumption increases, failing to incorporate in-depth analysis based on user profiles, leading to a high false alarm rate and weak targeting. Existing technologies cannot effectively distinguish between normal production electricity consumption increases and abnormal situations such as metering faults, line hazards, or electricity theft, making it difficult to support precise risk control and the implementation of differentiated service strategies. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a method, system, device, and medium for early warning of sudden power surges in low-voltage users to address the problems of existing monitoring methods, such as delayed early warning, high false alarm rate, inability to accurately identify the causes of anomalies, and difficulty in supporting proactive services and precise prevention and control.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for early warning of sudden power surges in low-voltage users, comprising: Acquire and standardize user electricity consumption data, equipment data, service data, and customer attribute data to form a multi-dimensional user profile; Core features are extracted from the multi-dimensional user profile and input into a two-layer machine learning model for analysis. The two-layer machine learning model includes a power surge identification model and a cause analysis model. Based on the output of the two-layer machine learning model, warning information of the corresponding level is generated and pushed. The corresponding service strategy is triggered based on the warning information to handle the situation, and the handling results are fed back as labeled samples for iterative optimization of the two-layer machine learning model.
[0007] As a preferred embodiment of the low-voltage user power surge warning method of the present invention, wherein: the power surge identification model is used to identify whether the power consumption has surged, and the cause analysis model is used to analyze the cause of the surge when it is identified as a surge; The power surge recognition model uses a random forest model to perform binary classification based on the input core features. The cause analysis model adopts a gradient boosting decision tree model to analyze the causes of sudden increases as normal electricity consumption growth, metering failure, line hazards, electricity theft, and other problems.
[0008] The beneficial effects of this preferred technical solution are that it achieves initial screening of anomalies through the high robustness of random forests, and performs cause subdivision by combining the high precision characteristics of gradient boosting decision trees, thereby effectively balancing the stability of power surge identification and the accuracy of attribution analysis.
[0009] As a preferred embodiment of the low-voltage user power surge warning method of the present invention, the step of generating and pushing warning information of corresponding levels based on the output of the two-layer machine learning model includes: Obtain the global anomaly probability output by the power surge identification model and the probability of various surge causes output by the cause analysis model; The global anomaly probability is concatenated with the probabilities of various sudden increase causes to form a feature vector, which is then input into the logistic regression model for calibration to obtain the comprehensive fusion probability. Based on the comprehensive fusion probability and the sudden increase causes output by the causal analysis model, early warning information of corresponding levels is generated.
[0010] The beneficial effect of this preferred technical solution is that by introducing a logistic regression model to splice and calibrate multi-level probabilistic features, the accuracy and reliability of the comprehensive assessment are effectively improved, thus providing solid data support for generating scientific and accurate hierarchical early warnings.
[0011] As a preferred embodiment of the low-voltage user power surge warning method of the present invention, the step of extracting core features from the multi-dimensional user profile includes: Initial features are extracted from the multidimensional user profile, and core features are obtained by joint screening using analysis of variance and mutual information. The core features include load volatility, equipment failure correlation, and historical request frequency. The load fluctuation rate is obtained by calculating the coefficient of variation of the load within a preset electricity consumption cycle; The correlation degree of the equipment failure is calculated by exponential decay weighting based on the severity and time distance of historical failures within a preset time window. The frequency of historical requests is calculated by weighting the request type and time.
[0012] As a preferred embodiment of the low-voltage user power surge warning method of the present invention, the step of generating warning information of corresponding levels based on the comprehensive fusion probability and the surge cause output by the causal analysis model includes: In response to the fact that the comprehensive fusion probability is not less than a first preset threshold and the cause of the sudden increase belongs to the first type of preset abnormal cause, a first-level early warning information is generated; In response to the fact that the comprehensive fusion probability is within the second preset threshold range and the cause of the sudden increase belongs to the second type of preset abnormal cause, a secondary warning message is generated; In response to the integrated probability falling within the third preset threshold range, a level-three early warning message is generated.
[0013] As a preferred embodiment of the low-voltage user power surge warning method of the present invention, the step of triggering the corresponding service strategy for handling based on the warning information includes: Upon receiving a Level 1 warning, an on-site handling strategy is triggered to conduct on-site verification within a preset response time limit; Upon receiving a Level 2 warning, a proactive reminder strategy is triggered to contact the individual by phone and schedule an investigation within a preset response time limit. Upon receiving a Level 3 early warning message, a tracking and guidance strategy for tracking user electricity consumption curves online is triggered.
[0014] As a preferred embodiment of the low-voltage user power surge warning method of the present invention, wherein: the step of using the processing results as labeled samples for iterative optimization of the two-layer machine learning model includes: Receive the labeled samples submitted after on-site handling, the labeled samples including fault labels and supporting materials; The labeled samples are verified, and the verified labeled samples are stored in the sample library; The samples in the sample library are periodically merged with the original dataset, and the two-layer machine learning model is iteratively trained based on the merged dataset.
[0015] Secondly, the present invention provides a low-voltage user power surge warning system, comprising: The user profile building module is used to acquire and standardize user electricity consumption data, equipment data, service data, and customer attribute data to form a multi-dimensional user profile. The dual-layer model analysis module is used to extract core features from the multi-dimensional user profile and input the core features into the dual-layer machine learning model for analysis. The dual-layer machine learning model includes a power surge identification model and a cause analysis model. The early warning generation and push module is used to generate and push early warning information of the corresponding level based on the output results of the two-layer machine learning model. The strategy handling and model iteration module is used to trigger the corresponding service strategy to handle the situation based on the warning information, and to feed back the handling results as labeled samples for iterative optimization of the two-layer machine learning model.
[0016] Thirdly, the present invention provides a computer device, comprising: Memory, used to store programs; A processor for executing computer-executable instructions, which, when executed by the processor, implement the steps of the low-voltage user power surge warning method.
[0017] Fourthly, the present invention provides a computer-readable storage medium, comprising the steps of implementing the low-voltage user power surge warning method when the program is executed by a processor.
[0018] The beneficial effects of this invention are as follows: By constructing a two-layer machine learning model that includes an identification layer and a cause analysis model, this invention utilizes random forests for initial binary classification screening of power surges and combines gradient boosting decision trees for refined classification of the causes of surges, achieving high robustness in anomaly identification and high accuracy in attribution analysis. This effectively overcomes the technical shortcomings of single models in balancing screening stability and diagnostic depth. By acquiring the output probabilities of the two-layer model and performing feature concatenation, and introducing a logistic regression model for probability calibration and comprehensive fusion, this invention achieves scientific classification and precise delivery of early warning information, solving the technical problem of difficulty in uniformly quantifying and evaluating the output results of multi-level models for decision-making. By establishing a closed-loop feedback mechanism from early warning triggering, strategy handling to result labeling and feedback, and periodically storing the labeled samples from on-site handling into a sample library and merging them with the original dataset for iterative model training, this invention achieves continuous self-optimization of the model based on real business data, significantly improving the system's adaptability to complex power consumption scenarios and its long-term operational accuracy. Attached Figure Description
[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1This is a basic flowchart illustrating a low-voltage user power surge warning method according to an embodiment of the present invention. Figure 2 A flowchart illustrating the judgment process of a low-voltage user power surge warning method according to an embodiment of the present invention; Figure 3 This is a timing diagram of a low-voltage user power surge warning method provided in one embodiment of the present invention. Detailed Implementation
[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0021] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for early warning of sudden power surges in low-voltage users is provided, comprising: S100: Acquire and standardize user electricity consumption data, equipment data, service data, and customer attribute data to form a multi-dimensional user profile; S200: Extract core features from the multi-dimensional user profile and input the core features into a two-layer machine learning model for analysis. The two-layer machine learning model includes a power surge identification model and a cause analysis model. S300: Based on the output of the two-layer machine learning model, generate and push warning information of the corresponding level; S400: Based on the warning information, the corresponding service strategy is triggered to handle the situation, and the handling result is fed back as a labeled sample for iterative optimization of the two-layer machine learning model.
[0022] It should be noted that the area has a large number of low-voltage users and complex electricity consumption scenarios. During operation, a sudden increase in user electricity consumption may pose multiple risks, such as equipment failure and electricity theft. However, existing monitoring methods mostly rely on manual periodic meter reading or single fixed thresholds, resulting in a serious lag in early warning response and an inability to accurately identify the causes of anomalies, making it difficult to support proactive services and precise prevention and control.
[0023] Therefore, in response to the problems of delayed early warning, high false alarm rate, and inability to accurately identify the causes of anomalies in existing monitoring methods, which make it difficult to support proactive services and precise prevention and control, the S100-S400 steps were used to construct multi-dimensional data profiles and a two-layer machine learning model. This enabled accurate identification and cause analysis of sudden increases in power consumption by low-voltage users. A closed-loop system covering the entire process from early warning and handling to model iteration and optimization was also established. This effectively solved the problems of delayed early warning, high false alarm rate, and passive service in existing technologies, and significantly improved the proactivity and accuracy of power supply services.
[0024] Example 2, refer to Figures 2-3 As an embodiment of the present invention, based on the previous embodiment, a method for early warning of sudden power surges in low-voltage users is provided, including: like Figure 2 As shown, the process begins with data collection standardization and core feature extraction, which are then input into a two-layer machine learning model consisting of a random forest (for identifying sudden increases) and a gradient boosting decision tree (for analyzing causes). Subsequently, a comprehensive fusion probability is obtained through logistic regression calibration, which is used to generate a level 1, 2, or 3 warning. Finally, differentiated handling strategies such as on-site verification, telephone appointment, or online monitoring are triggered according to the warning level, and the handling results are used as samples to iteratively optimize the model.
[0025] like Figure 3 As shown, the data layer transmits user data and features to the analysis layer for two-layer model calculations, and the resulting comprehensive fusion probability is pushed to the early warning layer for level determination. The early warning layer issues instructions to the service layer based on the determination results, triggering service actions such as on-site handling, proactive reminders, or follow-up guidance. Finally, the service layer feeds back the handling results to the optimization layer, completing the full-link data interaction of sample return and model iteration optimization.
[0026] In this embodiment, the method is implemented based on an overall architecture, which is mainly divided into four modules: a data layer, a model layer, an early warning layer, and a service layer. These modules operate collaboratively. Specifically, the data layer is used to collect and standardize all user data, providing a data foundation for backend algorithm calculations; the model layer is used to build machine learning models to identify abnormal power surges and automatically determine the underlying causes; the early warning layer is used to generate different levels of early warnings based on the model output; and the service layer is used to connect all power supply business processes, completing a closed-loop management system from order dispatch to handling and archiving.
[0027] In this embodiment of the application, step S100, which involves acquiring and standardizing data to form a multi-dimensional user profile, specifically includes: Four types of core multidimensional data are collected through a grid-level power energy data platform, metering automation system, customer service system, and marketing management system. Among them, electricity consumption data includes daily / hourly electricity consumption of low-voltage users, time-period load curves, power factor, load rate, historical electricity consumption during the same period, and electricity consumption fluctuation data for the past 3 months; equipment data includes the operating status of user metering devices, fault records of dedicated transformers / low-voltage lines, equipment inspection logs, and historical data on frequent power outages / low voltage; service data includes user historical request records, service work order handling ledgers, offline repair / complaint content, and service evaluation information; customer attribute data includes user type, electricity address, arrears status, payment records, and electricity usage scenarios.
[0028] After data collection, the multi-source heterogeneous data undergoes standardized cleaning, including: missing value imputation, outlier removal, and format standardization. Specifically, for missing values, short-term power consumption gaps are filled with the average of adjacent time periods, while long-term missing values are marked as invalid. For outliers, extreme erroneous data is removed using the 3σ principle. Data exported from various systems is uniformly converted into standard timestamps and numerical formats. Based on the user's unique account number, various data types are linked and integrated to form a unique data profile for each low-voltage user, which is then stored in a distributed database to obtain a complete standardized dataset. The entire system synchronizes all data in real-time on an hourly basis to ensure the timeliness of electricity consumption, equipment, and service information.
[0029] In an optional implementation, the abnormal data cleaning method in step S200 can also calculate the interquartile range by calculating the first and third quartiles of the data, thereby defining a reasonable data range consisting of "the first quartile minus 1.5 times the interquartile range" to "the third quartile plus 1.5 times the interquartile range", and then automatically identify, remove or correct outliers outside the range to improve data quality.
[0030] In an optional implementation, the abnormal data cleaning method in step S200 can also isolate the data samples by constructing multiple random trees, calculate the abnormal score by taking advantage of the characteristics that outliers are easy to isolate and have short paths, and then automatically identify, remove or correct outliers based on the score threshold to improve data quality.
[0031] In this embodiment of the application, the overall architecture has also been optimized for scenario adaptation. The model parameters are fine-tuned for two types of low-voltage users: residential and small industrial and commercial users, so that the whole system can adapt to different power consumption conditions.
[0032] In this embodiment of the application, the power surge identification model in step S200 is used to identify whether the power surge is sudden, and the cause analysis model is used to analyze the cause of the surge when it is identified as a surge. The power surge recognition model uses a random forest model to perform binary classification based on the input core features. In an optional implementation, the power surge identification model in step S200 can also construct a binary classification discrimination model based on SVM. By selecting key features from historical power data as input vectors, the kernel function is used to map the data to a high-dimensional space to find the optimal classification hyperplane, thereby achieving accurate identification and judgment of the power surge state.
[0033] In an optional implementation, the power surge identification model in step S200 can also construct a deep neural network architecture containing multiple hidden layers, and directly perform end-to-end training and learning on massive historical power time series data, thereby achieving high-precision automatic identification of complex power surge patterns.
[0034] The cause analysis model adopts the gradient boosting decision tree (GBDT) model, which analyzes the causes of sudden increases into five basic categories: normal electricity consumption growth, metering failure, line hazards, electricity theft, and other problems.
[0035] In an optional implementation, the surge cause analysis model in step S200 can also employ an efficient ensemble learning algorithm based on gradient boosting decision tree (GBDT). By introducing regularization terms or histogram optimization mechanisms, iterative training and node splitting optimization are performed on multidimensional feature data, thereby achieving high-precision and high-efficiency classification and discrimination of the cause of surge in power.
[0036] In an optional implementation, the surge cause analysis model in step S200 can also construct a fully connected neural network architecture containing an input layer, a hidden layer, and an output layer. Parameter optimization is performed using a nonlinear activation function and a backpropagation algorithm. Through the synergistic effect of multiple neurons, complex nonlinear relationships between features are automatically extracted, thereby achieving accurate classification and intelligent diagnosis of the cause of surge in power.
[0037] Furthermore, to ensure seamless alignment between the model output and the downstream early warning grading and manual handling system, the detailed early warning causes involved in the early warning grading stage (such as major line hazards, general line hazards, equipment aging, etc.) and the standard labels in the manual handling stage are all based on the five basic labels (normal electricity consumption growth, metering failure, line hazards, electricity theft, and other issues) output by the cause analysis model, and are carried out through business-level subdivision and classification mapping, thereby ensuring a one-to-one correspondence between the model's judgment results and the business handling rules.
[0038] In this embodiment of the application, step S200, which involves extracting core features from the multi-dimensional user profile, specifically includes: Initial features are extracted from the multidimensional user profile, and the analysis of variance and mutual information are used for joint screening. The core features that contribute the most to the model prediction are extracted from the initial features by dimensionality reduction. The initial features include 12 items, which are divided into electricity consumption behavior related features and auxiliary correlation features according to their attributes. Among them, the electricity consumption behavior related features cover the proportion of electricity consumption during a certain period, the deviation of the power factor, the duration of abnormal load rate, and the seasonal electricity consumption deviation value. The auxiliary correlation features include the correlation of equipment failure, the frequency of users' historical requests, the status of overdue payment, and the user type.
[0039] After the above joint screening, the core features obtained include at least eight key features such as load volatility, equipment failure correlation, and historical request frequency. Furthermore, the specific calculation rules for each key feature are as follows: The load volatility is used to characterize the degree of dispersion and fluctuation in a user's electricity consumption throughout the day. The specific calculation process is as follows: obtain 24-hour time-of-use load sampling points within a day, calculate the daily average load and the daily load standard deviation, and divide the daily load standard deviation by the daily average load (i.e., calculate the coefficient of variation of the load within the preset electricity consumption cycle) to obtain the load volatility. The specific calculation formula is as follows: The daily average load is expressed as: The standard deviation of daily load is expressed as follows: Load volatility is expressed as: N is fixed as the number of load sampling points in the 24 hours of the day. Let be the load value for the i-th hour.
[0040] The equipment fault correlation is a continuous value ranging from 0 to 1. The specific calculation rule is as follows: All line and meter fault records for the user within the past 90 days are statistically analyzed, and a base score is assigned based on the severity of the fault. Major faults have a base score of 1.0, and general faults have a base score of 0.2. An exponential attenuation coefficient is calculated based on the fault occurrence interval. Where T represents the number of days since the fault occurred, and α is the attenuation factor obtained by fitting the time distribution pattern of historical fault data through an empirical formula (in this preferred embodiment, α is set to 0.01, but in practical applications it can be adaptively adjusted between 0.005 and 0.02 according to the half-life requirement of historical fault records); the score of a single fault is the product of the base score and the attenuation coefficient; all single fault scores are summed up, and the maximum summed score is limited to 3. The final equipment fault correlation degree is equal to the summed score divided by 3.
[0041] The frequency of historical requests is a continuous value ranging from 0 to 1. The specific calculation rule is as follows: All customer service orders within the past 365 days are statistically analyzed, and a base score is assigned based on the request type. Specifically, electricity bill and metering issues have a base score of 1.0, line voltage faults have a base score of 0.7, power outage repairs have a base score of 0.3, and general business inquiries have a base score of 0.1. An exponential decay coefficient is then calculated based on the time of the request. Where D represents the number of days since the request occurred, and β is the attenuation factor obtained by fitting the time distribution pattern of historical request data through an empirical formula (in this preferred embodiment, β is set to 0.004, but in actual applications it can be adaptively adjusted between 0.002 and 0.008 according to the length of the business attention cycle); the score of a single request is the product of the type base score and the attenuation coefficient; the scores of all single requests are summed, and the upper limit of the summed score is set to 5. The final frequency of historical requests is equal to the summed score divided by 5.
[0042] In this embodiment of the application, step S200 further includes the division of the sample dataset, specifically including: labeling the multi-dimensional user profile according to past business records, classifying verified surges in electricity consumption and fault data as positive samples, classifying normal electricity consumption increases caused by seasonal changes or business expansion as negative samples, and classifying data with unclear boundaries as neutral samples; then splitting the dataset into a training set, a validation set, and a test set in a 7:2:1 ratio, with 70% of the data used as the training set for model training, 20% of the data used as the validation set for simultaneous verification of the effect during training, and 10% of the data used as the test set for final detection and not participating in training parameter tuning.
[0043] In this embodiment, the training process of the random forest model specifically includes: setting the number of decision trees to preferably 100 (which can be adjusted between 50 and 200 trees in practical applications depending on the sample size and feature dimension), the maximum depth of the trees to preferably 10 (which can be adjusted between 8 and 15 depending on the feature complexity), and using cross-entropy as the loss function; using grid search traversal to debug the core hyperparameters to select the globally optimal parameter combination; and introducing an early stopping mechanism during training, automatically terminating the iteration when the validation set accuracy does not improve for five consecutive rounds to prevent overfitting, until the validation accuracy stabilizes at 95% or higher to end the training.
[0044] In this embodiment of the application, the training process of the gradient boosting decision tree model for cause analysis specifically includes: setting the model learning rate to preferably 0.1 (which can be adjusted between 0.05 and 0.2 in practical applications depending on the model convergence speed), the overall number of iteration rounds to preferably 50 rounds (which can be dynamically adjusted between 30 and 100 rounds depending on the downward trend of the loss function on the validation set), and continuously iterating and optimizing the prediction effect based on the loss function until the cause identification accuracy stabilizes at over 90%.
[0045] In this embodiment of the application, step S300, which generates and pushes warning information of the corresponding level based on the output of the two-layer machine learning model, includes: Obtain the global anomaly probability output by the power surge identification model and the probability of various surge causes output by the cause analysis model; The global anomaly probability is concatenated with the probabilities of various sudden increase causes to form a feature vector, which is then input into the logistic regression model for calibration to obtain the comprehensive fusion probability. Based on the comprehensive fusion probability and the sudden increase causes output by the causal analysis model, early warning information of corresponding levels is generated.
[0046] In an optional implementation, the model probability fusion and calibration method in step S300 can also assign differentiated weights to each base model by comprehensively evaluating its performance or business importance, and perform a weighted summation of the probability values output by each model to generate the final prediction probability, thereby achieving effective integration of the advantages of multiple models and improving prediction accuracy.
[0047] In an optional implementation, the model probability fusion and calibration method in step S300 can also update the posterior probability by introducing a prior distribution and combining it with the historical performance or confidence of each model, and dynamically adjust the model weights using a Bayesian inference mechanism to output the final prediction probability, thereby effectively smoothing the prediction results and reducing the risk of overfitting while fusing information from multiple models.
[0048] In this embodiment, the global anomaly probability is concatenated with the probabilities of various surge causes to form a feature vector. Specifically, this includes: concatenating the one-dimensional global anomaly probability output by the identification layer with the probabilities corresponding to the five surge causes output by the causal analysis model to construct a six-dimensional feature vector; specifically, the six-dimensional feature vector is represented as follows: ,in, P1 represents the global anomaly probability output by the identification layer, and P5 represents the probabilities corresponding to the five types of sudden increase causes (normal electricity consumption growth, metering failure, line hidden danger, electricity theft, and other problems) output by the causal analysis model, respectively. The six-dimensional feature vector is input into the logistic regression model, and the model automatically learns the optimal weight of each feature dimension for linear fitting, and outputs the calibrated comprehensive fusion probability.
[0049] In this embodiment of the application, step S300 generates a corresponding level of early warning information based on the comprehensive fusion probability and the sudden increase cause output by the causal analysis model, including: In response to the fact that the comprehensive fusion probability is not less than a first preset threshold and the cause of the sudden increase belongs to the first type of preset abnormal cause, a first-level early warning information is generated; In response to the fact that the comprehensive fusion probability is within the second preset threshold range and the cause of the sudden increase belongs to the second type of preset abnormal cause, a secondary warning message is generated; In response to the integrated probability falling within the third preset threshold range, a level-three early warning message is generated.
[0050] In this embodiment, the specific rules for generating the graded early warning information include: the system automatically compares the comprehensive fusion probability output by the model with the determined cause, and automatically generates the corresponding handling work order without human intervention throughout the process. Specifically, the system generates early warning information of the corresponding level based on the preset threshold range of the comprehensive fusion probability and the preset abnormality category to which the sudden increase cause belongs.
[0051] In this embodiment, the specific settings for the preset threshold range and the causes of anomalies are as follows: When the overall fusion probability is greater than or equal to the first preset threshold (e.g., set to 90%), and the cause of the sudden increase is determined to be a first-class preset anomaly cause (e.g., electricity theft, major line hazards after mapping refinement, or serious anomalies that directly affect metering accuracy or power grid safety, such as metering device failure), a first-level early warning message is generated; when the overall fusion probability is within the second preset threshold range (e.g., set to...), a first-level early warning message is generated. When the sudden increase is determined to be a second-class preset abnormal cause (such as general line hazards, equipment aging, or suspected illegal electricity use, which have little impact on metering accuracy after mapping refinement), a second-level warning message is generated; when the comprehensive fusion probability is within the third preset threshold range (e.g., set to...), a second-level warning message is generated. When the overall fusion probability is less than the third preset threshold range (i.e., less than 50%), the system determines it as normal power consumption fluctuation, does not generate a warning message, and only retains the data for subsequent observation and training of the model.
[0052] It should be understood that the specific values (such as 90%, 70%) and specific cause classifications mentioned above are merely illustrative examples. Those skilled in the art will understand that the specific values or contents of the above thresholds and cause classifications can be adaptively adjusted according to the accuracy requirements of actual business scenarios, historical data statistical distribution, or equipment operating environment. In practical applications, adjusting the first preset threshold to 85%, or classifying a specific type of fault into the first preset abnormal cause, falls within the scope of the technical concept disclosed in this application and should not be considered a limitation on the scope of protection of this application.
[0053] In this embodiment of the application, the early warning information is generated in the form of an early warning work order, which clearly lists the user's basic information, abnormal power consumption, the identified cause of the fault, the early warning level, and the corresponding risk warning.
[0054] In this embodiment of the application, step S400, which triggers the corresponding service policy based on the warning information for processing, includes: Upon receiving a Level 1 warning, an on-site handling strategy is triggered to conduct on-site verification within a preset response time limit; Upon receiving a Level 2 warning, a proactive reminder strategy is triggered to contact the individual by phone and schedule an investigation within a preset response time limit. Upon receiving a Level 3 early warning message, a tracking and guidance strategy for tracking user electricity consumption curves online is triggered.
[0055] In this embodiment, the triggering of the corresponding service strategy for handling includes automatically matching a one-to-one handling service plan based on the warning level, the cause of the sudden increase in electricity consumption, and the user's multi-dimensional profile. Specifically, for Level 1 warning information, an immediate on-site verification strategy is triggered, requiring an on-site inspection within 24 hours, focusing on checking the lines and metering devices and conducting anti-theft checks. For Level 2 warning information, a proactive reminder and scheduled inspection strategy is triggered, requiring the customer to be contacted by phone within 12 hours to explain the abnormal situation and schedule an on-site inspection time. For Level 3 warning information, no on-site inspection is conducted; instead, the user's electricity consumption curve is continuously tracked online, and electricity consumption analysis and energy-saving suggestions are pushed out simultaneously. The warning level is adjusted based on changes in electricity consumption after continuous observation for a preset period.
[0056] In this embodiment of the application, the triggering of the corresponding service strategy for handling specifically includes targeted push logic: the system automatically matches the user type and the area to which the warning work order belongs based on the warning work order, and sends the warning work order to the mobile terminal of the corresponding maintenance personnel or account manager; for the first-level warning information, the system simultaneously pushes it to the dispatch and control platform and pops up an emergency reminder.
[0057] In this embodiment, the step of triggering the corresponding service strategy based on the early warning information also includes a cross-position collaboration and closed-loop management mechanism: the system connects the data channels of multiple positions such as marketing, operation and maintenance, customer service and auditing, and clarifies the job responsibilities and completion deadlines of each position; front-line staff carry out on-site inspections, fault repairs, customer communication and overdue payment reminders as required, and synchronize the processing progress in the system; the system automatically verifies the processing results, and automatically initiates a second supervision if the hidden danger has not been eliminated; after all processing work is completed and approved, the work order is automatically archived and the user's electricity consumption file is updated synchronously.
[0058] In this embodiment of the application, step S400 involves using the processing result as a labeled sample for backflow to iteratively optimize the two-layer machine learning model, including: Receive the labeled samples submitted after on-site handling, the labeled samples including fault labels and supporting materials; The labeled samples are verified, and the verified labeled samples are stored in the sample library; The samples in the sample library are periodically merged with the original dataset, and the two-layer machine learning model is iteratively trained based on the merged dataset.
[0059] In this embodiment of the application, receiving the labeled samples submitted after on-site handling specifically includes: staff completing sample labeling through a mobile application, with five preset optional tags in the system: normal electricity consumption growth, metering fault, line hazard, electricity theft, and others; before submitting the label, on-site photos and meter readings must be uploaded as supporting materials, and submission cannot be completed if the attachments are missing; if the on-site conditions are complex and cannot be classified into the first four standard situations, the operator selects the "other" option and records the complex on-site problems, marking them as difficult samples.
[0060] In this embodiment, the verification of the labeled samples and the storage of the verified labeled samples in the sample library specifically includes: pushing the completed work order to the team leader for secondary manual review; if the label selection is accurate and the supporting materials are complete, the review is approved and the sample is officially stored in the library for model training; if the label selection is incorrect or the materials are incomplete, the work order is rejected and sent to the staff's mobile terminal with modification suggestions, requiring a re-on-site review, correction of the labeling, and resubmission; for work orders marked as difficult, the back-end technical staff will make a unified assessment and store the samples that cannot be standardized and classified in the short term in a neutral sample partition, which will not participate in the regular model training.
[0061] In this embodiment, the step of periodically merging the samples in the sample library with the original dataset and iteratively training the two-layer machine learning model based on the merged dataset specifically includes: the system automatically triggering a round of model iteration according to a preset cycle (e.g., quarterly), merging the newly added backflow samples with the original dataset, and re-dividing the training set, validation set, and test set according to a preset ratio; during training, the parameter combination that performed best in the previous round is used first, and fine-tuning and retraining the two-layer recognition model based on this; after the iteration is completed, the accuracy is verified using an independent test set, and if all indicators meet the standards, the online model is directly replaced; if the accuracy does not meet the standards, the parameters are adjusted and retrained to adapt to changes in user electricity consumption habits and power grid equipment status, thereby maintaining the model's judgment accuracy.
[0062] In this embodiment of the application, the continuous upgrade of the overall service system also includes: regularly reviewing early warning work orders and customer feedback information, analyzing the actual effects of handling solutions in different scenarios, and continuously optimizing early warning classification standards, work order push logic and on-site handling procedures; at the same time, supplementing exclusive early warning rules for new electricity consumption scenarios such as distributed photovoltaics and energy storage to improve the scenario adaptability of the entire system.
[0063] In this embodiment of the application, the method also includes a fallback judgment mechanism for special situations: in response to the acquisition of abnormal data that cannot be fed into the model for calculation, such as missing data or data distortion, the system automatically activates the fallback rule to complete the preliminary judgment. Specifically, it extracts the user's normal electricity consumption data for the past three months as a benchmark. When the current electricity consumption exceeds the benchmark value by a preset positive or negative ratio (such as 50%), it is directly judged as a suspected abnormality and a level three warning is generated. The samples selected by the fallback rule are classified and stored separately. They must be manually reviewed before it can be decided whether to include them in the formal sample library for subsequent iterations, thereby ensuring that the overall analysis process is not interrupted when the model cannot be calculated.
[0064] Example 3 is an embodiment of the present invention. This embodiment provides a method for early warning of sudden power surges in low-voltage users. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through specific implementation methods and implementation effects.
[0065] The specific details of this embodiment are as follows: I. Experimental Environment This embodiment relies on a low-voltage user power surge early warning system built on a four-layer architecture. In terms of hardware and software deployment: the data layer deploys a distributed database (HBase) and a data cleaning tool (DataX) to connect to the grid-level power energy data platform, metering automation system, customer service system, and marketing management system, enabling multi-dimensional data collection and storage; the model layer deploys a machine learning algorithm platform (Spark MLlib) and configures algorithm components such as random forest and gradient boosting tree (GBDT) for model training and optimization; the early warning layer deploys an early warning rule engine and message push module, configures three levels of early warning rules, and connects to the terminals of grassroots staff and customer managers; the service layer connects to the power supply service work order system and collaborative operation platform, and configures a customized service strategy library to realize the flow and closed-loop tracking of early warning work orders.
[0066] II. Experimental Procedure This embodiment uses data from 100,000 low-voltage users within the jurisdiction of a power supply bureau over the past year as the research object. The specific experimental steps are as follows: Step 1: Multidimensional Data Acquisition and Standardization. Data is collected from the aforementioned systems, including user electricity consumption data (hourly consumption, load curves, etc.), equipment data (metering device status, line fault records, etc.), service data (historical requests, repair records, etc.), and customer attribute data (user type, electricity address, historical arrears status, etc.). Subsequently, a unified data cleaning process is conducted to fill in missing electricity consumption data, remove extreme anomalies caused by meter malfunctions, and standardize the scattered data into a standardized format. Finally, the data is integrated using the user account number as the unique primary key to construct a multidimensional data profile of 100,000 low-voltage users, which is then stored in an HBase database.
[0067] Step 2: Feature Engineering and Sample Splitting. Twelve initial features were extracted from the multidimensional data profile. After analysis of variance and mutual information, eight core features were retained: electricity consumption increase, month-on-month increase, load fluctuation rate, power factor deviation, equipment failure correlation, historical complaint frequency, overdue payment status, and user type. Samples were labeled based on historical complaints and verification records, resulting in 5,000 positive samples, 90,000 negative samples, and 5,000 neutral samples. The data was then divided into a training set (70,000 samples), a validation set (20,000 samples), and a test set (10,000 samples) in a 7:2:1 ratio.
[0068] Step 3: Machine Learning Early Warning Model Training and Optimization. A power surge detection model (random forest) was built on the Spark MLlib platform, with 8 core features input, 100 decision trees, a maximum depth of 10, and 30 iterations of training. A causal analysis model (GBDT) was also built, with abnormal sample features input, a learning rate of 0.1, 50 iterations, and 20 iterations of training. In the model optimization phase, grid search was used for optimization, and logistic regression was introduced to complete multi-model fusion. Specifically, the single anomaly probability output by the random forest and the five causal probabilities output by the GBDT were concatenated into a six-dimensional feature vector, which was then input into the logistic regression module for anomaly confidence calibration (this module does not perform separate anomaly judgment but rather corrects the original score to distinguish between normal fluctuations and real potential risks).
[0069] Step 4: Dynamic Tiered Early Warning and Closed-Loop Response. Every day at midnight, the model is automatically invoked to perform batch analysis of 100,000 user data points. Based on the comprehensive probability output by the model, different levels of early warning work orders are generated and accurately pushed to the terminals of maintenance personnel or account managers in the corresponding jurisdiction (Level 1 early warnings are simultaneously pushed to the service dispatch platform to trigger emergency alerts). Corresponding service strategies are triggered: Level 1 early warnings trigger 24-hour on-site verification; Level 2 early warnings trigger 12-hour telephone reminders and scheduled inspections; Level 3 early warnings trigger online electricity usage guidance. Multiple departments collaborate on the response and upload results in real time. After the response is completed, the validity is verified, the work order is archived, and the user profile is updated.
[0070] Step 5: Model Iteration and Optimization. Each quarter, closed-loop processing data is summarized, and approximately 3,000 new labeled samples are added to expand the dataset. The model is then retrained and parameters are fine-tuned to adapt to changes in user electricity consumption behavior within the jurisdiction.
[0071] III. Experimental Results After the complete experiment and system deployment described above, this solution has achieved significant practical application results: 1. Model performance indicators: After verification with the test set, the accuracy rate of the power surge alarm of the whole solution reached 95.8%, the cause identification accuracy rate was 91.7%, and all indicators met the preset standards.
[0072] 2. Operational performance indicators: The system's early warning accuracy rate remained stable at 95.8%, with a false alarm rate of only 4.2%. The response time for work orders pushed to staff was controlled within 30 minutes, with an information distribution accuracy rate of 100% and a timeliness rate of 99.8%.
[0073] 3. Closed-loop handling and user feedback: The completion rate of this early warning handling reached 100%, and the overall customer satisfaction reached 98.5%. After the system was implemented, the number of complaints related to abnormal electricity use in the area decreased by 35%, and the overall customer satisfaction increased by 22%, effectively solving the pain points of front-line inspection and maintenance personnel in their daily work.
[0074] Example 4 is an embodiment of the present invention. This embodiment provides a method for early warning of sudden power surges in low-voltage users. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through specific implementation methods and implementation effects.
[0075] The specific details of this embodiment are as follows: I. Experimental Environment 1. Experimental Subjects and Scenarios: A total of 23,000 low-voltage industrial and commercial users, including small processing and retail stores, within the industrial parks of Changshun and Libo counties under the jurisdiction of Duyun, were selected as pilot projects. This group exhibits characteristics such as seasonal work start-ups and phased production expansions, and the causes of electricity consumption fluctuations are complex. The aim is to verify the adaptability of the early warning method to differentiated industrial and commercial scenarios.
[0076] 2. Hardware and software infrastructure: The existing hardware and software infrastructure will be used with some targeted modifications. Three new business fields will be added to the data collection terminal: start-up ledger, business category, and shutdown filing. A separate industrial and commercial-specific policy library will be built in the service layer to distinguish between four types of situations: production expansion, equipment leakage, metering failure, and electricity theft, and match them with corresponding handling procedures.
[0077] 3. Model Algorithm Architecture: Retain the two-layer basic structure of random forest combined with GBDT, and use industrial and commercial-specific samples to fine-tune the weights of various features to adapt to the differentiated electricity consumption patterns of merchants.
[0078] II. Experimental Procedure 1. Multidimensional Data Collection and Standardized Processing: Collect all data from the pilot area for the past 18 months, and additionally access merchant suspension and resumption registration data and industry seasonal electricity consumption patterns in addition to the basic data. Fill in the missing electricity consumption during short-term shutdowns using the average of similar businesses, and use industry electricity consumption thresholds to remove distorted data caused by equipment failures. Associate all fields with the account number to construct a multidimensional profile dataset exclusive to industrial and commercial users.
[0079] 2. Feature Engineering and Sample Splitting: Utilizing the eight core features, the weights of load rate, time-of-use electricity consumption ratio, business attributes, and load fluctuation rate (intraday load variation coefficient) were increased for industrial and commercial users; the weights of equipment failure correlation features were also increased. A dataset containing 2100 positive samples, 19500 negative samples, and 1400 neutral samples was constructed by combining offline inspection history and merchant complaint records. The dataset was then divided into training, validation, and test datasets in a 7:2:1 ratio.
[0080] 3. Model Fine-tuning Training and Validation: Using the basic model parameters as a baseline, a secondary training was conducted using a dataset specifically designed for industrial and commercial applications. The decision tree depth of the random forest and the learning rate of the GBDT model were fine-tuned. A grid search algorithm was used to traverse and optimize the parameters, selecting the optimal combination suitable for industrial and commercial scenarios, and then testing and validation were completed.
[0081] 4. Tiered Early Warning Generation and Automatic Dispatch: The system performs batch processing on 23,000 user data points daily, automatically dispatching orders based on transformer area affiliation and customer manager responsibility zones. Level 1 early warnings (electricity theft / major metering failures) are simultaneously pushed to county / district power supply inspection teams.
[0082] 5. Closed-loop processing of differentiated customized services: Level 1 warning: Joint inspection and maintenance teams will conduct on-site verification 24 hours a day, and complete the fault meter replacement or electricity usage inspection on the spot; Level 2 warning: Contact the business owner by phone within 12 hours to arrange on-site inspection during the business closure period; Level 3 early warning: Automatically pushes time-of-use electricity pricing optimization and peak-shifting electricity consumption suggestions, and continuously tracks the electricity consumption curve for 7 natural days online.
[0083] All processing results are transmitted back for verification and archiving in real time, and user profiles are updated simultaneously.
[0084] 6. Periodic model iteration and optimization: Newly labeled samples are collected and processed in a closed loop every month (800-1200 samples per month), and the overall model is iterated every quarter in conjunction with the residential samples, dynamically correcting the peak season electricity consumption threshold rules for industrial and commercial enterprises.
[0085] III. Experimental Results 1. Model accuracy meets standards: After iterative optimization and testing, the accuracy rate of the power surge alarm reaches 95.3%, and the accuracy rate of cause identification reaches 91.2%. Both core accuracy indicators meet the standards for practical use.
[0086] 2. The system operates efficiently: it takes only 22 minutes to process 23,000 user data in a single batch; on average, it generates 7 Level 1 warnings, 36 Level 2 warnings, and 97 Level 3 warnings per day within the cycle, with an information distribution accuracy rate of 100% and a timeliness rate of 99.9%.
[0087] 3. Significant Business Results: After the pilot program was implemented, the false alarm rate of sudden surge in electricity consumption warnings in the county's industrial and commercial sectors was controlled within 4.7%; more than 100 potential problems with power lines and meters were identified and resolved in advance; and electricity-related complaints to industrial and commercial users decreased by 32.7% year-on-year, successfully achieving a transformation from passively accepting complaints to proactively providing services, and improving the supervision of five or more customer service behavior scenarios.
[0088] Example 5 is an embodiment of the present invention. This embodiment differs from the first embodiment in that it provides a low-voltage user power surge warning system.
[0089] It should be noted that the technical solution of the low-voltage user power surge warning system and the technical solution of the low-voltage user power surge warning method mentioned above belong to the same concept. For details not described in detail in the technical solution of the low-voltage user power surge warning system in this embodiment, please refer to the description of the technical solution of the low-voltage user power surge warning method mentioned above.
[0090] This embodiment provides a low-voltage user power surge warning system, comprising: The user profile building module is used to acquire and standardize user electricity consumption data, equipment data, service data, and customer attribute data to form a multi-dimensional user profile. The dual-layer model analysis module is used to extract core features from the multi-dimensional user profile and input the core features into the dual-layer machine learning model for analysis. The dual-layer machine learning model includes a power surge identification model and a cause analysis model. The early warning generation and push module is used to generate and push early warning information of the corresponding level based on the output results of the two-layer machine learning model. The strategy handling and model iteration module is used to trigger the corresponding service strategy to handle the situation based on the warning information, and to feed back the handling results as labeled samples for iterative optimization of the two-layer machine learning model.
[0091] This embodiment also provides a computer device applicable to a low-voltage user power surge warning method, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a method for providing early warning of sudden power surges for low-voltage users, as proposed in the above embodiments.
[0092] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a method for early warning of sudden power surges in low-voltage users as proposed in the above embodiments.
[0093] The storage medium proposed in this embodiment and the method for implementing a low-voltage user power surge warning proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.
[0094] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.
[0095] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A low voltage consumer power surge early warning method, characterized in that, include: Acquire and standardize user electricity consumption data, equipment data, service data, and customer attribute data to form a multi-dimensional user profile; Core features are extracted from the multi-dimensional user profile and input into a two-layer machine learning model for analysis. The two-layer machine learning model includes a power surge identification model and a cause analysis model. Based on the output of the two-layer machine learning model, warning information of the corresponding level is generated and pushed. The corresponding service strategy is triggered based on the warning information to handle the situation, and the handling results are fed back as labeled samples for iterative optimization of the two-layer machine learning model.
2. The low voltage consumer power surge pre-alarm method of claim 1, wherein: The power surge identification model is used to identify whether the power has surged, and the cause analysis model is used to analyze the cause of the surge when it is identified as a surge. The power surge recognition model uses a random forest model to perform binary classification based on the input core features. The cause analysis model adopts a gradient boosting decision tree model to analyze the causes of sudden increases as normal electricity consumption growth, metering failure, line hazards, electricity theft, and other problems.
3. The low-voltage consumer power surge pre-warning method according to claim 1 or 2, characterized in that: The generation and push of corresponding level early warning information based on the output of the two-layer machine learning model includes: Obtain the global anomaly probability output by the power surge identification model and the probability of various surge causes output by the cause analysis model; The global anomaly probability is concatenated with the probabilities of various sudden increase causes to form a feature vector, which is then input into the logistic regression model for calibration to obtain the comprehensive fusion probability. Based on the comprehensive fusion probability and the sudden increase causes output by the causal analysis model, early warning information of corresponding levels is generated.
4. The low voltage consumer power surge pre-alarm method of claim 3, wherein: The extraction of core features from the multi-dimensional user profile includes: Initial features are extracted from the multidimensional user profile, and core features are obtained by joint screening using analysis of variance and mutual information. The core features include load volatility, equipment failure correlation, and historical request frequency. The load fluctuation rate is obtained by calculating the coefficient of variation of the load within a preset electricity consumption cycle; The correlation degree of the equipment failure is calculated by exponential decay weighting based on the severity and time distance of historical failures within a preset time window. The frequency of historical requests is calculated by weighting the request type and time.
5. The low voltage consumer power surge pre-alarm method of claim 4, wherein: The generation of corresponding levels of early warning information based on the integrated probability and the cause of the surge output by the causal analysis model includes: In response to the fact that the comprehensive fusion probability is not less than a first preset threshold and the cause of the sudden increase belongs to the first type of preset abnormal cause, a first-level early warning information is generated; In response to the fact that the comprehensive fusion probability is within the second preset threshold range and the cause of the sudden increase belongs to the second type of preset abnormal cause, a secondary warning message is generated; In response to the integrated probability falling within the third preset threshold range, a level-three early warning message is generated.
6. The low voltage consumer power surge pre-alarm method of claim 5, wherein: The step of triggering the corresponding service strategy based on the warning information to handle the situation includes: Upon receiving a Level 1 warning, an on-site handling strategy is triggered to conduct on-site verification within a preset response time limit; Upon receiving a Level 2 warning, a proactive reminder strategy is triggered to contact the individual by phone and schedule an investigation within a preset response time limit. Upon receiving a Level 3 early warning message, a tracking and guidance strategy for tracking user electricity consumption curves online is triggered.
7. The low voltage consumer power surge pre-alarm method of claim 6, wherein: The step of using the processing results as labeled samples for backfeeding to iteratively optimize the two-layer machine learning model includes: Receive the labeled samples submitted after on-site handling, the labeled samples including fault labels and supporting materials; The labeled samples are verified, and the verified labeled samples are stored in the sample library; The samples in the sample library are periodically merged with the original dataset, and the two-layer machine learning model is iteratively trained based on the merged dataset.
8. A low voltage consumer power surge early warning system applying the method according to any one of claims 1 to 7, characterized in that, include: The user profile building module is used to acquire and standardize user electricity consumption data, equipment data, service data, and customer attribute data to form a multi-dimensional user profile. The dual-layer model analysis module is used to extract core features from the multi-dimensional user profile and input the core features into the dual-layer machine learning model for analysis. The dual-layer machine learning model includes a power surge identification model and a cause analysis model. The early warning generation and push module is used to generate and push early warning information of the corresponding level based on the output results of the two-layer machine learning model. The strategy handling and model iteration module is used to trigger the corresponding service strategy to handle the situation based on the warning information, and to feed back the handling results as labeled samples for iterative optimization of the two-layer machine learning model.
9. A computer device, comprising: include: Memory, used to store programs; A processor for loading the program to perform the steps of the method as claimed in any one of claims 1-7.
10. A computer-readable storage medium storing a program, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.
