District line loss abnormity diagnosis method and system based on large model

Abstract

The invention relates to the technical field of electric power fault diagnosis, and discloses a transformer area line loss abnormity diagnosis method and system based on a large model, and the method comprises the steps: collecting multi-dimensional data used for supporting transformer area line loss abnormity diagnosis, carrying out the cleaning, correlation fusion and standardization processing of the multi-dimensional data, and obtaining a comprehensive data set; a multi-layer diagnosis system based on a rule model, a random forest model and a large model is constructed, and the rule model identifies the simple and conventional anomalies of the transformer area according to a preset anomaly diagnosis rule based on the basic attribute data and the power operation state data in the comprehensive data set; the random forest model locates complex anomalies and novel anomalies by mining a coupling relationship among energy access condition data, external environment influence data and line loss fluctuation in the comprehensive data set; the big model carries out fusion verification on diagnosis results of the rule model and the random forest model, and outputs a final transformer area abnormity diagnosis result; and based on the diagnosis result, a differential loss reduction strategy adaptive to the actual data characteristics of the transformer area is recommended. The working efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of power fault diagnosis technology, specifically to a method and system for diagnosing abnormal line losses in transformer substations based on a large model. Background Technology

[0002] In the wave of digital transformation in the power industry, the construction of smart grids has entered a critical phase. With the continuous expansion of the power grid and the increasing complexity of line topologies, massive amounts of power operation data, equipment status data, and user electricity consumption data are growing exponentially. Currently, in the field of abnormal line loss diagnosis in distribution transformer areas, existing technologies are mostly based on rule-based methods and rely heavily on manual operation. Rule-based methods involve manually setting fixed thresholds and judgment conditions, such as current fluctuation ranges, voltage deviation values, and line loss rate ranges. When monitored data exceeds or meets these predetermined rules, it is judged as an abnormal line loss. Simultaneously, line loss personnel need to monitor line losses daily through synchronization or data acquisition systems, manually collect current, voltage, and other operational data, check and analyze anomalies on a household-by-household basis, and then go to the site for verification. However, this traditional technical system is gradually revealing many prominent problems under the urgent needs of new power systems for loss reduction, efficiency improvement, and lean management: 1. Extremely low efficiency in line loss diagnosis and analysis, significant limitations in data processing capabilities, and one-sided coverage of diagnostic analysis. Summary of the Invention

[0003] This invention provides a method and system for diagnosing abnormal line loss in transformer substations based on a large model, in order to solve the above-mentioned problems.

[0004] This invention is achieved through the following technical solution:

[0005] A method for diagnosing abnormal line loss in transformer substations based on a large model includes the following steps:

[0006] Multi-dimensional data is collected to support the diagnosis of abnormal line loss in transformer substations. The multi-dimensional data is then cleaned, correlated, fused, and standardized to obtain a comprehensive dataset covering transformer substation attributes, operating status, energy access, and external environment. The multi-dimensional data includes: basic attribute data to characterize the inherent features of the transformer substations, power operating status data reflecting the real-time and historical operating patterns of the power system, energy access data reflecting the impact of new energy access on the transformer substations, and external environmental impact data that may interfere with or correlate with line loss fluctuations.

[0007] A multi-layered diagnostic system based on a rule-based model, a random forest model, and a large-scale model is constructed. The rule-based model identifies simple, routine anomalies in transformer substations based on the basic attribute data and power operation status data in the comprehensive dataset, according to preset anomaly diagnosis rules. The random forest model locates complex and novel anomalies by mining the coupling relationship between energy access data, external environmental impact data, and line loss fluctuations in the comprehensive dataset. The large-scale model fuses and verifies the diagnostic results of the rule-based model and the random forest model, outputting the final anomaly diagnosis result for the transformer substation.

[0008] Based on the diagnostic results, the loss reduction strategy knowledge base is invoked, and the large model intelligently recommends differentiated loss reduction strategies that are adapted to the actual data characteristics of the transformer area.

[0009] As an optimization, the basic attribute data includes fixed parameters used to characterize the physical characteristics of the transformer area; the power operation status data includes real-time measurement data and historical statistical data, the real-time measurement data includes voltage, current, power factor, load rate and three-phase imbalance, and the historical statistical data includes daily line loss rate, hourly load curve and voltage qualification rate; the external environmental impact data includes meteorological data and time characteristic data, the meteorological data includes weather type, temperature and humidity, and the time characteristic data includes holiday identification and peak and valley electricity consumption periods; the energy access status data includes the installed capacity, output power curve, grid-connected electricity and access location information of distributed power sources, and the distributed power sources include photovoltaic, wind power and energy storage equipment.

[0010] As an optimization, the rule model, based on the basic attribute data and power operation status data in the comprehensive dataset, identifies simple and routine anomalies in the transformer substation according to preset anomaly diagnosis rules. The specific process includes the following sub-steps: First, an anomaly classification system is obtained by sorting through the basic attribute data and power operation status data in the comprehensive dataset. Second, an initial rule base is constructed based on the anomaly classification system. The initial rule base stores the judgment logic for various anomalies in an IF-THEN structure and is iteratively optimized through a dynamic update mechanism. Third, with the dynamically updated rule base as the core, power industry standards and expert experience corresponding to the anomaly classification system are integrated. A knowledge base containing dynamic indexes is constructed based on a knowledge graph. The knowledge base is a semantic association network based on basic attribute data, operation status data, rules, standards, and experience. Fourth, the basic attribute data and power operation status data of the transformer substation to be diagnosed are extracted from the comprehensive dataset. Matching rule base judgment conditions are located through dynamic indexing. Feature thresholds are verified using industry standards in the knowledge base. After large-scale model verification of the consistency between the basic attribute data and power operation status data of the transformer substation to be diagnosed and the rules, a diagnostic result containing anomaly category, feature basis, and rule version is output.

[0011] As an optimization, the dynamic update mechanism is as follows: when the number of identified cases of the same new type of anomaly reaches 5, the basic attribute data features and power operation status data features associated with this type of new anomaly are extracted, and the Apriori algorithm is used to mine the association rules between features and anomalies to generate candidate rules. When the accuracy of the candidate rule is ≥90%, the candidate rule is included in the rule base, and the corresponding level of the anomaly classification system is updated at the same time.

[0012] As an optimization, the random forest model locates complex and novel anomalies by mining the coupling relationship between energy access data, external environmental impact data, and line loss fluctuations in the comprehensive dataset. The specific process involves extracting time-series features characterizing line loss fluctuation patterns, correlation features representing the relationship between energy access and line loss, static features reflecting the inherent attributes of transformer substations, and dynamic features reflecting real-time operating status from the comprehensive dataset, forming a multi-dimensional feature set covering time evolution, correlation coupling, equipment attributes, and operating status. From these time-series, correlation, static, and dynamic features, principal component analysis is used to screen key features that contribute up to 85% to the cumulative anomaly diagnosis. Simultaneously, an attention mechanism is introduced to analyze the time-series features during peak electricity consumption periods. Features are assigned dynamic weights to achieve dimensionality reduction optimization of key features and enhancement of features in high-value periods. The key features are fused to form a feature vector, and a random forest model is constructed using a hybrid splitting criterion that weights the Gini coefficient and the contribution of time-series features. The model parameters are optimized by maximizing the F1 score of the large model, enabling the random forest model to discover the correlation between multi-dimensional features and line loss anomalies. Using the trained random forest model, complex anomalies and novel anomalies are located by identifying the coupling patterns between abrupt changes in energy access parameters, anomalies in external environmental factors, and line loss fluctuations. The diagnostic results include anomaly type, feature contribution, and associated influencing factors. The performance of the random forest model is verified by accuracy, recall, and F1 score.

[0013] As an optimization, an attention mechanism is introduced to assign dynamic weights to the temporal characteristics of peak electricity consumption periods. Specifically, a dynamic weight allocation strategy based on time windows is constructed, assigning an initial weight of 1.5 times to the temporal characteristics of peak electricity consumption periods. Combined with sliding window technology, the weight coefficient is adaptively adjusted by calculating the feature similarity between the current peak period and historical peak periods, based on the fluctuation trend of line loss data during the same period in history. Specifically, when the similarity between the fluctuation trend of the current peak period and the abnormal pattern of the typical historical peak period is ≥80%, the weight coefficient is increased to 1.5-2.0 times; when the similarity is in the range of 50%-80%, the initial weight of 1.5 times is maintained; and when the similarity is <50%, the weight coefficient is reduced to 1.2-1.5 times.

[0014] As an optimization, the key features are fused to form a feature vector. A random forest model is constructed using a hybrid splitting criterion weighted by the Gini coefficient and the contribution of time-series features. The model parameters are optimized by maximizing the F1 score of the large model, enabling the random forest model to have the ability to discover the correlation between multi-dimensional features and line loss anomalies. The specific process is as follows: Static features and enhanced time-series features from the key features are fused with energy access data, meteorological data, external environmental impact data, and holiday data to form a multi-dimensional feature vector; a hybrid evaluation index S weighted by the Gini coefficient and the contribution of time-series features, and a hybrid splitting criterion are constructed, S=α×G+(1-α)×C, where G is the Gini coefficient representing the spatial dispersion of line loss data, C is the contribution of time-series features representing the dynamic change trend of the time series, and α is an adaptive weighting coefficient. The hybrid splitting criterion is: at the decision tree node... During splitting, the influence of spatial distribution characteristics and temporal evolution on the classification results is dynamically balanced using the hybrid evaluation index. Multiple sample subsets are obtained by randomly sampling with replacement of the multidimensional feature vectors, and each sample subset is used to train a decision tree. During the node splitting process of each decision tree, some key features are randomly selected from the sample subsets, and the optimal splitting features are determined using the aforementioned hybrid splitting criteria to reduce inter-tree correlation and improve the model's generalization ability. The sample subsets are divided into training and testing sets, and the training set is used to train the random forest model. Based on the reasoning ability of the large model, with maximizing the F1 score as the optimization objective, the parameters of the random forest model are adaptively adjusted. The trained random forest model is evaluated using the testing set, and the performance of the random forest model is judged by calculating accuracy, recall, and F1 score evaluation indicators.

[0015] As an optimization, the specific process of fusing and verifying the diagnostic results of the rule model and the random forest model to output the final anomaly diagnosis result of the transformer area is as follows: The independent diagnostic results of the rule model and the random forest model are obtained separately. The independent diagnostic result of the rule model is the anomaly judgment result and corresponding rule basis output based on manually defined business rules. The independent diagnostic result of the random forest model is the anomaly judgment result, feature contribution, and anomaly category probability output based on multi-dimensional feature combinations. A phased fusion strategy is formulated and implemented by combining the rule model and the random forest model. The phased fusion strategy is as follows: A first preset time period is designated as the early application phase. In this early application phase, a result union fusion strategy is adopted to integrate all anomaly transformer area data determined by the rule model and the random forest model as anomaly samples. These anomaly samples are then used to optimize the... The system employs a rule base and a random forest model. A second preset time period is designated as the mid-application phase. During this phase, based on the abnormal samples, the anomaly detection results identified by the random forest model are analyzed, and anomaly detection results exceeding a set threshold are converted into rules to supplement the rule base. A third preset time period is designated as the late-application phase. During this phase, based on the different scenarios of different substations, a result intersection fusion strategy is used for substations associated with core indicators. Only results deemed abnormal by both the rule model and the random forest model are retained and output to ensure diagnostic accuracy. For substations in edge scenarios, a result union fusion strategy is used to avoid missing potential anomalies, retaining results deemed abnormal by either the rule model or the random forest model and outputting them. The output results are verified, and the verified results are used as the final substation anomaly diagnosis results.

[0016] As an optimization, based on the diagnostic results, the specific process of calling the loss reduction strategy knowledge base and having the large model intelligently recommend differentiated loss reduction strategies adapted to the actual data characteristics of the transformer area is as follows: The large model analyzes the final transformer area anomaly diagnosis results and extracts key information, including: anomaly type, anomaly characteristic parameters, basic attribute data of the transformer area, power operation status data, and energy access status data; based on the analyzed anomaly type, it calls the loss reduction strategy knowledge base constructed by a classification method matching the anomaly classification system to retrieve candidate loss reduction strategies directly associated with the current anomaly type; it performs a differentiated adaptability judgment and screening of candidate loss reduction strategies in combination with the actual data characteristics of the transformer area; and it outputs the optimal loss reduction strategy from the screened candidate loss reduction strategies according to the urgency and stability of the effect.

[0017] The present invention also discloses a transformer substation line loss anomaly diagnosis system based on a large model, which is used to perform the aforementioned transformer substation line loss anomaly diagnosis method based on a large model.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0019] 1. A multi-model fusion diagnostic system for transformer substation line loss anomalies was constructed, covering all types of anomalies and breaking through the limitations of traditional single-model diagnostics to improve the accuracy of transformer substation line loss anomaly diagnosis. A multi-layer diagnostic architecture of rule-based model + random forest + large-scale model was built. The rule-based model quickly matches simple, common anomalies, such as basic file issues like missing transformer substation master tables; the random forest model uncovers complex correlations involving multiple coupled features, such as line loss fluctuations caused by the superposition of load changes and equipment aging; the large-scale model intelligently provides loss reduction suggestions based on knowledge hints from a loss reduction measure knowledge base through semantic analysis technology. Leveraging the powerful parallel computing and rapid data processing capabilities of the large-scale model, the anomaly identification accuracy jumped from 70%-80% of traditional methods to over 90%, significantly improving diagnostic precision.

[0020] 2. Establish a streamlined intelligent diagnostic process for line loss anomalies across transformer substations, encompassing anomaly identification, anomaly cause diagnosis, and loss reduction recommendations. This process achieves intelligent anomaly diagnosis and improves efficiency. A fully automated workflow is constructed, integrating anomaly identification, anomaly cause diagnosis, and loss reduction recommendations. Utilizing a multi-layered diagnostic architecture combining rule-based models, random forests, and large-scale models, line loss anomalies are diagnosed, and specific loss reduction recommendations are automatically generated based on the cause diagnosis results. This changes the traditional diagnostic process where identification, diagnosis, and recommendations are fragmented and reliant on manual intervention. The diagnostic time for 100 transformer substations has been reduced from one day to 2-3 hours, significantly improving work efficiency and achieving intelligent anomaly diagnosis.

[0021] 3. A new abnormal cause diagnosis process for transformer substation line losses was introduced, upgrading the process from anomaly diagnosis to line loss prevention. When training the random forest model to identify abnormal line losses in transformer substations, a new cause for line loss anomaly identification was introduced, establishing a correlation model between the new anomaly cause and potential line loss risks. For example, when a new abnormal cause of excessive fluctuations in renewable energy output is identified in a transformer substation, the potential voltage deviation line loss risk can be predicted. This improves the accuracy of identifying new line loss anomalies, helping maintenance personnel to identify potential line loss risks in advance, and effectively achieving the leap from anomaly diagnosis to line loss prevention.

[0022] 4. The established cross-regional diagnostic resource sharing platform breaks down regional information barriers, enabling the sharing of loss reduction experiences across different transformer substations and improving the effectiveness of line loss management. In the process of line loss management, loss reduction measures and cases are scattered and lack integration across different regions. Remote areas often face difficulties in developing management plans and delays in handling complex problems due to insufficient case reserves. By establishing a loss reduction measures knowledge base sharing platform, information silos are broken down, enabling the interconnection of successful loss reduction cases, management strategies, and technical solutions across different transformer substations. When a region encounters complex line loss anomalies, it can quickly retrieve and access resources from the entire network's loss reduction knowledge base, referencing practical experience and optimization paths from other regions to accurately formulate solutions and improve the effectiveness of line loss management in the affected area. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0024] Figure 1 This is a flowchart of a method for diagnosing abnormal line loss in transformer substations based on a large model, as described in this invention. Detailed Implementation

[0025] This embodiment 1 provides a method for diagnosing abnormal line loss in transformer substations based on a large model, such as... Figure 1 As shown, it includes the following steps:

[0026] S1. Collect multi-dimensional data to support the diagnosis of abnormal line loss in the transformer substation. Clean, correlate, fuse, and standardize the multi-dimensional data to obtain a comprehensive dataset covering the attributes, operating status, energy access, and external environment of the transformer substation. The multi-dimensional data includes: basic attribute data to characterize the inherent features of the transformer substation, power operating status data to reflect the real-time and historical operating patterns of the power system, energy access data to reflect the impact of new energy access on the transformer substation, and external environmental impact data that may interfere with or be associated with line loss fluctuations.

[0027] S2. Construct a multi-layered diagnostic system based on a rule-based model, a random forest model, and a large-scale model. The rule-based model identifies simple, routine anomalies in distribution transformer areas based on the basic attribute data and power operation status data in the comprehensive dataset, according to preset anomaly diagnosis rules. The random forest model locates complex and novel anomalies by mining the coupling relationship between energy access data, external environmental impact data, and line loss fluctuations in the comprehensive dataset. The large-scale model fuses and verifies the diagnostic results of the rule-based model and the random forest model, outputting the final anomaly diagnosis result for the distribution transformer area.

[0028] S3. Based on the diagnostic results, the loss reduction strategy knowledge base is invoked, and the large model intelligently recommends differentiated loss reduction strategies that are adapted to the actual data characteristics of the transformer area.

[0029] In essence, this method mainly comprises three parts: a data collection and processing layer, a transformer area line loss anomaly diagnosis layer, and a loss reduction strategy layer. The data collection and processing layer is responsible for collecting multi-dimensional data such as transformer area master meter information, user-side load data, distributed power output data (e.g., renewable energy generation power), and equipment operating status. Through data cleaning (removing outliers and filling missing items) and format standardization, it provides high-quality data input for subsequent diagnosis, ensuring the accuracy and reliability of the diagnostic process. The transformer area line loss anomaly diagnosis layer achieves comprehensive diagnosis of transformer area line loss anomalies by constructing a multi-layered diagnostic system of "rule model + random forest + large model." The rule model quickly identifies simple, routine anomalies in the transformer area based on preset anomaly diagnosis rules; the random forest model accurately locates complex anomalies by mining the coupling relationships between transformer area data and automatically identifies potential new anomalies; the large model fuses and verifies the diagnostic results of the first two layers and establishes a connection between the anomaly causes and the loss reduction strategy layer; the loss reduction strategy layer constructs a loss reduction strategy knowledge base and uses the large model to intelligently propose differentiated loss reduction strategies for transformer area anomalies.

[0030] S1 is divided into two sub-processes: data acquisition and data preprocessing.

[0031] The data acquisition sub-process mainly collects basic attribute data, power operation status data, energy access data, and external environmental impact data. Basic attribute data includes fixed parameters characterizing the physical properties of the transformer substation, specifically covering substation capacity, power supply radius, main line type, main line cross-sectional area, number of users, meter configuration information, and operating years. Power operation status data includes real-time measurement data and historical statistical data. The real-time measurement data includes voltage, current, power factor, load rate, and three-phase imbalance. The historical statistical data includes daily line loss rate, hourly load curve, and voltage qualification rate. External environmental impact data includes meteorological data and time characteristic data. The meteorological data includes weather type, temperature, and humidity. The time characteristic data includes holiday markers and peak / valley electricity consumption periods. Energy access data includes the installed capacity, output power curve, grid-connected electricity volume, and access location information of distributed power sources, including photovoltaic, wind power, and energy storage devices.

[0032] (1) Data collection: Based on the data requirements for abnormal line loss diagnosis in the transformer area, collect transformer area archives, topology data, operation data, line loss data, historical diagnosis results, distributed power supply access status, and meteorological data.

[0033] 1) Archive data: Key information includes transformer area name (character type), transformer area number (character type), transformer area model (character type), transformer area capacity (numerical type), power supply radius (numerical type), main line model (character type), main line cross-sectional area (numerical type), number of users (numerical type), number of single-phase meters (numerical type), number of three-phase meters (numerical type), operating status (character type), meter replacement information (character type), and years of operation (numerical type).

[0034] 2) Topology data: Key information includes the transformer area name (character type), transformer-point relationship data (character type), and transformer-transformer relationship data (character type).

[0035] 3) Operational Data: Key information includes data acquisition success rate (numerical), reverse power consumption (numerical), total installed capacity of distributed power sources (numerical), active power consumption of distributed grid connection (numerical), reactive power consumption of distributed grid connection (numerical), A-phase voltage (numerical), B-phase voltage (numerical), C-phase voltage (numerical), A-phase current (numerical), B-phase current (numerical), C-phase current (numerical), daily maximum power factor of the distribution area (numerical), daily minimum power factor of the distribution area (numerical), daily average power factor of the distribution area (numerical), and daily average power factor of the distribution area (numerical). Load factor (numerical), daily maximum load factor of the transformer area (numerical), three-phase imbalance of the transformer area (numerical, calculated as (maximum current - minimum current) / maximum current × 100%), voltage qualification rate of the transformer area (numerical, the voltage of the low-voltage outlet side of the transformer area is qualified within the range of rated voltage × [0.95, 1.05], voltage qualification rate = number of qualified moments in the transformer area in a day / total number of transformer areas × 100%), whether the voltage is lost in phase (numerical, sourced from the data acquisition system (loss of voltage and loss of phase judgment)), upper surface of the transformer area meter (numerical), lower surface of the transformer area meter (numerical).

[0036] 4) Line loss data: Key information includes input power (numerical), output power (numerical), sales power (numerical), daily line loss rate (numerical), theoretical line loss rate (numerical), and double rate deviation (numerical, daily line loss rate - theoretical line loss rate).

[0037] 5) Historical diagnostic results data: Key information includes transformer area name (character type), transformer area number (character type), time (character type), line loss rate (numerical type, daily line loss rate for the same period), first-level abnormality cause (character type), second-level abnormality cause (character type), third-level abnormality cause (character type), problem analysis (character type), and rectification suggestions (character type).

[0038] 6) Distributed photovoltaic grid connection status: Key information includes name (character type), substation area (character type), and total installed capacity (numerical type).

[0039] 7) Meteorological data: Key information includes weather type (character type, such as sunny, cloudy and rainy) and maximum temperature (numerical type).

[0040] (2) Data processing:

[0041] 1) Data cleaning techniques are used to remove noise, duplicates and errors from the collected data. For a small number of missing data, the mean is used to fill in the missing data.

[0042] 2) Based on timestamps and device identifiers, multi-source data are correlated and integrated to construct a comprehensive dataset containing information such as power operation, meteorological environment, user behavior and device topology.

[0043] 3) Use data standardization methods to normalize data of different formats and magnitudes.

[0044] Step S2 is the transformer area line loss anomaly diagnosis. The transformer area line loss anomaly diagnosis layer mainly includes three parts: rule-based transformer area line loss anomaly diagnosis, random forest-based transformer area anomaly diagnosis, and fusion output of diagnosis results. The cloud server, relying on a large model, establishes a data channel with the terminal processor. The large model is responsible for controlling the parallel execution of the rule-based and random forest-based transformer area anomaly diagnosis tasks, and then fusing the diagnosis results to form the final transformer area anomaly diagnosis result.

[0045] In some embodiments, the specific process by which the rule model identifies simple routine anomalies in a distribution area based on the basic attribute data and power operation status data in the comprehensive dataset, according to preset anomaly diagnosis rules, includes the following sub-steps:

[0046] S2.1. Based on the basic attribute data (including transformer capacity, power supply radius, main line model, and years of operation) and power operation status data (including voltage, current, line loss rate, and three-phase imbalance) in the comprehensive dataset, an anomaly classification system is obtained. The anomaly classification system includes four types of primary anomalies: file relationship problems, equipment metering problems, customer electricity consumption problems, and power grid technical problems, as well as 12 types of secondary anomalies and 40 types of tertiary anomalies. Each type of anomaly corresponds to a set of judgment conditions formed by the combination of basic attribute characteristics and operation status characteristics (e.g., transformer operating years > 20 years + monthly line loss rate fluctuation > 8% corresponds to transformer aging anomaly).

[0047] More specifically, Section S2.1 categorizes abnormal line losses in transformer substations into 4 Level 1 anomalies, 12 Level 2 anomalies, and 40 Level 3 anomalies, and establishes diagnostic rules for diagnosing abnormal line losses in transformer substations. The specific basic rules are as follows:

[0048] Level 1 Anomaly: Issue with Record Relationships

[0049] Secondary category: Marketing system file issues:

[0050] Level 3 Classification: **Missing Transformer Area Archives (Same Period):** Rule Description: Using the marketing transformer area as the standard, compare the details of the transformer area archives for the same period on the same day. No details for the same period is considered abnormal. **Dual-Region Transformer Area Status Not Synchronized (Same Period):** Rule Description: Using the marketing transformer area as the standard, compare the status of the transformer area archives for the same period on the same day. Inconsistency is considered abnormal. **Dual-Region Transformer Area Master Table Missing (Same Period):** Rule Description: Using the marketing transformer area as the standard, compare the information in the digital transformer area master table for the same period on the same day. No information for the same period is considered abnormal. **Dual-Region Distribution Transformer Archives Missing (Same Period):** Rule Description: Using the same-source transformer ledger as the standard, compare the details of the distribution transformer archives for the same period on the same day. No details for the same period is considered abnormal. **Dual-Region Transformer-Household Relationship Not Synchronized (Same Period):** Rule Description: Using the same-source distribution transformer-household as the standard, compare the transformer-household relationship table for the same period on the same day. Inconsistency is considered abnormal. **Dual-Region Distribution Transformer Status Not Synchronized (Same Period):** Rule Description: Using the same-source transformer ledger as the standard, compare the details of the distribution transformer status for the same period on the same day. Inconsistency is considered abnormal. Missing table data in the middle platform or concurrent data collection. Rule description: An anomaly is determined when all of the following conditions are met (difficulty: Currently, concurrent data collection involves transmitting table data back to the network via retesting and aggregation. The tables corresponding to the retesting and aggregation do not have table data entry time information. Accurate judgment requires access to the middle platform database for comparison): 1. Complete data collection; 2. Table data entry time (time of the scheduled task synchronized to the middle platform); 3. Total table electricity consumption in the user acquisition area > 0; 4. No table data for the concurrent data collection day. Missing table replacement records in the concurrent data collection. Rule description: Using the marketing system's table replacement records as the standard, compare the table replacement details for the concurrent data collection day. No details for the concurrent data collection day are considered an anomaly.

[0051] Secondary category: Distribution and logistics integration issues:

[0052] Level 3 Classification: No corresponding transformer for the marketing area. Rule description: An anomaly occurs when no corresponding transformer can be found for the marketing area. One-to-many or many-to-many relationships between the marketing area and corresponding transformers. Rule description: This anomaly occurs when multiple corresponding transformers are found in the marketing area or multiple areas find the same transformer. Inconsistent details between the marketing user and corresponding transformer user records. Rule description: Compare the marketing user details with the corresponding transformer user details for the day; inconsistencies indicate an anomaly. Inconsistent status of the corresponding transformer for the marketing area. Rule description: Compare the status of the corresponding transformer with the marketing area's records for the day; both are operational, which is normal; otherwise, it is an anomaly.

[0053] Secondary category: Marketing system file issues:

[0054] Level 3 Classification: Marketing system transformer area operation status is unreasonable. Rule 1 description: The following conditions must be met simultaneously to be considered abnormal: 1. The marketing transformer area's operation status is not "operating" (i.e., under maintenance, dismantled, out of service, awaiting installation, not yet operational, etc.); 2. The total electricity consumption of the same period's meter on the same day is >0. Rule 2 description: The following conditions must be met simultaneously to be considered abnormal: 1. The transformer area has no master meter file; 2. The transformer area has no customer relationship; 3. The marketing transformer area is in operation; 4. The same source transformer area is in operation. Marketing transformer area new commissioning is not timely. Rule description: The following conditions must be met simultaneously to be considered abnormal: 1. The transformer area is connected to a distribution transformer; 2. There are no low-voltage customers in the same period's transformer area; 3. The new commissioning date on the same day is >7 working days. Marketing transformer area connected to high-voltage customers. Rule description: The following conditions must be met simultaneously to be considered abnormal: 1. The marketing transformer area is connected to a distribution transformer; 2. Both the marketing transformer area and the distribution transformer are in operation; 3. There are high-voltage customers in the marketing customer relationship. The marketing meter replacement process is abnormal. The rule description is as follows: An abnormality is determined when all of the following conditions are met: 1. The meter bottom line is reversed after replacement (the bottom line of the main meter and user meters is reversed; the bottom line before replacement (forward / reverse) is less than the top line before replacement (forward / reverse) or the bottom line after replacement (forward / reverse) is less than the top line after replacement). 2. The replacement multiplier of the main meter / user meter is inconsistent with the actual value. 3. The line loss rate is greater than 10% or less than -1%, or greater than the sum of the median line loss and fluctuation threshold of the past two months. 4. The absolute value of the Pearson correlation coefficient for newly installed or replaced high-voltage users and all electricity consumption and power loss in the past week is greater than 0.8. The marketing area main meter record creation is abnormal. The rule description is as follows: An abnormality is determined when all of the following conditions are met: 1. The marketing area has a connected distribution transformer. 2. Both the area and the distribution transformer are in operation. 3. There is no main meter record for the area. 4. The area's electricity sales are greater than 0. The main purpose type of the metering point is incorrect. Rule description: It is abnormal if the following conditions are met simultaneously: 1. There is no main meter in the same period of the transformer area; 2. The main purpose type of the metering point for the marketing transformer area assessment includes "transformer area power supply assessment". The office electricity metering point has disappeared. Rule description: It is abnormal if the following conditions are met simultaneously: 1. High loss in the transformer area on the day; 2. No metering relationship for office electricity. The metering information is unreasonable. Rule description: It is abnormal if the following conditions are met simultaneously: 1. The transformer area main meter multiplier configuration is unreasonable (high supply, low metering, current transformer multiplier < capacity / 0.4 / 1.732 / 5*1 / 3 or multiplier greater than capacity / 0.4 / 1.732 / 5*3); 2. The relationship between the metering point and the meter has changed (the relationship between the metering point and the meter has changed compared to the previous day).

[0055] Level 1 Anomaly: Equipment and Quantity Issues

[0056] Secondary category: Meter malfunction issues:

[0057] Level 3 Classification: User Power Consumption Sudden Change, Rule Description: Whether the daily power consumption deviation from the previous day exceeds 1.5 times the standard deviation (calculated using one month's historical data) or: Power consumption month-on-month change greater than 20%; Meter Failure, Rule Description: If the following conditions are met simultaneously, it is judged as abnormal: 1. Negative loss in the transformer area on the day; 2. The BC phase current at each point on the day is approximately 0; Data Acquisition Failure, Rule Description: If the following conditions are met simultaneously, it is judged as abnormal: 1. The line loss rate on the day is > 2 * the line loss rate of the previous day; 2. The user's daily power consumption is 0, but the user's power consumption on the previous day was 0. Quantity > 0; Clock anomaly, rule description: Daily line loss is high loss or negative loss, the cumulative line loss value in the past 6 days is within 1.5 standard deviations and the line loss meets 0.5, the transformer area line loss meets 0.7; Meter bottom freezing anomaly, rule description: if the following conditions are met simultaneously, it is judged as an anomaly: 1. Transformer area electricity sales > 0, 2. Transformer area main meter bottom is intact and the daily electricity consumption is 0; Metering error is large, rule description: if the following conditions are met simultaneously, it is judged as an anomaly: 1. Meter bottom is intact; 2. Transformer area negative loss on the day (-1 < line loss rate < 0); 3. No other anomalies.

[0058] Level 2 Anomaly: Meter Reading Time Issue

[0059] Level 3 Classification: Late meter reading time. Rule description: If the following conditions are met simultaneously, it is judged as abnormal: 1. Meter reading time > push to the middle platform (the middle platform is generally executed on a timer); 2. High loss in the distribution area on the same day; 3. Deviation in electricity sales between the two days before and after < adjustment value (a specific adjustment value needs to be defined).

[0060] Secondary category: User omission issues:

[0061] Level 3 classification: User omission, rule description: If the following conditions are met simultaneously, it is judged as abnormal: 1. The line loss of the transformer area is acceptable the day before, 2. The line loss of the transformer area is high on the day of the incident, 3. The deviation of the electricity sales volume in the two days before and after is less than the adjustment value (a specific adjustment value needs to be defined).

[0062] Level 1 Anomaly: User Power Supply Issues

[0063] Secondary classification: Abnormal correlation:

[0064] Level 3 classification: Abnormal correlation. Rule description: If the following conditions are met, it is judged as abnormal: 1. Calculate the Pearson correlation coefficient by taking the electricity consumption of the user or transformer area in the previous 21 days and the line loss electricity; 2. If the absolute value of the correlation coefficient is greater than 0.8, it is judged as high correlation.

[0065] Secondary category: Defaulted electricity use:

[0066] Level 3 classification: Electricity theft or meter malfunction, rule description: abnormal K value; after excluding power generation users, the Pearson correlation coefficient is the correlation between the transformer area loss rate and single-phase household electricity consumption in the previous 21 days (including the query day): -1≤R≤-0.7, which is suspected driver-related electricity theft; 0.7≤R≤1, which is suspected continuous electricity theft; Overcapacity electricity consumption, rule description: active power absolute value / operating capacity >100%, exceeding the threshold operating time for more than 10% of the total operating time, current >0.2A is considered as operating, and those with a capacity of 0 are excluded.

[0067] Level 1 Anomaly: Power Grid Technical Problem

[0068] Secondary category: Transmission area operation issues:

[0069] Level 3 Classification: Low Voltage for Users: Three-phase user voltage < 342V; Single-phase user voltage < 198V; Three-phase Imbalance: The three-phase imbalance of the main meter in the distribution area is calculated using 96 points of current. Current imbalance = [(maximum current - minimum current) / maximum current] * 100%. If current imbalance > 25%, and load rate = positive active power / rated capacity * 100, and load rate > 60%, both conditions must be met for more than 8 consecutive points to be considered abnormal; Low Power Factor: Calculate the daily average power factor of the distribution area. If it is less than 0.85, it is considered under-compensation; if it is greater than 1 or negative, it is considered over-compensation. Heavy Overload in the Distribution Area: Average load rate of the distribution area > 80%, and user voltage < 198V. Light Load in the Distribution Area: With the main meter intact, if one of the following conditions is met, it is considered light load: 1. Daily absolute value of electricity supplied / sold < capacity / 30; 2. Electricity loss is between -3 and 5 kWh.

[0070] Secondary classification: Dilapidated buildings in the area:

[0071] Level 3 classification: Aging of transformer substations, rule description: operating years of transformer substations > 20 years. Aging of power grid structure, rule description: operating years of conductors > 20 years.

[0072] Secondary category: Equipment malfunction issues:

[0073] Level 3 classification: Equipment cable fault, rule description: The following conditions must be met simultaneously: 1. Voltage fluctuation > 10%; 2. Zero-sequence current > zero; 3. Three-phase imbalance > 25%; 4. Cable temperature > 80℃.

[0074] Secondary category: Long power supply distance:

[0075] Level 3 classification: Long power supply distance, rule description: Power supply radius of the transformer area > 2km.

[0076] S2.2. Based on the aforementioned anomaly classification system, an initial rule base is constructed. The initial rule base stores the judgment logic of various anomalies in an IF-THEN structure and is iteratively optimized through a dynamic update mechanism: When the large model identifies a total of 5 new anomaly cases (not matched with the initial rule base but verified as real anomalies on-site) from the diagnostic results, the basic attribute data features (such as distributed power supply access capacity) and power operation status data features (such as the sudden drop in line loss rate) associated with this type of new anomaly are extracted. The Apriori algorithm is used to mine the association rules between features and anomalies to generate candidate rules. After the accuracy of the validation set is ≥90%, the candidate rule is included in the rule base, and the corresponding level of the anomaly classification system is updated at the same time.

[0077] When the large model reports a cumulative total of 5 new anomaly cases, rule updates are initiated. A deep analysis is performed on the key features, occurrence scenarios, and related data of these anomaly cases. Machine learning algorithms are used to integrate these newly discovered anomaly patterns into the existing diagnostic rule base, optimizing the recognition logic of the dynamically enhanced model. This improves the accuracy and timeliness of subsequent transformer substation line loss anomaly diagnosis. The large model extracts features from new anomaly cases (e.g., time characteristics and power characteristics of a sudden drop in line loss rate during photovoltaic grid connection); it uses an association rule mining algorithm (Apriori) to generate candidate rules (e.g., line loss rate < -3% during photovoltaic grid connection → reverse line loss anomaly); and it verifies the accuracy of the candidate rules (those with a verification set accuracy ≥ 90% are retained) and incorporates them into the basic rule base.

[0078] S2.3. Taking the dynamically updated rule base as the core, integrate the power industry standards and expert experience corresponding to the anomaly classification system, and construct a knowledge base with dynamic index based on the knowledge graph. The knowledge base is a semantic association network based on basic attribute data, operation status data, rules, standards and experience.

[0079] This step involves building a knowledge base with a dynamic index layer, centered on a dynamically updated rule base. It integrates power industry standards (such as GB / T13462-2008 and DL / T448-2016) and expert experience (diagnostic weight rules quantified by the analytic hierarchy process) corresponding to the anomaly classification system. The dynamic index layer, implemented using knowledge graph technology, establishes a semantic association network of basic attribute data, operational status data, rules, standards, and experience through entity association (e.g., associating three-phase imbalance data entities with power grid technical problem rule entities) and attribute mapping (e.g., mapping sudden increases in line loss rate characteristics to electricity theft anomaly judgment conditions). This network supports millisecond-level retrieval of cross-level knowledge. More specifically, this part involves the construction of a knowledge base for diagnosing line loss anomalies in transformer substations.

[0080] We collected power industry standards, line loss calculation theories, and expert experience, and combined them with established rules for diagnosing abnormal line losses in transformer substations to construct a knowledge base for diagnosing abnormal line losses. The specific contents are as follows:

[0081] (a) Integration of Power Industry Standards: This involves reviewing national standards (such as GB / T13462-2008 "Economic Operation of Power Transformers"), industry standards (such as DL / T448-2016 "Technical Management Regulations for Electricity Metering Devices"), and enterprise standards (such as a power grid company's "Transformer Area Line Loss Management Measures"). This provides a comprehensive basis for line loss management and anomaly diagnosis from multiple dimensions, including loss calculation, metering management, and transformer area definition. Specific standards are as follows: National Standard (GB / T): Article 5.3 of GB / T13462-2008 "Economic Operation of Power Transformers" stipulates that when the transformer load rate is between 30% and 75%, the standard coefficient method should be used for loss calculation; Appendix C clarifies the no-load loss and load loss thresholds for different transformer models, serving as a benchmark for loss anomaly judgment. GB / T34933-2017, "Guidelines for Calculation and Analysis of Line Losses in Power Systems," Article 6.2 requires that the line loss rate of low-voltage distribution networks should be ≤12%, medium-voltage line loss rate ≤7%, and ultra-high-voltage line loss rate ≤3%, serving as a quantitative standard for judging regional line loss anomalies. Industry standard (DL / T): DL / T448-2016, "Technical Management Regulations for Electricity Metering Devices," Article 5.4.3 stipulates that the comprehensive error of Class I and II metering devices should be ≤0.7%, and that of Class III should be ≤1.2%. Errors exceeding these ranges must be included in line loss anomaly investigations. Enterprise standard (taking a provincial power grid as an example): Article 3.2.1 of the "Specifications for Lean Management of Distribution Area Line Losses" defines a distribution area with a daily line loss rate >10% (>12% in rural areas) as a high-loss distribution area; fluctuations in the daily line loss rate exceeding 5 percentage points require triggering an early warning. Article 4.1.2 of the "Administrative Measures for Online Monitoring of Metering Devices" requires that if the clock error of a smart meter exceeds ±0.5 seconds / day, the impact of the metering data on the total loss must be recalculated.

[0082] (b) Expert Experience Transformation: A knowledge engineering methodology is used to construct a transformation system from tacit experience to structured rules. The specific process is as follows: Experience Collection Stage: Based on the core scenarios of line loss anomaly diagnosis, a semi-structured interview outline is designed for 20 typical operating conditions (including seasonal load changes, distributed power source access fluctuations, grid topology changes, equipment aging and failures, metering device errors, the impact of electricity theft, abnormal load curve fluctuations, unreasonable reactive power compensation, voltage deviation exceeding limits, line parameter changes, three-phase load imbalance, new energy grid connection impact, transformer tap adjustment, distribution line overload, environmental temperature influence, natural disasters such as lightning strikes, grid maintenance operations, changes in user electricity consumption behavior, harmonic interference influence, and tie line power fluctuations). Through in-depth interviews with 15 senior line loss experts, their diagnostic logic, key judgment nodes, and decision-making basis under complex operating conditions are systematically recorded. Element Decomposition Stage: Mind mapping tools are used to deconstruct the collected experience, abstracting the expert decision-making process into a three-layer transformation model of data feature extraction, weight quantification, and rule combination. Taking suspected electricity theft diagnosis as an example, experts typically consider key characteristics such as nighttime zero-load duration >8 hours (reflecting abnormal electricity consumption behavior) and a sudden increase in line loss rate >5% (reflecting electricity imbalance). Using the Analytic Hierarchy Process (AHP), the weights of these characteristics are calculated to be 0.6 and 0.4 respectively, constructing a joint judgment rule. In the rule modeling stage, the decomposed feature parameters, weight system, and decision logic are encoded into structured rules, forming a knowledge base containing 40 empirical rules. Three rounds of cross-validation are conducted using the Delphi method with a group of power industry experts, and rule accuracy tests are performed based on actual business scenario data to ensure that the rules match real-world operating conditions with a degree >90%.

[0083] (c) Knowledge base structure design: A layered hybrid architecture of rule base + dynamic index is adopted to realize the structured management and efficient retrieval of knowledge, and a full life cycle operation and maintenance mechanism is provided to ensure the continuous optimization of the knowledge base.

[0084] The knowledge-layered storage architecture comprises a rule base and a dynamic index layer, categorizing and storing various types of knowledge for easy and rapid retrieval and retrieval. The rule base employs a production rule notation (IF-THEN structure) to store standardized power industry clauses, expert experience rules, and diagnostic algorithms, and supports SQL queries and fuzzy matching. For example, when searching for diagnostic rules for abnormal line losses caused by three-phase imbalance, relevant rules can be quickly retrieved, ensuring accurate rule retrieval. The dynamic index layer, based on knowledge graph technology, constructs a semantic association network of "data-rules." When abnormal voltage deviation data is detected, it automatically associates with fault rule nodes of metering devices, achieving cross-database retrieval within seconds and intelligent recommendations. For example, when voltage deviation is abnormal, not only can the corresponding rules be found, but comprehensive references can also be quickly provided for diagnosis.

[0085] Full-cycle operation and maintenance mechanism: The full-cycle operation and maintenance mechanism safeguards the accuracy, timeliness and efficiency of the knowledge base: Dynamic updates: A dual review mechanism is established, with business experts and system administrators collaborating to modify or incrementally update the rules base every month; when industry standards or technical specifications change, manual knowledge maintenance is carried out to ensure that the knowledge base always keeps up with the forefront of the industry.

[0086] S2.4 Extract the basic attribute data and power operation status data of the transformer substation to be diagnosed from the comprehensive dataset, locate the matching rule base judgment conditions through dynamic indexing, verify the feature thresholds in combination with the industry standards in the knowledge base, and after verifying the consistency between the basic attribute data and power operation status data of the transformer substation to be diagnosed and the rules through the large model, output the diagnosis results including the anomaly category, feature basis and rule version.

[0087] The implementation process is as follows: basic attribute data and power operation status data of the transformer area to be diagnosed are extracted from the comprehensive dataset to form a standardized feature vector; the matching rule base judgment conditions are quickly located through the dynamic index layer (such as mapping the feature vector of three-phase imbalance > 25% + load rate > 60% to the three-phase imbalance anomaly rule); the feature threshold is verified in combination with the industry standard in the knowledge base (such as verifying whether the voltage qualification rate meets the requirement of ±5% of the rated voltage); after the large model verifies the consistency between the data and the rules (excluding misjudgments caused by data noise), the diagnostic results containing the anomaly category (up to three-level classification), feature basis and rule version are output.

[0088] The Qwen-2.5Instruct model (large model) possesses powerful language understanding and data analysis capabilities, accurately identifying abnormal line loss conditions in transformer substations. This improves the accuracy, efficiency, and intelligence of line loss anomaly diagnosis, providing strong support for optimizing line loss management. First, processed data (core data directly related to transformer substation line loss anomaly diagnosis extracted from the comprehensive dataset, after cleaning, association, fusion, and standardization, specifically focusing on basic attribute data and power operation status data required for rule-based model diagnosis) is input into the Qwen-2.5Instruct model in text form. Leveraging the model's powerful language understanding capabilities, the semantic relationships and logical connections between the data are analyzed. For example, the potential link between three-phase imbalance and line loss rate changes is analyzed. Then, by inputting a large amount of transformer substation line loss anomaly diagnosis case data, the model learns the diagnostic logic under different conditions. Through continuous iterative training, the model gradually refines and improves these diagnostic rules, enabling it to adapt to complex and ever-changing real-world scenarios and significantly improving diagnostic accuracy. For example, the model automatically identifies the mapping relationship between data features and diagnostic conclusions in different cases. When the electricity consumption of a certain line is consistently zero at night but the main meter still outputs electricity, there is a high probability of electricity theft. This diagnostic logic is transformed into quantifiable rules. Finally, the diagnostic rules adjusted and refined by the Qwen-2.5Instruct model are used to diagnose the line loss of the transformer area and determine the cause of the anomaly. The "quantifiable rules" extracted by the model through learning a large number of anomaly diagnostic cases (such as zero electricity consumption at night but output from the main meter → electricity theft) are the core input for the dynamic updating of the rule base in the knowledge base. These rules, automatically generated through data mining, complement traditional rules based on industry standards and expert experience, especially in covering new anomaly scenarios (such as reverse line loss caused by distributed power source access). For example, the judgment rules summarized by the model from 100 cases of sudden drop in line loss rate during photovoltaic grid connection periods will be incorporated into the rule base of the knowledge base after accuracy verification, becoming new nodes associated in the dynamic index layer.

[0089] The existing basic attribute data, operational status data, rules, standards, and empirical semantic network in the knowledge base provide knowledge anchors for model learning: when the model analyzes data correlations (such as the relationship between three-phase imbalance and line loss rate), it will refer to the limit requirements of three-phase imbalance in standards such as DL / T448-2016 in the knowledge base to avoid generating rules that conflict with industry standards; in the early stage of model training, expert empirical rules in the knowledge base (such as "prioritize checking meter faults in high-loss transformer areas") will be used as initial diagnostic logic inputs to accelerate model convergence.

[0090] Candidate rules generated during model training must pass a dynamic update mechanism of the knowledge base (e.g., a verification standard of accuracy ≥90%) before being added to the knowledge base. Meanwhile, the continuously enriched rules and related knowledge in the knowledge base serve as new training data input into the model, helping it refine its diagnostic logic (e.g., distinguishing the line loss rate threshold differences between rural and urban transformer substations). For example, a rule generated by the model that indicates a sudden increase in line loss rate will be associated with basic attribute data such as transformer substation capacity and power supply radius in the knowledge base, forming a more accurate rule for transformer substations with a capacity below 500kVA where a sudden increase in line loss rate >5% → an anomaly, which then feeds back into the model to improve diagnostic accuracy. In short, the Qwen-2.5Instruct model is an automatic generator of dynamic rules, and the knowledge base is a structured verification and storage carrier for rules. Together, they achieve data-driven rule evolution, ensuring that rules conform to industry standards and expert experience while quickly adapting to new anomaly scenarios.

[0091] In some embodiments, the random forest model locates complex and novel anomalies by mining the coupling relationship between energy access data, external environmental impact data, and line loss fluctuations in the comprehensive dataset.

[0092] S2.5 Multi-dimensional feature extraction: Extract time-series features to characterize the fluctuation pattern of line loss, correlation features to characterize the relationship between energy access and line loss, static features to reflect the inherent attributes of the transformer area, and dynamic features to reflect the real-time operating status from the comprehensive dataset, forming a multi-dimensional feature set covering time evolution, correlation coupling, equipment attributes, and operating status.

[0093] First, the time-series characteristics of line loss rate are calculated. For daily line loss and hourly load data of the distribution area, a sliding window (window size = 7) is used to extract trend features, including mean features (average line loss rate within the window), dispersion features (standard deviation of line loss rate), and change trend features (maximum increase in line loss rate within the window). Second, the correlation characteristics between load and line loss rate are analyzed, i.e., the correlation coefficient between load change rate and line loss rate change rate. Third, inherent equipment attribute characteristics (i.e., static characteristics) are extracted, including: distribution area capacity, power supply radius, main line type, main line cross-sectional area, number of users, number of single-phase meters, number of three-phase meters, and years of operation, to help identify technical problems in the distribution area's power grid. Finally, equipment operating status-related characteristics (i.e., dynamic characteristics) are extracted, including: distribution area load rate, power factor, three-phase imbalance, and user low voltage conditions.

[0094] S2.6 Key Feature Enhancement: From the aforementioned time-series features, correlation features, static features, and dynamic features, principal component analysis is used to screen key features that contribute up to 85% to the cumulative contribution rate of anomaly diagnosis. At the same time, an attention mechanism is introduced to assign dynamic weights to the time-series features during peak electricity consumption periods, thereby achieving dimensionality reduction optimization of key features and enhancement of high-value period features.

[0095] Considering the numerous factors influencing abnormal line losses in transformer substations, this invention employs Principal Component Analysis (PCA) to extract key features of line loss anomalies to reduce the impact of redundant features on the model. PCA effectively extracts key features of line loss anomalies and calculates their contribution. Therefore, PCA is used to calculate the contribution of various features to the line loss rate of transformer substations, selecting features with high contribution to the diagnosis of abnormal line losses. This invention retains the top k principal components with a cumulative contribution rate of 85% as key features. Simultaneously, an attention mechanism is introduced to construct a dynamic weight allocation strategy based on a time window, assigning a 1.5-fold weight to time-series features during peak electricity consumption periods (e.g., 18:00-22:00). During this period, residential electricity demand and commercial office electricity demand surge, resulting in drastic fluctuations in power data and a wealth of abnormal features. By strengthening the weight proportion of data during this period, key signals of abnormal events such as peak load mutations, electricity theft, and equipment failures can be accurately captured, significantly improving the sensitivity and accuracy of the anomaly diagnosis model in highly dynamic scenarios. Simultaneously, by combining sliding window technology, the weight coefficients are adaptively adjusted according to the fluctuation trend of historical peak data to ensure that the model has dynamic adaptability to peak load changes in different seasons and date types.

[0096] In some embodiments, an attention mechanism is introduced to assign dynamic weights to the temporal characteristics of peak electricity consumption periods. Specifically, this is done as follows:

[0097] A dynamic weight allocation strategy based on time windows is constructed, assigning an initial weight of 1.5 times to the temporal characteristics of peak electricity consumption periods (e.g., 18:00-22:00). Combined with sliding window technology, and using the fluctuation trend of line loss data during historical peak periods as a benchmark, the weight coefficients are adaptively adjusted by calculating the similarity of features between the current peak period and historical peak periods (including load fluctuation amplitude or the slope of line loss rate change). Specifically: when the similarity between the current peak period data fluctuation trend and the historical typical peak period anomaly pattern is ≥80%, the weight coefficient is increased to 1.5-2.0 times; when the similarity is in the 50%-80% range, the initial weight is maintained at 1.5 times; when the similarity is <50%, the weight coefficient is reduced to 1.2-1.5 times. Through this dynamic weight adjustment mechanism, the characteristic proportion of abnormal signals such as sudden load changes, electricity theft, and equipment failures during peak periods in highly dynamic scenarios is strengthened, improving the model's adaptability to peak load changes in different seasons and date types, and enhancing its sensitivity to anomaly diagnosis.

[0098] S2.7 Random Forest Model Construction: Key features (including static features, enhanced time-series features, energy access data, and external environmental impact data) are integrated to form feature vectors. A random forest model is constructed using a hybrid splitting criterion weighted by the Gini coefficient and the contribution of time-series features. The model parameters are optimized by maximizing the F1 score of the large model, enabling the random forest model to discover the correlation between multi-dimensional features and line loss anomalies. This step utilizes feature-engineered data to construct the random forest model. A random forest is an ensemble learning model composed of multiple decision trees. During construction, multiple sample subsets are randomly extracted from the original dataset with replacement, and each sample subset is used to train a decision tree. When splitting at each node of the decision tree, a subset of features is randomly selected for optimal splitting features to reduce the correlation between trees and improve the model's generalization ability.

[0099] In some embodiments, the specific process of S2.7 is as follows:

[0100] S2.7.1 Feature Vector Construction: Static features (including basic transformer attributes such as rated capacity, number of users, and power supply radius) and enhanced time-series features (including real-time monitoring data processed by a sliding window, such as voltage curve fluctuations, load peak-valley differences, time-of-use power factors, and daily trends in line loss rate over the past 30 days) are integrated with energy access data, meteorological data (temperature and humidity), and external influencing factors such as holidays to form a multi-dimensional feature vector. First, static features (transformer capacity and number of users) and enhanced time-series features are integrated. Multi-dimensional feature vectors are constructed by integrating multi-dimensional data. Static features cover basic transformer attributes such as rated capacity, number of users, and power supply radius, used to characterize the inherent characteristics of the transformer area; dynamic enhanced time-series features include real-time monitoring data such as voltage curve fluctuations, load peak-valley differences, time-of-use power factors, and daily trends in line loss rate over the past 30 days, capturing the temporal patterns of power operation through sliding window technology. In addition, meteorological data (temperature and humidity) and external influencing factors such as holidays are introduced to form a feature set that comprehensively reflects the operating status of the transformer area.

[0101] S2.7.2 Hybrid Splitting Criterion Design: A hybrid evaluation index (S=α×G+(1-α)×C) is constructed, weighted by the Gini coefficient and the contribution of time-series features, where G is the Gini coefficient representing the spatial dispersion of line loss data, C is the contribution of time-series features representing the dynamic change trend of the time series, and α is an adaptive weighting coefficient of 0.4-0.6. During decision tree node splitting, this index dynamically balances the influence of spatial distribution characteristics and temporal evolution on the classification results. This step proposes a hybrid splitting criterion weighted by the Gini coefficient and the contribution of time-series features. By quantifying the spatial dispersion (Gini coefficient) and the dynamic change trend of the time series (time-series feature contribution) of the transformer area line loss data, a multi-dimensional evaluation index system is constructed. During the decision tree node splitting process, an adaptive weight allocation mechanism dynamically balances the influence of spatial distribution characteristics and temporal evolution on the classification results. Compared with the traditional single splitting criterion, this can more accurately capture abnormal fluctuation patterns in line loss data, significantly improving the timeliness and accuracy of abnormal line loss diagnosis in transformer areas.

[0102] Adaptive weights can be adjusted using information entropy and anomaly patterns, as follows:

[0103] S2.7.2.1 Introducing Information Entropy to Measure Feature Stability: Calculating Feature Information Entropy: For spatially distributed features (measured by the Gini coefficient) and features exhibiting temporal evolution (measured by the contribution of time-series features), calculate their information entropy within different time windows. Information entropy reflects the uncertainty or stability of a feature; the higher the information entropy, the more complex the feature's changes and the higher its uncertainty. Let the spatially distributed feature sequence be... Its probability distribution is The information entropy of spatial distribution characteristics Similarly, for the characteristic sequence of time evolution patterns... Calculate its information entropy . , These are the i-th spatial distribution feature and the j-th temporal evolution feature, respectively. The weighting coefficient α is initially adjusted based on the relationship between the information entropy of the two features. This indicates that the spatial distribution characteristics are more complex. In this case, appropriately increasing α (within the range of 0.5-0.6) will make the decision tree more inclined to base its judgments on the spatial distribution characteristics when splitting nodes, as these characteristics contain more relevant information. If α is reduced, its value range is adjusted to 0.4-0.5 to highlight the influence of time evolution on the classification results.

[0104] S2.7.2.2 Optimizing Weights Based on Abnormal Pattern Changes: Constructing an Abnormal Pattern Library: Through analysis of historical line loss anomaly data, different types of abnormal patterns are summarized, such as the three-phase imbalance anomaly pattern under high load and the line loss fluctuation anomaly pattern caused by new energy access, etc., and corresponding feature weight preferences are set for each abnormal pattern. Real-time Matching of Abnormal Patterns: During model operation, when new line loss anomaly data is detected, machine learning algorithms (such as clustering algorithms and pattern recognition algorithms) are used to match it with patterns in the abnormal pattern library. If an abnormal pattern is matched, and the spatial distribution characteristics of that pattern are more critical for anomaly judgment, even if the current... Alternatively, α can be temporarily adjusted to the range of 0.5-0.6 to more accurately identify this type of anomaly. Conversely, if the matched anomaly pattern emphasizes the importance of temporal evolution characteristics, even... Alternatively, α can be temporarily adjusted to the range of 0.4-0.5.

[0105] S2.7.2.3 Long-term optimization based on feedback mechanism: Model performance evaluation: The performance of the random forest model is evaluated periodically using test set data, focusing on indicators such as accuracy, recall, and F1 score. Weight coefficient feedback adjustment: Based on the model performance evaluation results, the rationality of the current weight coefficient α setting is analyzed. If the model has low accuracy and recall when identifying certain types of anomalies, the actual impact of the spatial distribution characteristics and temporal evolution patterns of such anomaly data is compared to determine whether α needs to be adjusted. For example, if it is found that the identification effect of line loss anomalies caused by seasonal load changes is not good, and analysis shows that the temporal evolution patterns are more critical, but the current α value is too large, then α is appropriately reduced. After multiple evaluations and adjustments, a relatively stable adaptive α adjustment strategy is formed for different data characteristics and anomaly types. The innovation of this adaptive weight evaluation method is that it comprehensively considers the uncertainty of features (through information entropy) and the dynamic changes of anomaly patterns, and can more flexibly and accurately balance the impact of spatial distribution characteristics and temporal evolution patterns on classification results according to the actual data situation, which is more adaptable than traditional fixed-range adjustment. By introducing a feedback mechanism, the weight coefficients are optimized over a long period based on model performance, so that the weight coefficient settings continuously approach the optimal solution, thereby continuously improving the ability of the random forest model to discover the correlation between multi-dimensional features and abnormal line loss.

[0106] S2.7.3 Random Forest Ensemble Construction: Multiple sample subsets are extracted from the feature vector using random sampling with replacement. Each subset is used to train a decision tree. During the node splitting process of each decision tree, some features are randomly selected from the sample subsets and the optimal splitting features are determined through the above-mentioned hybrid splitting criteria to reduce the correlation between trees and improve the generalization ability of the model.

[0107] This step is the core design of the random forest model to reduce the correlation between trees and improve generalization ability. It can be understood by breaking down the principle and using analogies and examples, with a focus on explaining the meaning and role of randomly selecting some features.

[0108] The core advantage of random forests is that they use multiple different decision trees to 'vote' for judgments. The greater the difference between the trees (the lower the correlation), the more reliable the final voting result, effectively avoiding the problem of overfitting a single decision tree (only fitting the training data and making inaccurate judgments on new data). Randomness, with its sampling with replacement (i.e., bootstrap sampling) and random selection of some features, are two key means of creating differences between trees: the former ensures that each tree is trained on different data (e.g., tree A uses samples 1, 3, and 5, while tree B uses samples 2, 3, and 6); the latter ensures that each tree makes judgments based on different 'dimensional criteria' (e.g., tree A uses line loss rate increase + transformer capacity for judgment, while tree B uses three-phase imbalance + weather temperature for judgment). The multidimensional feature vector contains all the features that are the key feature set enhanced in step S2.7.2, such as: static features: transformer area capacity, power supply radius, and trunk line cross-sectional area; enhanced time-series features: the daily trend of line loss rate in the past 30 days and the peak-valley difference of load during peak hours; external features: daily average temperature and holiday identification; correlation features: the correlation coefficient between distributed power output and line loss rate.

[0109] Assume there are 10 features in total, denoted as "F1 (transformer capacity), F2 (power supply radius), ..., F10 (holiday indicator)". When training a decision tree, before splitting at each node (e.g., determining if a node is an anomaly due to high line loss), not all 10 features are used. Instead, a subset (e.g., 3 features) is randomly selected from these 10 features, and the feature that best distinguishes between anomalies and normal conditions is chosen based solely on these 3 features. For example, when training the first decision tree, node 1 might randomly select "F1 (transformer capacity), F3 (daily trend of line loss rate), F8 (temperature)," and the feature that best distinguishes anomalies (e.g., the daily trend of line loss rate) is chosen as the splitting criterion. When training the second decision tree, node 1 might randomly select "F2 (power supply radius), F5 (peak load peak-valley difference), F10 (holidays)," and the optimal splitting feature is chosen from these 3. Typically, the number of features randomly selected each time is set (e.g., the square root of all features; if there are 10 features, 3-4 are selected) to ensure that the selected feature subsets are different each time, yet still cover key information. If each tree splits using all 10 features, the judgment logic of all trees will be highly similar. For example, all trees will prioritize splitting based on the increase in line loss rate (the most intuitive anomaly feature), ultimately resulting in multiple trees making the same judgment repeatedly, thus negating the purpose of multi-tree voting.

[0110] Randomly selecting some features forces each tree to focus on different feature combinations: some trees rely on static features + temporal features for judgment, while others rely on temporal features + external features for judgment; even if a tree makes an inaccurate judgment because it did not extract key features, other trees can make up for it through different feature combinations, and the overall judgment is more robust (with stronger generalization ability).

[0111] S2.7.4 Model Training and Parameter Optimization: The sample subset is divided into a training set and a test set. The random forest model is trained using the training set. Leveraging the inference capabilities of the Qwen-2.5Instruct model, with maximizing the F1 score as the optimization objective, the parameters of the random forest model (including the number of decision trees (100-300), the maximum tree depth (10-30 layers), and the minimum number of sample splits) are adaptively adjusted to enable the random forest model to discover multi-dimensional features and the correlation between line loss anomalies. The feature-engineered data is divided into training and test sets, and the constructed random forest model is trained using the training set. During training, the parameters of the random forest (such as the number of decision trees, the maximum tree depth, and the minimum number of sample splits) are tuned using the inference capabilities of the Qwen-2.5Instruct model, with maximizing the F1 score as the optimization objective, to ensure the model can accurately classify and predict the input data. Finally, the trained random forest model is evaluated using the test set. Performance metrics such as accuracy, recall, and F1 score are calculated to determine the performance of the random forest model.

[0112] S2.8 Anomaly Localization Output: Using the trained random forest model, complex and novel anomalies are located by identifying the coupling patterns between abrupt changes in energy access parameters, anomalies in external environmental factors, and fluctuations in line loss. The diagnostic results include anomaly type, feature contribution, and associated influencing factors. The model performance is verified by accuracy, recall, and F1 score.

[0113] The innovation of the random forest in this invention lies in its deep integration of the conventional model framework with the spatiotemporal coupling of transformer line loss and the characteristics of multi-factor interference scenarios, thus solving the core pain points of traditional random forests in this scenario:

[0114] The limitations of traditional random forests in scene adaptation: When applied to line loss diagnosis, traditional random forests often suffer from the following problems: Undifferentiated feature processing: Static features (such as transformer capacity), time-series features (such as daily changes in line loss rate), and external features (such as temperature) are treated equally, ignoring the strong signal and high temporal correlation of abnormal line loss peak periods; Simplified splitting criteria: Using only the Gini coefficient (spatial dispersion) or information gain (feature discrimination), it is impossible to balance the spatial influence of the inherent attributes of the transformer area and the temporal evolution of power data (such as the difference in abnormal line loss patterns in different time periods of the same transformer area); Empirical parameter tuning: Relying on manual trial and error to adjust parameters such as the number of decision trees and tree depth, it is difficult to adapt to the differences in line loss fluctuations in different seasons and different transformer area types (rural / urban).

[0115] This solution addresses the aforementioned pain points through scenario-based modifications: At the feature level, an attention mechanism and sliding window are introduced to dynamically weight (1.2-2.0 times) the temporal features during peak electricity consumption periods (18:00-22:00), amplifying abnormal signals during peak hours (such as electricity theft and equipment overload) and resolving the issue of "undifferentiated temporal features." At the criterion level, a hybrid splitting criterion (S=α×G+(1-α)×C) is designed, combining the Gini coefficient and the contribution of temporal features. The weight coefficient α (0.4-0.6) is dynamically adjusted through information entropy and abnormal pattern matching to balance the influence of spatial attributes and temporal patterns, addressing the issue of a singular splitting criterion. At the parameter tuning level, the Qwen-2.5Instruct large model is introduced, adaptively optimizing parameters with the goal of maximizing the F1 score, replacing manual parameter tuning based on experience and resolving the problem of poor parameter adaptability. This customized modification targeting specific scenario pain points far surpasses the simple combination of conventional random forests and line loss data, forming a creative core foundation.

[0116] In some embodiments, the specific process by which the large model fuses and verifies the diagnostic results of the rule model and the random forest model to output the final anomaly diagnostic result for the transformer area is as follows:

[0117] S2.9 Preliminary acquisition of diagnostic results: Obtain independent diagnostic results for the two types of models respectively. The rule model outputs anomaly judgment results and corresponding rule basis based on manually defined business rules (including transformer area line loss rate threshold, voltage qualification rate, etc.). The random forest model outputs anomaly judgment results, feature contribution and anomaly category probability based on multi-dimensional feature combination.

[0118] S2.10. Phased Integration Strategy Development: Combining the advantages and disadvantages of the two models and diagnostic objectives, develop differentiated integration strategies according to the application stage:

[0119] S2.10.1, Preliminary Application: With the goal of expanding the scope of anomaly screening and accumulating anomaly samples, a result union fusion strategy is adopted to integrate all abnormal station data (including anomaly results determined by a single model) identified by the rule model and the random forest model. The resulting samples are used to optimize the rule base and train the random forest model.

[0120] S2.10.2 Mid-term application: Based on the abnormal samples accumulated in the early stage, analyze the high-frequency abnormal patterns identified by the random forest model and transform them into interpretable business rules (such as transforming the combination of features extracted by the model, such as load peak-to-valley difference exceeding 30% and line loss rate suddenly increasing by 5%, into rules), and supplement them to the rule base of the rule model to reduce the dependence on the random forest model. When fusion, the judgment results of the updated rule model are used first, and the results of the random forest model are used as auxiliary verification.

[0121] S2.10.3 Application Maturity Stage: Differentiated fusion based on transformer area scenarios. For transformer areas associated with core indicators (such as high-loss transformer areas), a result intersection fusion strategy is adopted, retaining only the results that are judged as abnormal by both models to ensure diagnostic accuracy; for transformer areas in edge scenarios (such as low-loss transformer areas), a result union fusion strategy is adopted to avoid missing potential anomalies.

[0122] S2.11 Optimization of Fusion Result Verification: For abnormal distribution areas output by the union of results, the rationality of the single model's judgment result is verified by combining the Qwen-2.5Instruct large model with the historical operating data of the distribution area (the trend of line loss fluctuation in the past 30 days and equipment failure records), and false alarms caused by overfitting of the random forest model are eliminated; for abnormal distribution areas output by the intersection of results, the business rule basis of the rule model and the feature contribution of the random forest model are combined to generate dual judgment basis, thereby improving the interpretability of the results.

[0123] S2.12 Final Result Output: The output includes the abnormal station area number, abnormal category, application stage, fusion strategy type, judgment basis (rule basis + feature basis), and final diagnostic result with confidence level. The confidence level is set according to the fusion strategy: the confidence level of the early stage union result is set to 60%-70%, the confidence level of the mid-term auxiliary verification result is set to 75%-85%, the confidence level of the mature stage intersection result is set to 90%-100%, and the confidence level of the edge scene union result is set to 70%-80%.

[0124] Rule-based methods primarily rely on manually defined business rules (such as transformer area line loss rate thresholds and voltage qualification rates) to identify anomalies. They are logically transparent and highly interpretable, making them suitable for scenarios with clearly defined business rules. However, their dependence on human experience may lead to missed anomaly patterns not covered by the rules, resulting in a high risk of missed detections. Random forest methods utilize machine learning models trained on historical data to automatically identify anomaly patterns through feature combinations. They can capture nonlinear relationships and complex feature interactions, exhibiting strong generalization ability and making them suitable for data-driven, refined diagnosis. However, their results have weaker interpretability, requiring sufficient high-quality data support, and they may face the risk of model overfitting, leading to a high risk of false alarms. For both anomaly diagnosis methods, taking the intersection of diagnostic results can prioritize diagnostic accuracy but increases the risk of missed detections. Taking the union can improve the comprehensiveness of anomaly identification but may increase the false alarm rate.

[0125] Therefore, to ensure the accuracy and comprehensiveness of abnormal line loss diagnosis in transformer substations, and considering the advantages and disadvantages of both methods, in the early stages of application, the union of diagnostic results is used to expand the screening scope and accumulate abnormal samples to optimize rules and models. In the mid-term, through sample analysis, high-frequency abnormal patterns identified by the random forest method are gradually transformed into rules (e.g., combining features extracted by the model into interpretable business rules), reducing model dependence. In the mature stage, the intersection of core indicators (such as high-loss transformer substations) is taken to ensure accuracy, while the union of peripheral scenarios (such as low-loss transformer substations) is taken to avoid missed detections.

[0126] The specific implementation process of S3 is as follows:

[0127] S3.1 Structured Analysis of Diagnostic Information: The large model analyzes the final abnormality diagnosis results of the transformer substation and extracts key information including: abnormality type (including primary, secondary and tertiary classifications), abnormality characteristic parameters (such as line loss rate exceeding the standard value, load change amplitude, three-phase imbalance), basic attributes of the transformer substation (capacity, power supply radius, main line model), real-time operating data (current load rate, power factor) and energy access characteristics (distributed power source type and installed capacity).

[0128] S3.2 Knowledge Base Strategy Matching: Based on the parsed anomaly type, the loss reduction strategy knowledge base constructed using a three-level classification method (4 types of first-level anomalies, 12 types of second-level anomalies, and 40 types of third-level anomalies) is invoked to retrieve and match candidate loss reduction strategies directly related to the current anomaly type. Candidate loss reduction strategies must include power industry standards, historical implementation cases, and quantifiable execution indicators (such as response time and loss reduction target value).

[0129] S3.3, Distribution Area Feature Adaptation Screening: The large model combines the actual data characteristics of the distribution area to make differentiated adaptation judgments on candidate loss reduction strategies: For small-capacity distribution areas with high load rate characteristics, priority is given to retaining adaptation measures such as load shifting and transformer capacity expansion; for distribution areas with distributed power supply access, special strategies such as reverse power limitation and reactive power compensation optimization are selected; referring to the implementation effects of similar historical cases with the current distribution area characteristics, strategies with loss reduction efficiency <10% or implementation costs exceeding the budget threshold are eliminated.

[0130] S3.4 Strategy Priority Ranking and Output: The selected strategies are ranked and output according to two dimensions: Urgency dimension: Fault-related strategies that require a response within 24 hours (such as emergency line repair) are prioritized over transformation strategies planned within 7 days (such as line phase switching); Effectiveness stability dimension: Strategies with a historical implementation success rate ≥80% are prioritized over strategies with a low success rate; The final output includes a differentiated loss reduction plan that includes anomaly type, adaptation characteristics, recommended strategy, standard basis, execution indicators, and matching degree rating. The matching degree rating is divided into three levels: "high (≥80%), medium (50%-80%), and low (<50%)" based on the fit between the strategy and the characteristics of the transformer area.

[0131] Based on the abnormal line loss diagnosis results of the transformer area, and combined with the loss reduction strategy knowledge base, the model intelligently recommends line loss reduction strategies for the transformer area according to the specific abnormal situation. Specific strategies are described in the loss reduction strategy knowledge base. For example, for situations where the line loss rate is too high and the transformer area has excessive three-phase imbalance, optimizing the load distribution may be recommended; for problems caused by high losses due to line aging, the model may suggest specific line modifications or replacement of high-energy-consuming equipment.

[0132] In some embodiments, the process of constructing and maintaining the loss reduction strategy knowledge base is as follows:

[0133] Knowledge base content collection and standardization:

[0134] The system systematically reviews existing and valid power industry standards, removes obsolete versions, and stores the loss reduction technical requirements specified in the standards according to the type of anomaly.

[0135] Collect historical loss reduction cases, screen and retain complete records containing four elements: anomaly type, handling measures, implementation effect, and transformer area characteristics; sort cases with multiple measures for the same anomaly by loss reduction efficiency (effectiveness / cost ratio);

[0136] Industry experts are organized to transform empirical statements (such as timely handling) into quantifiable implementation standards (such as response time ≤ 24 hours and implementation cycle ≤ 7 days).

[0137] The three-level classification system is constructed as follows: Based on a hierarchical structure of 4 primary anomaly categories (e.g., abnormal line loss, abnormal equipment operation, abnormal load characteristics, and abnormal energy access), 12 secondary categories, and 40 tertiary categories, loss reduction strategies are classified and mapped to ensure that each anomaly category corresponds to at least 3 candidate strategies. Internal review and cross-validation are used to verify the accuracy of the matching between strategies and anomaly types, and the standardization of the measures' descriptions, completing the initialization of the knowledge base. Finally, the loss reduction strategies are systematically integrated, and after internal review and cross-validation, the accuracy of the data and the rationality of the strategy classification are ensured, completing the initial content construction of the knowledge base.

[0138] Dynamic maintenance and version updates: Establish a monthly spot check mechanism to verify the timeliness of standards, the completeness of cases, and the applicability of strategies, and correct errors or outdated content; adopt a quarterly version update mechanism to synchronize the latest industry standards, add loss reduction strategies corresponding to new anomalies (such as reverse line loss caused by a high proportion of new energy access), and supplement innovative measures that have been proven effective in practice (such as intelligent load control algorithms) to ensure the comprehensive coverage of line loss anomalies in the distribution area and the advanced nature of the strategies in the knowledge base.

[0139] A routine maintenance mechanism has been established to conduct regular spot checks on the knowledge base content, focusing on verifying the timeliness of standard texts, the completeness of case studies, and the applicability of expert experience. Errors or omissions discovered are promptly corrected to ensure the quality of the knowledge base's basic data. Furthermore, a version maintenance mechanism is employed, synchronizing with the latest industry standards every quarter to continuously enrich the knowledge base content and optimize its version.

[0140] By systematically building, maintaining, and updating a knowledge base for loss reduction strategies, we ensure that the knowledge base accurately matches loss reduction strategies and provides effective loss reduction measures suggestions for anomaly diagnosis in transformer areas.

[0141] The final loss reduction strategy knowledge base mainly consists of three parts: anomaly types, loss reduction suggestions, and implementation methods. The specific content is as follows:

[0142] Level 1 Anomaly: Issue with Record Relationships

[0143] Secondary Category: Marketing System File Issues; Tertiary Category: Missing Transformer Area Files. Loss Reduction Recommendation: Dynamically complete files to ensure accuracy and completeness. Implementation Method: Weekly, automatically compare historical file data with real-time transformer area operation data through the system, marking the transformer area number and its corresponding power supply area for missing files. Organize transformer area customer managers to conduct on-site verification with mobile terminals, collecting basic information such as transformer name, transformer capacity, number of customers served, and line route, and uploading it to the marketing system on-site.

[0144] Level 3 Classification: Status of the same transformer area is not synchronized. Loss Reduction Recommendation: Establish a real-time channel for automatic status synchronization. Implementation Method: Deploy a data synchronization interface based on the MQTT protocol between the marketing system and the synchronization system, and set a status change trigger mechanism (e.g., synchronization should begin within 10 seconds of a transformer area commissioning / shutdown command being issued). The system automatically checks the consistency of the status at both ends every 5 minutes, automatically traces discrepancies, and has maintenance personnel verify and correct them within 1 hour.

[0145] Level 3 classification: Missing master meter for the same area. Loss reduction suggestion: Investigate and replace the master meter, and connect to the system for data collection. Implementation method: Based on the load calculation results of the area (calculated according to Article 5.1 of DL / T447-2010 "Technical Specification for Rural Low Voltage Power"), determine the master meter specifications (e.g., a 1.5(6)A master meter with CT for a 500kVA area). The metering team shall complete the meter installation within 72 hours, and use infrared debugging tools to configure the communication parameters on site to ensure a communication success rate of ≥99% with the data collection system. Enter the master meter asset number, meter constant and other information into the marketing system, and link the area files to form a data closed loop.

[0146] Level 3 Classification: Missing Transformer Records. Loss Reduction Recommendation: Complete Transformer Information. Implementation Method: Based on the GIS system's distribution network topology map, compare each transformer in the ledger, marking the transformer number and location of missing records. Organize operation and maintenance personnel to collect technical parameters such as transformer model, manufacturing date, and short-circuit impedance on-site (according to Appendix A of GB / T10228-2015 "Technical Parameters and Requirements for Dry-Type Power Transformers"). Establish an "Electronic Transformer Record Ledger," linking it to transformer area numbers and line information to ensure consistency with marketing system data.

[0147] Level 3 Classification: The user relationships for the same electricity distribution unit (TDU) are not synchronized. Loss Reduction Recommendation: Regularly check and adjust to ensure accurate matching. Implementation Method: Before the 5th of each month, the system automatically exports the user list for the distribution unit area and the user list from the data collection system. A fuzzy matching algorithm (matching degree ≥ 95% is considered consistent) is used to identify discrepancies. Account managers, equipped with smart terminals, verify the electricity address and meter affiliation on-site. For incorrectly associated users, the "TDU Relationship Change Process" is initiated, and system adjustments are completed within 24 hours.

[0148] Level 3 Classification: Distribution transformer status not synchronized. Loss Reduction Recommendation: Optimize the interface to transmit synchronization status in real time. Implementation Method: Upgrade the firmware of the distribution transformer monitoring terminal (TTU) to enable it to actively report status changes (reporting delay ≤ 5 seconds). Deploy a status verification engine in the synchronization system to compare the data uploaded by the TTU with the system records every minute, triggering audible and visual alarms if there are discrepancies.

[0149] Level 3 Classification: Missing data in the middle station or concurrent meter data. Loss Reduction Recommendation: Complete the data and fix data collection vulnerabilities. Implementation Method: For areas with missing meter data, prioritize recovering historical data through the data collection system (following the recovery process in DL / T698.41-2010 "Part 4-1: Communication Protocol - Communication between Master Station and Energy Meter"). If recovery fails, meter readers must manually re-read the data within 24 hours, with double verification required when entering the data into the system. Technical personnel will analyze the cause of the missing data and replace meters with faulty communication modules to ensure a meter data collection success rate of ≥99.8%.

[0150] Level 3 Classification: Missing records of meter replacement during the same period. Loss Reduction Recommendation: Standardize procedures and promptly enter records. Implementation Method: Meter replacement personnel use mobile work terminals to enter information such as the old and new meter numbers, replacement time, and meter data on-site (in accordance with Article 9.2 of Q / GDW1373-2013 "Functional Specification for Electricity User Information Collection System"). An electronic work order is automatically generated upon upload. The following day, the metering specialist verifies the completeness of the records, returning any records missing key fields such as the "old meter recycling number" for re-entry. Each month, 20% of meter replacement records are randomly selected for comparison with on-site photos to ensure 100% consistency.

[0151] Secondary Classification: Distribution and Operation Connection Issues; Tertiary Classification: Marketing areas do not correspond to transformers of the same source. Loss Reduction Recommendation: Investigate and sort out to establish correct correspondences. Implementation Method: Export the marketing system's distribution area-transformer association table and the GIS system topology map, and use "distribution area power supply radius ≤ 500 meters" (expert experience measurement value) to initially screen for incorrect associations. Operation and maintenance personnel carry appropriate testing instruments to verify on-site, and determine the correct association by comparing the distribution area master table with the transformer outgoing lines.

[0152] Level 3 classification: Marketing distribution areas and transformers of the same source in a one-to-many or many-to-many manner. Loss reduction recommendation: Adjust according to "one transformer per distribution area," ensuring a unique correspondence. Implementation method: For one-to-many distribution areas, split them according to the "load balancing" principle (refer to expert experience formula: load deviation of each distribution area after splitting ≤15%), and name the new distribution areas according to the "original number + branch number" rule. For many-to-many distribution areas, re-divide the power supply range through the GIS system to ensure that each distribution area corresponds to a unique transformer.

[0153] Level 3 Classification: Inconsistencies between marketing customer details and source customer details. Loss Reduction Recommendation: Compare and verify, correct to ensure consistency. Implementation Method: Before the 10th of each month, export the "Customer Details" from the marketing system and the "Customer Details" from the source system. Use a data comparison tool (supports fuzzy matching of key fields; a matching degree ≥98% is considered consistent) to generate a discrepancy list. Account managers verify the discrepancy list on-site for each customer, recording the actual electricity address, meter number, and other information, and upload it to the system via mobile terminal. The record administrator completes the correction of discrepancies within 3 working days. After correction, the accuracy is verified using the system's automatic verification tool (a pass rate ≥99.5% is considered acceptable).

[0154] Level 3 Classification: Inconsistent status of distribution transformers corresponding to the marketing area. Loss Reduction Recommendation: Regular verification and adjustments to ensure accuracy. Implementation Method: Every Monday at midnight, the system automatically triggers a "Distribution Area-Distribution Transformer Status Comparison Task," extracting the operating status (operating / out of service) of the marketing system distribution areas and the distribution transformer status of the corresponding system, generating a list of inconsistent statuses. Maintenance personnel must go to the site within 24 hours of receiving the list and verify the actual status by checking equipment operation indicator lights and collecting terminal data. Based on the verification results, the erroneous status is corrected in the marketing system or the corresponding system. After correction, the system automatically sends a status synchronization confirmation to the relevant departments.

[0155] Level 3 Classification: Marketing system transformer area operating status is unreasonable. Loss Reduction Recommendation: On-site inspection and adjustment to restore a reasonable status. Implementation Method: The system automatically screens "transformer areas with abnormal operating status" daily (e.g., zero electricity for 7 consecutive days but status is "operating," load rate >120% and status is "normal"), generating a checklist. The transformer area manager conducts an on-site inspection within 48 hours, recording the actual operating status of the transformer area (e.g., whether it is truly out of service, whether there is overload), and taking photos as evidence. For transformer areas confirmed to have incorrect status, a status adjustment application is initiated in the marketing system, which takes effect after approval by the operations supervisor. Adjustment records are kept for at least 3 years.

[0156] Level 3 Classification: Delayed commissioning of new transformer substations in marketing areas. Loss Reduction Recommendation: Optimize processes and ensure timely data entry. Implementation Method: Establish a "Pre-commissioning Mechanism for New Transformer Substations." Seven days before commissioning, the infrastructure department will submit the substation planning map, transformer parameters, and other information to the marketing department. The marketing department will assign a dedicated person to complete the pre-entry of the data within three working days and generate a temporary substation number. On the day of commissioning, after the metering personnel install the master meter on-site, they will update the data status to "Officially Operational" via a mobile terminal. The system will automatically link the master meter information to ensure normal data collection within 24 hours of commissioning.

[0157] Level 3 Classification: Marketing areas connected to high-voltage users. Loss Reduction Recommendation: Separate high-voltage users and manage them separately. Implementation Method: The system automatically filters "marketing areas connected to high-voltage users" monthly (high-voltage users refer to users with voltage levels >10kV), extracting user numbers, power capacity, and other information. Account managers, in conjunction with meter readers, verify on-site to confirm the user's power supply method (e.g., whether it is truly powered by a low-voltage area, and whether there are wiring errors). For those confirmed to be incorrectly connected, a dedicated high-voltage user file is created in the marketing system, the association with the low-voltage area is severed, and the metering points and data collection parameters are reconfigured.

[0158] Level 3 Classification: Marketing form replacement process anomaly. Loss Mitigation Recommendations: Investigate and optimize the process, clarify responsibilities, and strengthen monitoring. Implementation Method: Establish a form replacement process optimization team to analyze existing process nodes (application-approval-form replacement-acceptance), record the average time spent at each node, and identify bottlenecks exceeding 24 hours. The system will monitor the form replacement process progress in real time, automatically sending warning messages to responsible personnel for overdue nodes, and calculating the process compliance rate monthly (target ≥98%).

[0159] Level 3 Classification: Abnormal Master Meter File Creation in Marketing Areas. Loss Reduction Recommendation: Re-file the master meter and ensure accurate data upload. Implementation Method: For areas with abnormal master meter files (e.g., missing asset number, incorrect meter constants), the metering team should go to the site within 24 hours of receiving a system alert to collect the master meter nameplate information (model, number, constants, etc.). Initiate a "Master Meter File Correction Request" in the marketing system, uploading a photo of the nameplate taken on-site as an attachment. The correction will take effect after review by the metering specialist (completed within 12 hours). After correction, test the data upload function through the data collection system to ensure three consecutive successful meter readings with accurate data (error ≤ 0.5%).

[0160] Level 3 Classification: Incorrect primary use type of metering point. Loss Reduction Recommendation: Verify and correct to ensure accurate classification. Implementation Method: The system exports a "List of Primary Use Types of Metering Points" monthly, compares it with user application materials, and marks obvious errors such as "residential user type labeled as commercial". The account manager brings the user's electricity usage list (last 3 months) for on-site verification. If the user's actual electricity usage is mainly for residential purposes, a type change application is initiated on the mobile terminal. Marketing review personnel complete the approval within 3 working days, and the system automatically updates the type information and adjusts the electricity price standard accordingly.

[0161] Level 3 Classification: Office electricity metering points have disappeared. Loss Reduction Recommendation: Investigate, restore, or reinstall metering points. Implementation Method: Monthly, the logistics department, in conjunction with the metering team, will conduct a comprehensive survey of electricity usage points in office buildings, business halls, and other office areas, marking any missing points against the same period last year's metering point list. For metering points removed due to line renovations, relocation and installation will be carried out within 72 hours (preferably at the main incoming switch), configuring a Class 0.5 smart meter. After reinstallation, data uploads will be monitored continuously for 3 days through the electricity information collection system, ensuring a daily average data collection success rate of ≥99.9%.

[0162] Level 3 Classification: Inappropriate Meter Information. Loss Reduction Recommendation: Rectify meters to ensure accurate measurement. Implementation Method: The system automatically screens the "List of Abnormal Meter Information" weekly. Metering personnel conduct on-site verification within 3 working days and initiate replacement procedures for meters with mismatched specifications. Investigate the cause and take targeted measures such as supplementary or replacement meters.

[0163] Secondary Category: Meter Issue; Tertiary Category: Sudden Power Consumption Fluctuations. Loss Reduction Recommendations: Investigate the cause and take targeted measures such as replacement or meter replacement. Implementation Method: The system automatically issues warnings to users whose monthly power consumption fluctuations exceed 300%, generating a "Power Consumption Fluctuation List." The district manager conducts on-site verification within 24 hours: If it is a meter malfunction, replace the meter within 48 hours and retroactively compensate based on the average power consumption of the previous 3 months; if it is a newly added device for the user, record the cause and archive it. Establish a "Power Consumption Fluctuation Handling Ledger" and analyze the distribution of fluctuation causes monthly (fault percentage must be ≤30%).

[0164] Level 3 Classification: Main Meter Failure. Loss Reduction Recommendations: Regular inspections, timely repair or replacement. Implementation Method: Implement a "monthly inspection + quarterly full inspection" system: During inspections, use a clamp meter to compare the meter reading with the real-time load (deviation ≤2%); during full inspections, conduct error verification (according to JJ0596-2012 "Verification Procedure for Electronic Energy Meters"). Fault Handling After Discovery: For hardware faults, replace with a spare meter within 24 hours; for software faults, debug and repair on-site. Establish a fault log and initiate a replacement plan for meter brands with an annual failure rate exceeding 5%.

[0165] Level 3 Classification: Data Acquisition Failure. Loss Mitigation Recommendations: Troubleshoot and repair the fault to ensure data acquisition. Implementation Method: The system monitors the acquisition status every 15 minutes. For meters that fail three times consecutively, a diagnostic process is initiated: first, remotely restart the terminal; if this fails, a work order is dispatched to the data acquisition maintenance team. Maintenance personnel conduct on-site troubleshooting with a signal tester: communication module failures are replaced within 6 hours, line breaks are repaired within 24 hours, and repeaters are installed in areas with weak signals. Daily statistics on data acquisition success rate are maintained to ensure a district-level success rate of ≥99.5%.

[0166] Level 3 Classification: Time Anomaly. Loss Reduction Recommendations: Regular calibration, timely rectification and replacement of meters. Implementation Method: The system automatically synchronizes time monthly, marking meters with a clock deviation exceeding 30 seconds as abnormal. Calibration commands are remotely issued to smart meters; mechanical meters are calibrated on-site by meter readers carrying GPS time calibrators. Meters that still exceed the tolerance after two consecutive calibrations are replaced with new meters equipped with BeiDou time synchronization functionality within 72 hours.

[0167] Level 3 Classification: Meter Freeze Anomaly. Loss Reduction Recommendation: Investigate and adjust to ensure accurate freezing. Implementation Method: Automatically trigger freeze command at 00:00 daily. The system checks freeze results within 1 hour and marks the cause of failures (e.g., communication interruption, meter incompatibility). Technical personnel handle the issues accordingly: Repair communication problems according to the data acquisition failure process; upgrade firmware or replace unsupported meters. Calculate freeze success rate weekly, with a target of ≥99.8%.

[0168] Level 3 Classification: Significant Measurement Error. Loss Reduction Recommendations: Inspect meters; replace unqualified meters promptly; strengthen regular calibration. Implementation Method: Implement a "semi-annual random inspection + annual full inspection" approach: random inspection covers 20% of the distribution area; full inspection uses a 0.05-grade standard device. Replace meters with errors exceeding ±2% immediately; send out meters with errors exceeding tolerance to the laboratory for testing and trace the cause. Establish an "Error Exceedance Log"; initiate a recall procedure for batches with an error exceedance rate exceeding 3%.

[0169] Level 3 Classification: Untimely Meter Reading. Loss Reduction Recommendation: Optimize the plan and adopt automatic meter reading to improve efficiency. Implementation Method: Reconstruct the meter reading sequence: Divide the meter reading time into 5 periods (6:00-22:00) based on the geographical distribution of the transformer area, avoiding peak electricity consumption times. For the 10% of users whose automatic meter reading fails, configure mobile meter reading terminals to ensure supplementary reading is completed before 18:00. Calculate the meter reading completion rate monthly, requiring 100% completion before 19:00.

[0170] Level 3 Classification: User omission issues. Loss reduction recommendations: Conduct a comprehensive investigation, retroactively register users, and include them in management. Implementation method: Conduct two district-wide surveys annually in spring and autumn, using a "GIS map label + on-site visits" approach to compare the number of system users with the actual number of households. For users with omissions, complete the entire process of account opening, meter installation, and file creation within 3 working days, and retroactively collect electricity consumption based on actual usage duration. Establish an "omission rate assessment mechanism," requiring an omission rate of ≤1% for a single district.

[0171] Level 3 Classification: Abnormal Correlation. Loss Reduction Recommendation: Analyze and investigate to eliminate the abnormality and restore a reasonable correlation. Implementation Method: The system calculates the correlation coefficients of 12 key parameters, including "line loss rate - load" and "voltage - current," daily, marking those deviating from the historical range (±30°) as abnormal. Analysts complete the scrap location within 2 working days: correcting data collection points for data errors, dispatching repair orders for equipment failures, and initial user investigation. Follow up for 3 days after anomaly handling to ensure the correlation coefficient returns to the normal range.

[0172] Secondary Classification: Illegal Electricity Use; Tertiary Classification: Electricity Theft. Loss Reduction Recommendations: Strengthen inspections, repair and replace meters to reduce electricity loss. Implementation Method: The system automatically marks users with special circumstances such as "zero load at night + line loss" as suspected users, generating a weekly inspection checklist. Inspectors use electricity inspection equipment to check: those confirmed to be stealing electricity are immediately shut off and three times the amount of electricity is recovered; faulty meters are replaced within 24 hours. Monthly statistics on the detection rate are compiled, requiring a ≥95% completion rate for investigating suspected users.

[0173] Level 3 Classification: Overcapacity Electricity Consumption. Loss Reduction Recommendations: Urge rectification, replace equipment, and implement power rationing if necessary. Implementation Method: The system monitors user current in real time; an alert is issued if the current exceeds 1.2 times the rated value for 15 consecutive minutes. The area manager conducts on-site verification within 48 hours: If overcapacity is confirmed, a notification letter is issued (within 7 days), guiding the replacement of the meter with a higher capacity one and the incoming cable. If rectification is not completed within the deadline, power rationing measures will be implemented (current limited to the rated value).

[0174] Secondary Classification: Distribution Area Operation Issues; Tertiary Classification: Low Voltage for Users. Loss Reduction Recommendations: Grid Upgrade to Improve Voltage Quality. Implementation Method: The system will statistically analyze user-end voltage daily; voltages below 198V (220V system) for more than 4 hours will be flagged. The operation and maintenance team will conduct on-site testing within 2 hours: For excessively long lines, install voltage regulators; for excessively thin wires, replace them with newer wires; for three-phase imbalance, adjust the distribution system. Follow up for one month after the upgrade to ensure a voltage qualification rate ≥97%.

[0175] Level 3 Classification: Three-phase imbalance. Loss Reduction Recommendation: Monitor and adjust the load to reduce imbalance. Implementation Method: The system calculates the three-phase imbalance hourly, issuing a warning if it exceeds 15%. Maintenance personnel will make on-site adjustments within 24 hours: For residential users, phase switching will be performed (load deviation per phase ≤ 10%); for industrial and commercial users, balancing devices will be installed. Follow up for 3 days after adjustment to ensure imbalance ≤ 10%.

[0176] Level 3 Classification: Low Power Factor. Loss Reduction Recommendations: Urge the installation of compensation devices and implement delay measures. Implementation Method: Calculate the power factor of users monthly; those below 0.9 will be included in the rectification list. The district manager will visit users to explain the reward and penalty policy (power consumption will increase by 20%) and guide them in installing self-healing capacitors (capacity configured at 30% of the load). A retest will be conducted within one month of installation; users who meet the standards will have their normal billing restored.

[0177] Level 3 Classification: Overloaded transformer area. Loss Reduction Recommendations: Adjust load or upgrade capacity to avoid increased line losses. Implementation Method: The system monitors the transformer load rate in real time; an alert is issued if the load exceeds 80% for 2 consecutive hours, and an immediate alarm is triggered if it exceeds 100%. For lightly loaded transformer areas where load can be transferred, line upgrades should be completed within 24 hours; for areas requiring capacity upgrades, an upgrade plan (replacing with a larger capacity transformer) should be developed within 7 days. After the upgrade, ensure the load rate is controlled within the 30%-70% range.

[0178] Level 3 Classification: Lightly Loaded Transformer Area. Loss Reduction Recommendation: Assess merging or replacing with smaller capacity transformers to improve load factor. Implementation Method: The system will statistically analyze the transformer load factor monthly; transformers with a load factor below 30% for three consecutive months will be included in the lightly loaded list. On-site assessment by the operation and maintenance department: If adjacent transformer areas can be merged, the line modification will be completed within 30 days; if a smaller capacity transformer is required, it will be replaced with a matching model within 15 days. Follow up for 3 months after the modification to ensure that the load factor increases to above 40%.

[0179] Level 3 Classification: Aging Distribution Area. Loss Reduction Recommendation: Develop a plan to replace and upgrade line equipment. Implementation Method: Conduct annual assessments of distribution areas that have been in operation for over 15 years, testing line insulation resistance (≤0.5MΩ is considered aging) and equipment temperature rise (over 60K is considered abnormal). Included in the annual upgrade plan: Replace conductors with cross-linked cables and switches with intelligent circuit breakers; the upgrade cycle should not exceed 6 months. Post-upgrade acceptance must meet the requirement of a ≥3% reduction in line loss rate.

[0180] Level 3 Classification: Aging power grid structure. Loss Reduction Recommendations: Upgrade and optimize, shorten distances, replace equipment. Implementation Method: Analyze the power grid using a GIS system: Include circuitous power supply lines (length exceeding 1.5 times the straight-line distance) and conductor cross-sections ≤16mm² in the upgrade plan. Develop a "short-circuit loop, high-capacity" upgrade plan: shorten the straight-line power supply distance, upgrade conductors to 50mm² or higher, and complete construction within 3 months. After the upgrade, test the line loss rate, requiring a reduction of ≥4% compared to before the upgrade.

[0181] Secondary Category: Equipment Failure Issues; Tertiary Category: Equipment Cable Failures. Loss Reduction Recommendations: Strengthen inspection and maintenance, and promptly repair or replace faulty components. Implementation Methods: Perform monthly inspections + infrared thermography to check cable joints (temperature rise exceeding 40K is considered abnormal) and equipment noise (obvious abnormal noise is considered a fault). Fault Classification and Handling: Overheated joints should be re-crimped within 2 hours; damaged insulation should be replaced within 24 hours; severely aged joints should be completely replaced within 7 days. Establish a fault log, requiring a 100% timely fault repair rate.

[0182] Level 3 Classification: Long power supply distance. Loss Reduction Recommendations: Shorten the distance or use higher voltage levels to reduce losses. Implementation Methods: For distribution areas with a power supply radius exceeding 0.5 km (as defined in Article 3.2 of the "Rural Low-Voltage Power Technology Regulations"), calculate the optimal path using a GIS system and formulate a "short-distance upgrade plan": add a distribution transformer at the load center to reduce the original power supply radius from 1.5 km to within 0.5 km. For remote areas where adding a transformer is not feasible, implement a "voltage upgrade" to upgrade the 220V low-voltage power supply to 10kV high-voltage power supply, and install a miniaturized prefabricated substation (capacity configured at 1.2 times the maximum load). After the upgrade, continuously monitor the line loss rate for 3 months to ensure a reduction of ≥5% compared to before the upgrade, and improve the voltage qualification rate to over 98%.

[0183] The purpose, technical solution and beneficial effects of the present invention have been further described in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing abnormal line loss in transformer substations based on a large model, characterized in that, Includes the following steps: Multi-dimensional data is collected to support the diagnosis of abnormal line loss in transformer substations. The multi-dimensional data is then cleaned, correlated, fused, and standardized to obtain a comprehensive dataset covering transformer substation attributes, operating status, energy access, and external environment. The multi-dimensional data includes: basic attribute data for characterizing the inherent features of the distribution area, power operation status data reflecting the real-time and historical operation patterns of the power system, energy access data reflecting the impact of new energy access on the distribution area, and external environmental impact data that may interfere with or be associated with line loss fluctuations. A multi-layered diagnostic system based on a rule-based model, a random forest model, and a large-scale model is constructed. The rule-based model identifies simple, routine anomalies in transformer substations based on the basic attribute data and power operation status data in the comprehensive dataset, according to preset anomaly diagnosis rules. The random forest model locates complex and novel anomalies by mining the coupling relationship between energy access data, external environmental impact data, and line loss fluctuations in the comprehensive dataset. The large-scale model fuses and verifies the diagnostic results of the rule-based model and the random forest model, outputting the final anomaly diagnosis result for the transformer substation. Based on the diagnostic results, the loss reduction strategy knowledge base is invoked, and the large model intelligently recommends differentiated loss reduction strategies that are adapted to the actual data characteristics of the transformer area.

2. The method for diagnosing abnormal line loss in transformer substations based on a large model according to claim 1, characterized in that, The basic attribute data includes fixed parameters used to characterize the physical properties of the transformer area; The power operation status data includes real-time measurement data and historical statistical data. The real-time measurement data includes voltage, current, power factor, load rate and three-phase imbalance. The historical statistical data includes daily line loss rate, hourly load curve and voltage qualification rate. The external environmental impact data includes meteorological data and time characteristic data. The meteorological data includes weather type, temperature and humidity. The time characteristic data includes holiday identification and peak and valley electricity consumption periods. The energy access data includes the installed capacity, output power curve, grid-connected electricity, and access location information of distributed power sources, including photovoltaic, wind power, and energy storage equipment.

3. The method for diagnosing abnormal line loss in transformer substations based on a large model according to claim 1, characterized in that, The rule model, based on the basic attribute data and power operation status data in the comprehensive dataset, identifies simple and routine anomalies in the transformer substation according to preset anomaly diagnosis rules. The specific process includes the following sub-steps: An anomaly classification system was derived by analyzing the basic attribute data and power operation status data in the comprehensive dataset. An initial rule base is constructed based on the aforementioned anomaly classification system. The initial rule base stores the judgment logic of various anomalies in an IF-THEN structure and is iteratively optimized through a dynamic update mechanism. With the dynamically updated rule base as the core, power industry standards and expert experience corresponding to the anomaly classification system are integrated, and a knowledge base with dynamic indexes is constructed based on the knowledge graph. The knowledge base is a semantic association network based on basic attribute data, operation status data, rules, standards and experience. The basic attribute data and power operation status data of the transformer substation to be diagnosed are extracted from the comprehensive dataset. The matching rule base judgment conditions are located through dynamic indexing. The feature thresholds are verified in combination with the industry standards in the knowledge base. After the basic attribute data and power operation status data of the transformer substation to be diagnosed are verified by the large model to check the consistency of matching with the rules, the diagnostic results containing the anomaly category, feature basis and rule version are output.

4. The method for diagnosing abnormal line loss in transformer substations based on a large model according to claim 3, characterized in that, The dynamic update mechanism is as follows: when the number of identified cases of the same new type of anomaly reaches 5, the basic attribute data features and power operation status data features associated with this type of new anomaly are extracted, and the Apriori algorithm is used to mine the association rules between features and anomalies to generate candidate rules. When the accuracy of the candidate rule is ≥90%, the candidate rule is included in the rule base, and the corresponding level of the anomaly classification system is updated at the same time.

5. The method for diagnosing abnormal line loss in transformer substations based on a large model according to claim 1, characterized in that, The random forest model locates complex and novel anomalies by mining the coupling relationship between energy access data, external environmental impact data, and line loss fluctuations in the comprehensive dataset. The specific process is as follows: The dataset extracts time-series features to characterize line loss fluctuation patterns, correlation features to represent the relationship between energy access and line loss, static features to reflect the inherent attributes of transformer substations, and dynamic features to reflect real-time operating status, forming a multi-dimensional feature set covering time evolution, correlation coupling, equipment attributes, and operating status. From the aforementioned time-series features, correlation features, static features, and dynamic features, principal component analysis is used to screen key features that contribute up to 85% to the cumulative contribution rate of anomaly diagnosis. At the same time, an attention mechanism is introduced to assign dynamic weights to the time-series features during peak electricity consumption periods, thereby achieving dimensionality reduction optimization of key features and strengthening of features during high-value periods. The key features are fused to form a feature vector. A random forest model is constructed using a hybrid splitting criterion that weights the contribution of the Gini coefficient and the time series features. The model parameters are optimized by maximizing the F1 score of the large model, so that the random forest model has the ability to discover the correlation between multi-dimensional features and line loss anomalies. Using the trained random forest model, complex and novel anomalies are located by identifying the coupling patterns of abrupt changes in energy access parameters, abnormal external environmental factors, and fluctuations in line loss. The diagnostic results include anomaly type, feature contribution, and associated influencing factors. The performance of the random forest model is verified by accuracy, recall, and F1 score.

6. The method for diagnosing abnormal line loss in transformer substations based on a large model according to claim 5, characterized in that, The attention mechanism is introduced to assign dynamic weights to the temporal characteristics of peak electricity consumption periods. The specific method is as follows: A dynamic weighting strategy based on time windows is constructed, assigning an initial weight of 1.5 times to the temporal characteristics of peak electricity consumption periods. Combining this with sliding window technology, and using the fluctuation trend of line loss data during historical peak periods as a benchmark, the weighting coefficients are adaptively adjusted by calculating the feature similarity between the current peak period and historical peak periods. When the current peak data fluctuation trend has a similarity of ≥80% to the abnormal pattern of historical typical peaks, the weighting coefficient will be increased to 1.5-2.0 times. When the similarity is in the range of 50%-80%, maintain the initial weight of 1.5 times; When the similarity is less than 50%, the weighting coefficient is reduced to 1.2-1.5 times.

7. The method for abnormal line loss diagnosis of transformer substations based on a large model according to claim 5, characterized in that, The key features are fused to form a feature vector. A random forest model is constructed using a hybrid splitting criterion weighted by the Gini coefficient and the contribution of time-series features. The model parameters are optimized by maximizing the F1 score of the large model, enabling the random forest model to discover the correlation between multi-dimensional features and line loss anomalies. The specific process is as follows: The static features and enhanced time-series features in the key features are fused with energy access data, meteorological data, external environmental impact data, and holiday data to form a multi-dimensional feature vector; A hybrid evaluation index S, weighted by the Gini coefficient and the contribution of time series features, and a hybrid splitting criterion are constructed, S = α × G + (1 - α) × C, where G is the Gini coefficient characterizing the spatial dispersion of line loss data, C is the contribution of time series features characterizing the dynamic change trend of time series, and α is an adaptive weighting coefficient. The hybrid splitting criterion is: when splitting decision tree nodes, the influence of spatial distribution features and temporal evolution law on classification results is dynamically balanced by the hybrid evaluation index. Multiple sample subsets are obtained by randomly sampling with replacement of the multidimensional feature vector, and each sample subset is used to train a decision tree. During the node splitting process of each decision tree, some key features are randomly selected from the sample subsets and the optimal splitting features are determined by the above-mentioned hybrid splitting criteria to reduce the correlation between trees and improve the generalization ability of the model. The sample subset is divided into a training set and a test set. A random forest model is trained using the training set. The parameters of the random forest model are adaptively adjusted based on the reasoning ability of the large model and with maximizing the F1 score as the optimization objective. The trained random forest model is evaluated using the test set. The performance of the random forest model is judged by calculating the accuracy, recall, and F1 score evaluation metrics.

8. The method for diagnosing abnormal line loss in transformer substations based on a large model according to claim 1, characterized in that, The specific process of fusing and validating the diagnostic results of the rule model and the random forest model to output the final anomaly diagnosis result for the transformer area is as follows: Independent diagnostic results are obtained for the rule model and the random forest model respectively. The independent diagnostic results of the rule model are the anomaly judgment results and corresponding rule basis output based on manually defined business rules. The independent diagnostic results of the random forest model are the anomaly judgment results, feature contribution and anomaly category probability output based on multi-dimensional feature combination. Combining the aforementioned rule model and random forest, a phased fusion strategy was formulated, but the phased fusion strategy was not implemented as follows: The first preset time period is taken as the early stage of application. In the early stage of application, a result union fusion strategy is adopted to integrate all abnormal station data determined by the rule model and the random forest model as abnormal samples. The abnormal samples are used to optimize the rule base and the random forest model. The second preset time period is taken as the application mid-term. During the application mid-term, based on the abnormal samples, the anomaly judgment results identified by the random forest model are analyzed, and the anomaly judgment results that exceed the set threshold are converted into rules to supplement the rule base. The third preset time period is designated as the later stage of the application. In the later stage of the application, the results are fused according to the differences in the scene of the transformer area. For transformer areas associated with core indicators, the result intersection fusion strategy is adopted, and only the results that are judged as abnormal by both the rule model and the random forest model are retained and output to ensure diagnostic accuracy. For transformer areas in edge scenes, the result union fusion strategy is adopted to avoid missing potential anomalies, and the results that are judged as abnormal by either the rule model or the random forest model are retained and output. The output results are validated, and the results that pass the validation are used as the final anomaly diagnosis results for the transformer area.

9. The method for diagnosing abnormal line loss in transformer substations based on a large model according to claim 1, characterized in that, Based on the diagnostic results, the specific process of calling the loss reduction strategy knowledge base and having the large model intelligently recommend differentiated loss reduction strategies adapted to the actual data characteristics of the transformer area is as follows: The large model analyzes the final anomaly diagnosis results of the transformer area and extracts key information, including: anomaly type, anomaly characteristic parameters, basic attribute data of the transformer area, power operation status data and energy access status data; Based on the parsed anomaly type, a loss reduction strategy knowledge base constructed using a classification method that matches the anomaly classification system is invoked to retrieve candidate loss reduction strategies that are directly associated with the current anomaly type. Based on the actual data characteristics of the transformer area, the candidate loss reduction strategies are evaluated and screened for their suitability. The optimal loss reduction strategy is output from the selected candidate loss reduction strategies based on the urgency and stability of the effect.

10. A large-model-based system for diagnosing abnormal line losses in transformer substations, used to execute the large-model-based method for diagnosing abnormal line losses in transformer substations as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-dimensional data to support the diagnosis of abnormal line loss in the transformer area. The multi-dimensional data is cleaned, correlated and fused, and standardized to obtain a comprehensive dataset covering transformer area attributes, operating status, energy access, and external environment. The multi-dimensional data includes: basic attribute data for characterizing the inherent features of the distribution area, power operation status data reflecting the real-time and historical operation patterns of the power system, energy access data reflecting the impact of new energy access on the distribution area, and external environmental impact data that may interfere with or be associated with line loss fluctuations. The anomaly diagnosis module is used to construct a multi-layered diagnostic system based on a rule-based model, a random forest model, and a large-scale model. The rule-based model identifies simple, routine anomalies in distribution transformer areas based on the basic attribute data and power operation status data in the comprehensive dataset, according to preset anomaly diagnosis rules. The random forest model locates complex and novel anomalies by mining the coupling relationship between energy access data, external environmental impact data, and line loss fluctuations in the comprehensive dataset. The large-scale model fuses and verifies the diagnostic results of the rule-based model and the random forest model, outputting the final anomaly diagnosis result for the distribution transformer area. The loss reduction strategy recommendation module is used to call the loss reduction strategy knowledge base based on the diagnostic results, and the large model intelligently recommends differentiated loss reduction strategies that are adapted to the actual data characteristics of the transformer area.

Images