A power distribution transformer health state evaluation, life prediction and governance suggestion generation method and system based on multi-source time series data

Abstract

This invention provides a method and system for assessing the health status, predicting the lifespan, and generating remediation recommendations for distribution transformers based on multi-source time-series data. The system includes: fusing data from ledgers, operations, power outages, faults, defects, maintenance, and grid structures using the transformer ID as the primary key, mapping this data to 15-minute, daily, and monthly scales to form a time-series sample library; cleaning the data through power completion, voltage correction, and power outage verification; extracting eight-dimensional features including load pressure, low voltage, power outage reliability, faults and defects, maintenance, aging, capacity matching, and power supply radius, and accumulating a health index by combining severity, duration, and attenuation factor; constructing a lifespan model with censored samples to output fault probability and remaining lifespan; integrating rules such as heavy overload, low voltage, and frequent power outages with data-driven adjustments to the health index and fault probability; generating remediation recommendations based on risk contribution and updating them through a feedback loop. This invention improves the accuracy and business adaptability of health assessment and lifespan prediction.

Description

Technical Field

[0001] This invention relates to the field of power system automation and distribution network equipment condition assessment technology, specifically a method and system for assessing the health status, predicting the lifespan, and generating management suggestions for distribution transformers based on multi-source time-series data. Background Technology

[0002] Distribution transformers are crucial equipment in medium- and low-voltage distribution networks, connecting the power supply side and the user side. Their operating status directly affects the power supply capacity, voltage quality, power supply reliability, and user safety. With the continuous expansion of distribution networks and the rapid growth of distributed power sources, electric vehicles, and residential and commercial loads, distribution transformers have long faced problems such as large load fluctuations, significant peak-to-valley differences, localized overloads, low voltage, frequent power outages, equipment aging, and untimely maintenance. Failure to promptly identify trends in transformer health deterioration can easily lead to decreased voltage quality, increased equipment failure rates, and even operational risks such as power outages, burnouts, or tripping.

[0003] Existing methods for assessing the condition of distribution transformers often rely on a single data source or manual rule-based judgment, such as simple assessments based solely on maximum load rate, number of outages, fault records, or years of operation. While these methods can identify some obvious problems, they fail to comprehensively reflect the true health status of distribution transformers under the combined influence of multiple factors, including load, voltage, outages, faults, defects, maintenance, capacity configuration, and power supply radius. Furthermore, transformer ledgers, operating curves, outage records, fault and defect records, and maintenance records are typically scattered across different business systems, with inconsistent data definitions, sampling frequencies, and time granularities, making it difficult to form a unified time-series sample library. This results in a lack of continuity and comprehensiveness in the health assessment model. In addition, raw operating data commonly contains issues such as missing apparent power, abnormal voltage acquisition, false outage reports, null values, and extreme outliers. Directly using this data for condition assessment can easily lead to misjudgments of heavy overload, low voltage, or frequent outages, affecting the accuracy of the assessment results. Existing methods do not adequately consider data cleaning, anomaly correction, and verification of the authenticity of outage events, making it difficult to distinguish between actual operational anomalies and acquired anomalies.

[0004] In terms of feature modeling, traditional methods typically focus on current state assessments, lacking a comprehensive characterization of historical time-series trends, risk duration, event severity, and the impact of time decay, thus failing to effectively identify the accumulation process of fault risks. For transformer remaining life prediction, existing methods also rarely consider censored samples of equipment that has not experienced faults, has been decommissioned, or been replaced, leading to insufficient reliability of life prediction results. Furthermore, existing data-driven models are often independent of distribution network business rules, and model predictions may be inconsistent with business discrimination rules for heavy overload, low voltage, frequent power outages, aging transformers, and high-energy-consuming transformers, reducing the interpretability and applicability of model results. On the other hand, existing systems mostly remain at the problem identification level, lacking the ability to automatically generate governance suggestions based on risk sources, such as transformer capacity expansion, new transformer construction, transformer renovation, low-voltage switching, low-voltage renovation, and maintenance plan optimization. They also lack post-governance feedback loops, making it impossible to continuously adjust model parameters based on the effects of renovations or maintenance.

[0005] Therefore, there is an urgent need for a method for assessing the health status, predicting the lifespan, and generating governance suggestions for distribution transformers based on multi-source time-series data. This method should be able to integrate data from distribution transformer ledgers, operation, voltage, current, power outages, faults, defects, maintenance, and grid structure data to complete data cleaning, feature extraction, health assessment, status classification, lifespan prediction, rule constraint correction, governance suggestion generation, and closed-loop updating. This would improve the accuracy and adaptability of distribution transformer risk identification, fault prediction, and governance decision-making. Summary of the Invention

[0006] This invention aims to address the problems existing in current distribution transformer condition assessment methods, such as single data sources, insufficient utilization of time-series features, impact of abnormal data on assessment results, separation of rule-based discrimination and data-driven models, and inability to simultaneously output health status and remaining life prediction results. It proposes a method for health status assessment, life prediction, and governance suggestion generation for distribution transformers based on multi-source time-series data. This method achieves multi-source data fusion, abnormal data cleaning, indicator feature extraction, health status assessment, life prediction, rule constraint correction, governance suggestion generation, and closed-loop updating.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0008] A method for assessing the health status, predicting the lifespan, and generating remediation recommendations for distribution transformers based on multi-source time-series data includes the following steps:

[0009] S1: Establish a multi-source data fusion model for distribution transformers, using the unique identifier of the distribution transformer as the primary key, and associate the distribution transformer ledger data, rated capacity, commissioning date, equipment model, power supply radius, number of users connected, load curve, voltage curve, current curve, power outage record, fault record, defect record, maintenance record and grid topology data, and uniformly map data with different sampling frequencies to 15-minute, daily and monthly multi-scale time windows to form a multi-source time series sample library for distribution transformers;

[0010] S2: Perform data cleaning and quality control on the raw data in the multi-source time-series sample library of the distribution transformer, including: for data points with missing apparent power, calculate the apparent power based on active power and reactive power to complete them; mark data points with null values ​​or exceeding the reasonable range of rated capacity as outliers; correct or remove abnormal voltage data by combining data from adjacent times, adjacent users and adjacent transformer areas; verify power outage events based on voltage, current, duration, number of times per day and the proportion of users affected, and remove false power outage events caused by non-real power outages or abnormal data collection, to obtain the cleaned multi-source time-series sample.

[0011] S3: Extract the indicator features of the distribution transformer based on the cleaned multi-source time-series samples, including: extracting the maximum load rate, number of heavy loads, number of overloads, duration of heavy loads, duration of overloads, and load rate change trend as the underlying features of load pressure based on the load rate sequence; extracting the number of low-voltage events, cumulative duration of low voltage, longest duration of continuous low voltage, degree of voltage deviation, voltage fluctuation rate, and low voltage trend as the underlying features of low voltage based on voltage time-series data; extracting the number of power outages, cumulative duration of power outages, average duration of power outages, frequent power outage indicators, proportion of users affected by power outages, and number of power outages in the past 60 days as the underlying features of power outage reliability based on power outage event data; extracting fault and defect features and maintenance features based on fault, defect, and maintenance data; and extracting equipment aging features, capacity matching features, and power supply radius features based on ledger and grid data to form a multi-dimensional underlying feature vector.

[0012] S4: Based on the multi-dimensional underlying feature vector, the characteristics of load pressure, low voltage, power outage reliability, fault defects, maintenance, equipment aging, capacity matching, and power supply radius are normalized to form a health evaluation feature vector; and using the underlying features of load pressure, low voltage, and power outage reliability, combined with the event severity, duration influence factor, and time decay factor, the load pressure risk, low voltage risk, and power outage reliability risk are calculated respectively, and then a comprehensive risk accumulation model is constructed to calculate the comprehensive health risk and health index of the distribution transformer;

[0013] S5: Based on the health index, rule indicator triggering conditions and future fault probability, the health status of the distribution transformer is divided into normal, attention, warning, severe and emergency levels to obtain the health status classification result;

[0014] S6: Using the historical time series data of the multidimensional low-level feature vector as input, and taking the time of failure, decommissioning, replacement or major overhaul of the distribution transformer as the end of its lifespan, construct a remaining life prediction model with censored sample processing capability, and output the failure probability and predicted remaining lifespan for multiple future time windows.

[0015] S7: Establish a fusion mechanism between rule constraints and data-driven models. Use the rule indicators of distribution transformer overload, low voltage, frequent power outages, old distribution transformers and high energy consumption distribution transformers as model prior constraints or strong feature inputs. Penalize and correct the health index and risk gain correction the future failure probability to obtain the corrected health index and the final failure probability.

[0016] S8: Outputs the current health index, health level, main sources of risk, probability of future failure, predicted remaining life, risk contribution ranking and similar historical cases for each distribution transformer, and automatically generates governance suggestions based on the trigger indicator type and risk level.

[0017] S9: Establish a closed-loop update mechanism. When a distribution transformer experiences a new operational anomaly, power outage, low voltage, fault, defect, or maintenance event, automatically update the multi-source time series samples, characteristic indicators, health index, health level, fault probability, and remaining life prediction results. Use the changes before and after the treatment as feedback samples for continuous model training and parameter correction.

[0018] Furthermore, in step S1, the multi-source data set of the distribution transformer is represented as follows:

[0019] ;

[0020] in, This represents the data in the distribution transformer ledger. Represents the load or power curve. Represents the voltage curve. Represents the current curve. This indicates a power outage record. Indicates fault record, Indicates a defect record. This indicates maintenance records. Represents network topology data;

[0021] Data from different sampling frequencies are uniformly mapped as follows:

[0022] ;

[0023] in, Indicates the first Taiwan distribution transformer at time Time window Multi-source time series samples below, It includes 15-minute, day, and month timescales.

[0024] Furthermore, in step S2, when apparent power is missing, it is calculated according to the following formula:

[0025] ;

[0026] in, Apparent power, Active power Reactive power;

[0027] When satisfied ;

[0028] or

[0029] When this data point is an anomaly, it is marked as a power anomaly. This refers to the rated capacity of the distribution transformer.

[0030] When verifying a power outage event, a real power outage event must simultaneously meet the following conditions:

[0031] (1) Voltage and current are null values ​​at the same time;

[0032] (2) The power outage lasted between 15 minutes and 15 hours;

[0033] (3) The number of power outages in the same day shall not exceed 2;

[0034] (4) The number of users affected by the power outage shall not be less than 50% of the total number of users in the transformer area;

[0035] If a power outage event does not simultaneously meet the above conditions, it will be judged as a pseudo-power outage event and removed.

[0036] Furthermore, in step S4, the load pressure risk is calculated using an event accumulation method:

[0037] ;

[0038] in, Indicates the first The severity of each load anomaly event, Indicates the duration of influence factor. Indicates the time decay factor;

[0039] Low voltage risk The following formula is used for calculation:

[0040] ;

[0041] Among them, the severity of low voltage events for:

[0042] ;

[0043] in, This is the lower limit for acceptable low voltage. For the first The lowest voltage in a low-voltage event. This is the severe lower limit of low voltage.

[0044] Power outage reliability risk The following formula is used for calculation:

[0045] ;

[0046] in, The severity of the power outage;

[0047] ;

[0048] To determine the percentage of users affected by the power outage, ;

[0049] For frequent power outages, the amplification factor is... ;

[0050] in, For the first The number of users affected by this power outage This represents the total number of users in the Taiwan region. This refers to the cumulative number of power outages in the 60 days following the incident.

[0051] Furthermore, in step S4, the health evaluation feature vector for:

[0052] ;

[0053] in, This is the evaluation value for the equipment load level. This is the power supply voltage quality evaluation value. This is the evaluation value for the impact of the power outage event. This is the equipment failure risk assessment value. This is an evaluation value for the equipment's operation, maintenance, and repair status. This is the equipment aging evaluation value. To match the evaluation value to the equipment capacity, This is the risk assessment value for the power supply radius;

[0054] Comprehensive health risks for:

[0055] ;

[0056] Health index is :

[0057] ;

[0058] in, Risks are triggered by the rules.

[0059] Furthermore, in step S6, the remaining life prediction model uses the failure, decommissioning, replacement, or major overhaul of the distribution transformer as the life end event and sets an event identifier. ;

[0060] When a distribution transformer experiences a life-end event during the observation period:

[0061] ;

[0062] When the distribution transformer does not experience a life-end event at the end of the observation period:

[0063] ;

[0064] This sample was then used as a right-censored sample in model training.

[0065] Future No. Probability of life end event occurring within a time window for:

[0066] ;

[0067] in, For the first Taiwan distribution transformer at time Under the conditions of the state, in the future The conditional probability of a life-end event occurring within a time window. To be used to map the model output to to Activation function for the interval, For the first The bias parameters corresponding to each time window The total number of health assessment features. For the health assessment characteristics, For the first The health assessment feature in the first The influence coefficient under each time window For the first Taiwan power distribution transformer serial number, For the k-th time at time \(t) A health assessment characteristic value, For the rule to trigger risk in the first The influence coefficient under each time window For the first Taiwan distribution transformer at time The rules trigger risk, Numbering of distribution transformers For the current observation time, This refers to the sequence number of the future time window;

[0068] Survival probability for:

[0069] ;

[0070] in, For the first Taiwan distribution transformer at time Under the condition of the future, The probability of survival if no life-end event occurs within a time window; For the first Taiwan distribution transformer at time Under the conditions of the state, in the future The conditional probability of a life-end event occurring within a time window; For this distribution transformer in the future The conditional probability that no life-end event will occur within a time window; Indicates the future The first to the second Multiply the probabilities of not experiencing the life end event within each time window; Numbering of distribution transformers For the current observation time, This represents the time window number during the multiplication process. This refers to the number of time windows to predict or the sequence number of future time windows.

[0071] future The probability of failure within the time window is:

[0072] ;

[0073] Predicting remaining lifespan for:

[0074] ;

[0075] in, For the first Taiwan distribution transformer at time Predicted remaining lifespan; For the first Taiwan distribution transformer at time Under the condition of the future, The probability of survival if no life-end event occurs within a time window; The total number of future time windows covered by the remaining lifespan prediction; This refers to the sequence number of the future time window; The time length corresponding to a single forecast time window; For the first The expected survival time corresponding to the survival probability of each time window; Indicates the future The first to the second The expected survival time of each time window is accumulated; Numbering of distribution transformers This is the current observation time.

[0076] Furthermore, in step S7, the rule triggers risk. for:

[0077] ;

[0078] in, Indicates the first The severity of each rule indicator, Indicates the first Each rule indicator at time Whether to trigger, Indicates the rule trigger time. Indicates the time decay coefficient;

[0079] The fault probability after rule correction is:

[0080] ;

[0081] in, The base failure probability output by the data-driven model. This is the rule-based risk gain coefficient.

[0082] Furthermore, in step S8, the risk contribution rate... Calculate using the following formula:

[0083] ;

[0084] in, For the first A health assessment characteristic at time Contribution to the overall health risk of distribution transformers; For the first Risk weights corresponding to each health assessment feature; For the first A health assessment characteristic at time Normalized eigenvalues; For the first The weighted risk contribution value of each health assessment feature to the overall health risk; For the distribution transformer at time Comprehensive health risks; For health assessment feature serial numbers, This is the current evaluation moment. The larger the value, the more significant the number. The greater the contribution of each health assessment characteristic to the current overall health risk;

[0085] The system identifies primary risk sources based on their risk contribution and generates mitigation recommendations based on these sources and their health levels. Specifically, when the primary risk is transformer overload, heavy load, or heavy load warning, recommendations are generated for transformer capacity expansion, new transformer construction, or low-voltage switching modification. When the primary risk is low voltage in the distribution area, recommendations are generated for transformer renovation, new transformer construction, or low-voltage renovation. When the primary risk is frequent power outages, recommendations are generated for transformer renovation, transformer and low-voltage renovation, or low-voltage renovation. When the primary risk is aging transformers or high-energy-consuming transformers, recommendations are generated for transformer renovation or transformer and low-voltage renovation.

[0086] A system for assessing the health status and predicting the lifespan of distribution transformers based on multi-source time-series data, used to execute the method described above, includes:

[0087] The multi-source data fusion module is used to associate ledger data, operation data, power outage data, fault and defect data, maintenance data and grid data with the unique identifier of the distribution transformer as the primary key, and to uniformly map data with different sampling frequencies to a multi-scale time window of 15 minutes, day and month to form a multi-source time series sample library of distribution transformers.

[0088] The data cleaning and quality control module is used to complete missing data points of apparent power, identify and correct abnormal power and abnormal voltage, and verify power outage events to eliminate false power outage events, and output cleaned multi-source time series samples.

[0089] The feature extraction module is used to extract load pressure underlying features, low voltage underlying features, power outage reliability underlying features, fault and defect features, maintenance and repair features, equipment aging features, capacity matching features, and power supply radius features from the cleaned multi-source time series samples, forming a multi-dimensional underlying feature vector;

[0090] The health status evaluation module is used to normalize the multi-dimensional low-level feature vector to form a health evaluation feature vector; and to use the load pressure low-level feature, low voltage low-level feature and power outage reliability low-level feature, combined with the event severity, duration impact factor and time decay factor to calculate the load pressure risk, low voltage risk and power outage reliability risk respectively, and then calculate the comprehensive health risk and health index.

[0091] The health status classification module is used to classify the status of the distribution transformer into normal, attention, warning, serious and emergency levels according to the health index, rule indicator triggering conditions and future fault probability, and obtain the health status classification result.

[0092] The remaining lifetime prediction module is used to construct a lifetime prediction model with censored sample processing capability by taking historical multi-source time series features as input and the time of failure, decommissioning, replacement or major overhaul of distribution transformer as the lifetime endpoint, and outputting the future failure probability and predicted remaining lifetime.

[0093] The rule fusion module is used to use rule indicators such as heavy overload, low voltage, frequent power outages, old distribution transformers and high energy consumption distribution transformers as model constraints, to penalize and correct the health index and to correct the risk gain of future failure probability.

[0094] The results output module is used to output health index, health level, failure probability, remaining life, risk contribution and governance recommendations;

[0095] The closed-loop update module is used to update multi-source time series samples, feature indicators, and model parameters based on newly added running events and governance feedback samples.

[0096] Furthermore, the data cleaning and quality control module includes: a power completion unit for calculating apparent power based on active and reactive power; an anomaly identification unit for identifying power anomalies where the apparent power is null or exceeds the reasonable range of the rated capacity; a voltage correction unit for correcting or eliminating voltage anomaly data by combining data from adjacent times, adjacent users, and adjacent transformer areas; and a power outage verification unit for verifying power outage events and eliminating non-genuine power outage events based on voltage, current, duration, number of outages per day, and the proportion of users affected.

[0097] The rule fusion module is used to penalize and correct the health index based on the maximum load rate threshold, low voltage threshold, frequent power outage number threshold, old distribution transformer service life threshold, and high energy consumption equipment identification, and to correct the risk gain for the probability of future failures.

[0098] The closed-loop update module supports using changes in load rate, number of low voltage events, number of power outages, number of faults, and health index before and after the governance as feedback samples for continuous model training and parameter correction.

[0099] Compared with the prior art, the present invention has at least the following beneficial effects:

[0100] 1. This invention uses the unique identifier of the distribution transformer as the primary key and integrates multi-source data such as ledgers, operating curves, voltage and current, power outages, faults, defects, maintenance, and grid structure to comprehensively reflect the operating status of the distribution transformer.

[0101] 2. This invention improves the consistency and computability of time series samples by mapping data at different sampling frequencies to time scales such as 15 minutes, days, and months through multi-scale time window mapping.

[0102] 3. This invention reduces the impact of abnormal data acquisition, missing data, and false power outages on the assessment results by power compensation, voltage anomaly correction, and power outage event verification, thereby improving the accuracy of health assessment.

[0103] 4. This invention employs multi-dimensional risk characteristics such as load pressure, low voltage, power outage reliability, fault defects, maintenance, equipment aging, capacity matching, and power supply radius to identify potential transformer hazards under multiple overlapping risks.

[0104] 5. This invention integrates rule indicators with data-driven models, which not only preserves the interpretability of distribution network business rules, but also utilizes historical samples to learn fault patterns, thereby improving the reliability of fault probability prediction and lifetime prediction.

[0105] 6. This invention can output health index, health level, failure probability, predicted remaining life, main sources of risk and management suggestions, providing a basis for decision-making for distribution transformer capacity expansion, new construction, renovation, low-voltage switching, low-voltage renovation and maintenance plan optimization.

[0106] 7. This invention achieves closed-loop model updates through governance feedback samples, enabling the model to be continuously corrected based on operational data and governance effects, thereby improving its adaptability to different regions, different equipment models, and different operating conditions. Attached Figure Description

[0107] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0108] Figure 1 This is the overall flowchart of the method for assessing the health status and predicting the lifespan of distribution transformers based on multi-source time-series data according to the present invention.

[0109] Figure 2 This is a module architecture diagram of the distribution transformer health status assessment and life prediction system of the present invention;

[0110] Figure 3 This is a diagram of the multi-source data fusion and multi-scale time window mapping model of the present invention;

[0111] Figure 4 This is a diagram of the data cleaning and quality control model of the present invention;

[0112] Figure 5 This is a diagram of the health status assessment and grading model of the present invention;

[0113] Figure 6 This is a diagram of the remaining lifetime prediction model of the present invention;

[0114] Figure 7 This is a diagram of the rule constraints, governance suggestion generation, and closed-loop update model of this invention. Detailed Implementation

[0115] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0116] like Figure 1 As shown in the figure, this invention provides a method for assessing the health status and predicting the lifespan of distribution transformers based on multi-source time-series data, including the following steps:

[0117] Step 1. Multi-source data fusion of distribution transformers

[0118] In this embodiment, distribution transformers within a certain power supply area are selected as the evaluation object. Let the set of distribution transformers be:

[0119]

[0120] in, Indicates the first Taiwan distribution transformer.

[0121] The system uses a unique identifier for the distribution transformer. Primary key, such as Figure 3 As shown, multi-source data is extracted from the distribution transformer ledger system, electricity consumption information collection system, distribution automation system, power outage management system, fault repair system, defect management system, and maintenance management system.

[0122] For the The distribution transformer generates the following multi-source data set:

[0123]

[0124] in, This includes ledger data such as rated capacity, commissioning date, equipment model, power supply radius, and number of users connected; For load or power curves; The voltage curve; The current curve; For power outage records; For fault recording; Record defects; For maintenance records; This is the network topology data.

[0125] Because different systems have different data sampling frequencies, this embodiment sets three time windows: 15 minutes, daily, and monthly. For continuous data, the original sampling points are retained at the 15-minute scale, the maximum, minimum, average, 95th percentile, and volatility are calculated at the daily scale, and the cumulative number of occurrences, duration, and trend changes are calculated at the monthly scale. For event-based data, the number of events, cumulative duration, most recent occurrence time, and proportion of affected users are statistically analyzed at the daily and monthly scales.

[0126] This yields a multi-source timing sample for the distribution transformer:

[0127]

[0128] in, Indicates a 15-minute, day, or month time window. This represents the time window mapping function.

[0129] Step 2. Data Cleaning and Quality Control

[0130] like Figure 4As shown, the data cleaning and quality control model includes a power data processing module, a voltage data processing module, a power outage data processing module, and an abnormal data removal module, which respectively correspond to apparent power completion and anomaly identification, voltage anomaly identification and neighborhood correction, power outage duration, number of outages and the proportion of users affected, and output the cleaned data.

[0131] In this embodiment, the system first completes the power data. If the apparent power... Missing, but active power and reactive power If it exists, the apparent power is calculated according to the following formula:

[0132]

[0133] Assume the rated capacity of the distribution transformer is If the following conditions are met:

[0134]

[0135] or:

[0136]

[0137] This point will then be marked as a power anomaly and will not be directly included in the load rate and health risk calculations.

[0138] For voltage data, the system determines whether the voltage value is within an abnormal range, or whether it deviates significantly from data from adjacent times, adjacent users, or adjacent transformer areas. For correctable abnormal data points, a neighborhood-weighted correction is used.

[0139]

[0140] in, This is the corrected voltage value. and These are the voltages at adjacent time points. For adjacent user voltages, This represents the voltage of adjacent transformer substations.

[0141] Each weight satisfies:

[0142]

[0143] If the corrected voltage still does not meet the quality requirements, then the data point is discarded.

[0144] For power outage data, the system verifies suspected power outage events. An event is considered a genuine power outage only if the voltage and current are both empty at the same time, the outage duration is between 15 minutes and 15 hours, the number of outages on the same day does not exceed two, and the number of affected users is not less than 50% of the total number of users in the affected area. Otherwise, the event is considered a false outage and is removed.

[0145] Step 3. Feature Extraction

[0146] (1) Extraction of load pressure characteristics

[0147] The system calculates the load factor based on apparent power and rated capacity. :

[0148]

[0149] Extract the maximum load rate, average load rate, 95th percentile load rate, number of reloads, number of overloads, duration of reloads, duration of overloads, and load rate trend from the load rate sequence.

[0150] Load pressure risk Calculated using event accumulation:

[0151]

[0152] in, This is to evaluate the number of load anomalies identified during the evaluation period.

[0153] Severity of load events for:

[0154]

[0155] in, For the first Maximum load rate during a load event As the warning threshold, This is the overload threshold.

[0156] Duration Influence Factor for:

[0157]

[0158] Time decay factor for:

[0159]

[0160] Therefore, the load pressure risk can be further expressed as:

[0161]

[0162] Specifically, when the maximum load rate is not less than 100% and the duration of a single instance is not less than 2 hours, it is judged as overload; when the maximum load rate is not less than 80% and less than 100% and the duration of a single instance is not less than 2 hours, it is judged as heavy load; when the maximum load rate is not less than 70% and less than 80% and the duration of a single instance is not less than 2 hours, it is judged as heavy load warning.

[0163] (2) Low voltage feature extraction

[0164] The system extracts the number of low-voltage events, cumulative low-voltage duration, longest continuous low-voltage duration, low-voltage occurrence frequency, voltage deviation degree, voltage fluctuation rate, and low-voltage trend based on voltage time series data.

[0165] Low voltage risk for:

[0166]

[0167] Severity of low voltage events for:

[0168]

[0169] in, The lower limit for acceptable low voltage is preferably 207V; For the first The lowest voltage in a low-voltage event; This is the severe lower limit of low voltage.

[0170] Factors affecting the duration of low voltage events for:

[0171]

[0172] Low voltage event time decay factor for:

[0173]

[0174] Therefore, low voltage risk for:

[0175]

[0176] When the voltage is between 154V and 207V and there are 8 consecutive 15-minute sampling points, the system determines it as a low voltage event; when the voltage is less than 120V for more than 2600 points in total, or when the voltage is less than 93V and there are 8 consecutive 15-minute sampling points, it is also determined as a low voltage event.

[0177] (3) Extraction of power outage reliability features

[0178] The system extracts data such as the number of power outages, cumulative power outage duration, average power outage duration, frequent power outage indicators, the proportion of users affected by power outages, the number of power outages in the past 60 days, and the number of days since the most recent power outage based on actual power outage events.

[0179] Power outage reliability risk for:

[0180]

[0181] Percentage of users affected by power outage for:

[0182]

[0183] Frequent power outage amplification factor for:

[0184]

[0185] Severity of the power outage for:

[0186]

[0187] Therefore, power outage reliability risk for:

[0188]

[0189] When the number of actual power outages within the past 60 days is no less than 3, the system generates a frequent power outage flag.

[0190] (4) Extraction of fault defects and maintenance features

[0191] The system extracts fault count, defect level, fault interval, number of days since the last fault, number of days since the last maintenance, number of maintenance times, maintenance type, and maintenance response time after a fault based on fault, defect, and maintenance records.

[0192] Fault and Defect Risk for:

[0193]

[0194] in:

[0195]

[0196] Maintenance and repair risks for:

[0197]

[0198] If the fault occurred at the time The maintenance completion time is Then the repair response time after the fault for:

[0199]

[0200] The maintenance hysteresis amplification factor is:

[0201]

[0202] (5) Extraction of features related to equipment aging, capacity matching, and power supply radius

[0203] The service life is:

[0204]

[0205] in, For the commissioning date, This is the current evaluation time. When the service life is not less than 30 years, an old distribution transformer identifier is generated.

[0206] Capacity mismatch risk for:

[0207]

[0208] in:

[0209]

[0210] Power supply radius risk is :

[0211]

[0212] in:

[0213]

[0214] Step 4. Health Status Assessment and Classification

[0215] like Figure 5 As shown, the health status evaluation and classification model takes multiple features such as load, low voltage, power outage, fault defects, maintenance, aging, capacity and power supply radius as inputs, and performs feature normalization, comprehensive risk calculation, rule correction, health index calculation and health status classification in sequence, and outputs five status categories: normal, attention, warning, serious and emergency.

[0216] The system will load pressure Low voltage Power outage reliability Faults and defects Inspection and maintenance Equipment aging Capacity matching and power supply radius risk Normalization is performed to form a health assessment feature vector. :

[0217]

[0218] Comprehensive health risks for:

[0219]

[0220] Health Index for:

[0221]

[0222] Rule Penalty Items for:

[0223]

[0224] Revised Health Index for:

[0225]

[0226] The system classifies the status of distribution transformers into normal, watch, warning, critical, and emergency based on the corrected health index, rule triggering status, and future failure probability.

[0227] The normal state is when no rules such as heavy load, overload, low voltage, or frequent power outages are triggered and the health index is high; the watch state is when there is a heavy load warning, slight low voltage, or an increasing fault trend; the warning state is when heavy load, low voltage, or a high level of defect are triggered; the severe state is when overload, frequent power outages, severe low voltage, old distribution transformers with high load, or high-energy-consuming distribution transformers with faults and defects are triggered; the emergency state is when the health index is low and multiple high-risk rules are triggered at the same time, or the probability of a short-term failure in the future exceeds the preset threshold.

[0228] Step 5. Remaining life prediction

[0229] like Figure 6 As shown, the remaining useful life prediction model takes historical time series features as input, first identifies life end events such as failure, decommissioning, replacement or major overhaul, then performs censoring and life prediction, and outputs the probability of future failure, the probability of survival and the predicted remaining useful life.

[0230] The system takes the historical multi-source time-series characteristics of distribution transformers as input:

[0231]

[0232] The end of the distribution transformer's lifespan is defined as the time when it fails, is decommissioned, is replaced, or undergoes major maintenance. Let the end-of-life time be... The event identifier is .

[0233] When the distribution transformer experiences a life-end event during the observation period:

[0234]

[0235] When the distribution transformer does not experience a life-end event at the end of the observation period:

[0236]

[0237] Future No. The conditional risk probability for each time window is:

[0238]

[0239] The probability of survival is:

[0240]

[0241] future The probability of failure within the time window is:

[0242]

[0243] The predicted remaining lifespan is:

[0244]

[0245] During model training, paired variable samples that have not experienced a lifespan end event are used as right-censored samples for training. The loss function can be expressed as:

[0246]

[0247] Step 6. Integration of rule constraints and data-driven models

[0248] like Figure 7 As shown, the rule constraint, governance suggestion generation and closed-loop update model incorporates rule indicators such as heavy overload, low voltage, frequent power outages, old distribution transformers, and high energy consumption distribution transformers into the comprehensive risk calculation. After rule correction, health index calculation and status health classification, governance suggestions such as capacity expansion, new construction, renovation, low voltage switching and maintenance optimization are generated. The governance feedback samples are fed back to the comprehensive risk calculation and rule correction stage to support the closed-loop update.

[0249] The system uses rule-based indicators such as transformer overload, low voltage, frequent power outages, aging transformers, and high-energy-consuming transformers as strong feature inputs to the model. (Rule triggering risk) for:

[0250]

[0251] in, For the first Whether a rule indicator is triggered is set to 1 if triggered and 0 if not triggered.

[0252] The base failure probability output by the data-driven model is:

[0253]

[0254] The fault probability after rule correction is:

[0255]

[0256] The more rules are triggered, the higher the severity of the rules, and the closer the rule event is to the current time, the more likely it is to occur. The larger the value, the higher the probability of eventual failure.

[0257] Step 7. Output Results and Generation of Governance Suggestions

[0258] The system outputs the current health index, health level, main sources of risk, probability of failure in the next 7 days, 30 days, 90 days and 1 year, predicted remaining life, risk contribution ranking and similar historical cases for each distribution transformer.

[0259] No. The contribution of risk class is:

[0260]

[0261] The system sorts the main sources of risk from highest to lowest contribution.

[0262] The similarity between the current distribution transformer and the historical distribution transformer is:

[0263]

[0264] The system selects historical cases with high similarity and known governance results as the basis for generating the current governance plan.

[0265] The rules for generating governance recommendations are as follows:

[0266] (1) When the main risk is transformer overload, heavy load or heavy load warning, generate suggestions for transformer capacity expansion, transformer construction or low-voltage switching;

[0267] (2) When the main risk is low voltage in the distribution area, generate suggestions for transformer renovation, new transformer construction or low voltage renovation;

[0268] (3) When the main risk is frequent power outages, generate suggestions for transformer upgrades, transformer and low-voltage upgrades, or low-voltage upgrades;

[0269] (4) When the main risk is an old or high-energy-consuming distribution transformer, generate recommendations for transformer renovation or transformer and low-voltage renovation;

[0270] (5) When the main risks are faults, defects, delayed maintenance response or repeated faults, generate maintenance plan optimization, defect elimination or equipment replacement suggestions.

[0271] Step 8. Closed-loop update

[0272] When a new operational anomaly, power outage, low voltage, fault, defect, or maintenance event occurs in a distribution transformer, the system automatically writes the new event into the multi-source time-series sample library and recalculates various risks, health indices, health levels, future failure probabilities, and remaining lifespan.

[0273] The health characteristics prior to the event update were:

[0274]

[0275] The updated health characteristics are as follows:

[0276]

[0277] The characteristic change is:

[0278]

[0279] For distribution transformers that have already undergone remediation, the system extracts changes in load rate, number of low voltage occurrences, number of power outages, number of faults, and health index before and after the remediation to form feedback samples:

[0280]

[0281] The system adds feedback samples to the model training sample library and periodically updates feature weights, rule penalty coefficients, failure probability prediction parameters, and remaining life prediction parameters, thus forming a closed-loop update mechanism of "data fusion - feature extraction - health assessment - life prediction - governance suggestions - governance feedback - model correction".

[0282] This invention also provides a system for assessing the health status and predicting the lifespan of distribution transformers based on multi-source time-series data, corresponding to the above method, for executing the method described above, including:

[0283] The multi-source data fusion module is used to associate ledger data, operation data, power outage data, fault and defect data, maintenance data and grid data with the unique identifier of the distribution transformer as the primary key, and to uniformly map data with different sampling frequencies to a multi-scale time window of 15 minutes, day and month to form a multi-source time series sample library of distribution transformers.

[0284] The data cleaning and quality control module is used to complete missing data points of apparent power, identify and correct abnormal power and abnormal voltage, and verify power outage events to eliminate false power outage events, and output cleaned multi-source time series samples.

[0285] The feature extraction module is used to extract load pressure underlying features, low voltage underlying features, power outage reliability underlying features, fault and defect features, maintenance and repair features, equipment aging features, capacity matching features, and power supply radius features from the cleaned multi-source time series samples, forming a multi-dimensional underlying feature vector;

[0286] The health status evaluation module is used to normalize the multi-dimensional low-level feature vector to form a health evaluation feature vector; and to use the load pressure low-level feature, low voltage low-level feature and power outage reliability low-level feature, combined with the event severity, duration impact factor and time decay factor to calculate the load pressure risk, low voltage risk and power outage reliability risk respectively, and then calculate the comprehensive health risk and health index.

[0287] The health status classification module is used to classify the status of the distribution transformer into normal, attention, warning, serious and emergency levels according to the health index, rule indicator triggering conditions and future fault probability, and obtain the health status classification result.

[0288] The remaining lifetime prediction module is used to construct a lifetime prediction model with censored sample processing capability by taking historical multi-source time series features as input and the time of failure, decommissioning, replacement or major overhaul of distribution transformer as the lifetime endpoint, and outputting the future failure probability and predicted remaining lifetime.

[0289] The rule fusion module is used to use rule indicators such as heavy overload, low voltage, frequent power outages, old distribution transformers and high energy consumption distribution transformers as model constraints, to penalize and correct the health index and to correct the risk gain of future failure probability.

[0290] The results output module is used to output health index, health level, failure probability, remaining life, risk contribution and governance recommendations;

[0291] The closed-loop update module is used to update multi-source time series samples, feature indicators, and model parameters based on newly added running events and governance feedback samples.

[0292] Compared with the prior art, the present invention has the following characteristics and effects:

[0293] 1. By using the unique identifier of the distribution transformer as the primary key, the data of the distribution transformer ledger, operation curve data, power outage records, fault records, defect records, maintenance records, and grid topology data are uniformly associated. Data with different sampling frequencies and time granularities are mapped to multi-scale time windows of 15 minutes, days, and months to form a multi-source time series sample library of distribution transformers. This solves the problems of scattered data sources, inconsistent data caliber, and insufficient continuity of time series samples in the existing technology, and provides a complete, continuous, and computable data foundation for the health status assessment, life prediction, and management suggestions of distribution transformers.

[0294] 2. By establishing a data cleaning and quality control mechanism for distribution transformer operation data, missing apparent power data is supplemented using active and reactive power data, abnormal power data is identified and marked, and abnormal voltage data is corrected or removed by combining data from adjacent times, adjacent users, and adjacent transformer areas. Furthermore, the authenticity of outage events is verified based on voltage, current, outage duration, number of outages per day, and the proportion of users affected by outages. This effectively eliminates false outage events caused by abnormal data collection, missing data, and false alarms, reduces the interference of low-quality data on health assessment results, and improves the accuracy of identifying risks such as heavy overload, low voltage, and frequent outages in distribution transformers.

[0295] 3. Construct eight-dimensional health risk characteristics for distribution transformers, including load pressure, low voltage, power outage reliability, fault defects, maintenance, equipment aging, capacity matching, and power supply radius. This comprehensively reflects the health level of distribution transformers in terms of operating load, voltage quality, power supply reliability, equipment status, operation and maintenance response, capacity adaptation, and grid conditions. It overcomes the problem of insufficient assessment dimensions caused by existing technologies that rely solely on a single load rate, single fault record, or operating years for status judgment, and improves the comprehensiveness and accuracy of distribution transformer health status evaluation.

[0296] 4. The distribution network business rules such as heavy overload, low voltage, frequent power outages, old distribution transformers and high energy consumption distribution transformers are used as prior constraints or strong feature inputs of the model. When the rules are triggered, the health index is penalized and the future failure probability is risk-gained. This makes the model output results not only consistent with historical data patterns, but also meet the actual business judgment logic of the distribution network. It avoids the problem of inconsistency between pure data-driven models and engineering experience, and improves the interpretability and reliability of the model results.

[0297] 5. Based on the risk contribution ranking, automatically match differentiated governance suggestions such as distribution transformer capacity expansion, new construction, renovation, low-voltage switching, low-voltage renovation and maintenance plan optimization, and use the changes before and after governance as feedback samples for continuous model correction, forming a closed loop of "assessment-suggestion-governance-feedback", supporting the model to adaptively update with operating conditions.

[0298] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for assessing the health status of distribution transformers, predicting their lifespan, and generating remediation recommendations based on multi-source time-series data, characterized in that, Includes the following steps: S1: Establish a multi-source data fusion model for distribution transformers, using the unique identifier of the distribution transformer as the primary key, and associate the distribution transformer ledger data, rated capacity, commissioning date, equipment model, power supply radius, number of users connected, load curve, voltage curve, current curve, power outage record, fault record, defect record, maintenance record and grid topology data, and uniformly map data with different sampling frequencies to 15-minute, daily and monthly multi-scale time windows to form a multi-source time series sample library for distribution transformers; S2: Perform data cleaning and quality control on the raw data in the multi-source time-series sample library of the distribution transformer, including: for data points with missing apparent power, calculate the apparent power based on active power and reactive power to complete them; mark data points with null values ​​or exceeding the reasonable range of rated capacity as outliers; correct or remove abnormal voltage data by combining data from adjacent times, adjacent users and adjacent transformer areas; verify power outage events based on voltage, current, duration, number of times per day and the proportion of users affected, and remove false power outage events caused by non-real power outages or abnormal data collection, to obtain the cleaned multi-source time-series sample. S3: Extract the indicator features of the distribution transformer based on the cleaned multi-source time-series samples, including: extracting the maximum load rate, number of heavy loads, number of overloads, duration of heavy loads, duration of overloads, and load rate change trend as the underlying features of load pressure based on the load rate sequence; extracting the number of low-voltage events, cumulative duration of low voltage, longest duration of continuous low voltage, degree of voltage deviation, voltage fluctuation rate, and low voltage trend as the underlying features of low voltage based on voltage time-series data; extracting the number of power outages, cumulative duration of power outages, average duration of power outages, frequent power outage indicators, proportion of users affected by power outages, and number of power outages in the past 60 days as the underlying features of power outage reliability based on power outage event data; extracting fault and defect features and maintenance features based on fault, defect, and maintenance data; and extracting equipment aging features, capacity matching features, and power supply radius features based on ledger and grid data to form a multi-dimensional underlying feature vector. S4: Based on the multi-dimensional underlying feature vector, the characteristics of load pressure, low voltage, power outage reliability, fault defects, maintenance, equipment aging, capacity matching, and power supply radius are normalized to form a health evaluation feature vector; and using the underlying features of load pressure, low voltage, and power outage reliability, combined with the event severity, duration influence factor, and time decay factor, the load pressure risk, low voltage risk, and power outage reliability risk are calculated respectively, and then a comprehensive risk accumulation model is constructed to calculate the comprehensive health risk and health index of the distribution transformer; S5: Based on the health index, rule indicator triggering conditions and future fault probability, the health status of the distribution transformer is divided into normal, attention, warning, severe and emergency levels to obtain the health status classification result; S6: Using the historical time series data of the multidimensional low-level feature vector as input, and taking the time of failure, decommissioning, replacement or major overhaul of the distribution transformer as the end of its lifespan, construct a remaining life prediction model with censored sample processing capability, and output the failure probability and predicted remaining lifespan for multiple future time windows. S7: Establish a fusion mechanism between rule constraints and data-driven models. Use the rule indicators of distribution transformer overload, low voltage, frequent power outages, old distribution transformers and high energy consumption distribution transformers as model prior constraints or strong feature inputs. Penalize and correct the health index and risk gain correction the future failure probability to obtain the corrected health index and the final failure probability. S8: Outputs the current health index, health level, main sources of risk, probability of future failure, predicted remaining life, risk contribution ranking and similar historical cases for each distribution transformer, and automatically generates governance suggestions based on the trigger indicator type and risk level. S9: Establish a closed-loop update mechanism. When a distribution transformer experiences a new operational anomaly, power outage, low voltage, fault, defect, or maintenance event, automatically update the multi-source time series samples, characteristic indicators, health index, health level, fault probability, and remaining life prediction results. Use the changes before and after the treatment as feedback samples for continuous model training and parameter correction.

2. The method according to claim 1, characterized in that, In step S1, the multi-source data set of the distribution transformer is represented as follows: ； in, This represents the data in the distribution transformer ledger. Represents the load or power curve. Represents the voltage curve. Represents the current curve. This indicates a power outage record. Indicates fault record, Indicates a defect record. This indicates maintenance records. Represents network topology data; Data from different sampling frequencies are uniformly mapped as follows: ； in, Indicates the first Taiwan distribution transformer at time Time window Multi-source time series samples, It includes 15-minute, day, and month timescales.

3. The method according to claim 1, characterized in that, In step S2, when apparent power is missing, it is calculated according to the following formula: ； in, Apparent power, Active power Reactive power; When satisfied ; or When this data point is an anomaly, it is marked as a power anomaly. This refers to the rated capacity of the distribution transformer. When verifying a power outage event, a real power outage event must simultaneously meet the following conditions: (1) Voltage and current are null values ​​at the same time; (2) The power outage lasted between 15 minutes and 15 hours; (3) The number of power outages in the same day shall not exceed 2; (4) The number of users affected by the power outage shall not be less than 50% of the total number of users in the transformer area; If a power outage event does not simultaneously meet the above conditions, it will be judged as a pseudo-power outage event and removed.

4. The method according to claim 1, characterized in that, In step S4, the load pressure risk is calculated using an event accumulation method: ； in, Indicates the first The severity of each load anomaly event, Indicates the duration of influence factor. Indicates the time decay factor; Low voltage risk The following formula is used for calculation: ； Among them, the severity of low voltage events for: ； in, This is the lower limit for acceptable low voltage. For the first The lowest voltage in a low-voltage event. This is the severe lower limit of low voltage. Power outage reliability risk The following formula is used for calculation: ； in, The severity of the power outage; ； To determine the percentage of users affected by the power outage, ; For frequent power outages, ; in, For the first The number of users affected by this power outage This represents the total number of users in the Taiwan region. This refers to the cumulative number of power outages in the 60 days following the incident.

5. The method according to claim 1, characterized in that, In step S4, the health evaluation feature vector for: ； in, This is the evaluation value for the equipment load level. This is the power supply voltage quality evaluation value. This is the evaluation value for the impact of the power outage event. This is the equipment failure risk assessment value. This is an evaluation value for the equipment's operation, maintenance, and repair status. This is the equipment aging evaluation value. To match the evaluation value to the equipment capacity, This is the risk assessment value for the power supply radius; Comprehensive health risks for: ； Health index is : ； in, Risks are triggered by the rules.

6. The method according to claim 1, characterized in that, In step S6, the remaining life prediction model uses the failure, decommissioning, replacement, or major overhaul of the distribution transformer as the life end event and sets an event identifier. ; When a distribution transformer experiences a life-end event during the observation period: ； When the distribution transformer does not experience a life-end event at the end of the observation period: ； This sample was then used as a right-censored sample in model training. Future No. Probability of life end event occurring within a time window for: ； in, For the first Taiwan distribution transformer at time Under the conditions of the state, in the future The conditional probability of a life-end event occurring within a time window. To be used to map the model output to to Activation function for the interval, For the first The bias parameters corresponding to each time window The total number of health assessment features. For the health assessment characteristics, For the first The health assessment feature in the first The influence coefficient under each time window For the first Taiwan distribution transformer serial number, For the k-th time \(t) A health assessment characteristic value, For the rule to trigger risk in the first The influence coefficient under each time window For the first Taiwan distribution transformer at time The rules trigger risk, Numbering of distribution transformers For the current observation time, This refers to the sequence number of the future time window; Survival probability for: ； in, For the first Taiwan distribution transformer at time Under the condition of the future, The probability of survival if no life-end event occurs within a time window; For the first Taiwan distribution transformer at time Under the conditions of the state, in the future The conditional probability of a life-end event occurring within a time window; For this distribution transformer in the future The conditional probability that no life-end event will occur within a time window; Indicates the future The first to the second Multiply the probabilities of not experiencing the life end event within each time window; Numbering of distribution transformers For the current observation time, This represents the time window number during the multiplication process. For the number of forecast time windows or future time window numbers; future The probability of failure within the time window is: ； Predicting remaining lifespan for: ； in, For the first Taiwan distribution transformer at time Predicted remaining lifespan; For the first Taiwan distribution transformer at time Under the condition of the future, The probability of survival if no life-end event occurs within a time window; The total number of future time windows covered by the remaining lifespan prediction; This refers to the sequence number of the future time window. The time length corresponding to a single forecast time window; For the first The expected survival time corresponding to the survival probability of each time window; Indicates the future The first to the second The expected survival time of each time window is accumulated; Numbering of distribution transformers This is the current observation time.

7. The method according to claim 1, characterized in that, In step S7, the risk of rule triggering. for: ； in, Indicates the first The severity of each rule indicator, Indicates the first Each rule indicator at time Whether to trigger, Indicates the rule trigger time. Indicates the time decay coefficient; The fault probability after rule correction is: ； in, The base failure probability output by the data-driven model. This is the rule-based risk gain coefficient.

8. The method according to claim 1, characterized in that, In step S8, risk contribution Calculate using the following formula: ； in, For the first A health assessment characteristic at time Contribution to the overall health risk of distribution transformers; For the first Risk weights corresponding to each health assessment feature; For the first A health assessment characteristic at time Normalized eigenvalues; For the first The weighted risk contribution value of each health assessment feature to the overall health risk; For the distribution transformer at time Comprehensive health risks; For health assessment feature serial numbers, This is the current evaluation moment. The larger the value, the more significant the number. The greater the contribution of each health assessment feature to the current comprehensive health risk, the more the system determines the main risk sources based on the risk contribution ranking, and generates governance suggestions based on the main risk sources and health level: Specifically, when the main risk is distribution transformer overload, heavy load, or heavy load warning, suggestions are generated for distribution transformer capacity expansion, new distribution transformer construction, or low-voltage switching modification; when the main risk is low voltage in the distribution area, suggestions are generated for distribution transformer renovation, new distribution transformer construction, or low-voltage renovation; when the main risk is frequent power outages, suggestions are generated for distribution transformer renovation, distribution transformer and low-voltage renovation, or low-voltage renovation; when the main risk is old distribution transformers or high-energy-consuming distribution transformers, suggestions are generated for distribution transformer renovation, or distribution transformer and low-voltage renovation.

9. A system for assessing the health status and predicting the lifespan of distribution transformers based on multi-source time-series data, used to execute the method described in any one of claims 1-8, characterized in that, include: The multi-source data fusion module is used to associate ledger data, operation data, power outage data, fault and defect data, maintenance data and grid data with the unique identifier of the distribution transformer as the primary key, and to uniformly map data with different sampling frequencies to a multi-scale time window of 15 minutes, day and month to form a multi-source time series sample library of distribution transformers. The data cleaning and quality control module is used to complete missing data points of apparent power, identify and correct abnormal power and abnormal voltage, and verify power outage events to eliminate false power outage events, and output cleaned multi-source time series samples. The feature extraction module is used to extract load pressure underlying features, low voltage underlying features, power outage reliability underlying features, fault and defect features, maintenance and repair features, equipment aging features, capacity matching features, and power supply radius features from the cleaned multi-source time series samples, forming a multi-dimensional underlying feature vector; The health status evaluation module is used to normalize the multi-dimensional low-level feature vector to form a health evaluation feature vector; and to use the load pressure low-level feature, low voltage low-level feature and power outage reliability low-level feature, combined with the event severity, duration impact factor and time decay factor to calculate the load pressure risk, low voltage risk and power outage reliability risk respectively, and then calculate the comprehensive health risk and health index. The health status classification module is used to classify the status of the distribution transformer into normal, attention, warning, serious and emergency levels according to the health index, rule indicator triggering conditions and future fault probability, and obtain the health status classification result. The remaining lifetime prediction module is used to construct a lifetime prediction model with censored sample processing capability by taking historical multi-source time series features as input and the time of failure, decommissioning, replacement or major overhaul of distribution transformer as the lifetime endpoint, and outputting the future failure probability and predicted remaining lifetime. The rule fusion module is used to use rule indicators such as heavy overload, low voltage, frequent power outages, old distribution transformers and high energy consumption distribution transformers as model constraints, to penalize and correct the health index and to correct the risk gain of future failure probability. The results output module is used to output health index, health level, failure probability, remaining life, risk contribution and governance recommendations; The closed-loop update module is used to update multi-source time series samples, feature indicators, and model parameters based on newly added running events and governance feedback samples.

10. The system as described in claim 9, characterized in that, The data cleaning and quality control module includes: a power completion unit for calculating apparent power based on active and reactive power; an anomaly identification unit for identifying power anomalies where the apparent power is null or exceeds the reasonable range of the rated capacity; a voltage correction unit for correcting or removing voltage anomaly data by combining data from adjacent times, adjacent users, and adjacent transformer areas; and a power outage verification unit for verifying power outage events and removing non-genuine power outage events based on voltage, current, duration, number of outages per day, and the proportion of users affected. The rule fusion module is used to penalize and correct the health index based on the maximum load rate threshold, low voltage threshold, frequent power outage number threshold, old distribution transformer service life threshold, and high energy consumption equipment identification, and to correct the risk gain for the probability of future failures. The closed-loop update module supports using changes in load rate, number of low voltage events, number of power outages, number of faults, and health index before and after the governance as feedback samples for continuous model training and parameter correction.

Images