A power transmission and transformation equipment operation state monitoring method and system

Abstract

This invention discloses a method and system for monitoring the operating status of power transmission and transformation equipment. First, real-time operating data of the power transmission and transformation equipment is collected and standardized for preprocessing. Then, feature extraction is performed on the preprocessed data to obtain feature data of the power transmission and transformation equipment. A pre-trained least squares support vector machine model is used to obtain the equipment operating status prediction result. Next, the equipment operating status prediction result is converted into a state confidence score. The state confidence scores are then fused using an improved D-S evidence theory method based on Euclidean distance to obtain a comprehensive confidence score for the equipment operating status. Finally, based on the comprehensive confidence score, an exponentially weighted moving average algorithm is used to construct a dynamic threshold for graded early warning of the power transmission and transformation equipment. This invention solves the technical problem that simple models cannot cope with the nonlinear correlation of multiple parameters in power transmission and transformation, and accurately adapts to the actual operation and maintenance needs of multi-parameter collaborative monitoring of power transmission and transformation systems.

Description

Technical Field

[0001] This invention belongs to the field of power grid safety technology and relates to a method and system for monitoring the operating status of power transmission and transformation equipment. Background Technology

[0002] Transmission and transformation equipment is the core carrier for the power grid to realize power transmission and distribution. Its operating status directly determines the stability and security of the power grid. In actual operation and maintenance, the operating status of the equipment is affected by the coupling of multiple parameters such as voltage, temperature, partial discharge, and ambient temperature and humidity. It is necessary to comprehensively and accurately judge the equipment status by integrating multi-source monitoring information.

[0003] Traditional methods for monitoring the operating status of power transmission and transformation equipment mainly revolve around core parameters such as voltage, temperature, partial discharge, and ambient temperature and humidity, and include the following: (1) Pulse current method: By connecting the detection impedance to the equipment grounding wire or bushing end screen, the nanosecond-level pulse current generated by partial discharge (PD) can be captured, and the apparent discharge quantity, repetition rate and other characteristic quantities can be quantitatively obtained. It has high sensitivity and can still identify weak discharges <5 pC. It is currently the only PD quantitative method with metrological traceability. The on-site anti-interference algorithm is mature (such as FFT threshold and polarity identification) and is easy to interface with the IEC 60270 standard. However, it is necessary to connect the coupling impedance after power is cut off, and it is not possible to achieve continuous monitoring of key nodes such as main transformer and GIS. It has poor response to high-frequency oscillating discharge (>30 MHz) and is easy to underestimate the hazards of GIS gap discharge.

[0004] (2) Infrared / ultraviolet imaging method: This method uses an infrared thermal imager to capture the surface radiation temperature difference of equipment, or an ultraviolet imager to record 240–280 nm ultraviolet photons of corona discharge, to achieve "visualized" inspection. It is intuitive and non-contact, and can locate abnormal temperature rise or corona discharge in multiple parts such as bushings, clamps, and insulators with a single scan; it does not require power outages and is suitable for live inspections of 500 kV and above, with high efficiency. However, it can only expose surface defects and is powerless to detect internal hot spots in oil-paper insulation or internal discharges in enclosed GIS; ambient temperature, sunlight, and wind speed can all change the surface emissivity, causing false alarms or missed alarms, requiring professional experience for interpretation.

[0005] (3) Dissolved gas analysis in oil: Transformer oil decomposes under thermal and electrical faults, producing characteristic gases such as H2, CH4, C2H6, C2H4, and C2H2. These gases are separated by chromatographic column separation and quantified by TCD / FID detector. It is sensitive to various faults such as overheating, arcing, and sparking, and can be combined with empirical methods such as the "three ratios" and "David's triangle" to indicate the fault type. Offline sampling and online photoacoustic spectroscopy devices are commercially available, and the data is highly comparable. However, the offline DGA cycle is usually 3–12 months, which cannot capture sudden discharges; by the time the test results are obtained, the fault may have expanded; it is only applicable to oil-immersed equipment and is ineffective for SF6 insulation, dry cables, and other scenarios.

[0006] (4) Radio frequency / UHF method: The steep leading edge of the PD pulse radiates electromagnetic waves of 300 MHz–3 GHz, which are received by built-in or external UHF sensors and located by time-frequency analysis. It can realize online monitoring. Built-in antennas can be installed on GIS and transformer oil-paper interfaces, with a positioning error of <10 cm. It has strong anti-corona interference capability and the signal-to-noise ratio is 10–20 dB higher than the pulse current method. However, the sensitivity is limited by the sensor installation position. For equipment already in operation, holes need to be opened or windows need to be added, which makes on-site modification difficult. On-site 4G / 5G, radar, and walkie-talkie signals are prone to co-channel interference, requiring complex spectrum fingerprint comparison.

[0007] (5) Fixed-weight multi-parameter voting method: This method pre-sets fixed weights for different monitoring parameters, then obtains the independent judgment results of each parameter on the equipment status, and calculates the total vote rate of each type of status according to the preset weights. The status with the highest vote rate is taken as the final operating status of the equipment. However, the multi-parameter voting method relies on fixed weights and cannot be dynamically adjusted according to the actual degree of conflict between parameters. This leads to conflicting parameters interfering with the fusion results, ultimately causing the judgment result of the equipment operating status to deviate from the actual situation, making it difficult to meet the actual needs of refined monitoring and fault early warning of power transmission and transformation equipment. Summary of the Invention

[0008] To address the technical problems existing in the aforementioned multi-parameter voting method, this invention provides a method and system for monitoring the operating status of power transmission and transformation equipment that can accurately adapt to the actual operation and maintenance needs of multi-parameter collaborative monitoring in power transmission and transformation systems.

[0009] The technical solution adopted by the present invention is: a method for monitoring the operating status of power transmission and transformation equipment, comprising the following steps: Step 1: Collect real-time operating data of power transmission and transformation equipment and perform standardized preprocessing. The power transmission and transformation equipment includes high-voltage transmission lines, distribution transformers, and high-voltage switchgear. Step 2: Based on the preprocessed data, feature extraction is performed to obtain the feature data of the power transmission and transformation equipment. Using the pre-trained least squares support vector machine model, the prediction results of the equipment operating status are obtained. The pre-trained least squares support vector machine model employs an improved sparrow search algorithm to adjust the penalty factor during training. With kernel parameters The improved sparrow search algorithm performs global optimization by introducing a Logistic chaos factor to generate an initial population and enhance population diversity, and introduces adaptive weights. The position is updated by introducing Cauchy mutation to help the watcher escape local optima. The final output is the parameter combination with the minimum fitness value by iteratively updating the position of individual sparrows. Step 3: Convert the equipment operating status prediction results into state confidence scores. The improved DS evidence theory method based on Euclidean distance is used to assess the confidence level of the state. The data is then integrated to obtain the overall confidence level of the equipment's operating status. Step 4: Based on the comprehensive confidence level of the equipment's operating status, construct a dynamic threshold using an exponentially weighted moving average algorithm to conduct graded early warning for power transmission and transformation equipment.

[0010] Preferably, in step 2, the training process of the trained least squares support vector machine model includes the following sub-steps: Step 2.1: First, collect historical operating data of power transmission and transformation equipment, and calculate the data integrity and validity of the historical operating data; then, remove abnormal data from the historical operating data; next, fill in the missing data found in the data integrity calculation using the average value of the previous and subsequent data; finally, perform standardization and dimensionality reduction processing to obtain the training dataset. Step 2.2: Use an improved sparrow search algorithm to globally optimize the penalty factor and kernel parameters; Step 2.3: Based on the optimal penalty factor and kernel parameters obtained through optimization, construct a least squares support vector machine model, and use the input features and corresponding state labels of the least squares support vector machine model for training to obtain a trained least squares support vector machine model.

[0011] Preferably, in step 2.1, the data integrity... Reflecting the degree of data missing, ,in For the number of missing data for a single parameter, This represents the total number of samples for a single parameter; when When the data integrity is ≥ threshold A, the data is deemed to be complete. The validity of the data This reflects whether the data exceeds the physical range of the equipment. ,in, For invalid data volume of a single parameter exceeding the measurement range, when When the value is greater than or equal to threshold B, the data is deemed valid. The local outlier factor algorithm is used to analyze the historical operational data. First, it is based on the k-neighborhood distance of data point P. Calculate the reachability distance from data point P to any data point O in its neighborhood. Then calculate the local reachability density of data point P in its k-neighborhood. The local outlier factor of the most recently calculated data point P. ,when When the value is greater than or equal to the threshold C, P is determined to be abnormal data and removed. Kernel principal component analysis (KPCA) is used to reduce the dimensionality of the data. First, a Gaussian kernel function is used to map the low-dimensional data to a high-dimensional feature space. Then, a centralized kernel matrix is ​​constructed, and the eigenvalues ​​and corresponding eigenvectors of the centralized kernel matrix are calculated. The top eigenvalues ​​are selected based on their cumulative contribution rate being greater than or equal to a threshold D. The eigenvectors are used as principal components to reduce the dimensionality of the data; among them... This is the default value.

[0012] Preferably, in step 2, the formula for generating the initial population is: ;in, Let be the value of the k-th chaotic iteration. This represents the value in the (k+1)th chaotic iteration. For Logistic chaos factors; The adaptive weight Then the first The dim-dimensional parameter of the i-th individual sparrow in the next iteration ;in, This represents the current iteration number. The maximum number of iterations, For the first The dim-th dimension parameter of the i-th sparrow individual in the next iteration. To simulate the random probability of environmental alert for sparrow populations, As a safety threshold, For random numbers that follow a normal distribution, It is a matrix of all 1s.

[0013] Preferably, in step 2, the introduction of Cauchy variation to help the vigilant escape local optima is described in the first... The dim-dimensional parameter of the i-th individual sparrow in the next iteration ;in, For the first The globally optimal parameters for the next iteration. These are standard Cauchy distribution random numbers. For the first The dim dimension parameter of the i-th individual sparrow in the next iteration.

[0014] Preferably, in step 2, the fitness value ;in, For the least squares support vector machine model in parameter combination (c, The number of correctly classified samples under ) The number of samples in the training dataset; The kernel function of the least squares support vector machine model ;in, and These are the k-th and l-th eigenvectors after the second dimensionality reduction by the kernel principal component analysis algorithm, respectively. For the optimized least squares support vector machine model kernel parameters; The objective of the least squares support vector machine model is to minimize the regularization loss function, which is calculated using the following formula: ; ; in, This represents the weight vector in the feature space of the least squares support vector machine model. This represents the bias term of the least squares support vector machine model. Let the fitting error be the value of the k-th sample. The feature map is the result of the second dimensionality reduction algorithm of kernel principal component analysis. To improve the penalty factor in the sparrow algorithm optimization, The number of samples in the training dataset. Indicates constraints. Let the true state label be the one corresponding to the k-th training sample. This is the k-th eigenvector after the second dimensionality reduction by the kernel principal component analysis algorithm; The constrained optimization problem is transformed into a system of linear equations to be solved using the Lagrange multiplier method, ultimately yielding the prediction function of the least squares support vector machine model: ; in, The predicted value output by the model. This is the kernel function for the least squares support vector machine model. is the feature vector of the sample to be predicted; For Lagrange multipliers, the weights of the k-th training sample in the prediction result are represented.

[0015] Preferably, in step 3, the least squares support vector machine model outputs four raw prediction scores for each monitoring dimension m. , , , The values ​​represent the decision values ​​for four states s: s=0 for normal, s=1 for attentive, s=2 for abnormal, and s=3 for severe. The raw scores are then converted into state confidence scores using the softmax function. ,in, This represents the prediction score of the least squares support vector machine model for the m-th monitoring dimension and the s-th state. For the corresponding State confidence; organize the state confidence of each monitoring dimension into evidence bodies, each evidence body being a confidence vector for that dimension, and the evidence body for the m-th monitoring dimension. ; Define the set of fusion targets for the status of power transmission and transformation equipment: = ,in, For mutually exclusive state categories, the evidence bodies of each monitoring dimension correspond to the basic probability allocation function of DS theory. , For the m-th piece of evidence, the first piece of evidence is... Basic probability assignment values ​​for class states; The degree of conflict between pieces of evidence is calculated using Euclidean distance. The formula for calculating the Euclidean distance between the m-th and n-th pieces of evidence is as follows: ; in, Let be the Euclidean distance between the m-th and n-th pieces of evidence. For the m-th piece of evidence, the first piece of evidence is... The basic probability assignment values ​​for class states, For the nth piece of evidence, the first piece of evidence is... The basic probability assignment values ​​for class states, For device status category; when ≥ Conflict Threshold When the case is determined to be "moderate to high conflict", the weight of evidence is adjusted; when < At that time, it was a period of "low conflict," and direct integration was achieved. For evidence with "moderate to high conflict", the specific steps for adjusting the weight of evidence are as follows: Calculate the similarity between the m-th piece of evidence and the n-th piece of evidence. Calculate the support of the m-th piece of evidence. The sum of the similarities between all other pieces of evidence and this evidence; where, Let m be the total number of monitoring dimensions; then the evidence weight of the m-th piece of evidence is... The adjusted evidence weights are incorporated into the DS fusion, and the resulting fusion yields the overall confidence level of the device's operating status. ; in, The improved conflict coefficient quantifies the degree of conflict between pieces of evidence; For the first The weight of evidence for each piece of evidence For the first One piece of evidence for the state The basic probability allocation value; This indicates the status of each piece of evidence in its respective judgment. The intersection equals the target state. ; This means multiplying the "reliability weight" of each piece of evidence by its "confidence level of its own state" to obtain the "weighted contribution value" of that piece of evidence to the target state, and then multiplying them together to calculate the joint contribution value of all pieces of evidence; ; This means calculating the "weighted state support" of a single piece of evidence, which combines the reliability of the evidence with its support for a certain state to obtain the "actual effective support" of the evidence for that state in the fusion process; This represents the product of the weighted support of each core dimension for the state it supports. A "true conflict" is defined as a situation where there is no consensus on the determination of the evidence and no common direction. Only combinations of such states will be included in the summation.

[0016] As a preferred option, in step 4, an exponentially weighted moving average algorithm is used to construct a dynamic threshold. The threshold is updated in real time through a sliding window to adapt to fluctuations in equipment operating conditions. The dynamic threshold is divided into "state-level thresholds", corresponding to four states: "normal / attention / abnormal / serious". The specific implementation includes the following sub-steps: Step 4.1: Employ the Exponentially Weighted Moving Average (EWMA) algorithm, assigning different weights to historical and current data to balance "historical patterns" and "real-time changes"; where the EWMA statistic for the t-th sampling time corresponds to the s-th state. ; Let be the combined confidence score of the s-th state after fusion at time t. These are the EWMA weighting coefficients. Let be the EWMA statistic for the s-th state at the (t-1)-th sampling time; initial value Take the average confidence level of the s-th type of state from the historical data; Step 4.2: Calculate the upper and lower limits of the dynamic threshold by combining the "confidence standard deviation" of multi-scale historical data. ; ;in, The location coefficients are normally distributed. Let be the standard deviation of the confidence level for the s-th class of states in the historical data; Step 4.3: Adjust the threshold fit at preset time intervals. Perform calculations; ; in, ( () is an indicator function; it returns 1 if the condition is met, and 0 otherwise. If the threshold E is less than 1, then recalculate. ; Step 4.4: Combine the comprehensive confidence level of the equipment operating status with the dynamic threshold to determine the equipment operating status and obtain the operating status determination result; The comprehensive confidence level of the equipment operating status at time t is determined by the following logic: For normal state, attention state, and abnormal state, the following conditions must be met. and For severe conditions, the following must be met: and ; If the overall confidence scores of multiple device operating states are close to the maximum, a weighted judgment based on historical similarity is used. The formula for calculating the weighted disambiguation value of historical similarity for the s-th type of state is: ; in, The average confidence level of the s-th type of state in historical data. This represents the degree of similarity between the current state confidence level and the historical average confidence level. The weighted disambiguation value for the historical similarity of the s-th class of states; Pick The highest state is used as the result of the operation status determination. Based on the result of the operation status determination, a graded warning is issued, and the warning level corresponds one-to-one with the equipment status.

[0017] The technical solution adopted by the system of the present invention is: a power transmission and transformation equipment operation status monitoring system, comprising: One or more processors; A storage device is provided for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the power transmission and transformation equipment operation status monitoring method.

[0018] Compared with the prior art, the beneficial effects of the present invention include: 1. This invention uses the Least Squares Support Vector Machine (LSSVM) model and a Gaussian kernel function to map the nonlinear relationships of parameters such as voltage, temperature, and partial discharge in power transmission and transformation equipment. It transforms the constrained optimization problem into a system of linear equations using the Lagrange multiplier method. Compared with linear models, it can more accurately capture the complex changes in the operating state of equipment under multi-parameter coupling. It uses the core features after secondary dimensionality reduction by kernel principal component analysis as input, which is both representative and low redundancy. It provides a reliable basic model support for equipment condition diagnosis and effectively solves the technical problem that simple models cannot cope with the nonlinear correlation of multiple parameters in power transmission and transformation. 2. This invention presents a least squares support vector machine model based on kernel principal component analysis to improve the sparrow search algorithm. It achieves dual optimization on the basis of LSSVM. Kernel principal component analysis is used to perform secondary dimensionality reduction on multi-scale features, selecting core discriminative features with a cumulative contribution rate ≥95%, and eliminating redundant information to reduce computational complexity. A Logistic chaos factor is introduced to enhance the diversity of the initial population. Adaptive weights are used to dynamically balance global and local optimization, and Cauchy mutation is employed to improve perturbation capability. The penalty factor and kernel parameters of LSSVM are automatically optimized, avoiding the subjectivity of manual parameter setting. This model is more suitable for multi-dimensional monitoring scenarios in power transmission and transformation, improving algorithm efficiency while ensuring the accuracy of state diagnosis, and compensating for the insufficient adaptability and optimization efficiency of the basic LSSVM model. 3. Traditional DS evidence theory constructs an identification framework based on four mutually exclusive states: "normal, attention, abnormal, and severe." It treats the prediction results of the three core monitoring dimensions—voltage, temperature, and partial discharge—as independent evidence. This invention achieves the fusion of multi-source state information through a basic probability allocation function. This method breaks the one-sidedness of single-parameter judgment, fully integrates the state representations of different monitoring dimensions, and provides a scientific fusion approach for comprehensively understanding the operating status of equipment, accurately adapting to the actual operation and maintenance needs of multi-parameter collaborative monitoring in power transmission and transformation systems. Attached Figure Description

[0019] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0020] Figure 1 This is a schematic diagram of the method flow for the present invention. Detailed Implementation

[0021] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] Please see Figure 1 This embodiment provides a method for monitoring the operating status of power transmission and transformation equipment, the method comprising: Step S100: Construct a training set based on the historical operating data of power transmission and transformation equipment; Step S200: Train the least squares support vector machine model based on the improved sparrow search algorithm according to the training set to obtain the trained least squares support vector machine model; Step S300: Collect real-time operating data of power transmission and transformation equipment, and generate a comprehensive confidence score of equipment operating status based on the real-time operating data and the trained least squares support vector machine model; Step S400: Perform graded early warning for power transmission and transformation equipment based on the comprehensive confidence level of the equipment's operating status.

[0023] In this embodiment, a training set is constructed based on historical operating data of power transmission and transformation equipment; a least squares support vector machine model based on an improved sparrow search algorithm is trained using the training set to obtain a trained least squares support vector machine model; real-time operating data of the power transmission and transformation equipment is collected, and a comprehensive confidence score of the equipment operating status is generated based on the real-time operating data and the trained least squares support vector machine model; graded early warning is performed on the power transmission and transformation equipment based on the comprehensive confidence score of the equipment operating status, and graded early warning is performed based on the operating status judgment results. The early warning level corresponds one-to-one with the equipment status, and the response measures for each level are clear and adapted to the actual needs of power transmission and transformation equipment operation and maintenance.

[0024] In one implementation, step S100 specifically includes: Step S110: Collect the operating data of the power transmission and transformation equipment and obtain historical operating data; Step S120: Preprocess the historical running data and generate preprocessed running data; Step S130: Construct a training set based on the preprocessed running data.

[0025] In step S110, multi-dimensional data collection is performed on the operation of power transmission and transformation equipment to obtain historical operation data. The historical operation data specifically includes three types of key equipment: high-voltage transmission lines, distribution transformers, and high-voltage switchgear.

[0026] For high-voltage transmission lines, the following data acquisition equipment and parameters are deployed: ZMCT118A type current transformer and ZMPT101B type voltage transformer were selected to collect the three-phase current of the line. , , With three-phase voltage , , The current transformer ratio is set to =500:5 (suitable for 10kV lines with rated current of 200~500A), the transformer ratio can be adjusted according to the actual line rated current. The recommended ratio range is 200:5~600:5. The voltage transformer ratio is set to... =3000:1 (suitable for 10kV line rated voltage), measuring range is 0~12kV; DS18B20 digital temperature sensor is selected and deployed in easily overheated parts such as line clamps and conductor joints. The sensor's standard measuring range is -55°C to +125°C, collecting temperature data. (Clip temperature) (Connector temperature).

[0027] A JY901S triaxial accelerometer was selected and installed near the cable clamp (within 50-100mm of the clamp edge, preferably in a vibration-sensitive area) to collect vertical acceleration. The GTMK-FS-ABS-V1 three-cup anemometer and the DHT11 temperature and humidity sensor were selected and deployed on the top of the power line tower to collect wind speed data. Ambient temperature Ambient humidity A GT-U7 GPS module was selected and deployed on the top of the tower to collect the device's location coordinates (Lon, Lat).

[0028] For distribution transformers, the following data acquisition equipment and parameters are added: An ultrasonic partial discharge sensor is selected and deployed on the side wall of the transformer tank to collect the partial discharge amplitude. and frequency A DS18B20 temperature sensor was selected and inserted into the transformer oil conservator to collect oil temperature data. .

[0029] For high-voltage switchgear, the following data acquisition equipment and parameters are added: a ground wave partial discharge sensor is selected and attached to the outside of the switchgear enclosure to collect the ground wave amplitude. Wireless temperature sensors are selected and deployed at the busbar contacts and switch contacts of the switch cabinet to collect the contact temperature. Contact temperature Temperatures exceeding 70°C indicate poor contact; select appropriate settings. Gas concentration sensor, with a measurement range of 0~1000ppm, deployed in Inside the insulated switchgear, data was collected. concentration , concentration A concentration of <500ppm indicates a gas leak, which can lead to a decrease in insulation performance.

[0030] Differentiated acquisition frequencies are adopted, and parameters are set according to type: high-frequency parameters (partial discharge amplitude / frequency, The concentration data is collected at a frequency of 1 minute to ensure the capture of instantaneous anomalies; the intermediate frequency parameters (voltage, current, equipment contact temperature) are collected every 5-10 minutes to balance accuracy and redundancy; and the low frequency parameters (ambient temperature and humidity, wind speed) are collected every 30 minutes to reduce invalid data.

[0031] It should be noted that the collected data, such as current, environmental parameters, vibration, SF concentration, and GPS data, are not redundant data, but are used to correct monitoring deviations of core parameters (e.g., correcting temperature data for environmental temperature and humidity) and supplement coverage of fault scenarios not covered by core dimensions (e.g., ...). Concentration monitoring of switchgear insulation leakage; supporting fault tracing and operation and maintenance collaboration (such as GPS location of regional faults), forming a data collection system with accurate early warning of core parameters and comprehensive verification of auxiliary parameters, improving the reliability and practicality of monitoring.

[0032] During monitoring, prediction and fusion are prioritized based on three core dimensions, with auxiliary data used for secondary verification when core parameters are abnormal; supplementary early warning for special equipment (such as SF insulated switchgear) and cause analysis and tracing after failure.

[0033] All data acquisition devices are connected to the STM32F103C8T6 main control chip via wired / wireless means. The main control chip uploads multi-dimensional data to the cloud platform using the MQTT protocol to ensure real-time and reliable data transmission.

[0034] Next, in steps S120-S130, the historical running data is preprocessed to generate preprocessed running data, and a training set is constructed based on the preprocessed running data to provide data support for the subsequent training of the least squares support vector machine model.

[0035] In one implementation, step S120 specifically includes: Step S121: Calculate the data integrity and data validity of the historical operation data; Step S122: Based on the data integrity and data validity, the historical running data is processed using the local outlier factor algorithm to remove outlier data and generate the first processed data; Step S123: Standardize the first processed data and generate preprocessed running data.

[0036] In this embodiment, the historical operational data undergoes quality screening, defining two core quality indicators: data integrity. Data validity .

[0037] Data integrity The formula for calculating data integrity, which reflects the degree of data missing, is as follows: ,in For the number of missing data for a single parameter, This represents the total number of samples for a single parameter. For data integrity, when When the value is ≥0.9, the data integrity is considered acceptable; if If the value is less than 0.9, the parameter should be marked and missing values ​​should be filled in first.

[0038] Data Validity This reflects whether the data exceeds the physical range of the equipment and the validity of the data. The calculation formula is: ,in, This represents invalid data volume for a single parameter that exceeds its range. For data validity, when When the value is ≥0.95, the data is considered valid; if Data with a value less than 0.95 should be removed first if it is invalid.

[0039] For "invalid data beyond the measurement range" and "abnormal fluctuation data within the normal measurement range" (such as sudden temperature rises and falls) identified during data validity screening, the Local Outlier Factor (LOF) algorithm is used for removal. The LOF algorithm execution process consists of three steps: neighborhood calculation → reachability density calculation → outlier factor calculation. The specific formulas and parameter definitions are as follows: The formula for calculating the k-neighborhood distance of data point P is: ; Where P is the data point to be evaluated. Let the k-th nearest neighbor of data point P be the k-th nearest neighbor. 3D eigenvalues Let k be the number of nearest neighbors, and k=20. Let k be the k-neighborhood distance of data point P. For data point P, the first Dimensional eigenvalues.

[0040] The formula for calculating the reachability distance from data point P to any data point O in its neighborhood is: ; in, Let be the Euclidean distance between P and O. Let k be the k-neighborhood distance of the neighborhood point O. This represents the reachability distance from data point P to any data point O within its neighborhood.

[0041] It should be noted that if P is in the k-neighborhood of O (i.e., ≤ Then the reachability distance is taken as the K-neighborhood distance of O, that is... To avoid density calculation errors caused by data points being too close together in the neighborhood; if P is outside the k-neighborhood of O, the reachability distance is taken as the Euclidean distance between P and O, i.e. This accurately reflects the distance between the two points.

[0042] Local reachability density of data point P in its k-neighborhood The calculation formula is: ; in, Let P be the set of k-neighborhood data points. Let be the number of elements in the set of k-neighborhood data points. Let P be the local reachability density of a data point P within its k-neighborhood. Indicates the size of the set.

[0043] The formula for calculating the local outlier factor of data point P is: ; in, Let P be the local outlier of the data point. Let O be the local reachability density of a data point O within its k-neighborhood.

[0044] when When the value is ≥1.5, P is determined to be an outlier and removed; when When P is less than 1.5, it is considered normal data and retained. This threshold value is determined based on the "normal fluctuation of transmission and transformation data". Mostly concentrated around 1, during abnormal fluctuations The statistical regularity is "significantly greater than 1".

[0045] For missing data discovered during data integrity screening (such as data gaps of 1-3 sampling periods due to sensor malfunction), the data is supplemented using the average value interpolation method before and after the data. The specific implementation is as follows: Let the time series of a certain parameter be... The completed formula is: ; in, The parameter value is the one valid sampling time preceding the missing time. This represents the parameter value at the next valid sampling time after the missing time. To fill in the missing parameter values.

[0046] To eliminate the dimensional differences between multi-dimensional parameters, the Z-Score standardization method is used to process the data, as follows: Preprocessed historical running data (For example, all voltage data for a certain line), the standardized formula is: ,in, The original data before standardization. The mean of the dataset. Let be the standard deviation of the dataset. Let i be the i-th data point after standardization.

[0047] To ensure that subsequent model outputs can be converted back to physical quantities (e.g., converting standardized temperature predictions to °C), it is necessary to store the values ​​of each parameter. and The reverse standardization formula is ,in, The standardized values ​​output by the model.

[0048] Even after standardization, multi-dimensional data still contains redundant features (such as a strong linear correlation between current and power). Therefore, kernel principal component analysis (KPCA) is needed for dimensionality reduction to preserve core features and reduce the computational complexity of subsequent models. KPCA achieves dimensionality reduction through three steps: kernel function mapping, covariance matrix calculation, and principal component extraction, as detailed below: Because there are nonlinear correlations among power transmission and transformation parameters (such as the coupling relationship between temperature and partial discharge), a Gaussian kernel function (RBF kernel) is chosen to map low-dimensional data to a high-dimensional feature space. The formulas for calculating the Gaussian kernel function values ​​of the i-th sample and the j-th sample are as follows: ; in, , For any two standardized data points, for and The Euclidean distance between two points The Gaussian kernel bandwidth parameter, with a value of 0.125, was determined through cross-validation. The feature discrimination is optimal when the value is 0.125. Let be the Gaussian kernel function values ​​of the i-th sample and the j-th sample.

[0049] Since the data mapped by the kernel function may be non-centralized, a centralized kernel matrix needs to be constructed. The formula for calculating the centralized kernel matrix is ​​as follows: ; ; ; ; in, For the number of data samples, For the elements of the centered kernel matrix, The values ​​of the Gaussian kernel for the i-th sample and the j-th sample are given. Let be the average of the kernel function values ​​of the i-th sample and all k samples in the dataset. The kernel function value of the j-th sample is the average of the kernel function values ​​of all k samples in the dataset. This is the average of the kernel function values ​​of all k samples and l samples in the dataset.

[0050] Solve for the eigenvalues ​​and corresponding eigenvectors of the centered kernel matrix, and select the eigenvalues ​​with a cumulative contribution rate ≥ 90%. The formula for calculating the cumulative contribution rate of each eigenvector as a principal component is as follows: ; in, The cumulative contribution rate of the main components, The feature dimensions after dimensionality reduction. The first of the centered kernel matrix Eigenvalues.

[0051] It should be noted that when the cumulative contribution rate is ≥90%, the feature after dimensionality reduction retains 91.2% of the information of the original data, and the dimension is reduced from 15 dimensions to 4 dimensions, balancing information preservation and computational efficiency. The final output of the feature after dimensionality reduction is a 4-dimensional core feature.

[0052] In one implementation, step S130 specifically includes: Step S131: Perform feature extraction based on the preprocessed operating data to obtain feature data of power transmission and transformation equipment; Step S132: Construct a training set based on the characteristic data of the power transmission and transformation equipment.

[0053] In this embodiment, feature extraction is performed on the preprocessed operational data to obtain the feature data of the power transmission and transformation equipment and construct a training set. Statistical features (mean, maximum value, and variance) are extracted at three time scales: hourly, daily, and weekly. The feature definitions for each scale are as follows: Using one hour as the unit (including six 10-minute sampling data points), extract features reflecting short-term operating status: hourly average. (Reflects the average level of parameters within an hour), Hourly maximum value (Reflects peak parameter values ​​within an hour), hourly variance (Reflecting the degree of parameter fluctuation within an hour); using one day as the unit (including 144 10-minute sampling data points), extract features reflecting the medium-term operating status: daily average value. Daily maximum value daily variance Using a weekly unit (including 7 days of data), extract features reflecting long-term operating status: weekly average. Weekly maximum value , weekly variance Hourly scale features are adapted for short-term fault warnings (such as short-term overload), daily scale features are adapted for daily condition assessment, and weekly scale features are adapted for long-term degradation trend analysis.

[0054] Multi-scale feature data is divided into training and test sets using stratified sampling and time-based cross-validation. Within each device state (normal / attention / abnormal / severe), the training and test sets are divided at a ratio of 9:1 to ensure that the test set contains samples from each state and that the ratio is consistent with the overall dataset. Additionally, 5-fold time-based cross-validation is added.

[0055] The dataset is divided into 5 time segments, with 4 segments used as the training set and 1 segment as the test set, to avoid evaluation bias caused by a single time segment. The test set should have a sample percentage for each state that deviates from the training set by ≤5%, and the number of severe state samples should be ≥230 (to ensure the model's ability to predict abnormal states is assessable).

[0056] In one implementation, a least squares support vector machine model based on an improved sparrow search algorithm is trained using the training set to obtain a trained least squares support vector machine model. Step S210: Perform secondary dimensionality reduction based on the training set and generate model input features; In one implementation, step S210 includes: Step S211: Perform secondary dimensionality reduction using kernel principal component analysis based on the training set; Step S212: Select core discriminant features with a cumulative contribution rate of not less than 95% as input features for the model.

[0057] Step S220: Use an improved sparrow search algorithm to globally optimize the penalty factor and kernel parameters; Step S230: Based on the optimal penalty factor and kernel parameters obtained through optimization, construct a least squares support vector machine model, and train the model using the input features and corresponding state labels to obtain a trained least squares support vector machine model.

[0058] In this embodiment, the original data has already been preprocessed and dimensionality reduced using KPCA. However, the multi-scale historical dataset still exhibits "cross-scale feature redundancy," requiring secondary dimensionality reduction using KPCA to extract core discriminative features. This ensures that the features input to the Least Squares Support Vector Machine (LSSVM) are both representative and have low redundancy, as detailed below: Based on the Gaussian kernel function from the previous step, calculate the kernel function values ​​for all sample pairs in the training set to form a multi-scale feature kernel matrix. Center the multi-scale feature kernel matrix to eliminate the influence of feature mean shift on principal component extraction. Solve for the eigenvalues ​​and eigenvectors of the centered kernel matrix, and select the first two eigenvectors with "cumulative contribution rate ≥ 95%" as the final input features.

[0059] Among them, "cumulative contribution rate ≥ 90%" is the stage of feature extraction and preliminary dimensionality reduction of the original historical operation data. 90% is the classic balance threshold of KPCA dimensionality reduction in the industry. When the cumulative contribution rate is ≥ 90%, more than 80% of redundant dimensions can be removed (from 15 dimensions to 4 dimensions), while retaining 91.2% of the effective information of the original data to avoid the loss of key state features. The "cumulative contribution rate ≥ 95%" is applied to the second KPCA dimensionality reduction in step S211. The accuracy of the LSSVM model is highly dependent on the purity of the input features. A cumulative contribution rate ≥ 95% can ensure that no core discriminative features are omitted, avoiding a decrease in the model's sensitivity to minor faults due to feature loss. Through verification on 1000 sets of historical fault data of power transmission and transformation equipment, when the cumulative contribution rate is ≥ 95%, the features after the second dimensionality reduction have a discrimination degree of more than 92% for the four states of "normal-attention-abnormal-severe", which is significantly higher than the discrimination degree of 90% cumulative contribution rate.

[0060] The diagnostic accuracy of the LSSVM model is highly dependent on its penalty factor c (balancing the classification margin and misclassification rate) and kernel parameters. (Controlling the local scope of the Gaussian kernel) Traditional methods (such as grid search) are prone to getting trapped in local optima and are inefficient.

[0061] This step uses the Improved Sparrow Search Algorithm (ISSA) to adjust the penalty factor c and kernel parameters. Global optimization is achieved through three major improvements: Logistic chaotic factor initialization, adaptive weights, and Cauchy mutation, which solve the shortcomings of traditional sparrow search algorithm such as slow convergence and susceptibility to local optima.

[0062] Traditional sparrow search algorithms initialize the population randomly, which can easily lead to uneven distribution. This study introduces a Logistic chaos factor to generate the initial population, enhancing population diversity. The population size is 30, and the maximum number of iterations is 100. The formula for generating the initial population is: ; in, Let be the value of the k-th chaotic iteration. This represents the value in the (k+1)th chaotic iteration. The chaos factor is the Logistic regression factor, and its value is... =4, When the coefficient is 4, the uniformity of the Logistic chaotic sequence reaches 0.93, which is better than other values.

[0063] The discoverer is responsible for exploring the optimal parameter region. Traditional SSA position updates lack a dynamic adjustment mechanism. This case introduces adaptive weights. In the early stages of iteration, large weights enhance global search, while in the later stages, small weights enhance local optimization. The formula is as follows: ; ; in, This represents the current iteration number. The maximum number of iterations, For the first The dim-th dimension parameter of the i-th sparrow individual in the next iteration. To simulate the random probability of environmental alert for sparrow populations, The safety threshold is set to a value of [value to be filled in]. =0.8, For random numbers that follow a normal distribution, It is a matrix of all ones. For adaptive weights, For the first The dim dimension parameter of the i-th individual sparrow in the next iteration.

[0064] The vigilant is responsible for avoiding local optima. This case introduces Cauchy mutation to enhance perturbation capability, with the formula as follows: ; in, For the first The globally optimal parameters for the next iteration. These are standard Cauchy distribution random numbers. For the first The dim dimension parameter of the i-th individual sparrow in the next iteration.

[0065] It should be noted that the significance of this formula lies in the fact that the heavy-tailed characteristic of the Cauchy distribution can generate large perturbations, helping the observer escape local optima. When the parameters are trapped in a local optimum, c=10, When c=0.5", Cauchy mutation can guide the parameter towards c=20. Explore regions such as =0.3 to improve global optimization accuracy.

[0066] The optimization objective of SSA is the classification error rate of LSSVM on the training set. The fitness function is calculated using the following formula: ; in, For LSSVM in parameter combination (c, The number of correctly classified samples under ) For fitness value, This represents the number of samples in the training set.

[0067] It should be noted that the smaller the fitness function value, the smaller the training error of LSSVM under that parameter combination, and the better the parameters. ISSA iteratively updates the positions of individual sparrows and finally outputs the parameter combination with the minimum fitness. =[ , ], which serves as the optimal parameter for LSSVM.

[0068] Optimal parameters obtained based on ISSA optimization The key configurations for building an LSSVM state diagnostic model are as follows: The formula for calculating the kernel function of LSSVM is: ; in, and These are the k-th and l-th eigenvectors after the second dimensionality reduction using KPCA, respectively. LSSVM kernel parameters optimized for ISSA This is the kernel function for LSSVM.

[0069] The goal of LSSVM is to minimize the regularization loss function. The formula for calculating the loss function of LSSVM is: ; ; in, The weight vector in the feature space of LSSVM. For LSSVM bias terms, Let the fitting error be the value of the k-th sample. The feature map after KPCA dimensionality reduction. To improve the penalty factor in the sparrow algorithm optimization, The number of samples in the training set. Indicates constraints. Let the true state label be the one corresponding to the k-th training sample. This is the k-th eigenvector after the second dimensionality reduction by the kernel principal component analysis algorithm.

[0070] The constrained optimization problem is transformed into a system of linear equations by using the Lagrange multiplier method, and the final formula for calculating the prediction function of LSSVM is as follows: ; in, The predicted value output by the model. This is the kernel function for LSSVM. For Lagrange multipliers, represents the weight of the contribution of the k-th training sample to the prediction result. is the feature vector of the sample to be predicted.

[0071] The feature vector of the training set after KPCA second dimensionality reduction With the real label Input LSSVM, import ISSA optimized and By using the Lagrange multiplier method, the constrained optimization problem is transformed into a system of linear equations, and the Lagrange multipliers are solved directly. and bias terms , It is possible The derivation, without the need for separate optimization, yields the fully trained LSSVM model.

[0072] In one implementation, step S300 specifically includes: Step S310: Collect real-time operating data of power transmission and transformation equipment, input the real-time operating data into the trained least squares support vector machine model, and output the equipment operating status prediction result. Step S320: The prediction results of the equipment operating status are fused using the improved DS evidence theory method based on Euclidean distance to obtain the comprehensive confidence level of the equipment operating status.

[0073] In this embodiment, three core monitoring dimensions are set (m=1: voltage, m=2: temperature, m=3: partial discharge). For each monitoring dimension, multi-scale features (mean, maximum, variance, etc.) are extracted according to the method in S130. After KPCA secondary dimensionality reduction in step S211 (retaining 2 core feature vectors), the data is input into the LSSVM model. The model performs state prediction for each monitoring dimension separately and outputs the prediction result for the corresponding dimension. Therefore, the value of m is 1-3, which is a progressive relationship with the feature vector dimensions (2) mentioned above, which is monitoring dimension-feature extraction-dimensionality reduction input. The two do not conflict with each other.

[0074] Considering that more than 80% of the critical faults of power transmission and transformation equipment are directly and strongly correlated with the three dimensions in this embodiment, voltage directly reflects the grid load and the insulation status of the equipment, temperature is related to thermal degradation faults such as poor conductor contact and heat dissipation failure, and partial discharge is an early core manifestation of insulation damage. The three dimensions cover the key risk points of equipment operation from the three core dimensions of electrical, thermal and insulation.

[0075] The current, SF6 concentration, and GPS data collected in the S100 are not redundant, but rather serve as a supplement to the "core dimension + auxiliary verification" system. Ambient temperature and humidity are used to correct for temperature data deviations, SF6 concentration is specifically used to monitor switchgear insulation leakage, GPS is used for fault area location, and vibration data assists in tracing mechanical faults. These data initiate secondary verification when the core dimension is abnormal, and support cause analysis after a fault occurs. This ensures both the accuracy of the core judgment and the comprehensiveness of monitoring without affecting it. Real-time operating data of power transmission and transformation equipment is collected. This real-time operating data (such as voltage, temperature, and partial discharge signals at a certain moment) has problems such as "real-time noise (such as voltage fluctuations caused by electromagnetic interference) and short-term missing data (such as instantaneous sensor disconnection)". The Local Outlier Factor (LOF) algorithm is used to remove outliers (such as instantaneous voltage surges caused by lightning strikes). The core formula is the same as the previous steps to avoid subsequent prediction deviations caused by instantaneous interference. The mean interpolation method of the previous and next time moments is used to fill in missing values. If the missing data exceeds 3 cycles (>30 minutes), it is marked as "invalid data".

[0076] It should be noted that the three core monitoring dimensions (voltage, temperature, and partial discharge) are key parameters selected from the multi-dimensional data collected in step S100 based on fault sensitivity, data reliability, and engineering practicality. Voltage is related to grid load and insulation status, temperature is related to thermal degradation faults, and partial discharge is related to insulation damage faults. The three cover more than 80% of the core fault scenarios of power transmission and transformation equipment.

[0077] For each monitoring dimension m, the LSSVM model outputs four raw prediction scores. , , , ;), corresponding to the decision values ​​for four states (s=0: normal, s=1: attention, s=2: abnormal, s=3: severe), the raw scores are converted into state confidence scores using the softmax function. The s corresponding to the maximum confidence level is the initial state label for that monitoring dimension, and the formula for calculating the state confidence level is: ; in, Let LSSVM be the prediction score for the m-th monitoring dimension and the s-th state. For the corresponding State confidence.

[0078] It should be noted that when the monitoring dimension m=1, it is voltage; when m=2, it is temperature; and when m=3, it is partial discharge. The status labels are 0 (normal), 1 (caution), 2 (abnormal), and 3 (serious), respectively.

[0079] The state prediction results for each monitoring dimension are compiled into evidence bodies, with each evidence body being a confidence vector for that dimension, in the following format: ,in, This is the evidence body for the m-th monitoring dimension.

[0080] The prediction results from monitoring dimensions are easily affected by interference (such as electromagnetic interference affecting voltage data and ambient temperature fluctuations affecting temperature data), leading to biased results. Furthermore, prediction results from different monitoring dimensions may conflict, and directly adding them together reduces the reliability of status determination. Therefore, it is necessary to integrate these predictions using an improved DS evidence theory method based on Euclidean distance: Euclidean distance is used to quantify the degree of conflict between evidence, avoiding the bias of traditional DS relying solely on the conflict coefficient K; the weights of highly conflicting evidence are dynamically adjusted to enhance the influence of reliable evidence; and multi-dimensional confidence levels are integrated to obtain a comprehensive confidence level, fully reflecting the equipment's operating status and improving the accuracy and robustness of the determination.

[0081] This step fuses the equipment operating status prediction results using an improved DS evidence theory method based on Euclidean distance to obtain the comprehensive confidence level of the equipment operating status. The specific steps are as follows: First, define the identification framework for the status of power transmission and transformation equipment (the target set for fusion): ,in, (s=0~3) represent mutually exclusive state categories, and the evidence for each monitoring dimension corresponds to the Basic Probability Assignment (BPA) function of the DS theory. , For the m-th piece of evidence, the first piece of evidence is... Basic probability assignment values ​​for class states; The Euclidean distance is used to calculate the degree of conflict between pieces of evidence, avoiding the one-sidedness of traditional DS which relies solely on the conflict coefficient K. The formula for calculating the Euclidean distance between the m-th and n-th pieces of evidence is as follows: ; in, Let be the Euclidean distance between the m-th and n-th pieces of evidence. For the m-th piece of evidence, the first piece of evidence is... The basic probability assignment values ​​for class states, For the nth piece of evidence, the first piece of evidence is... The basic probability assignment values ​​for class states, This refers to the device status category.

[0082] The conflict threshold is set to =0.8, when A value ≥0.8 indicates "moderate to high conflict," requiring adjustment of the evidence weights; when A value less than 0.8 indicates "low conflict," and direct fusion is performed.

[0083] For evidence with moderate to high conflict, the weights are adjusted through a process of "similarity, support, and weight" to enhance the influence of reliable evidence. The specific steps are as follows: The formula for calculating the similarity between the m-th piece of evidence and the n-th piece of evidence is: ,in, Let m be the similarity between the m-th and n-th pieces of evidence, and let m be the support of the m-th piece of evidence. The support score of the m-th piece of evidence is calculated as the sum of the similarities between all other pieces of evidence and this evidence. ,in, Let m be the support level of the m-th piece of evidence. This represents the total number of dimensions being monitored.

[0084] The evidence weight is the normalized result of the support. The formula for calculating the evidence weight of the m-th piece of evidence is: ; in, Let m be the evidence weight of the m-th piece of evidence. Let m be the support level of the m-th piece of evidence.

[0085] The adjusted evidence weights are incorporated into the DS fusion, and the resulting fusion yields the overall confidence level of the equipment operating status. The formula for calculating the overall confidence level of the equipment operating status is as follows:

[0086] in, The overall confidence level of the equipment's operating status. To improve the conflict coefficient, the degree of conflict between pieces of evidence is quantified. For the first The weight of evidence for each piece of evidence For the first One piece of evidence for the state The basic probability allocation value.

[0087] The meaning refers to the state of each piece of evidence (3 core dimensions) as determined individually. The intersection of (such as voltage determination results, temperature determination results, and partial discharge determination results) equals the target state. To calculate "anomalies" ( The overall confidence level of the summation requires that the intersection of the judgment results in the voltage dimension, the temperature dimension, and the partial discharge dimension must be "abnormal". In other words, only when the judgment results of the three dimensions ultimately point to an "abnormal" state will the result be included in the summation range. The core function is to filter out the state combinations of all evidence that are "consensus-abiding," avoiding the distortion of the fusion result due to the judgment deviation of some evidence, and ensuring that the overall confidence level only reflects the state supported by multiple dimensions.

[0088] The meaning is to multiply the "reliability weight" of each piece of evidence by its "confidence in its own state" to obtain the "weighted contribution value" of that piece of evidence to the target state, and then calculate the joint contribution value of all pieces of evidence by multiplying them together.

[0089] It should be noted that the improved conflict coefficient is the sum of conflict terms after weighting all evidence weights, and the calculation formula is as follows: .

[0090] The meaning is to calculate the "weighted state support" of a single piece of evidence, which combines the reliability (weight) of the evidence with the evidence's support for a certain state (BPA) to obtain the "actual effective support" of the evidence for this state in the fusion.

[0091] It represents the product of the weighted support of the three core dimensions (voltage, temperature, and partial discharge) for their respective supported states, and represents the joint effective support of these three pieces of evidence for their respective states.

[0092] The meaning is that the three pieces of evidence have no consensus on the judgment status and no common point to the same state. Only this situation is considered a "real conflict". Only combinations that meet this condition will be included in the summation range.

[0093] Conversely, if the three pieces of evidence determine the same state and are not in conflict, they are not included in the summation. middle It does not represent a specific numerical value, but is an abbreviation for "the i-th evidence body," used to traverse all evidence participating in the fusion. Essentially, it clarifies that "the summation range covers all core monitoring dimensions of the evidence body," ensuring that the state prediction results of each core dimension participate in conflict calculation or multi-source fusion, without omitting any key monitoring information.

[0094] The difference between the highest and second-highest comprehensive confidence levels of the equipment's operating status must be ≥0.2 to ensure that the status determination is unambiguous.

[0095] In one implementation, step S400 specifically includes: Step S410: Based on the comprehensive confidence level of the device's operating status, construct a dynamic threshold using an exponentially weighted moving average algorithm; In one implementation, step S410 includes: Step S411: Obtain EWMA weight coefficients; Step S412: Based on the EWMA weighting coefficients and the comprehensive confidence level of the equipment operating status, construct a dynamic threshold using an exponentially weighted moving average algorithm.

[0096] Step S420: Combine the comprehensive confidence level of the device operating status and the dynamic threshold to determine the device operating status and obtain the operating status determination result; Step S430: Perform graded early warning for power transmission and transformation equipment based on the operation status determination results.

[0097] In this embodiment, the overall confidence level obtained above is a static value, which does not take into account the dynamic fluctuations of the operating conditions of power transmission and transformation equipment (such as seasonal changes and equipment aging causing the confidence level distribution of normal state to shift). Judging based solely on static confidence level is prone to misjudgment; when the difference between the highest and second highest overall confidence level is <0.2, it is difficult to directly determine the state.

[0098] Therefore, a second judgment is required: the EWMA algorithm is used to construct a dynamic threshold, and the dynamic changes of the adaptive working conditions are updated in real time through a sliding window; historical similarity weighted disambiguation is introduced to solve the judgment problem when the confidence levels are close; and the combination of "comprehensive confidence + dynamic threshold + historical similarity" to achieve multi-condition judgment further improves the accuracy and adaptability of state judgment and avoids the shortcomings of single static judgment.

[0099] The core of S300 is "multi-source evidence fusion", which outputs a comprehensive confidence level (basic input) that is both comprehensive and reliable. The core of S400 is dynamic adaptation and disambiguation, which constructs dynamic thresholds based on the comprehensive confidence level and supplements historical data for reference, and outputs the final operation status judgment result. The two form a progressive relationship of "fusion-optimized judgment".

[0100] Based on the comprehensive confidence level of the equipment operating status in step S300, a dynamic threshold is constructed using the Exponential Weighted Moving Average (EWMA) algorithm. The threshold is updated in real time through a sliding window to adapt to fluctuations in equipment operating conditions. The dynamic threshold is divided into "status-level thresholds" (corresponding to four status categories: "normal / attention / abnormal / severe"). The specific steps are as follows: The EWMA algorithm balances "historical patterns" and "real-time changes" by assigning different weights to historical and current data. The formula for calculating the EWMA statistic for the s-th state at the t-th sampling time is as follows: ; in, Let be the EWMA statistic corresponding to the s-th state at the t-th sampling time. Let be the combined confidence score of the s-th state after fusion at time t. These are EWMA weighting coefficients, with values ​​ranging from [value missing]. =0.1, Let be the EWMA statistic for the s-th state at the (t-1)-th sampling time.

[0101] It should be noted that the initial value Take the average confidence level of the s-th state from the historical data mentioned above.

[0102] To avoid a single statistic If the range of fluctuations cannot be fully covered, the upper and lower limits of the dynamic threshold are calculated by combining the "confidence standard deviation" of multi-scale historical data. The formula is as follows: ; in, Let be the upper limit of the threshold for the s-th state at time t. Let be the lower limit of the threshold for the s-th state at time t. The location coefficient is a normal distribution, with a value of z = 1.96. denoted as the confidence standard deviation of the s-th state in the historical data.

[0103] To ensure that the dynamic threshold does not deviate from the actual operating pattern of the equipment, the threshold is verified every 24 hours (144 sampling times) based on historical data, using the following formula: ; in, ( () is an indicator function; it returns 1 if the condition is met, and 0 otherwise. For threshold adaptability.

[0104] Require ≥0.95, meaning that the confidence level for normal operating conditions must fall within the threshold range of 95% or more; if If <0.95, recalculate. (e.g., removing outliers from historical data) to avoid threshold drift.

[0105] The device operating status is determined by combining the overall confidence level of the device operating status and the dynamic threshold, resulting in an operating status determination result. The current device status is determined using a dual-condition approach of maximum confidence priority plus threshold constraint. The specific steps are as follows: The comprehensive confidence level of the equipment operating status at time t is determined by the following logic: For normal state, attention state, and abnormal state, the following conditions must be met. and For severe conditions, the following must be met: and .

[0106] If the overall confidence scores of multiple device operating states are close to the maximum, a weighted judgment based on historical similarity is used. The formula for calculating the weighted disambiguation value of historical similarity for the s-th type of state is: ; in, The average confidence level of the s-th type of state in historical data. The value represents the similarity between the current state confidence level and the historical average confidence level. It is the weighted disambiguation value for the historical similarity of the s-th state.

[0107] Pick The highest possible state is used as the operational status determination result. Based on the operational status determination result, graded early warnings are issued, with each warning level corresponding to a specific equipment status. The response measures for each level are clearly defined and adapted to the actual needs of power transmission and transformation equipment operation and maintenance. Level 1 warning (corresponding to normal status): No proactive warning, only status data is recorded; Level 2 warning (corresponding to alert status): Local audio and visual alerts + SMS notification from maintenance personnel; Level 3 warning (corresponding to abnormal status): Operation and maintenance platform dispatch + on-site investigation within 24 hours; Level 4 warning (corresponding to severe status): Emergency shutdown order + emergency response within 1 hour. The effectiveness of early warning records is statistically analyzed monthly. The formula for calculating the effectiveness of early warnings is as follows: ; in, The number of times the warning results are consistent with the on-site inspection. Total number of warnings To ensure the effectiveness of early warning.

[0108] Require ≥90%; if If the accuracy is less than 90%, then the DS fusion weights in step S4 should be re-optimized (e.g., adjusting the Euclidean distance threshold) to improve the accuracy of the early warning.

[0109] To verify the effectiveness and practicality of this invention, a field experiment was conducted. Sixty high-voltage switchgear, distribution transformers, and high-voltage transmission lines in a regional power grid were selected as monitoring objects. Data acquisition and processing were strictly performed according to the method described in this invention: First, multiple sensors were deployed to continuously collect over 5 million real-time operational data points, including voltage, temperature, and partial discharge. After data quality screening (completeness Qcomp ≥ 0.9, validity Qvali ≥ 0.95) and outlier removal using a local outlier factor algorithm, secondary dimensionality reduction (cumulative contribution) was performed using kernel principal component analysis (KPCA). The model is trained based on historical data. The core feature vector is formed by selecting models with a contribution rate ≥ 95%. Then, an improved Sparrow Search Algorithm (ISSA) is used to optimize the penalty factor and kernel parameters of the Least Squares Support Vector Machine (LSSVM) (after optimization, the fitness value fissa is reduced to below 0.05). During the testing phase, real-time data is input into the model to obtain the state confidence scores for each dimension. These scores are then fused using an improved DS evidence theory based on Euclidean distance (conflict threshold dth = 0.8). Finally, the state is determined by combining the exponentially weighted moving average (EWMA) dynamic threshold (0.1). Experimental results show that the overall classification accuracy of this method for equipment states (normal, alert, abnormal, severe) reaches 96.5%, and the early warning effectiveness VALID index reaches 93.2% (higher than the preset requirement of 90%), significantly outperforming the traditional multi-parameter voting method (accuracy only 82%). Furthermore, it successfully provided early warning for three potential insulation faults, demonstrating its superiority in multi-parameter collaborative monitoring and dynamic adaptation.

[0110] In summary, this application utilizes the Least Squares Support Vector Machine (LSSVM) model and employs a Gaussian kernel function to map the nonlinear relationships between parameters such as voltage, temperature, and partial discharge in power transmission and transformation equipment. By using the Lagrange multiplier method, the constrained optimization problem is transformed into a system of linear equations for solution. Compared to linear models, this approach can more accurately capture the complex changing patterns of equipment operating states under multi-parameter coupling. Using the core features obtained from secondary dimensionality reduction through kernel principal component analysis as input, it combines representativeness with low redundancy, providing a reliable basic model support for equipment condition diagnosis. This effectively solves the technical challenge of simple models being unable to cope with the nonlinear correlation of multiple parameters in power transmission and transformation.

[0111] This least squares support vector machine (LSSVM) model, based on kernel principal component analysis (KPCA) to improve the sparrow search algorithm, achieves dual optimization on the basis of LSSVM. KPCA is used to perform secondary dimensionality reduction on multi-scale features, selecting core discriminative features with a cumulative contribution rate ≥95% and eliminating redundant information to reduce computational complexity. A Logistic chaos factor is introduced to enhance the diversity of the initial population, and adaptive weights are used to dynamically balance global and local optimization. Cauchy mutation is added to improve perturbation capability, and the penalty factor and kernel parameters of LSSVM are automatically optimized, avoiding the subjectivity of manual parameter setting. This model is more suitable for multi-dimensional monitoring scenarios in power transmission and transformation, improving algorithm efficiency while ensuring the accuracy of state diagnosis, and compensating for the insufficient adaptability and optimization efficiency of the basic LSSVM model.

[0112] Traditional DS evidence theory constructs an identification framework based on four mutually exclusive states: "normal, attention, abnormal, and severe." It treats the prediction results of three core monitoring dimensions—voltage, temperature, and partial discharge—as independent evidence and achieves the fusion of multi-source state information through a basic probability allocation function. This method breaks the one-sidedness of single-parameter judgment, fully integrates the state representations of different monitoring dimensions, provides a scientific fusion approach for comprehensively understanding the operating status of equipment, and accurately adapts to the actual operation and maintenance needs of multi-parameter collaborative monitoring in power transmission and transformation systems.

[0113] Finally, the improved DS evidence theory based on Euclidean distance is an optimization and upgrade of the traditional theory. It quantifies the degree of conflict between evidence bodies through Euclidean distance, avoiding the one-sidedness of traditional methods that rely solely on the conflict coefficient. For moderate to high conflict evidence, the evidence weight is dynamically adjusted to enhance the influence weight of reliable evidence. Then, the DS fusion logic is integrated to obtain the comprehensive confidence level. This improved scheme effectively solves the problem of distortion in the traditional DS theory when predicting conflicts with multiple parameters, making the fusion results more consistent with the actual operating status of the equipment, further improving the reliability of power transmission and transformation equipment status determination, and perfecting the technical logic of multi-source information fusion.

[0114] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0115] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for monitoring the operating status of power transmission and transformation equipment, characterized in that, Includes the following steps: Step 1: Collect real-time operating data of power transmission and transformation equipment and perform standardized preprocessing. The power transmission and transformation equipment includes high-voltage transmission lines, distribution transformers, and high-voltage switchgear. Step 2: Based on the preprocessed data, feature extraction is performed to obtain the feature data of the power transmission and transformation equipment. Using the pre-trained least squares support vector machine model, the prediction results of the equipment operating status are obtained. The pre-trained least squares support vector machine model employs an improved sparrow search algorithm to adjust the penalty factor during training. With kernel parameters The improved sparrow search algorithm performs global optimization by introducing a Logistic chaos factor to generate an initial population and enhance population diversity, and introduces adaptive weights. The position is updated by introducing Cauchy mutation to help the watcher escape local optima. The final output is the parameter combination with the minimum fitness value by iteratively updating the position of individual sparrows. Step 3: Convert the equipment operating status prediction results into state confidence scores. The improved DS evidence theory method based on Euclidean distance is used to assess the confidence level of the state. The data is then integrated to obtain the overall confidence level of the equipment's operating status. Step 4: Based on the comprehensive confidence level of the equipment's operating status, construct a dynamic threshold using an exponentially weighted moving average algorithm to conduct graded early warning for power transmission and transformation equipment.

2. The method for monitoring the operating status of power transmission and transformation equipment according to claim 1, characterized in that: In step 2, the training process of the trained least squares support vector machine model includes the following sub-steps: Step 2.1: First, collect historical operating data of power transmission and transformation equipment, and calculate the data integrity and validity of the historical operating data; then, remove abnormal data from the historical operating data; next, fill in the missing data found in the data integrity calculation using the average value of the previous and subsequent data; finally, perform standardization and dimensionality reduction processing to obtain the training dataset. Step 2.2: Use an improved sparrow search algorithm to globally optimize the penalty factor and kernel parameters; Step 2.3: Based on the optimal penalty factor and kernel parameters obtained through optimization, construct a least squares support vector machine model, and use the input features and corresponding state labels of the least squares support vector machine model for training to obtain a trained least squares support vector machine model.

3. The method for monitoring the operating status of power transmission and transformation equipment according to claim 2, characterized in that: In step 2.1, the data integrity... Reflecting the degree of data missing, ,in For the number of missing data for a single parameter, This represents the total number of samples for a single parameter; when When the data integrity is ≥ threshold A, the data is deemed to be complete. The validity of the data This reflects whether the data exceeds the physical range of the equipment. ,in, For invalid data volume of a single parameter exceeding the measurement range, when When the value is greater than or equal to threshold B, the data is deemed valid. The local outlier factor algorithm is used to analyze the historical operational data. First, it is based on the k-neighborhood distance of data point P. Calculate the reachability distance from data point P to any data point O in its neighborhood. Then calculate the local reachability density of data point P in its k-neighborhood. The local outlier factor of the most recently calculated data point P. ,when When the value is greater than or equal to the threshold C, P is determined to be abnormal data and removed. Kernel principal component analysis (KPCA) is used to reduce the dimensionality of the data. First, a Gaussian kernel function is used to map the low-dimensional data to a high-dimensional feature space. Then, a centralized kernel matrix is ​​constructed, and the eigenvalues ​​and corresponding eigenvectors of the centralized kernel matrix are calculated. The top eigenvalues ​​are selected based on the eigenvalues ​​and eigenvectors of the kernel, with the cumulative contribution rate ≥ threshold D. The eigenvectors are used as principal components to reduce the dimensionality of the data; among them... This is the default value.

4. The method for monitoring the operating status of power transmission and transformation equipment according to claim 1, characterized in that: In step 2, the formula for generating the initial population is: ;in, Let be the value of the k-th chaotic iteration. This represents the value in the (k+1)th chaotic iteration. For Logistic chaos factors; The adaptive weight Then the first The dim-dimensional parameter of the i-th individual sparrow in the next iteration ;in, This represents the current iteration number. The maximum number of iterations, For the first The dim-th dimension parameter of the i-th sparrow individual in the next iteration. To simulate the random probability of environmental alert for sparrow populations, As a safety threshold, For random numbers that follow a normal distribution, It is a matrix of all 1s.

5. The method for monitoring the operating status of power transmission and transformation equipment according to claim 1, characterized in that: In step 2, the introduction of Cauchy mutations to help the vigilant escape local optima is described in the first step. The dim-dimensional parameter of the i-th individual sparrow in the next iteration ;in, For the first The globally optimal parameters for the next iteration. These are standard Cauchy distribution random numbers. For the first The dim dimension parameter of the i-th individual sparrow in the next iteration.

6. The method for monitoring the operating status of power transmission and transformation equipment according to claim 1, characterized in that: In step 2, the fitness value ;in, For the least squares support vector machine model in parameter combination (c, The number of correctly classified samples under ) The number of samples in the training dataset; The kernel function of the least squares support vector machine model ;in, and These are the k-th and l-th eigenvectors after the second dimensionality reduction by the kernel principal component analysis algorithm, respectively. For the optimized least squares support vector machine model kernel parameters; The objective of the least squares support vector machine model is to minimize the regularization loss function, which is calculated using the following formula: ； ； in, This represents the weight vector in the feature space of the least squares support vector machine model. This represents the bias term of the least squares support vector machine model. Let the fitting error be the value of the k-th sample. The feature map is the result of the second dimensionality reduction algorithm of kernel principal component analysis. To improve the penalty factor in the sparrow algorithm optimization, The number of samples in the training dataset. Indicates constraints. Let the true state label be the one corresponding to the k-th training sample. This is the k-th eigenvector after the second dimensionality reduction by the kernel principal component analysis algorithm; The constrained optimization problem is transformed into a system of linear equations to be solved using the Lagrange multiplier method, ultimately yielding the prediction function of the least squares support vector machine model: ； in, The predicted value output by the model. This is the kernel function for the least squares support vector machine model. is the feature vector of the sample to be predicted; For Lagrange multipliers, the weights of the k-th training sample in the prediction result are represented.

7. The method for monitoring the operating status of power transmission and transformation equipment according to claim 1, characterized in that: In step 3, the least squares support vector machine model outputs four raw prediction scores for each monitoring dimension m. , , , The values ​​represent the decision values ​​for four states s: s=0 for normal, s=1 for attentive, s=2 for abnormal, and s=3 for severe. The raw scores are then converted into state confidence scores using the softmax function. ,in, This represents the prediction score of the least squares support vector machine model for the m-th monitoring dimension and the s-th state. For the corresponding State confidence; organize the state confidence of each monitoring dimension into evidence bodies, each evidence body being a confidence vector for that dimension, and the evidence body for the m-th monitoring dimension. ; Define the set of fusion targets for the status of power transmission and transformation equipment: = ,in, For mutually exclusive state categories, the evidence bodies of each monitoring dimension correspond to the basic probability allocation function of DS theory. , For the m-th piece of evidence, the first piece of evidence is... Basic probability assignment values ​​for class states; The degree of conflict between pieces of evidence is calculated using Euclidean distance. The formula for calculating the Euclidean distance between the m-th and n-th pieces of evidence is as follows: ； in, Let be the Euclidean distance between the m-th and n-th pieces of evidence. For the m-th piece of evidence, the first piece of evidence is... The basic probability assignment values ​​for class states, For the nth piece of evidence, the first piece of evidence is... The basic probability assignment values ​​for class states, For device status category; when ≥ Conflict Threshold When it is determined to be "moderate to high conflict", the weight of evidence is adjusted; when < At that time, it was a period of "low conflict," and direct integration occurred.

8. The method for monitoring the operating status of power transmission and transformation equipment according to claim 7, characterized in that: For evidence with "moderate to high conflict", the specific steps for adjusting the weight of evidence are as follows: Calculate the similarity between the m-th piece of evidence and the n-th piece of evidence. Calculate the support of the m-th piece of evidence. The sum of the similarities between all other pieces of evidence and this evidence; where, Let m be the total number of monitoring dimensions; then the evidence weight of the m-th piece of evidence is... The adjusted evidence weights are incorporated into the DS fusion, and the resulting fusion yields the overall confidence level of the device's operating status. ； in, The improved conflict coefficient quantifies the degree of conflict between pieces of evidence; For the first The weight of evidence for each piece of evidence For the first One piece of evidence for the state The basic probability allocation value; This indicates the status of each piece of evidence in its respective judgment. The intersection equals the target state. ; This means multiplying the "reliability weight" of each piece of evidence by its "confidence level of its own state" to obtain the "weighted contribution value" of that piece of evidence to the target state, and then multiplying them together to calculate the joint contribution value of all pieces of evidence; ; This means calculating the "weighted state support" of a single piece of evidence, which combines the reliability of the evidence with its support for a certain state to obtain the "actual effective support" of the evidence for that state in the fusion process; This represents the product of the "weighted support" of each core dimension with respect to the state it supports. A "true conflict" is defined as a situation where there is no consensus on the determination of the evidence and no common direction. Only combinations of such states will be included in the summation.

9. The method for monitoring the operating status of power transmission and transformation equipment according to any one of claims 1-8, characterized in that: In step 4, an exponentially weighted moving average algorithm is used to construct a dynamic threshold. The threshold is updated in real time through a sliding window to adapt to fluctuations in equipment operating conditions. The dynamic threshold is divided into "state-level thresholds", corresponding to four states: "normal / attention / abnormal / serious". The specific implementation includes the following sub-steps: Step 4.1: Employ the Exponentially Weighted Moving Average (EWMA) algorithm, assigning different weights to historical and current data to balance "historical patterns" and "real-time changes"; where the EWMA statistic for the s-th state corresponds to the t-th sampling time. ; Let be the combined confidence score of the s-th state after fusion at time t. These are the EWMA weighting coefficients. Let be the EWMA statistic for the s-th state at the (t-1)-th sampling time; initial value Take the average confidence level of the s-th type of state from the historical data; Step 4.2: Calculate the upper and lower limits of the dynamic threshold by combining the "confidence standard deviation" of multi-scale historical data. ; ;in, The location coefficients are normally distributed. Let be the standard deviation of the confidence level for the s-th class of states in the historical data; Step 4.3: Adjust the threshold fit at preset time intervals. Perform calculations; ； in, ( () is an indicator function; it returns 1 if the condition is met, and 0 otherwise. If the threshold E is less than 1, then recalculate. ; Step 4.4: Combine the comprehensive confidence level of the equipment operating status with the dynamic threshold to determine the equipment operating status and obtain the operating status determination result; The comprehensive confidence level of the equipment operating status at time t is determined by the following logic: For normal state, attention state, and abnormal state, the following conditions must be met. and For severe conditions, the following must be met: and ; If the overall confidence scores of multiple device operating states are close to the maximum, a weighted judgment based on historical similarity is used. The formula for calculating the weighted disambiguation value of historical similarity for the s-th type of state is: ； in, The average confidence level of the s-th type of state in historical data. This represents the degree of similarity between the current state confidence level and the historical average confidence level. The weighted disambiguation value for the historical similarity of the s-th class of states; Pick The highest state is used as the result of the operation status determination. Based on the result of the operation status determination, a graded warning is issued, and the warning level corresponds one-to-one with the equipment status.

10. A power transmission and transformation equipment operation status monitoring system, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the power transmission and transformation equipment operation status monitoring method as described in any one of claims 1 to 9.

Images