A method and system for intelligent condition assessment of a substation capacitor DC power supply system

Abstract

A method and system for intelligent state assessment of a substation capacitor DC power supply system are disclosed. The method includes the following steps: collecting bus voltage, load current, ambient temperature, ripple voltage signal, and ripple current signal of the substation DC power supply system; calculating the equivalent series resistance and actual capacitance value of the capacitors; extracting multi-dimensional features to generate a comprehensive feature vector; classifying the state levels and setting corresponding baseline health scores, discretizing and correcting them to generate true health score labels; constructing a state assessment model, calculating a hierarchical adaptive dynamic penalty function, and training the model; using the trained state assessment model to predict the health score in real time and verify its physical rationality, obtaining the final health score and making corrections. This invention achieves accurate assessment of the health status and early anomaly identification of the capacitor DC power supply system through online identification of capacitor parameters, extraction of physical degradation features, a state assessment model, and a multi-source correction mechanism.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring technology, and relates to an intelligent condition assessment method and system for a substation capacitor DC power supply system. Background Technology

[0002] The DC power supply system with capacitor banks in substations is the "heart" of relay protection and automatic devices, and its reliability directly affects the safe and stable operation of the power grid. Among the many DC power supply solutions, capacitor-based DC power supplies or DC systems that include capacitors as important energy storage / filtering elements are still widely used in small and medium-sized substations, distribution stations, and the renovation of old substations due to their advantages such as fast response speed, strong instantaneous discharge capacity, and relatively simple maintenance. In this system, the state of the capacitor bank (such as filter capacitors and energy storage capacitors) directly determines the stability of the DC bus voltage, the load-carrying capacity, and the fault discharge capacity.

[0003] With the advancement of smart grid and ubiquitous power Internet of Things construction, substations are placing higher demands on power systems, including full lifecycle management, condition-based maintenance, and fault early warning. Currently, existing methods for DC power system condition assessment or monitoring include: Patent application CN118316157A discloses a method for assessing and predicting the condition of a substation DC power system. This method involves a verification discharge test after the DC power system is removed from the float charging state, collecting online monitoring status quantities such as voltage, resistance, and temperature, as well as operational record status quantities. Based on a condition assessment model and the analytic hierarchy process (AHP), the weight coefficients of each assessment index are determined to obtain a comprehensive score for the DC power system. A prediction model is then constructed using variational mode decomposition and recurrent neural networks to achieve condition prediction. However, this method relies on a verification discharge test for condition assessment, requires removal from the float charging state, and cannot achieve online real-time assessment. Furthermore, the weights of the assessment indicators depend on subjective weighting using the AHP, making it insensitive to subtle changes in the early, slow degradation of capacitors. The patent application with publication number CN120870944A discloses a method for monitoring the status of DC power supply in a substation. By calculating the fluctuation coefficients and weighting coefficients of the output voltage and current under pure discharge and simultaneous charge and discharge conditions, it assesses the faults in the charging and discharging function of the DC power supply and constructs a linear regression equation for the change of internal resistance with the duration of use to predict the remaining service time of the DC power supply before the occurrence of an internal resistance fault. This monitoring method is essentially still a fixed threshold method, which is difficult to adapt to the dynamic process of capacitor degradation. Moreover, it only relies on single operating parameters such as voltage, current, and temperature, resulting in an incomplete state characterization and weak interpretability.

[0004] Existing substation DC power system condition assessment methods are mostly based on external operating parameters such as voltage, current, and temperature for condition analysis, lacking effective perception of key degradation parameters such as equivalent series resistance and actual capacitance of capacitors, resulting in insufficient ability to identify the early slow degradation process of capacitors. In addition, existing assessment methods lack a comprehensive correction mechanism that combines the physical degradation law of equipment, historical operation and maintenance information, and condition change trend, which can easily lead to inaccurate assessment results and unreasonable condition changes. Summary of the Invention

[0005] To address the technical problems in existing technologies, such as poor adaptability of fixed thresholds, difficulty in identifying early degradation, susceptibility to false alarms due to transient interference, insufficient fusion of multi-dimensional features, inability to coordinate the identification of long-cycle aging and short-term mutations, and lack of refined grading in state assessment, this invention provides an intelligent state assessment method and system for substation capacitor DC power supply systems. This method enables multi-source feature extraction, time-series health assessment, four-level state determination, and early warning, effectively improving the operational reliability and intelligent maintenance level of capacitor DC power supply systems. The method includes the following steps: collecting bus voltage, load current, ambient temperature, ripple voltage signal, and ripple current signal of the substation DC power supply system; calculating the equivalent series resistance and actual capacitance value of the capacitors; extracting multi-dimensional features to generate a comprehensive feature vector; classifying state levels and setting corresponding baseline health scores, discretizing and correcting them to generate true health score labels; constructing a state assessment model, calculating the graded adaptive dynamic penalty function, and training the model; using the trained state assessment model to predict real-time health scores and verify physical rationality, obtaining the final health score and making corrections. This invention achieves accurate assessment of the health status and early anomaly identification of a capacitor DC power supply system through online identification of capacitor parameters, extraction of physical degradation features, state assessment model, and multi-source correction mechanism.

[0006] The present invention adopts the following technical solution: The first aspect of the present invention provides an intelligent state assessment method for a substation capacitor DC power supply system, comprising: S1. Collect the bus voltage, load current, ambient temperature, ripple voltage signal and ripple current signal of the DC power supply system of the substation, and calculate the equivalent series resistance and actual capacitance value of the capacitor in each time period using the ripple vector decomposition method. S2. Extract time-domain statistical features, frequency-domain spectral entropy features, and physical degradation slope features from the collected data to generate a comprehensive feature vector; S3. Based on the equivalent series resistance of the capacitor and the actual capacitance value, classify the state levels and set the corresponding baseline health score for each level; discretize and correct the baseline health score for each level to generate a real health score label; construct a state assessment model, take the comprehensive feature vector as the input of the model, take the corresponding real health score label as the label, calculate the hierarchical adaptive dynamic penalty function and train the model. S4. Predict real-time health scores using a trained state assessment model, and verify the physical rationality of the predicted health scores for the current period by combining the final health scores for the corresponding time period of the previous day, thus obtaining the final health score; acquire historical operation and maintenance data and cluster it, and correct the final health score based on the clustering results.

[0007] Preferably, in S1, Fourier transform is performed on the ripple voltage signal and ripple current signal for each time period to extract the effective values ​​of the fundamental components of the ripple voltage and ripple current, which are used as the amplitude of the ripple voltage and the amplitude of the ripple current for the corresponding time period. For each time period, the difference between the current ambient temperature and the reference temperature for the corresponding time period is multiplied by a predefined temperature compensation coefficient and then 1 is added to obtain the temperature compensation term. The ratio of the ripple voltage amplitude to the ripple current amplitude for the corresponding time period is multiplied by the cosine of the phase difference between the ripple voltage and the ripple current at the ripple fundamental frequency and the temperature compensation term to obtain the equivalent series resistance of the capacitor after temperature compensation. For each time period, the ratio of the ripple voltage amplitude to the ripple current amplitude for that time period is multiplied sequentially by the sine of the phase difference between the ripple voltage and ripple current at the ripple fundamental frequency, the ripple fundamental frequency, and 2. Take the reciprocal as the actual capacitance value for the corresponding time period.

[0008] Preferably, in S2, the time-domain statistical features include the mean voltage, voltage variance, peak voltage, voltage kurtosis, and voltage margin factor of the bus voltage. Frequency domain spectral entropy characteristics include the power spectral entropy and wavelet packet energy entropy of the ripple voltage signal; The physical degradation slope characteristics include the equivalent series resistance change rate, capacitance decay rate, equivalent series resistance temperature drift rate, equivalent series resistance deviation coefficient, and capacitance deviation coefficient; among which, the equivalent series resistance change rate is calculated based on the initial equivalent series resistance of the capacitor at the time of manufacture and the cumulative operating time of the capacitor; the capacitance decay rate is calculated by dividing the difference between the actual capacitance value and the predefined initial capacitance value in each time period by the corresponding time period; The extracted features are time-aligned and then concatenated to form a comprehensive feature vector.

[0009] Preferably, the process of calculating the equivalent series resistance temperature drift rate, the equivalent series resistance deviation factor, and the capacitance deviation factor is as follows: For each time period, the difference between the equivalent series resistance of the capacitor after temperature compensation and the equivalent series resistance of the capacitor before temperature compensation is divided by the difference between the current ambient temperature and the reference temperature in the corresponding time period to obtain the equivalent series resistance temperature drift rate. The ratio of the difference between the current equivalent series resistance of the capacitor and the initial equivalent series resistance of the capacitor to the initial equivalent series resistance of the capacitor is calculated to obtain the equivalent series resistance deviation coefficient. The difference between the actual capacitance value and the initial capacitance value during the corresponding time period is divided by the initial capacitance value to obtain the capacitance deviation coefficient.

[0010] Preferably, the process of discretizing the baseline health score for each level in S3 is as follows: Calculate the ratio of the current equivalent series resistance of the capacitor to the initial equivalent series resistance of the capacitor, and subtract 1 to obtain the equivalent resistance influence coefficient; take the opposite of the ratio of the current actual capacitance value to the initial capacitance value and add 1 as the numerator, and use the predefined proportional coefficient as the denominator to calculate the capacitance influence coefficient; take the maximum value of the equivalent resistance influence coefficient and the capacitance influence coefficient as the degradation index; divide the degradation index into multiple state levels according to the predefined state division rules, and discretize the baseline health score under each state level through linear interpolation.

[0011] Preferably, the process of correcting the discretized baseline health score in S3 is as follows: Obtain device status information from historical operation and maintenance records. If the operation and maintenance records confirm that the device is operating normally and without faults, the baseline health score is used as the true health score label. If the operation and maintenance records confirm that the device has defects or has been replaced, the baseline health score is subtracted from the predefined score threshold and the maximum value between the threshold and 0 is used as the true health score label. If the maintenance records confirm that the equipment has undergone power outage maintenance and has been repaired, for the pre-maintenance scheduled period, obtain the number of defect records within the scheduled maintenance duration and take the maximum value between the number of records and the scheduled upper limit; subtract the product of the predefined adjustment coefficient and the maximum value from the predefined upper limit of low health scores to obtain the true health score label for the pre-maintenance scheduled period; for the post-maintenance scheduled period, obtain the number of days between the current maintenance and the previous maintenance of the same type, take the ratio of this interval to the scheduled upper limit of days and take the minimum value between this ratio and the predefined upper limit of maintenance days, subtract the corresponding minimum value from the predefined upper limit of high health scores and multiply it by the predefined amplification coefficient to obtain the true health score label for the post-maintenance scheduled period.

[0012] Preferably, the process of calculating the hierarchical adaptive dynamic penalty function in S3 is as follows: The comprehensive feature vector of each time period is taken as a sample. Based on the predicted health score and the real health score label of each sample, the mean absolute error during training is calculated as the absolute error term. For each sample, the maximum value between 0, the predefined alarm threshold and the corresponding sample's real health score label is taken as the numerator, the predefined alarm threshold is taken as the denominator, and the ratio of the numerator to the denominator is multiplied by the predefined weight amplification factor and then added to 1 as the dynamic sample importance weight of the corresponding sample. The mean squared error between the predicted health score and the real health score label of each sample is calculated, multiplied by the dynamic sample importance weight, accumulated and multiplied by the predefined weighted squared loss balance coefficient as the weighted squared loss term. The absolute error term and the weighted squared loss term are added together to obtain the hierarchical adaptive dynamic penalty function.

[0013] Preferably, the process of verifying the physical rationality of the predicted health score for the current time period in S4 is as follows: The difference between the predicted health score output by the state assessment model in the current time period and the final health score in the corresponding time period of the previous day is taken as the daily score difference. The daily score difference is input into the sign function to obtain the sign coefficient. The absolute value of the daily score difference is subtracted from the predefined maximum allowable jump threshold of the daily score, and the maximum value between it and 0 is taken. This maximum value is then multiplied by the sign coefficient and the predefined constraint strength coefficient in sequence, the inverse number is taken, and then added to the predicted health score output by the state assessment model in the current time period to obtain the final health score in the current time period.

[0014] Preferably, the process of correcting the final health score in S4 is as follows: Collect historical operation and maintenance data of the target equipment and convert it into operation and maintenance feature vectors; among which, operation and maintenance features include cumulative number of defects, cumulative number of maintenance, number of days since the last maintenance, level of the most recent defect, and years of equipment operation; Set a predetermined number of clusters, and cluster the maintenance feature vectors of all historical maintenance records. Each cluster corresponds to a status level. The status levels include alarm status, abnormal status, status requiring attention, and healthy status. For the maintenance feature vector of the day, calculate its Euclidean distance to each cluster center, and determine the status level corresponding to the cluster center with the smallest Euclidean distance as the maintenance status level of the current day. Based on the final health score for the current period, determine the corresponding status level as the initial level; If the operation and maintenance status level is the same as or higher than the initial level, it means that the operation and maintenance status level considers the current day's status to be better than the initial level. In this case, the initial level will remain unchanged to prevent operation and maintenance data from masking the true degradation data. If the operation and maintenance status level is lower than the initial level, the level will be downgraded to the operation and maintenance status level.

[0015] Preferably, the secondary correction of the status level after the maintenance data has been corrected is carried out as follows: If the growth rate of the equivalent series resistance of the capacitor in the current time period of the day compared with the equivalent series resistance of the capacitor in the corresponding time period of the previous day exceeds the predetermined first growth rate threshold, then the rule correction amount is set to 1; If the actual capacitance value during the current time period of the day is less than the predetermined capacitance threshold, the rule correction amount will be set to 2; If the current ambient temperature at the current time of day exceeds the upper limit of the rated operating temperature of the equipment and the growth rate of the equivalent series resistance of the capacitor is greater than the predetermined second growth rate threshold, the rule correction amount will be set to 1. When multiple rules are triggered simultaneously, the rule correction amounts under the triggering rules are summed to obtain the final rule correction amount; Set the corresponding level number according to the degree of abnormality of the status level from high to low; subtract the final rule correction amount from the level number corresponding to the status level after the operation and maintenance data is corrected, and take the maximum value between it and 1 to obtain the final level number and determine the corresponding final status level.

[0016] A second aspect of the present invention provides an intelligent state assessment system for a substation capacitor DC power supply system, which executes an intelligent state assessment method for a substation capacitor DC power supply system, including: The data acquisition module collects the bus voltage, load current, ambient temperature, ripple voltage signal, and ripple current signal of the DC power supply system of the substation, and calculates the equivalent series resistance and actual capacitance value of the capacitor for each time period using the ripple vector decomposition method. The multi-dimensional feature extraction module extracts time-domain statistical features, frequency-domain spectral entropy features, and physical degradation slope features from the collected data to generate a comprehensive feature vector. The state assessment model training module divides the state levels and sets the corresponding baseline health scores based on the equivalent series resistance and actual capacitance value; it discretizes and corrects the baseline health scores for each level to generate real health score labels; it constructs a state assessment model, using the comprehensive feature vector as the model input and the corresponding real health score labels as labels, calculates the hierarchical adaptive dynamic penalty function, and trains the model. The status assessment module uses a trained status assessment model to predict real-time health scores and combines the final health scores for the corresponding time period of the previous day to perform physical rationality verification on the predicted health scores for the current time period, thus obtaining the final health score; it also acquires historical operation and maintenance data and clusters it, and corrects the final health score based on the clustering results.

[0017] A third aspect of the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of a smart state assessment method for a substation capacitor DC power supply system.

[0018] A fourth aspect of the invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of an intelligent state assessment method for a substation capacitor DC power supply system.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1. This invention jointly collects bus voltage, load current, ambient temperature, ripple voltage, and ripple current during the operation of a substation capacitor DC power supply system, and simultaneously obtains the equivalent series resistance and actual capacitance value of the capacitor based on the ripple vector decomposition method. Compared with methods that rely solely on external operating conditions such as operating voltage and current for judgment, this invention can directly characterize the internal performance changes of the capacitor, effectively improving the detection capability of early degradation states such as capacitance decay and internal resistance increase, and enhancing the accuracy and real-time performance of substation DC power supply system condition assessment.

[0020] 2. This invention integrates time-domain statistical features, frequency-domain spectral entropy features, and degradation slope features constructed based on the physical degradation laws of capacitors to form a comprehensive feature vector that includes operational fluctuation features, ripple energy change features, and performance degradation trend features. Simultaneously, it constructs state level labels based on the equivalent series resistance and actual capacitance value of the capacitor, and corrects the health score by incorporating historical maintenance records, generating training labels that better reflect the actual operating state of the equipment. Furthermore, it enhances the model's attention to abnormal degradation samples through a hierarchical adaptive dynamic penalty function, reducing the impact of uneven distribution between healthy and abnormal samples on the evaluation results, thereby improving the state assessment model's ability to identify and generalize capacitor devices at different degradation stages.

[0021] 3. This invention verifies physical rationality by combining the health score change trend of the corresponding time period of the previous day, and performs multi-level state correction using historical operation and maintenance data clustering results and real-time degradation rules. This ensures that the final evaluation result simultaneously meets the constraints of data-driven laws and the actual degradation mechanism of the equipment, avoiding unreasonable jumps in model output due to instantaneous data disturbances. Furthermore, by enhancing rules for typical abnormal situations such as sudden increases in equivalent series resistance, abnormal decreases in capacitance, and accelerated degradation at high temperatures, potential fault risks can be identified in advance, providing a reliable basis for preventive maintenance, maintenance strategy formulation, and operational safety assurance of substation DC power supply systems. Attached image description: Figure 1 The flowchart of an intelligent state assessment method for a substation capacitor DC power supply system provided by the present invention is shown. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0023] Example 1 Embodiment 1 of the present invention provides an intelligent state assessment method for a substation capacitor DC power supply system, comprising the following steps: S1. Collect bus voltage, load current, ambient temperature, ripple voltage signal and ripple current signal of the DC power supply system of the substation, and calculate the equivalent series resistance and actual capacitance value of the capacitor in each time period using the ripple vector decomposition method.

[0024] A smart data acquisition terminal is configured on the DC power supply panel to collect bus voltage, load current, ambient temperature, ripple voltage signal, and ripple current signal of the DC power supply system at a frequency of 1kHz. Preliminary calculations are made of aging characteristics such as equivalent series resistance (ESR) and capacitance value, and these data are uploaded to the cloud data center along with the raw data.

[0025] A temperature-compensated ripple vector decoupling algorithm is proposed to address the ESR measurement error problem over a wide temperature range. Fourier transforms are performed on the ripple voltage and ripple current signals for each time period to extract the RMS values ​​of the fundamental components of the ripple voltage and current, which are then used as the amplitudes of the ripple voltage and current for that time period. For each time period, the difference between the current ambient temperature and the reference temperature is multiplied by a predefined temperature compensation coefficient and then incremented by 1 to obtain the temperature compensation term. The ratio of the ripple voltage amplitude to the ripple current amplitude for that time period is multiplied sequentially by the cosine of the phase difference between the ripple voltage and ripple current at the ripple fundamental frequency, and the temperature compensation term, to obtain the temperature-compensated equivalent series resistance of the capacitor. Specifically, this is expressed as follows: ; In the formula, The equivalent series resistance of the capacitor after temperature compensation during time period t; Let t be the amplitude of the ripple voltage during time period t. Take the effective value of the fundamental component of the ripple separated by the 2048-point Fourier transform within a 30-second sampling window. It is the steady-state estimate within this window. The ripple current amplitude during time period t shares the same sampling window and synchronization clock as the aforementioned ripple voltage. It is the phase difference between the ripple voltage and the ripple current at the ripple fundamental frequency, used to separate the resistive component from the total impedance mode formed by dividing the ripple voltage amplitude by the ripple current amplitude. The temperature compensation coefficient is a predefined parameter, determined by the capacitor manufacturer's parameters or actual measurements during the initial operation. In this embodiment, it is taken as 0.002 / ℃. Let t be the current ambient temperature during time period t; For reference temperature, take 25℃.

[0026] This embodiment introduces the temperature drift coefficient into the ripple decoupling calculation for the first time, establishing a quantitative relationship between ESR and ambient temperature. Within a wide temperature range of -40℃ to +85℃, the ESR measurement error is reduced from ±15% of the conventional method to within ±3%, significantly improving the consistency of evaluation results under different environmental conditions.

[0027] For each time period, the ratio of the ripple voltage amplitude to the ripple current amplitude for that time period is multiplied sequentially by the sine of the phase difference between the ripple voltage and ripple current at the ripple fundamental frequency, the ripple fundamental frequency, and 2. Take the reciprocal as the actual capacitance value for the corresponding time period; the specific process is as follows: ; in, This represents the actual capacitance value during time period t. Let t be the fundamental frequency of the ripple during time period t.

[0028] S2. Extract time-domain statistical features, frequency-domain spectral entropy features, and physical degradation slope features from the collected data to generate a comprehensive feature vector.

[0029] Using a predetermined number of samples as the time window, the mean voltage, voltage variance, peak voltage, voltage kurtosis, and voltage margin factor (i.e., time domain statistical characteristics) of the bus voltage are extracted within each sampling window; specifically represented as follows: ; In the formula, The first characteristic is the average voltage. N represents the predetermined number of samples within the time window; For the i-th sampling time Bus voltage; The second characteristic is voltage variance. ; The third characteristic is the voltage peak value within the time window; The standard deviation of voltage within the time window; The fourth characteristic is voltage kurtosis; The fifth characteristic is the voltage margin factor; The ripple voltage signal is subjected to a Fourier Transform (FFT) to obtain the power spectrum; the frequency axis is divided into predetermined frequency bands (16 frequency bands in this embodiment), and the cumulative power of each frequency band is calculated; based on the cumulative power, the proportion of the cumulative power of each frequency band to the total power is calculated to determine the power spectral entropy of the ripple voltage signal; specifically expressed as: ; In the formula, The sixth characteristic is the power spectral entropy of the ripple voltage signal. ; For the first k The proportion of cumulative power in each frequency band to the total power; For the predetermined frequency band, this embodiment uses 16; For the first k Cumulative power of each frequency band; For the first i frequency Power at the location; , The first k The starting and ending frequencies of each frequency band are numbered; The ripple voltage signal is decomposed using wavelet packets, and the sub-band energy of each sub-band is calculated. The proportion of each sub-band energy to the total energy is calculated to determine the wavelet packet energy entropy of the ripple voltage signal. Specifically, it is expressed as follows: ; In the formula, It consists of seven features, namely wavelet packet energy entropy. ; The amount of energy of the subbands obtained from the decomposition; This represents the proportion of the energy of the m-th subband to the total energy. Let m be the energy of the m-th subband.

[0030] As a preferred implementation, based on the physical laws of capacitance, degradation features are extracted from historical time-series data of the equivalent series resistance and actual capacitance value of the capacitor. The rate of change of the equivalent series resistance is calculated based on the initial equivalent series resistance of the capacitor at the time of manufacture and the cumulative operating time of the capacitor; specifically expressed as: ; In the formula, The eighth characteristic is the rate of change of the equivalent series resistance. ; Let be the equivalent series resistance of the capacitor during time period t; This is the initial equivalent series resistance value of the capacitor when it leaves the factory; Cumulative runtime; The capacitance decay rate is calculated based on the initial capacitance value at the time of manufacture and the actual capacitance value over the cumulative operating time. Specifically, the capacitance decay rate is obtained by dividing the difference between the actual capacitance value and the predefined initial capacitance value for the corresponding time period by the corresponding time period. This is expressed as follows: ; in, The ninth characteristic is the capacitance decay rate. ; C0 represents the actual capacitance value during time period t; C0 is the initial capacitance value of the capacitor when it leaves the factory. Based on the equivalent series resistance of the capacitor at the current ambient temperature and reference temperature during the current time period, the temperature drift rate of the equivalent series resistance is calculated. Specifically, for each time period, the difference between the temperature-compensated equivalent series resistance and the temperature-uncompensated equivalent series resistance during the corresponding time period is divided by the difference between the current ambient temperature and the reference temperature during the corresponding time period to obtain the temperature drift rate of the equivalent series resistance. This is specifically expressed as follows: ; In the formula, The tenth characteristic is the temperature drift rate of the equivalent series resistance. This reflects the drift effect of ambient temperature changes on the equivalent series resistance, and is constrained by the temperature characteristics of the capacitor material. Let t be the current ambient temperature during time period t; For reference temperature; Based on the initial equivalent series resistance of the capacitor at the time of manufacture, the equivalent series resistance deviation factor is calculated. Specifically, the ratio of the difference between the current equivalent series resistance and the initial equivalent series resistance to the initial equivalent series resistance is calculated to obtain the equivalent series resistance deviation factor. This is expressed as follows: ; In the formula, The eleventh feature is the equivalent series resistance deviation coefficient, which reflects the degree of deviation of the current equivalent series resistance of the capacitor from the initial value. It is constrained by the capacitor series equivalent circuit model: the ESR deviation coefficient reflects the degree of deviation of the current equivalent series resistance from the initial value. This is the initial equivalent series resistance of the capacitor when it leaves the factory; Based on the initial capacitance value at the time of manufacture, the capacitance deviation coefficient is calculated; the difference between the actual capacitance value and the initial capacitance value in the corresponding period is divided by the initial capacitance value to obtain the capacitance deviation coefficient; specifically expressed as: ; In the formula, The twelfth feature is the capacitance deviation coefficient, which reflects the degree of decay of the current capacitance value relative to the initial value. Since the energy storage capacity of a capacitor is directly proportional to its capacitance value, the capacitance deviation coefficient directly corresponds to the decay ratio of the energy storage capacity and is one of the core indicators for assessing the health status of a capacitor. The extracted features are time-aligned and then concatenated to form a comprehensive feature vector. .

[0031] It should be further explained that each dimension of the above-mentioned comprehensive feature vector is a scalar: x1-x5 are single scalars obtained from the bus voltage time series (containing approximately 300,000 sampling points) within each 30-second time window through statistical operations of mean / variance / peak value / kurtosis / margin factor; x6 is the single value of power spectral entropy calculated according to 16 frequency bands after FFT of the ripple voltage within the corresponding window; x7 is the single value of energy entropy calculated after wavelet packet decomposition of the ripple voltage; x8-x 12 All values ​​are obtained by normalizing the temperature-compensated ESR / capacity sequence for each 30-second sampling window; at the end of each 30-second sampling window, a 12-dimensional scalar vector is output with a unified timestamp; then the temporal buffer unit stacks the 12-dimensional vectors of the most recent 30 windows into a 30×12 temporal feature matrix for use by the LSTM-Attention model.

[0032] As an optional implementation, based on the equivalent series resistance of the capacitor in each time period, its first and second order differences are calculated to determine the degradation acceleration factor in each time period; specifically expressed as: ; In the formula, The denominator is the degradation trend acceleration factor at time t, with the numerator being the second difference (i.e., discrete acceleration) of the temperature-compensated ESR sequence and the denominator being the first difference (i.e., discrete velocity). The equivalent series resistance of the capacitor in the period preceding time t; The duration of the time period; for The equivalent series resistance of the capacitor during a given time period.

[0033] It should be noted that, considering that this quantity is mainly used for operation and maintenance auxiliary monitoring in the evaluation process and does not participate in the core training input of the LSTM-Attention model, it does not affect the composition of the aforementioned 12-dimensional comprehensive feature vector, and is only used as an optional supplementary monitoring quantity output to provide inflection point warnings to operation and maintenance personnel.

[0034] S3. Based on the degradation degree of the equivalent series resistance and actual capacitance value of the capacitor, classify the state levels and set the corresponding baseline health score for each level; discretize and correct the baseline health score for each level to generate a true health score label; construct a state assessment model, take the comprehensive feature vector as the input of the model, take the corresponding true health score label as the label, calculate the hierarchical adaptive dynamic penalty function and train the model; wherein, the state levels include alarm state, abnormal state, state requiring attention and healthy state.

[0035] As a preferred implementation method, the specific process of classifying the status levels and setting the corresponding baseline health scores is as follows: when or hour, The status is determined to be an alarm. Otherwise when or hour, This is determined to be an abnormal state; Otherwise when or hour, This is determined to be a state requiring attention; Otherwise when and hour, The condition is determined to be healthy. Furthermore, linear interpolation is used to discretize the baseline health score. Specifically, the ratio of the current equivalent series resistance of the capacitor to the initial equivalent series resistance of the capacitor is calculated and then subtracted by 1 to obtain the equivalent resistance influence coefficient. The inverse of the ratio of the current actual capacitance value to the initial capacitance value is taken and 1 is added to obtain the numerator. A predefined proportional coefficient (0.2 in this embodiment) is used as the denominator to calculate the capacitance influence coefficient. The maximum value of the equivalent resistance influence coefficient and the capacitance influence coefficient is taken as the degradation index. According to the predefined state division rules, the degradation index is divided into multiple state levels, and the baseline health score under each state level is discretized using linear interpolation. Specifically, it is expressed as follows: ; In the formula, As a baseline health score; The degradation index, Time corresponds to health status. The corresponding status needs to be monitored. Corresponding to abnormal states, Corresponding alarm status; Obtain manually confirmed equipment status information from historical maintenance records and correct the baseline health score: If the maintenance records confirm that the equipment is operating normally and without faults, then the baseline health score will be used as the true health score label. ; If the maintenance records confirm that the equipment is defective or has been replaced, the baseline health score is subtracted from a predefined scoring threshold (30 in this example), and the maximum value between the threshold and 0 is taken as the true health score label. ; If the maintenance records confirm that the equipment underwent a power outage for maintenance and recovered well afterward, then the tags before and after maintenance will be marked as the corresponding low score and the high score after recovery, respectively; specifically, if the maintenance records indicate that maintenance was performed and repaired on day k, then: For the day before maintenance, obtain the number of defect records within the predetermined maintenance period and take the maximum value between this number and the predetermined upper limit (5 in this embodiment); subtract the product of the predefined adjustment coefficient (2 in this embodiment) and the maximum value from the predefined lower health score upper limit (40 in this embodiment) as the true health score label for the day before maintenance; specifically expressed as: ; In the formula, This is the actual health score label for the day before maintenance (k-1 day). The more defects there are, the lower the baseline score before maintenance. The value range is [30, 40]. This refers to the number of defect records within the scheduled maintenance period before the maintenance (i.e., 90 days before the maintenance on day k); For the day following maintenance, the interval between the current maintenance and the previous maintenance of the same type is obtained. The ratio of this interval to a predetermined upper limit of days (365 in this embodiment) is taken, and the minimum of this ratio and the predefined upper limit of maintenance days is taken. The product of the corresponding minimum value and a predefined amplification factor (5 in this embodiment) is then subtracted from the predefined upper limit of high-score health score to obtain the true health score label for the day following maintenance; specifically expressed as:

[0036] In the formula, This is the true health score label one day after maintenance (k+1 day). The longer the interval, the higher the degree of thoroughness of the equipment maintenance and repair. The value range is [90, 95]. This represents the number of days between the current maintenance and the previous maintenance of the same type. In this embodiment, linear interpolation is performed only on adjacent day labels during stable operation days without major faults or maintenance. Once maintenance or a sudden change in state occurs on a certain day, that day and the two days before and after it are directly assigned values ​​according to the above discrete step formula and do not participate in the linear interpolation window. Linear interpolation is only used to fill the missing day labels during stable periods far from maintenance events. At this time, ESR degradation is approximately linear, label changes are gradual, and the interpolation error is within the acceptable range for engineering (within ±3 minutes).

[0037] Furthermore, we can ultimately obtain the true health score label for each device and each sampling moment.

[0038] In this embodiment, historical operating data of capacitor DC power supply systems from different manufacturers, models, and years of operation are collected, covering the following scenarios: Normal operation samples: The equipment operates smoothly with no fault records; Samples requiring attention: The equivalent series resistance or actual capacitance value of the capacitor shows a slight degradation trend; Abnormal state samples: The equivalent series resistance or actual capacitance value of the capacitor deviates significantly from the normal range, but the equipment can still operate; Alarm state samples: The equipment has failed or has been replaced. A training dataset is constructed, containing 12,000 normal samples, 3,000 attention samples, 800 abnormal samples, and 200 alarm samples, divided into training, validation, and test sets in a ratio of 7:1.5:1.5.

[0039] As a preferred implementation, a state evaluation model is constructed, including an LSTM layer, an Attention layer, and a fully connected layer; the LSTM layer has a hidden layer dimension of 64 and a number of layers of 2, and is used to extract temporally dependent features; the Attention layer has an attention dimension of 32 and calculates attention weights for the hidden states at each time step of the LSTM; the fully connected layer maps the attention-weighted features to a scalar health score.

[0040] During training, the temporal feature matrices of all training samples are input into the state assessment model, and the predicted health score S∈[0, 100] is output. Five-fold cross-validation is used to adjust the model hyperparameters, with a learning rate of 0.001 and 100 training epochs. A dynamic penalty function is used to optimize the loss function to solve the sample imbalance problem. This embodiment proposes a hierarchical adaptive dynamic penalty function to replace the traditional fixed-weight loss function. The comprehensive feature vector of each time period is treated as a sample. Based on the predicted health score and the actual health score label of each sample, the mean absolute error during training is calculated as the absolute error term. For each sample, the maximum value between 0, a predefined alarm threshold, and the difference between the corresponding sample's actual health score label is taken as the numerator, and the predefined alarm threshold is taken as the denominator. The ratio of the numerator to the denominator is multiplied by a predefined weight amplification factor and then added to obtain the dynamic sample importance weight for the corresponding sample. The mean squared error between the predicted health score and the actual health score label of each sample is calculated, multiplied by the dynamic sample importance weight, summed, and then multiplied by a predefined weighted squared loss balance coefficient to obtain the weighted squared loss term. The absolute error term and the weighted squared loss term are added together to obtain the hierarchical adaptive dynamic penalty function as the total training loss. Specifically, it is expressed as follows: ; ; In the formula, The total training loss (the smaller the loss, the more accurate the prediction); Predicted health score for the i-th sample; The weight amplification factor is a predefined value; in this embodiment, it is set to 2.0. The true health score label for the i-th sample; The predefined alarm threshold is set to 50 points in this embodiment; The predefined weighted squared loss balance coefficient is set to 0.5 in this embodiment; is the dynamic sample importance weight; N is the total number of training samples, and i is the sample index; in this embodiment, the comprehensive feature vector under each time period is used as the training sample; It should be noted that the balance coefficient λ=0.5 and the weight amplification coefficient k=2.0 before the weighted squared loss term are empirical values ​​obtained after repeated optimization in this embodiment for the training sample distribution (12,000 normal samples, 3,000 attention samples, 800 abnormal samples, and 200 alarm samples): at this time, the alarm area ( <50) Dynamic weights of samples The maximum value can reach 3.0, and the weighted squared loss term ranges from approximately 2 to 4 times that of the absolute error term, without exhibiting a magnitude imbalance where the penalty term is much larger than the absolute error term. If the device is deployed in environments with significantly different sample distributions, it can be validated through 5-fold cross-validation. ∈[0.1,1.0]、 Recalibrate within the range of [0.5, 3.0], using the two values ​​being of the same order of magnitude and having the highest F1 score on the validation set as the calibration criterion.

[0041] Traditional methods either employ fixed-weighted MSE loss or simply perform class weighting in classification problems. This invention is the first to propose an adaptive loss mechanism for regression problems, which uses graded weighting based on score intervals, giving higher training attention to samples closer to the fault threshold.

[0042] In this embodiment, new operational data and fault samples are received every 30 days, and the LSTM-Attention model parameters are updated through incremental learning. The update process does not affect the normal operation of the system.

[0043] After training, the model is deployed on the FPGA neural network accelerator of the edge intelligent terminal. During online evaluation: the 12-dimensional comprehensive feature vectors of the most recent 30 time points are automatically extracted each day to form a time-series feature matrix; input to the state evaluation model, and output the device's daily health score S through forward inference; S ∈ [0, 100], the higher the score, the better the device's health status.

[0044] S4. Predict real-time health scores using a trained state assessment model, and verify the physical rationality of the predicted health scores for the current period by combining the final health scores for the corresponding time period of the previous day, thus obtaining the final health score; acquire historical operation and maintenance data and cluster it, and correct the final health score based on the clustering results.

[0045] As an optional implementation, a weighted normalization method is used to calculate the output of the comprehensive health score alternative state assessment model and map it to four levels of health status; the formula for calculating the comprehensive health score is: ; in, Let be the weight coefficient of the i-th feature, satisfying The attention weights output by the Attention layer during the training of the LSTM-Attention model are automatically obtained after normalization, reflecting the contribution of each feature to the health status assessment. is the normalized value of the i-th feature; This is the baseline value for this feature (the average value over 30 periods during the initial operation of the equipment).

[0046] When S≥90, it is judged as a normal state; when 70≤S<90, it is judged as a state requiring attention; when 50≤S<70, it is judged as an abnormal state; when S<70, an early warning is issued; when S<50, it is judged as an alarm state.

[0047] It should be noted that the comprehensive health score S calculated in this step is an explicit normalized fusion result of the original health score output by the LSTM-Attention model, which is weighted by the Attention weights and the deviation of each dimension of the device's current measured features. This result is used to provide the operation and maintenance system with a final score that has physical feature traceability.

[0048] As a preferred implementation, the physical rationality of the real-time output health score is verified; the difference between the predicted health score output by the state assessment model in the current time period and the final health score in the corresponding time period of the previous day is taken as the daily score difference; the daily score difference is input into a sign function to obtain a sign coefficient; the absolute value of the daily score difference is subtracted from the predefined maximum allowable jump threshold of the daily score, and the maximum value between the absolute value and 0 is taken. This maximum value is then multiplied by the sign coefficient and the predefined constraint strength coefficient, the inverse of which is taken, and added to the predicted health score output by the state assessment model in the current time period to obtain the final health score in the current time period; specifically expressed as:

[0049] In the formula, The corresponding time period for the current day, i.e. The final health score; The current time period of day k output by the state assessment model Predicting health scores; The corresponding time period of the previous day, i.e. The final health score; The constraint strength coefficient is a predefined value, and in this embodiment, the value ranges from 0.8 to 1.0. The maximum permissible jump threshold for a single day's score is predefined; in this embodiment, it is set to 5 points. It is a symbolic function; This embodiment introduces physical continuity constraints in the health assessment of capacitor DC power supplies; capacitor aging is a gradual electrochemical process, and the health status cannot change abruptly beyond a certain threshold in a single day; this algorithm effectively filters false alarms caused by data fluctuations and noise interference by limiting the maximum jump variable in daily scores.

[0050] As a preferred implementation method, historical operation and maintenance data clustering is used to correct the health score; the specific process is as follows: Collect the target equipment's historical operation and maintenance records for the past 3 years, including: Inspection records: inspection time, inspection personnel, visual inspection results (whether there are bulges, leaks, overheating, etc.), measurement data (voltage, internal resistance, temperature); Defect records: defect discovery time, defect type (ESR exceeding the standard, capacity decay, thermal runaway, etc.), defect level, and handling measures; Maintenance records: maintenance time, maintenance content (replacing capacitors, cleaning terminals, tightening connections, etc.), and retest data after maintenance; Fault records: fault time, fault phenomenon, fault cause analysis, and replacement records. Furthermore, cluster analysis is performed on the operation and maintenance data: Each operation and maintenance record undergoes feature extraction and encoding, transforming it into a 5-dimensional operation and maintenance feature vector. y1: Cumulative number of defects (normalized to the [0,1] interval); y2: Cumulative number of maintenance visits (normalized to the [0,1] interval); y3: Number of days since the last maintenance visit (normalized to the [0,1] interval); y4: Defect level of the most recent maintenance visit (coded as: no defects = 0, normal = 0.33, severe = 0.67, critical = 1.0); y5: Equipment service life (normalized to the [0,1] interval); Input the 5-dimensional maintenance feature vector Y of all historical maintenance records into the KMeans clustering algorithm, setting the number of clusters K=4, corresponding to four equipment maintenance status levels: Cluster center C1: Low defects, low maintenance, stable operation, corresponding to normal state; Cluster center C2: Occasional defects, regular maintenance, and corresponding status needs to be monitored; Cluster center C3: Frequent defects, multiple repairs, corresponding to abnormal conditions; Cluster center C4: Defects are concentrated and recently replaced, corresponding to the alarm status.

[0051] It should be noted that the operation and maintenance status mode corresponding to each cluster is obtained by manually semantically annotating the clustering results based on the operation and maintenance experience of the power industry, and is not part of the output of the clustering algorithm itself; in actual implementation, this semantic annotation can be readjusted according to the distribution of on-site operation and maintenance data.

[0052] For the daily operation and maintenance feature vector, calculate its Euclidean distance to each cluster center, and determine the status level corresponding to the cluster center with the smallest Euclidean distance as the current day's operation and maintenance status level; Based on the final health score for the current period, determine the corresponding status level as the initial level; If the operation and maintenance status level is the same as or higher than the initial level, it means that the operation and maintenance status level considers the current day's status to be better than the initial level. In this case, the initial level will remain unchanged to prevent operation and maintenance data from masking the true degradation data. If the operation and maintenance status level is lower than the initial level, the level will be downgraded to the operation and maintenance status level. In this embodiment, corresponding level numbers are set from high to low according to the degree of abnormality of the status level; level number 1 corresponds to alarm status, level number 2 corresponds to abnormal status, level number 3 corresponds to status requiring attention, and level number 4 corresponds to normal status; that is, the higher the level number, the better the corresponding status.

[0053] The following three operational experience rules are introduced to perform a secondary correction of the status level of the corrected operational data; specifically expressed as follows: Rule 1: If the equivalent series resistance of a capacitor in the current time period of a given day increases compared to the equivalent series resistance of a capacitor in the corresponding time period of the previous day, then... If the predetermined first growth rate threshold (i.e., 10%) is exceeded, the status level will be downgraded by one level (e.g., normal downgrade to "need attention") to reflect the urgency of the rapid deterioration of the capacitor's equivalent series resistance, and the rule correction amount ΔR will be set to 1. Rule 2: If the actual capacitance value in the current time period of the day is less than the predetermined capacitance threshold, i.e., C(t) < 0.7 × C0, then regardless of the model score, the rule correction amount ΔR will be set to 2. Rule 3: If the current ambient temperature at the current time of day exceeds the upper limit of the rated operating temperature of the equipment (85℃ in this embodiment) and the growth rate of the equivalent series resistance of the capacitor is greater than the predetermined second growth rate threshold of 5%, then the status level is downgraded by one level to reflect the accelerated coupling effect of thermal aging, and the rule correction amount ΔR is set to 1. When multiple rules are triggered simultaneously, the rule correction amounts under the triggering rules are summed to obtain the final rule correction amount.

[0054] Subtract the final rule correction amount from the level number corresponding to the corrected status level in the maintenance data, and take the maximum value between the final correction amount and 1 to obtain the final level number and determine the corresponding final status level; the specific formula for calculating the final level number is as follows: ; In the formula, This is the final grade number; The status level corresponding to the corrected status level for operation and maintenance data; This represents the final rule adjustment amount.

[0055] It should be noted that the final level number must be limited to the valid level number [1,4]. If the final level number is <1, it will be forced to be 1 (alarm) to ensure that the corrected level does not exceed the definition range of the four-level classification.

[0056] Example 2 Embodiment 2 of the present invention provides an intelligent state assessment system for a substation capacitor DC power supply system, which executes an intelligent state assessment method for a substation capacitor DC power supply system, including: The data acquisition module collects the bus voltage, load current, ambient temperature, ripple voltage signal, and ripple current signal of the DC power supply system of the substation, and calculates the equivalent series resistance and actual capacitance value of the capacitor for each time period using the ripple vector decomposition method. The multi-dimensional feature extraction module extracts time-domain statistical features, frequency-domain spectral entropy features, and physical degradation slope features from the collected data to generate a comprehensive feature vector. The state assessment model training module divides the state levels and sets the corresponding baseline health scores based on the equivalent series resistance and actual capacitance value; it discretizes and corrects the baseline health scores for each level to generate real health score labels; it constructs a state assessment model, using the comprehensive feature vector as the model input and the corresponding real health score labels as labels, calculates the hierarchical adaptive dynamic penalty function, and trains the model. The status assessment module uses a trained status assessment model to predict real-time health scores and combines the final health scores for the corresponding time period of the previous day to perform physical rationality verification on the predicted health scores for the current time period, thus obtaining the final health score; it also acquires historical operation and maintenance data and clusters it, and corrects the final health score based on the clustering results.

[0057] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0058] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0059] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0060] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for intelligent condition assessment of substation capacitor DC power system, characterized in that, include: S1. Collect the bus voltage, load current, ambient temperature, ripple voltage signal and ripple current signal of the DC power supply system of the substation, and calculate the equivalent series resistance and actual capacitance value of the capacitor in each time period using the ripple vector decomposition method. S2. Extract time-domain statistical features, frequency-domain spectral entropy features, and physical degradation slope features from the collected data to generate a comprehensive feature vector; S3. Based on the equivalent series resistance of the capacitor and the actual capacitance value, classify the state levels and set the corresponding baseline health score for each level; discretize and correct the baseline health score for each level to generate a real health score label; construct a state assessment model, take the comprehensive feature vector as the input of the model, take the corresponding real health score label as the label, calculate the hierarchical adaptive dynamic penalty function and train the model. S4. Predict real-time health scores using a trained state assessment model, and combine the final health scores for the corresponding time period of the previous day to verify the physical rationality of the predicted health scores for the current time period, thus obtaining the final health score. Historical operation and maintenance data is acquired and clustered, and the final health score is corrected based on the clustering results.

2. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 1, characterized in that: In S1, Fourier transform is performed on the ripple voltage signal and ripple current signal for each time period to extract the effective values ​​of the fundamental components of the ripple voltage and the fundamental components of the ripple current, which are used as the amplitude of the ripple voltage and the amplitude of the ripple current for the corresponding time period. For each time period, the difference between the current ambient temperature and the reference temperature in the corresponding time period is multiplied by a predefined temperature compensation coefficient and then 1 is added to obtain the temperature compensation term. The ratio of the ripple voltage amplitude to the ripple current amplitude in the corresponding time period is multiplied by the cosine of the phase difference between the ripple voltage and the ripple current at the ripple fundamental frequency and the temperature compensation term in sequence to obtain the equivalent series resistance of the capacitor after temperature compensation. For each time interval, the ratio of the ripple voltage amplitude to the ripple current amplitude at the corresponding time interval is multiplied in sequence by the sine of the phase difference between the ripple voltage and the ripple current at the ripple fundamental frequency, the ripple fundamental frequency, and 2 Taking the reciprocal of the result as the actual capacitance value at the corresponding time interval.

3. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 1, characterized in that: In S2, the time-domain statistical features include the mean voltage, voltage variance, peak voltage, voltage kurtosis, and voltage margin factor of the bus voltage. Frequency domain spectral entropy characteristics include the power spectral entropy and wavelet packet energy entropy of the ripple voltage signal; The physical degradation slope characteristics include the equivalent series resistance change rate, capacitance decay rate, equivalent series resistance temperature drift rate, equivalent series resistance deviation coefficient, and capacitance deviation coefficient; among which, the equivalent series resistance change rate is calculated based on the initial equivalent series resistance of the capacitor at the time of manufacture and the cumulative operating time of the capacitor; the capacitance decay rate is calculated by dividing the difference between the actual capacitance value and the predefined initial capacitance value in each time period by the corresponding time period; The extracted features are time-aligned and then concatenated to form a comprehensive feature vector.

4. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 3, characterized in that: The process for calculating the equivalent series resistance temperature drift rate, equivalent series resistance deviation factor, and capacitance deviation factor is as follows: For each time period, the difference between the equivalent series resistance of the capacitor after temperature compensation and the equivalent series resistance of the capacitor before temperature compensation is divided by the difference between the current ambient temperature and the reference temperature in the corresponding time period to obtain the equivalent series resistance temperature drift rate. The ratio of the difference between the current equivalent series resistance of the capacitor and the initial equivalent series resistance of the capacitor to the initial equivalent series resistance of the capacitor is calculated to obtain the equivalent series resistance deviation coefficient. The difference between the actual capacitance value and the initial capacitance value during the corresponding time period is divided by the initial capacitance value to obtain the capacitance deviation coefficient.

5. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 1, characterized in that: The process of discretizing the baseline health score for each level in S3 is as follows: Calculate the ratio of the current equivalent series resistance of the capacitor to the initial equivalent series resistance of the capacitor, and subtract 1 to obtain the equivalent resistance influence coefficient; take the opposite of the ratio of the current actual capacitance value to the initial capacitance value and add 1 as the numerator, and use the predefined proportional coefficient as the denominator to calculate the capacitance influence coefficient; take the maximum value of the equivalent resistance influence coefficient and the capacitance influence coefficient as the degradation index; divide the degradation index into multiple state levels according to the predefined state division rules, and discretize the baseline health score under each state level through linear interpolation.

6. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 1, characterized in that: The process of correcting the discretized baseline health score in S3 is as follows: Obtain device status information from historical operation and maintenance records. If the operation and maintenance records confirm that the device is operating normally and without faults, the baseline health score is used as the true health score label. If the operation and maintenance records confirm that the device has defects or has been replaced, the baseline health score is subtracted from the predefined score threshold and the maximum value between the threshold and 0 is used as the true health score label. If the maintenance records confirm that the equipment has undergone power outage maintenance and has been repaired, for the pre-maintenance scheduled period, obtain the number of defect records within the scheduled maintenance duration and take the maximum value between the number of records and the scheduled upper limit; subtract the product of the predefined adjustment coefficient and the maximum value from the predefined upper limit of low health scores to obtain the true health score label for the pre-maintenance scheduled period; for the post-maintenance scheduled period, obtain the number of days between the current maintenance and the previous maintenance of the same type, take the ratio of this interval to the scheduled upper limit of days and take the minimum value between this ratio and the predefined upper limit of maintenance days, subtract the corresponding minimum value from the predefined upper limit of high health scores and multiply it by the predefined amplification coefficient to obtain the true health score label for the post-maintenance scheduled period.

7. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 1, characterized in that: The process of calculating the hierarchical adaptive dynamic penalty function in S3 is as follows: The comprehensive feature vector of each time period is taken as a sample; based on the predicted health score and the real health score label of each sample, the average absolute error during the training process is calculated as the absolute error term; for each sample, the maximum value between 0, the predefined alarm threshold and the real health score label of the corresponding sample is taken as the numerator, the predefined alarm threshold is taken as the denominator, and the ratio of the numerator to the denominator is multiplied by the predefined weight amplification factor and then 1 is added to obtain the dynamic sample importance weight of the corresponding sample. The mean squared error between the predicted health score and the true health score label for each sample is calculated, multiplied by the dynamic sample importance weight, and then summed and multiplied by a predefined weighted squared loss balance coefficient to obtain the weighted squared loss term. The absolute error term and the weighted squared loss term are added together to obtain the hierarchical adaptive dynamic penalty function.

8. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 1, characterized in that: The process of verifying the physical reasonableness of the predicted health score for the current time period in S4 is as follows: The difference between the predicted health score output by the state assessment model in the current time period and the final health score in the corresponding time period of the previous day is taken as the daily score difference. The daily score difference is input into the sign function to obtain the sign coefficient. The absolute value of the daily score difference is subtracted from the predefined maximum allowable jump threshold of the daily score, and the maximum value between it and 0 is taken. This maximum value is then multiplied by the sign coefficient and the predefined constraint strength coefficient in sequence, the inverse number is taken, and then added to the predicted health score output by the state assessment model in the current time period to obtain the final health score in the current time period.

9. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 1, characterized in that: The process of revising the final health score in S4 is as follows: Collect historical operation and maintenance data of the target equipment and convert it into operation and maintenance feature vectors; among which, operation and maintenance features include cumulative number of defects, cumulative number of maintenance, number of days since the last maintenance, level of the most recent defect, and years of equipment operation; Set a predetermined number of clusters, and cluster the maintenance feature vectors of all historical maintenance records. Each cluster corresponds to a status level. The status levels include alarm status, abnormal status, status requiring attention, and healthy status. For the maintenance feature vector of the day, calculate its Euclidean distance to each cluster center, and determine the status level corresponding to the cluster center with the smallest Euclidean distance as the maintenance status level of the current day. Based on the final health score for the current period, determine the corresponding status level as the initial level; If the operation and maintenance status level is the same as or higher than the initial level, it means that the operation and maintenance status level considers the current day's status to be better than the initial level. In this case, the initial level will remain unchanged to prevent operation and maintenance data from masking the true degradation data. If the operation and maintenance status level is lower than the initial level, the level will be downgraded to the operation and maintenance status level.

10. The intelligent state assessment method for a substation capacitor DC power supply system according to claim 9, characterized in that: The secondary correction of the status level after the operation and maintenance data has been corrected is performed as follows: If the growth rate of the equivalent series resistance of the capacitor in the current time period of the day compared with the equivalent series resistance of the capacitor in the corresponding time period of the previous day exceeds the predetermined first growth rate threshold, then the rule correction amount is set to 1; If the actual capacitance value during the current time period of the day is less than the predetermined capacitance threshold, the rule correction amount will be set to 2; If the current ambient temperature at the current time of day exceeds the upper limit of the rated operating temperature of the equipment and the growth rate of the equivalent series resistance of the capacitor is greater than the predetermined second growth rate threshold, the rule correction amount will be set to 1. When multiple rules are triggered simultaneously, the rule correction amounts under the triggering rules are summed to obtain the final rule correction amount; Set the corresponding level number according to the degree of abnormality of the status level from high to low; subtract the final rule correction amount from the level number corresponding to the status level after the operation and maintenance data is corrected, and take the maximum value between it and 1 to obtain the final level number and determine the corresponding final status level.

11. An intelligent condition assessment system for substation capacitor DC power systems, performing the method of any one of claims 1-10, characterized by, include: The data acquisition module collects the bus voltage, load current, ambient temperature, ripple voltage signal, and ripple current signal of the DC power supply system of the substation, and calculates the equivalent series resistance and actual capacitance value of the capacitor for each time period using the ripple vector decomposition method. The multi-dimensional feature extraction module extracts time-domain statistical features, frequency-domain spectral entropy features, and physical degradation slope features from the collected data to generate a comprehensive feature vector. The state assessment model training module divides the state levels and sets the corresponding baseline health scores based on the equivalent series resistance and actual capacitance value; it discretizes and corrects the baseline health scores for each level to generate real health score labels; it constructs a state assessment model, using the comprehensive feature vector as the model input and the corresponding real health score labels as labels, calculates the hierarchical adaptive dynamic penalty function, and trains the model. The status assessment module uses a trained status assessment model to predict health scores in real time. It then combines the final health scores for the corresponding time period of the previous day to verify the physical rationality of the predicted health scores for the current time period, thus obtaining the final health score. Historical operation and maintenance data is acquired and clustered, and the final health score is corrected based on the clustering results.

12. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-10.

13. A computer readable storage medium having stored thereon a computer program, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-10.

Images