A power distribution network operation state evaluation method and system

Abstract

The application discloses a power distribution network operation state evaluation method and system, relates to the technical field of power distribution network operation state evaluation and fault diagnosis, and comprises the following steps: collecting system original signals and extracting fault characteristic parameters; screening the fault characteristic parameters based on an improved characteristic selection algorithm to obtain an optimized characteristic set; performing operation state determination based on the optimized characteristic set, starting a positioning process if the determination result is a fault, and preliminarily positioning a fault area; performing multi-level accurate positioning of a node level and a branch level based on the preliminary positioning result; and fusing the multi-level positioning results to generate a fault positioning report. Through multi-scale time-frequency fusion feature extraction, optimized characteristic selection and a multi-level fault positioning strategy, the application realizes rapid detection and high-precision positioning of faults of a light-storage-straight-flexible power distribution system, and solves the problems that fault signals are difficult to detect and fault areas are difficult to accurately position.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution network operation state evaluation and fault diagnosis, more particularly, the present application relates to a power distribution network operation state evaluation method and system. BACKGROUND

[0002] With the large-scale access of new energy and the development of distributed power supply, energy storage units and flexible load, the operation mode of power distribution network gradually evolves from the traditional one-way power supply to a complex system with high coupling of source, network, load and storage. In particular, the direct current flexible power distribution system combined with photovoltaic and energy storage (referred to as photovoltaic and energy storage direct flexible power distribution system) plays an important role in improving renewable energy consumption capacity and energy utilization efficiency and power quality. However, due to its direct current characteristics and multi-terminal access mode, this kind of system faces new challenges in operation state evaluation and fault protection.

[0003] At present, the operation state evaluation method of power distribution network mostly borrows from the protection and monitoring technology of high-voltage direct-current transmission system or low-voltage alternating-current system. However, due to the characteristics of low voltage level, short line length, small impedance and complex dynamic characteristics of multiple energy units of the photovoltaic and energy storage direct flexible power distribution system, the above-mentioned methods have insufficient applicability in application, especially in operation state evaluation and fault protection, which faces the following outstanding problems:

[0004] On the one hand, it is difficult to detect fault signals. In the alternating current system, the voltage and current signals usually contain power frequency components and rich harmonic information, which can be extracted by Fourier transform, wavelet analysis and other methods to extract fault characteristics. However, in the direct current system, the voltage and current signals do not have power frequency characteristics, the waveform changes suddenly and smoothly, and it is difficult to extract effective features through traditional signal processing methods. When the photovoltaic and energy storage direct flexible power distribution system fails, the voltage drop and current mutation propagate at a very fast speed. If the protection device cannot accurately detect the fault characteristics in a short time, the protection action will be delayed or even failed, thus amplifying the safety risk of the system.

[0005] On the other hand, it is difficult to locate the fault area. In the high-voltage direct-current transmission system, the fault point is usually located by relying on line impedance calculation or traveling wave analysis. However, the photovoltaic and energy storage direct flexible power distribution system has short line length and small impedance value, making it difficult to obtain accurate positioning results by using the above-mentioned methods. At the same time, the multi-source access and bidirectional power flow characteristics make the fault current direction complex and variable, and simply relying on current amplitude or direction criterion cannot reliably determine the fault area, which is easy to cause misjudgment or refusal to act.

[0006] Based on the above problems, the existing technology has defects such as insufficient response speed, imperfect feature extraction and limited positioning accuracy when comprehensively and accurately evaluating the operation state of the photovoltaic and energy storage direct flexible power distribution system, which is difficult to meet the requirements of safe and stable operation of the system. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, embodiments of the present application provide a power distribution network operation state evaluation method and system, which realizes rapid detection and high-precision positioning of faults of a photovoltaic energy storage direct flexible power distribution system through multi-scale time-frequency fusion feature extraction, optimized feature selection and multi-level fault positioning strategies, thereby solving the problems of difficulty in detecting fault signals and difficulty in accurately positioning fault areas.

[0008] In order to achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0009] In a first aspect, the present application provides a power distribution network operation state evaluation method, which comprises: collecting system original signals and extracting fault feature parameters; screening the fault feature parameters based on an improved feature selection algorithm to obtain an optimized feature set; performing operation state determination based on the optimized feature set, and if the determination result is a fault, starting a positioning process to preliminarily position the fault area; based on the preliminary positioning result, performing multi-level accurate positioning at the node level and branch level; and fusing the multi-level positioning results to generate a fault positioning report.

[0010] In one of the embodiments, the system original signals are collected and the fault feature parameters are extracted, specifically as follows: the original signals are preprocessed; a self-adaptive multi-resolution decomposition method is used to perform multi-scale time-domain decomposition on the preprocessed signals to obtain time-domain components; frequency domain analysis is performed on the time-domain components to extract frequency domain features at each time scale and screen out key components; and signal reconstruction is performed based on the key components to generate the fault feature parameters.

[0011] In one of the embodiments, the self-adaptive multi-resolution decomposition method is used to perform multi-scale time-domain decomposition on the preprocessed signals to obtain time-domain components, specifically as follows:

[0012] A plurality of candidate wavelet basis functions are used to preliminarily decompose the preprocessed signals, and evaluation indexes of each decomposition result are calculated, wherein the evaluation indexes include a correlation coefficient, a minimum entropy and an energy aggregation index;

[0013] Based on the weighted comprehensive score of the evaluation indexes, an optimal wavelet basis function is selected from the candidate wavelet basis functions, and the number of wavelet decomposition layers is calculated;

[0014] The preprocessed signals are wavelet packet decomposed according to the optimal wavelet basis function and the number of wavelet decomposition layers to obtain the time-domain components.

[0015] In one of the embodiments, the frequency domain analysis is performed on the time-domain components to extract the frequency domain features at each time scale and screen out the key components, specifically as follows:

[0016] The time-domain components are subjected to overlapping window processing to obtain a plurality of windowed components;

[0017] Performing frequency domain analysis on each windowed component to obtain a frequency domain amplitude spectrum and corresponding frequency points;

[0018] Obtaining frequency domain features from the frequency domain amplitude spectrum and corresponding frequency points, including energy features, frequency center features and harmonic features;

[0019] Based on the frequency domain features, screening each windowed component, and marking the windowed component meeting the preset condition as a key component.

[0020] In one embodiment, the improved feature selection algorithm is used to screen the fault feature parameters to obtain an optimized feature set, specifically:

[0021] Taking the equipment operating state as a target variable and the fault feature parameters as input independent variables, a decision tree model is constructed;

[0022] Based on the decision tree model, the preliminary contribution of each feature is calculated;

[0023] The correlation matrix between the fault feature parameters is calculated;

[0024] According to the correlation matrix and the preliminary contribution, the redundant features are removed;

[0025] The remaining features are re-integrated to form an optimized feature set.

[0026] In one embodiment, the operating state is determined based on the optimized feature set, specifically:

[0027] An operating state evaluation model is constructed, which includes a time series analysis unit and a pattern recognition unit;

[0028] The optimized features are subjected to time series feature extraction to obtain time series features reflecting the dynamic evolution trend of the equipment;

[0029] The time series features and the optimized features are fused to form a fusion feature set containing static features and dynamic features;

[0030] The fusion feature set and the operating state label data are used to train the operating state evaluation model;

[0031] The optimized feature set of the to-be-tested equipment is input into the trained operating state evaluation model to output the operating state category result.

[0032] In one embodiment, if the determination result is a fault, a positioning process is started to preliminarily locate the fault area, specifically:

[0033] Based on the structural layout and sensor distribution of the equipment, the equipment is divided into a plurality of initial areas;

[0034] For each initial region, obtain the internal sensor density and generate a density distribution matrix;

[0035] Based on the density distribution matrix, recursively divide the initial region using a spatial segmentation algorithm to obtain an optimized region division;

[0036] Take the sensors in each optimized region as the center point to construct a spatial cell;

[0037] Map the optimized features to the spatial cell, and distribute the features across multiple cells according to the distance from the center point;

[0038] Based on the mapped results, implement preliminary positioning of the fault region.

[0039] In one embodiment, based on the mapped results, preliminary positioning of the fault region is implemented, specifically:

[0040] Calculate the anomaly index of each spatial cell;

[0041] Construct a spatial cell adjacency matrix and perform neighborhood weighting processing on the anomaly index of each cell;

[0042] Based on the neighborhood weighted anomaly index, construct a probability distribution model;

[0043] According to the preset anomaly threshold and the probability distribution model, identify potential abnormal spatial cells;

[0044] Spatially cluster the potential abnormal spatial cells to form a preliminary positioning fault region set.

[0045] In one embodiment, the fault location report is generated as follows:

[0046] Analyze the anomaly index of each node in the fault region to obtain a node-level fault location list;

[0047] Based on the node-level fault location list, divide the device running path into several branch units;

[0048] Map the anomaly index of the node to the corresponding branch unit and analyze the overall impact of the branch;

[0049] Determine the branch-level fault region according to the overall impact of the branch;

[0050] Fuse the node-level and branch-level fault regions to generate the final fault location and generate the fault location report.

[0051] In a second aspect, the present application provides a power distribution network operation state evaluation system, which comprises: a feature extraction module for collecting system original signals and extracting fault feature parameters;

[0052] A feature screening module is configured to screen fault feature parameters based on an improved feature selection algorithm to obtain an optimized feature set;

[0053] A running state determination and preliminary positioning module is configured to determine a running state based on the optimized feature set, and if the determination result is a fault, to start a positioning process to preliminarily position a fault area;

[0054] A deep positioning module is configured to perform multi-level accurate positioning at a node level and a branch level based on the preliminary positioning result;

[0055] A report generation module is configured to fuse the multi-level positioning result to generate a fault positioning report.

[0056] As can be seen from the above technical solutions, the embodiments of the present application have the following advantages:

[0057] 1. By performing multi-scale time domain and frequency domain fusion analysis on the original signals of the optical storage direct flexible power distribution system, high-confidence fault feature parameters are extracted, and an improved feature selection algorithm is used to screen an optimized feature set. Time series analysis unit and pattern recognition unit are combined to perform dynamic modeling, realizing a complete closed-loop process from feature extraction, optimization, time series modeling to state determination. This scheme can simultaneously capture the transient and steady-state characteristics of device operation, quantify the importance of each fault feature and eliminate redundancy, fuse static and dynamic features to improve the recognition ability of complex operation modes, thereby significantly improving the accuracy, stability and robustness of fault diagnosis, and providing reliable and operable data support for real-time state evaluation and operation and maintenance decision-making, with higher precision, efficiency and practicality than traditional methods.

[0058] 2. By segmenting the fault preliminary positioning of the optical storage direct flexible power distribution system into areas, combining spatial unit mapping, neighborhood weighting and probability model analysis, fine identification of the fault area is realized. Further, through multi-level positioning at the node level and the branch level, and fusing the results of each layer to generate the final fault positioning report, the fault positioning is extended from local anomalies to the overall impact of the system, which can accurately identify key nodes and affected branches, reasonably assess the severity of the fault and the priority processing order, and provide visual analysis and operation and maintenance reference. Compared with traditional single-layer or coarse-grained positioning methods, this scheme significantly improves the positioning accuracy, adaptability and reliability, and provides an efficient and operable solution for real-time fault diagnosis and maintenance decision-making of complex power distribution systems. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A power distribution network running state evaluation method flowchart provided by an embodiment of the present application.

[0060] Figure 2 A power distribution network running state evaluation system structure diagram provided by an embodiment of the present application.

[0061] Figure 3 A multi-scale time-domain component diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0063] Referring to Figure 1 The present application provides a power distribution network operation state evaluation method and a flowchart thereof, which comprises the following steps:

[0064] S1, collecting original signals of the optical storage direct flexible power distribution system, and performing multi-scale time-domain and frequency-domain fusion analysis on the signals to extract fault feature parameters.

[0065] The original signals include node voltage signals and branch current signals, which are obtained by real-time collection through sensors arranged at each node of the optical storage direct flexible power distribution system.

[0066] In the embodiments, the original signals of the optical storage direct flexible power distribution system are collected, and the signals are subjected to multi-scale time-domain and frequency-domain fusion analysis to extract fault feature parameters, specifically as follows:

[0067] The original signals are preprocessed, and the preprocessing includes filtering, denoising and normalization to ensure the reliability and analyzability of the signals.

[0068] The preprocessed signals are subjected to multi-scale time-domain decomposition by an adaptive multi-resolution wavelet decomposition method to obtain time-domain components, and the time-domain components include approximate components and detail components at different time scales.

[0069] The time-domain components are subjected to frequency-domain analysis to extract frequency-domain features at each time scale, and the key components are obtained through screening.

[0070] The key components are subjected to signal reconstruction to obtain fault feature parameters, and the reconstruction method is inverse transformation superposition of the key components, and the fault feature parameters include reconstructed signals, i.e., superposition results of the key components in the time domain, which retain the components that best reflect the fault features and remove irrelevant or noise components, and each key component has feature parameters (energy, energy proportion, frequency center, harmonic amplitude and phase, and harmonic number).

[0071] It should be noted that while retaining the main fault information of the signal, the noise and irrelevant components are effectively removed, and the multi-scale, multi-time scale and multi-frequency dimension comprehensive analysis of the node voltage and branch current signals in the optical storage direct flexible power system is realized. Through adaptive multi-resolution wavelet decomposition to obtain the approximation component and the detail component, and combined with frequency domain feature extraction and key component screening, the signal components related to the fault can be accurately identified, so as to obtain the reconstructed signal with high reliability and key feature parameters (such as energy, frequency center and harmonic information). These fault feature parameters not only reflect the transient and steady state characteristics of the system, but also provide reliable input data for subsequent fault diagnosis and operation state evaluation, making the fault identification more accurate, fast and robust.

[0072] Further, as shown in Figure 3 The preprocessed signal is subjected to multi-scale time domain decomposition by an adaptive multi-resolution wavelet decomposition method to obtain time domain components, specifically:

[0073] A library containing a plurality of candidate wavelet basis functions is established, and the wavelet basis functions include Symlet, Coiflet and Daubechies;

[0074] The preprocessed signal is subjected to preliminary wavelet decomposition using each candidate wavelet basis function in the library;

[0075] The evaluation index of each preliminary wavelet decomposition result is obtained, and the evaluation index includes the correlation coefficient, the minimum entropy and the energy aggregation index;

[0076] The comprehensive score is obtained by weighted summation of each data in the evaluation index, and the candidate wavelet basis functions are comprehensively sorted according to the comprehensive score, and the wavelet basis function with the highest comprehensive score is selected as the optimal wavelet basis function;

[0077] According to the selected optimal wavelet basis function, the wavelet decomposition layer number is calculated based on the signal sampling frequency and the lowest fault frequency;

[0078] The wavelet decomposition layer number is calculated according to the following formula:

[0079]

[0080] In the formula, is the wavelet decomposition layer number, is the signal sampling frequency, is the lowest fault frequency.

[0081] The preprocessed signal is subjected to wavelet packet decomposition according to the optimal wavelet basis function and the decomposition layer number to obtain time domain components, and the time domain components include approximation components and detail components at different time scales.

[0082] The correlation coefficient is used to measure the similarity between the decomposed signal and the original signal, and the specific calculation formula is as follows:

[0083]

[0084] In the formula, is the correlation coefficient, is the value of the original signal at the i th sampling point, is the mean value of the original signal, is the value of the signal component obtained after the preliminary wavelet decomposition at the i th sampling point, is the mean value of the decomposed signal component.

[0085] The minimum entropy is used to measure the concentration degree of energy distribution, and the specific calculation formula is as follows:

[0086]

[0087] In the formula, is the minimum entropy, is the proportion of the energy contained in the decomposed signal component to the total energy of the entire signal.

[0088] The energy aggregation index is used to measure the ability of the component to carry fault feature information, and the specific calculation formula is as follows:

[0089]

[0090] In the formula, is the energy aggregation index, is the coefficient value of the i th sampling point of the component after the preliminary wavelet decomposition, is the square of the coefficient, indicating the energy contained in the signal component at the point, and N is the total number of component coefficients, i.e. the total number of sampling points contained in the component.

[0091] Further, the time domain component is analyzed in the frequency domain to extract the frequency domain features under each time scale, and the key components are obtained by screening, specifically:

[0092] The time domain component is processed by overlapping window to obtain each windowed component, and the window type is selected according to the frequency characteristics of the component, the Hanning window is used for low frequency component, and the Hamming window is used for high frequency component;

[0093] Each windowed component is analyzed in the frequency domain to obtain the frequency domain amplitude spectrum and the corresponding frequency spectrum point, the frequency domain analysis includes fast Fourier transform and variational mode decomposition, the fast Fourier transform is used to obtain the amplitude spectrum of the full frequency band of the component, and the variational mode decomposition is used to further decompose the component into narrowband modes, which is convenient for extracting instantaneous frequency;

[0094] The frequency domain features are obtained according to the frequency domain amplitude spectrum and corresponding frequency points, and the frequency domain features include energy features, frequency center features, and harmonic features (amplitude and phase);

[0095] Based on the frequency domain features, each windowed component is screened, and the screening conditions are specifically as follows:

[0096] 1) The ratio of energy to total frequency domain energy exceeds a preset proportion threshold;

[0097] 2) The frequency center falls within a preset fault frequency band;

[0098] 3) The harmonic amplitude is significantly higher than the background noise (more than 10% of the fundamental frequency amplitude);

[0099] The windowed component satisfying any condition is marked as a key component.

[0100] The energy features are specifically calculated according to the following formula:

[0101]

[0102] In the formula, is the energy of the i-th windowed component, is the complex amplitude of the i-th windowed component at the k-th frequency spectrum point in the frequency domain, and M is the total number of frequency spectrum points.

[0103] The frequency center features are specifically calculated according to the following formula:

[0104]

[0105] In the formula, is the frequency center, is the actual frequency corresponding to the k-th frequency point, wherein, .

[0106] The harmonic amplitude and phase are extracted by the fundamental frequency to extract the amplitude and phase of the integer multiple frequency The extraction method is to locate the point of in the amplitude spectrum, and read the amplitude and phase information, wherein, is the amplitude of the n-th harmonic, reflecting the strength of the harmonic component, is the phase of the n-th harmonic, reflecting the relative position of the harmonic in time.

[0107] S2, based on the improved feature selection algorithm, the fault feature parameters are screened to obtain an optimized feature set, and input to a running state evaluation model for state determination, and the evaluation model includes a time series analysis unit and a pattern recognition unit.

[0108] In this embodiment, the improved feature selection algorithm is used to screen the fault feature parameters to obtain an optimized feature set, specifically as follows:

[0109] The device running state is taken as a target variable, the fault feature parameters are taken as input independent variables, and a decision tree model is constructed based on a random forest algorithm;

[0110] The model training parameter setting includes setting the maximum depth of the tree to 3-10 layers to control the model complexity and prevent overfitting, and setting the minimum sample split number to 5-20 samples to ensure that the data amount of each node is sufficient for reliable splitting. The node splitting condition adopts information gain, Gini coefficient or mean square error (MSE) reduction, and is selected according to the classification or regression task. The training sample division method includes randomly dividing the data into a training set and a validation set with a ratio of 70:30 to ensure the independence of model training and performance verification. For the case of few abnormal samples, oversampling, undersampling or weighting method can be used to balance the training data;

[0111] In the decision tree model, the information gain of the feature splitting of each decision tree node is obtained and accumulated to obtain a preliminary contribution degree;

[0112] The correlation coefficient between any two features is calculated based on the fault feature parameters to obtain a feature correlation matrix, and the correlation coefficient calculation method is Pearson correlation coefficient;

[0113] The feature correlation matrix is analyzed, and if the correlation coefficient between two features is greater than a preset correlation threshold, it is determined that the two features have a redundant relationship;

[0114] For the feature pair with a redundant relationship, the preliminary contribution degrees of the two features are compared, and if the contribution degree score of one of the features is lower than a preset contribution degree threshold, the feature with a low contribution degree is removed;

[0115] The remaining features are re-integrated to form an optimized feature set.

[0116] The information gain of the feature splitting of each decision tree node is obtained and accumulated to obtain a preliminary contribution degree, specifically as follows:

[0117] For the sample set contained in the decision tree node , the parent node entropy Hparent of the node is calculated, which is used to measure the uncertainty of the sample categories in the node;

[0118] The specific calculation formula of the parent node entropy Hparent is as follows:

[0119]

[0120] In the formula,​​ Total number of possible state categories for the target variable, such as normal, abnormal, fault, is the jth class label, is the proportion of samples in the node belonging to the class .

[0121] For each candidate fault feature F in the node , judge its splitting ability, and split the node into sub-node sample number;

[0122] Calculate the sub-node entropy of each sub-node, and obtain the weighted entropy of the fault feature F after the node splitting according to the sub-node sample number ;

[0123] The weighted entropy , the specific calculation formula is as follows:

[0124]

[0125] In the formula, is the sub-node sample number, is the parent node, i.e. the total sample number of the decision tree node, is the total sample number of the sub-node.

[0126] Based on the parent node entropy and the weighted entropy, calculate the information gain of the fault feature F in the node , to quantify the degree of reduction of the uncertainty of the feature in the current node splitting for the determination of the running state;

[0127] The information gain , the specific calculation formula is as follows:

[0128] .

[0129] Cumulate the information gain of each feature in all nodes of the tree to obtain the preliminary contribution degree.

[0130] It should be noted that the improved feature selection algorithm refers to an algorithm optimized by combining information gain accumulation and feature correlation analysis on the basis of traditional feature importance evaluation methods based on decision tree or random forest: first, the uncertainty reduction of each feature in the node splitting for the running state determination is calculated by constructing a random forest decision tree model, and the information gain in the whole tree is accumulated to obtain the preliminary contribution of the feature; then, the Pearson correlation coefficient is used to construct a feature correlation matrix, and the feature pairs with high correlation are analyzed, and the low-contribution redundant features are removed by combining the contribution comparison, thereby forming an optimized feature set. The advantage of this algorithm is that it can not only quantify the actual contribution of each fault feature to the running state determination, realize the importance ranking of the features, but also effectively remove redundant features, improve the model training efficiency and determination accuracy, and at the same time ensure that the selected features are representative and independent of each other, thereby improving the stability and generalization ability of the evaluation model.

[0131] Further, input to the running state evaluation model for state determination, specifically:

[0132] The running state evaluation model is constructed to include a time series analysis unit and a pattern recognition unit, wherein the time series analysis unit is used to extract the dynamic change law of the features, and the pattern recognition unit is used to distinguish different running state categories;

[0133] It should be noted that the time series analysis unit uses a sliding time window segmentation, a difference operation and an exponential weighted moving average method to dynamically process the input features, which are used to extract time sequence features reflecting the trend of device state change, and the pattern recognition unit includes at least one classification model in random forest, support vector machine and deep neural network, which is used to model and identify different running state categories;

[0134] The optimized feature set is input to the time series analysis unit in time sequence, and the optimized feature set is processed by recursive sliding window and difference analysis to extract time sequence features reflecting the dynamic evolution trend of the device;

[0135] Among them, the recursive sliding window processing and difference analysis can be understood as inputting the optimized feature set to the time series analysis unit in the order of sampling time, segmenting each feature parameter, calculating the mean, standard deviation, skewness and kurtosis, and removing the trend item by difference or filtering method to obtain the dynamic change feature sequence of the device;

[0136] The time sequence features are fused with the optimized feature set to form a fusion feature set containing static features and dynamic features, which provides more comprehensive input for the pattern recognition unit;

[0137] The fusion feature set and the running state label data are input into a pattern recognition unit, and a model is trained through a supervised learning algorithm, so that the model can learn the feature patterns of different running states.

[0138] The running state label data refers to the real class information of the running state manually or automatically labeled in the model training stage, so that the recognition model can learn the distribution law of the features in different states. For example, the normal running state: the parameters of the device are in the standard range, and are labeled as "normal"; the slight abnormal state: some parameters deviate but do not affect the running, and are labeled as "abnormal"; the fault states A, B and C: specific labeling according to the actual fault type, such as "bearing wear", "motor overheating" and "sensor failure". According to the device running log and maintenance record, the data segments corresponding to the normal running, slight abnormality and various faults are labeled as different state labels; for samples lacking labels, the corresponding running state label data can be generated by setting a threshold or manually testing.

[0139] The optimized feature set of the device to be tested is input into the trained pattern recognition unit, the running state of the current device is determined based on the trained classification boundary, and the corresponding state class label is output.

[0140] It should be noted that the introduction of the time series analysis unit can capture the dynamic evolution law of the running parameters changing with time through sliding window, difference and exponential weighted average processing methods, avoid the loss of time sequence information caused by only using static features, realize information complementation in state recognition by fusing static features and dynamic features, improve the perception ability of the model to complex running modes, and form a closed loop process from feature optimization-time series modeling-pattern recognition-state determination, realize higher accuracy, stability and practicality, and have obvious creativity and application value.

[0141] S3, if the determination result is a fault, a positioning process is started, and the fault region is preliminarily positioned based on a region segmentation method.

[0142] In this embodiment, if the determination result is a fault, a positioning process is started, and the fault region is preliminarily positioned based on a region segmentation method, specifically;

[0143] In the case that the determination result of the device running state evaluation model is a fault, the optimized feature set and its related sensor distribution data of the corresponding time period are collected to provide basic information for fault positioning;

[0144] Based on the structure layout and sensor distribution of the device, the device is divided into a plurality of initial regions, and each initial region contains at least one sensor measurement point.

[0145] For each initial region, obtain its internal sensor density, and generate a density distribution matrix to quantify the partitioning requirement of each region;

[0146] Take the initial region and its density distribution matrix as input, select the quadtree algorithm, and divide each initial region as a root node;

[0147] For each region to be partitioned, determine whether to perform recursive partitioning according to its sensor density;

[0148] If the internal sensor density is higher than the preset density threshold, the region is divided into four sub-regions, and each sub-region is recursively judged;

[0149] If the internal sensor density is lower than the preset density threshold, the region is retained and no longer divided;

[0150] Among them, by dynamically adjusting the partitioning depth, high-density concentrated regions are fine grids, and low-density regions are coarse grids;

[0151] After completing the recursive partitioning of all regions, the optimized region partitioning is obtained, the sensors within each optimized region are taken as the center points of the Voronoi diagram, and the spatial units covering the entire device are constructed, so that each spatial unit corresponds to the nearest sensor, thereby improving the spatial partitioning accuracy of the irregular layout device;

[0152] Map the fault-related optimized feature set to the spatial unit, and distribute the features across multiple units according to the distance from the center point by weight to obtain the feature parameters and their weight information contained in each unit;

[0153]

[0154] wherein, is the feature The weight of the spatial unit j is is the Euclidean distance of the feature to the center of the unit, is the sum of the inverse of the distance of feature i to all related spatial units, used for normalization.

[0155] Based on the mapped spatial unit, the fault region is preliminarily located.

[0156] It should be noted that Voronoi is a spatial division method, and the spatial unit constructed by the Voronoi diagram is a spatial region automatically divided according to the "nearest distance" principle with the sensor position as the center, and any point in each region is closer to its corresponding sensor than other regions. In this way, in the case of complex device structure and uneven sensor distribution, spatial division covering the whole device can be adaptively generated, so that high-density areas form fine grids, sparse areas form coarse grids, and fault features can be weighted and mapped to the corresponding area according to the distance from the center point, thereby realizing accurate preliminary positioning of the fault area, thereby providing quantifiable, continuous and operable input for subsequent regional fault positioning, and improving the accuracy, reliability and adaptability of fault positioning as a whole.

[0157] Further, based on the mapped spatial unit, the fault area is preliminarily positioned, specifically:

[0158] For each spatial unit, the abnormal index is calculated by fusing the statistical features (mean, variance, kurtosis, skewness) and dynamic features (sliding window trend, difference features) and frequency domain features (FFT energy, wavelet energy).

[0159] The specific calculation formula of the abnormal index is as follows:

[0160]

[0161] In the formula, is a weighted abnormal index, is the i-th feature of unit j, is a feature weight, which is calculated by PCA or random forest feature importance and standardized, and comprehensively reflects the abnormal degree of the unit relative to the normal state.

[0162] The spatial unit adjacency matrix is constructed, and the abnormal index of each unit is weighted by the neighborhood.

[0163] The specific calculation formula of the neighborhood weighting is as follows:

[0164]

[0165] In the formula, is the abnormal index after neighborhood weighting, is the proportion of the contribution of the abnormal index of the unit itself and the neighborhood unit, is the neighborhood unit set of the j-th unit, is the weight of the neighborhood unit u to the unit j, is the weighted sum of the neighborhood abnormal index.

[0166] Based on the neighborhood weighted abnormal index, a probability distribution model of the abnormal index distribution of each spatial unit is established through a Gaussian mixture model, and the model can fit the probability distribution characteristics of the unit abnormal index;

[0167] An abnormal threshold is obtained based on the probability model, and the abnormal threshold is determined by the statistical distribution of the historical operation data or the expected value and standard deviation of the model prediction, so that the threshold can dynamically adapt to the device state change;

[0168] The neighborhood weighted abnormal index of each spatial unit is compared with the abnormal threshold, the fault probability is obtained by using the probability model, and the potential abnormal spatial unit is determined according to a preset probability threshold. The potential abnormal spatial unit is the potential abnormal spatial unit whose fault probability exceeds the probability threshold;

[0169] The fault probability is calculated according to the following formula:

[0170]

[0171] In the formula, yc is the abnormal threshold.

[0172] The potential abnormal spatial unit is spatially clustered to form a preliminary positioning fault region set, and the fault region set includes the spatial position of each region, the units contained therein and the comprehensive abnormal index thereof.

[0173] S4, based on the preliminary positioning fault region, multi-level positioning is performed at the node level and the branch level, and the positioning results of each layer are fused to generate a final fault positioning report.

[0174] In this embodiment, based on the preliminary positioning fault region, multi-level positioning is performed at the node level and the branch level, and the positioning results of each layer are fused to generate a final fault positioning report. Specifically,

[0175] Based on each preliminary positioning fault region, the abnormal index of each unit in the region is analyzed, the abnormal degree of each key node (including a sensor point, a control node or a functional node) is determined through weighted fusion and sorting, and a node-level fault position list is formed. The weight can be set in combination with the abnormal index value of the node in the region and the spatial or functional connection relationship between the nodes;

[0176] Based on the node-level fault position list, the device running path is divided into branch units, each branch unit contains one or more key nodes, and the nodes maintain physical connection or functional association relationship;

[0177] The abnormal index of each key node is mapped to the branch unit to which it belongs, and the overall affected degree of the branch is analyzed according to the number, distribution density and abnormal index of the node abnormality in the branch;

[0178] Wherein, judging the overall affected degree of the branch can be understood as the proportion of the number of abnormal nodes to the total nodes, and the overall affected degree of the branch is judged according to the proportion: the proportion is high and the first threshold value is preset → high proportion impact, the proportion is low and the second threshold value is preset → low proportion impact, and the proportion is between the first threshold value and the second threshold value → medium proportion impact; the distribution density of abnormal nodes in the branch, abnormal nodes adjacent or continuous → high concentration, overall impact is large, abnormal nodes are large and the preset interval → local impact, overall impact is small, abnormal nodes are low and the preset interval and are not adjacent or continuous → medium concentration; the abnormal index is large and the first intensity threshold value is preset for high intensity impact, and the abnormal index is less than the second intensity threshold value for low intensity impact, and the abnormal index is between the first intensity threshold value and the second intensity threshold value for medium intensity impact; high proportion + high concentration + high intensity → high impact, medium proportion + medium concentration + medium intensity → medium impact, and low proportion + dispersion + low intensity → low impact;

[0179] According to the overall affected degree of the branch, the branches are sorted according to high, medium and low, the branch-level fault area is determined, and the priority processing order is determined;

[0180] The node-level fault position list and the branch-level fault area are fused to generate a final fault position set;

[0181] According to the final fault position set, a fault positioning report is generated, and the report includes: fault position and its corresponding relationship with the preliminary fault area; node-level abnormality analysis description, including abnormal trend and priority; branch-level abnormality analysis description, including abnormal node distribution and potential impact range; maintenance priority suggestion and operation reference.

[0182] It should be noted that by extending the analysis of the preliminary fault area to multi-level positioning of the node level and the branch level, and fusing the positioning results of each layer to generate a final fault positioning report, a comprehensive evaluation from local anomaly to overall system impact is realized. It not only can accurately identify key nodes and affected branches, reasonably judge the fault severity and priority processing order, but also can provide clear visual report and maintenance reference, effectively improve the accuracy, efficiency and operability of fault positioning, and significantly enhance the comprehensiveness, reliability and practicality of system diagnosis compared with traditional single-layer fault analysis method.

[0183] Referring to Figure 2 The power distribution network operation state evaluation system structure schematic diagram provided by the application includes a feature extraction module, a feature screening module, an operation state determination and preliminary positioning module, a deep positioning module and a report generation module, and there is a connection between the modules:

[0184] The feature extraction module is used for collecting system original signals and extracting fault feature parameters;

[0185] a feature screening module configured to screen the fault feature parameters based on an improved feature selection algorithm to obtain an optimized feature set;

[0186] a running state determination and preliminary positioning module configured to determine a running state based on the optimized feature set, and if the determination result is a fault, start a positioning process to preliminarily position a fault area;

[0187] a deep positioning module configured to perform multi-level accurate positioning at a node level and a branch level based on the preliminary positioning result;

[0188] a report generation module configured to fuse the multi-level positioning result to generate a fault positioning report.

[0189] The above formulas are all dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the nearest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0190] The above embodiments can be realized wholly or partially by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product.

[0191] Those skilled in the art can realize that the modules and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solutions. A person skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0192] In addition, the functional modules in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0193] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0194] Finally, the above is only the preferred embodiments of the present application, and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A power distribution network operating state assessment method, characterized by, The method comprises the following steps: Collecting system original signals and extracting fault characteristic parameters; Screening the fault characteristic parameters based on an improved feature selection algorithm to obtain an optimized feature set, specifically: Taking the equipment operating state as a target variable and the fault characteristic parameters as input independent variables, a decision tree model is constructed; Calculating the preliminary contribution of each feature based on the decision tree model; Calculating the correlation matrix between the fault characteristic parameters; According to the correlation matrix and the preliminary contribution, redundant features are removed; The remaining features are re-integrated to form an optimized feature set; Based on the optimized feature set, the operating state is determined, and if the determination result is a fault, the positioning process is started to preliminarily locate the fault area; Based on the preliminary positioning result, multi-level accurate positioning of the node level and branch level is performed; The multi-level positioning results are fused to generate a fault positioning report.

2. The power distribution network operating state assessment method according to claim 1, characterized by, The collecting system original signals and extracting fault characteristic parameters are specifically: Pretreating the original signals; Using an adaptive multi-resolution decomposition method to perform multi-scale time domain decomposition on the pretreated signals to obtain time domain components; Performing frequency domain analysis on the time domain components, extracting frequency domain features at each time scale, and screening out key components; Based on the key components, signal reconstruction is performed to generate fault characteristic parameters.

3. The power distribution network operating state assessment method according to claim 2, characterized in that, The adaptive multi-resolution decomposition method is used to perform multi-scale time domain decomposition on the pretreated signals to obtain time domain components, specifically: A plurality of candidate wavelet basis functions are used to preliminarily decompose the pretreated signals, and evaluation indexes of each decomposition result are calculated, the evaluation indexes including correlation coefficients, minimum entropies, and energy aggregation indexes; Based on the weighted comprehensive score of the evaluation indexes, the optimal wavelet basis function is selected from the candidate wavelet basis functions, and the wavelet decomposition level is calculated; According to the optimal wavelet basis function and the wavelet decomposition level, wavelet packet decomposition is performed on the pretreated signals to obtain time domain components.

4. The power distribution network operating state assessment method according to claim 3, characterized by, The frequency domain analysis on the time domain components, the extraction of frequency domain features at each time scale, and the screening out of key components are specifically: The time domain components are subjected to overlapping window processing to obtain a plurality of windowed components; Each windowed component is subjected to frequency domain analysis to obtain a frequency domain amplitude spectrum and a corresponding frequency point; According to the frequency domain amplitude spectrum and the corresponding frequency point, frequency domain features are obtained, including energy features, frequency center features, and harmonic features; Based on the frequency domain features, each windowed component is screened, and the windowed components meeting the preset conditions are marked as key components.

5. The power distribution network operating state assessment method according to claim 1, characterized by, The operating state determination based on the optimized feature set is specifically: An operating state evaluation model is constructed, which includes a time series analysis unit and a pattern recognition unit; Time series features are extracted from the optimized features to obtain time series features reflecting the dynamic evolution trend of the equipment; The time series features and the optimized features are fused to form a fusion feature set containing static features and dynamic features; The fusion feature set and the operating state label data are used to train the operating state evaluation model; The optimized feature set of the equipment to be tested is input into the trained operating state evaluation model to output an operating state category result.

6. The power distribution network operating state assessment method according to claim 5, characterized by, If the determination result is a fault, the positioning process is started to preliminarily locate the fault area, specifically: Based on the structural layout and sensor distribution of the device, the device is divided into several initial regions; For each initial region, obtain its internal sensor density and generate a density distribution matrix; Based on the density distribution matrix, use a spatial segmentation algorithm to recursively divide the initial region and obtain an optimized region division; Taking the sensors in each optimized region as the center point, construct a spatial cell; Map the optimized features to the spatial cell and distribute the features across multiple cells according to the distance from the center point; Based on the mapping results, the preliminary positioning of the fault region is realized.

7. The power distribution network operating state assessment method according to claim 6, characterized by, The preliminary positioning of the fault region based on the mapping results is specifically: Calculate the anomaly index of each spatial cell; Construct a spatial cell adjacency matrix and perform neighborhood weighting processing on the anomaly index of each cell; Based on the neighborhood weighted anomaly index, construct a probability distribution model; According to the preset anomaly threshold and the probability distribution model, identify the potential abnormal spatial cell; Spatially cluster the potential abnormal spatial cells to form a preliminary positioning fault region set.

8. The power distribution network operating state assessment method according to claim 7, characterized by, The fault location report is generated in the following steps: Analyze the anomaly index of each node in the fault region to obtain a node-level fault location list; Based on the node-level fault location list, divide the device running path into several branch units; Map the anomaly index of the node to the corresponding branch unit and analyze the overall impact of the branch; Determine the branch-level fault region according to the overall impact of the branch; Fuse the node-level and branch-level fault regions to generate the final fault location and generate the fault location report.

9. A system for operating state evaluation of a power distribution network using the method according to any one of claims 1 to 8, characterized in that It includes: A feature extraction module for collecting system raw signals and extracting fault feature parameters; A feature selection module for filtering fault feature parameters based on an improved feature selection algorithm to obtain an optimized feature set; An operating state determination and preliminary positioning module for determining the operating state based on the optimized feature set, and if the determination result is a fault, starting the positioning process to preliminarily locate the fault region; A deep positioning module for performing multi-level accurate positioning of node-level and branch-level based on the preliminary positioning results; A report generation module for fusing multi-level positioning results to generate a fault location report.

Images