Search

CN-122000839-A - Intelligent identification linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network

CN122000839ACN 122000839 ACN122000839 ACN 122000839ACN-122000839-A

Abstract

The invention discloses an intelligent identification linkage protection method for equipment of 35kV and below based on multi-source adaptive fusion and deep-shallow dual-path network. The method comprises the steps of collecting multi-source data of target equipment, splitting the multi-source data into instantaneous and steady-state feature flows, respectively extracting features by utilizing a deep-shallow dual-path network, carrying out dynamic self-adaptive fusion according to a real-time operation scene to obtain comprehensive features, inputting a time sequence formed by continuous comprehensive features into an LSTM network which is subjected to equipment exclusive gating regulation, combining a layered time sequence matching mechanism to output a combined identification result of equipment types and working states and initial confidence, calibrating the confidence according to feature fluctuation and sample difficulty, and finally grading, linkage and adjustment of a protection strategy according to the calibrated confidence. The invention improves the adaptability of dynamic scenes and the recognition accuracy of the mass equipment, effectively reduces the risk of misjudgment and missed judgment through the confidence calibration and dynamic constant value linkage, and enhances the reliability of a protection system.

Inventors

  • Shi Bendong
  • Luo Qiluo
  • MA GUIBIN
  • LIAN ZHIWEI
  • XU MAOLIN
  • Xing Mengxiao
  • ZHANG PENGFEI
  • WEI KEHUA

Assignees

  • 烟台东方威思顿电气有限公司

Dates

Publication Date
20260508
Application Date
20260123

Claims (10)

  1. 1. The intelligent identification linkage protection method for the equipment of 35kV and below based on multi-source adaptive fusion and deep-shallow dual-path networks is characterized in that steps S1 to S5 are circularly executed according to preset sampling frequency, the comprehensive characteristics of target equipment at the current moment are respectively obtained in each execution period, and the comprehensive characteristics of a plurality of continuous execution periods form a time sequence characteristic sequence for subsequent time sequence identification and protection judgment; S1, collecting multi-source data of target equipment to obtain multi-dimensional characteristics; S2, performing dimension optimization and feature splitting treatment on the multidimensional features to obtain an instantaneous feature stream and a steady-state feature stream; S3, inputting the instantaneous characteristic flow and the steady-state characteristic flow into a deep-shallow dual-path network, extracting instantaneous mutation characteristics from the instantaneous characteristic flow through a shallow path, and extracting steady-state global characteristics from the steady-state characteristic flow through a deep path; S4, according to a real-time operation scene of the target equipment, carrying out dynamic self-adaptive fusion on the transient mutation characteristic and the steady global characteristic to obtain the comprehensive characteristic of the target equipment at the current moment; S5, inputting a time sequence feature sequence formed by continuous comprehensive features at a plurality of moments into an LSTM network, enhancing time sequence relevance through a layered time sequence matching mechanism, and outputting a combined recognition result of 'equipment type-working state' and corresponding initial confidence; S6, judging whether the initial confidence coefficient is over-confident or under-sensitive according to the fluctuation degree of the transient mutation characteristic and the variance of the steady global characteristic, dynamically calculating a penalty coefficient by combining a sample difficulty coefficient, and calibrating the initial confidence coefficient to obtain a calibrated confidence coefficient; and S7, dynamically adjusting the protection strategy of the target equipment according to the calibrated confidence coefficient.
  2. 2. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 1, wherein the multi-source data collected in the step S1 is 18-dimensional in total: (1) Electrical characteristics, 10 dimensions total; Wherein the static characteristic comprises a three-phase current effective value of the target device 、 、 Effective value of line voltage 、 、 Power factor of Zero sequence current Dynamic characteristics including three-phase current change rate maximum value Coefficient of voltage fluctuation ; Wherein the voltage fluctuation coefficient The calculation formula of (2) is as follows: ; In the above-mentioned method, the step of, And The maximum value and the minimum value of the line voltage of a single target device for a plurality of continuous sampling periods respectively, A nominal line voltage for the target device; (2) Test data, 4-dimensional in total, including insulation resistance Dielectric loss tangent value DC resistor No-load loss ; (3) Tour data, 4-dimensional in total, including device appearance status Sound of operation State of oil leakage And meter display status 。
  3. 3. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 2, wherein the step S2 specifically comprises: S2.1, dividing the 18-dimensional features obtained in the step S1 into a maximum index and a minimum index according to the influence direction of the feature indexes on the equipment state, and respectively carrying out standardization treatment; S2.2, performing nonlinear dimension reduction on the normalized 18-dimensional features by adopting kernel principal component analysis, and reserving principal components with accumulated contribution rate not lower than a preset proportion to obtain KPCA fusion features; S2.3, obtaining an instantaneous characteristic stream according to the standardized characteristic index and the principal components obtained by analysis of the nuclear principal components And steady state feature flow ; Instantaneous feature flow In total 7 dimensions, comprising normalized 、 、 Also comprises a start feature entropy 3 Transient principal components with highest contribution rate; Wherein, start-up characteristic entropy Is obtained according to the following calculation formula in the equipment starting stage: ; In the above-mentioned method, the step of, For the sampling sequence number of the start-up phase, Is the first Current normalized probability for each sampling period, which is obtained by averaging three-phase current effective values of the period Calculated to be, i.e ; Steady state feature flow In total 7 dimensions, comprising normalized 、 、 、 、 、 And a steady-state principal component with the highest contribution rate.
  4. 4. The intelligent recognition linkage protection method for the equipment of 35kV and below based on the multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 1, wherein the feature extraction process of the step S3 is as follows: step S3.1, the instantaneous characteristic stream obtained in the step S2 is processed The shallow path of the input depth dual path network consists of 3 light convolution layers, and the final output transient abrupt change characteristic is recorded as ; Step S3.2, the steady-state characteristic stream obtained in the step S2 is processed Inputting deep path of deep-shallow dual-path network, first making preliminary treatment by 2-layer depth separable convolution, and output is recorded as And then Deep feature extraction is carried out through a 3-layer pre-activated residual block, and output is recorded as Finally, passing through a feature filter pair Weighting to obtain steady global features 。
  5. 5. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 4, wherein the importance weight formula of the steady-state feature calculated by the feature filter is as follows: ; In the above-mentioned method, the step of, For the sigmoid activation function, And To train the learned parameters, weight vectors Dimension and dimension In accordance with the method, the device and the system, Representation of Is the first of (2) Individual elements, i.e. the first Maintaining importance weights of steady-state features; weighted steady state global features Calculated by the element product: Wherein Representing element-wise multiplication.
  6. 6. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 3, wherein the step S4 specifically comprises: s4.1, calculating the load density of the real-time operation scene parameters And degree of interference Determining the current scene type; Load density The calculation formula of (2) is as follows: ; The value range of (2) is 0, 2; Interference level The calculation formula of (2) is as follows: ; the value range of (2) is 0, 1; According to the load density And degree of interference Determining the current scene type; step S4.2, scene-based template library Dynamic calculation of instantaneous feature fusion weights ; The calculation formula is as follows: ; In the above-mentioned method, the step of, Representing the calculation of the entropy of the mutual information, Characterization of the current 7-dimensional transient mutations With the current scene type In scene template library Corresponding 7-dimensional instantaneous template in (3) Is used for the mutual information entropy of (1), Representing current 7-dimensional steady-state global features With the current scene type In scene template library Corresponding 7-dimensional steady-state template Mutual information entropy of (2); S4.3, fusing weights according to the instantaneous characteristics For transient mutation characteristics And steady-state global features Feature fusion is carried out to generate comprehensive features ; The fusion mode is as follows: ; In the above-mentioned method, the step of, Representing a1 x 1 convolution operation for unifying feature dimensions; Is a residual term, and the calculation formula is ; The scene template library The acquisition mode of (a) is as follows: For each type of scene, 7-dimensional instantaneous characteristic flows and 7-dimensional steady-state characteristic flows of at least 1000 groups of target devices are collected, then each group of mutually corresponding instantaneous characteristic flows and steady-state characteristic flows are fused into 7-dimensional pseudo-mixed characteristics by weighting, then the pseudo-mixed characteristics are lifted to 128 dimensions by a 1X 1 convolution layer to obtain 128-dimensional pseudo-fusion characteristics, finally, K-means clustering is respectively carried out on the 128-dimensional pseudo-fusion characteristics, 7-dimensional instantaneous characteristic flows and 7-dimensional steady-state characteristic flows in each type of scene, and clustering parameters K=1 to obtain 128-dimensional fusion templates, 7-dimensional instantaneous templates and 7-dimensional steady-state templates corresponding to each type of scene, thereby forming a scene template library 。
  7. 7. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 1, wherein the step S5 specifically comprises: s5.1, constructing a time sequence characteristic sequence; continuous acquisition target device 128-Dimensional integrated features of frames Constitute a time sequence feature sequence Wherein the superscript is a frame number; s5.2, prejudging the equipment type of the current target equipment by using an equipment type classifier; The integrated features of any frame are first added before the time series is input into LSTM Inputting a lightweight device type classifier which is a fully connected network comprising two hidden layers, and finally outputting probability distribution of the type of the target device through a Softmax layer Taking out The type corresponding to the maximum value in the middle is used as a device type pre-judging result; step S5.3, sequence the time sequence characteristics Inputting an LSTM network, and performing equipment exclusive gating adjustment according to the type of the prejudged equipment; The main structure of the LSTM network is unified, but the internal control parameters of the LSTM network are dynamically adjusted according to the device type pre-judging result obtained in step S5.2, and the specific adjustment mode is as follows: (1) If the motor is prejudged, strengthening a forgetting door and introducing a load adaptation factor , Reinforced forgetting door for corresponding moment load density The calculation formula of (2) is as follows: ; (2) If the transformer is pre-determined, the input gate is strengthened, and an interference adaptation factor is introduced , For interference level, the reinforced input door The calculation formula of (2) is as follows: ; In the above-mentioned method, the step of, The LSTM hidden state at time t-1, As an input feature at the time t, 、 、 、 In order to train the learned parameters, For the purpose of batch normalization, Activating a function for sigmoid; (3) For the capacitor and the reactor, universal gating parameters are adopted; S5.4, introducing a hierarchical time sequence matching mechanism in the LSTM network processing process; Specifically, the length is Dividing the time sequence feature sequence of the frame into a plurality of time window levels of short term, medium term, long term and the like, calculating cosine similarity between frames by adopting an attention mechanism in each level, and only reserving strong association relations with the similarity not lower than a preset relationship threshold value so as to reduce cross-level interference; s5.5, obtaining a combined recognition result and initial confidence coefficient based on the LSTM network; The final output layer of the LSTM network is a Softmax layer whose output is the probability distribution of the "device type-active state" combination class, the probability distribution comprising The value is taken as the initial confidence coefficient The corresponding equipment type-working state combination is the combination identification result of the current target equipment.
  8. 8. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 6, wherein the step S6 specifically comprises: s6.1, performing over-confidence and under-sensitivity judgment; Setting a sliding window containing continuous 5 frames of data, and calculating 7-dimensional instantaneous mutation characteristics in the window Degree of fluctuation of (a) And 7-dimensional steady state global features Variance of (2) ; Wherein the degree of temporal characteristic fluctuation The calculation formula of (2) is as follows: ; In the above-mentioned method, the step of, 、 、 Respectively in the sliding window Maximum, minimum and mean of (a); the rules of the over-confidence and under-sensitivity judgment are as follows, if Greater than a first threshold and initial confidence If the value is greater than the second threshold value, the instantaneous characteristic is judged to be confidence, if Less than a third threshold value and If the threshold value is smaller than the fourth threshold value, judging that the steady-state characteristic is not sensitive; Step S6.2, comprehensive characteristics and scene template library based on the current moment Calculating a sample difficulty coefficient : ; In the above-mentioned method, the step of, For the 128-dimensional composite feature at the current time, For the current scene type In scene template library In the corresponding 128-dimensional fusion template, Representing mutual information entropy calculation; step S6.3, based on the over-confidence and under-sensitivity determination result and the sample difficulty coefficient Calculating penalty coefficients And calibrating the initial confidence; Penalty coefficient The calculation rule of (2) is as follows: If the instant characteristic confidence condition is: ; If the condition is "steady state feature undersensitivity": ; if it is normal condition: ; In the above-mentioned method, the step of, The calculation method comprises the steps of setting a sliding window containing continuous 5 frames of data, calculating standard deviations of all dimensions of the sliding window based on KPCA fusion features obtained by using kernel principal component analysis in step S2, and taking the average value of the standard deviations of all dimensions as the overall characteristic fluctuation degree ; Obtaining penalty coefficients After that, for the initial confidence And (3) performing calibration: 。
  9. 9. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network as claimed in claim 1, wherein the step S7 comprises the following steps: (1) If the working state in the combined recognition result is normal operation, the protection fixed value of the target equipment for triggering the protection measures is not adjusted; (2) If the working state in the combined recognition result is a fault state, the working state is calibrated according to the confidence coefficient To execute a hierarchical linkage strategy: (2.1) when The dynamic constant value is used as a protection constant value of target equipment for triggering protection measures, and the dynamic constant value is calculated by the product of the reference constant value, the equipment health index and the scene coefficient; Wherein: the reference fixed value is a rated protection threshold value set when the equipment leaves the factory; Device health index The calculation formula of (2) is as follows: ; In the above-mentioned method, the step of, And Respectively obtaining time attenuation weighted fusion values of insulation resistance and dielectric loss tangent in test data in multi-source data; Scene coefficient The calculation formula of (2) is as follows: ; In the above-mentioned method, the step of, Is the device load density; The action logic in the situation is that for the overcurrent and overvoltage faults, if the corresponding electric quantity exceeds the time of the dynamic fixed value by not less than 2ms, the tripping is immediately carried out, and for the overload and insulation faults, if the corresponding electric quantity exceeds the time of the dynamic fixed value by not less than 5 seconds, the tripping is carried out; (2.2) when When the method is used, a default setting value is adopted as a protection setting value for triggering protection measures, wherein the default setting value is a threshold value preset by on-site electric personnel and is higher than a reference setting value; the specific measures are as follows: firstly loading a default setting value, then checking quantized inspection data, triggering an acousto-optic alarm if the inspection data prompts faults, and automatically switching to a general setting value if the inspection data is not updated for more than 24 hours, wherein the general setting value is a preset universal protection threshold value in the industry; The action logic under the condition is that for the overcurrent and overvoltage faults, if the corresponding electric quantity exceeds the default value for not less than 5ms, the tripping is immediately carried out; (2.3) when When the emergency alarm is triggered, locking the protection fixed value triggering the protection measure into a general fixed value, and triggering the emergency alarm; If the corresponding electric quantity exceeds the general constant value, the operation and maintenance personnel need to trip after remote or on-site confirmation within 30 seconds, and trip automatically after 30 seconds.
  10. 10. The intelligent recognition linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path network as claimed in claim 1, wherein the training process of the deep and shallow dual-path network is supervised by adopting a multi-task joint loss function, and the loss function is used The calculation formula of (2) is as follows: ; In the above-mentioned method, the step of, In order to classify the loss of the device, In order to characterize the loss of importance of a feature, Calibrating the loss for the confidence; 、 、 the task weight coefficient is; (1) Classification loss The multi-classification cross entropy loss is combined by adopting equipment types and equipment states, and the expression is as follows: ; In the above-mentioned method, the step of, The number of samples for a single training; For the total number of federated categories of device types and device states, For the sample Belonging to the combined category Adopts single-hot coding; predictive samples output for the LSTM network final Softmax layer Belonging to the combined category Is a raw probability of (1); (2) Loss of feature importance The expression of (2) is: ; In the above-mentioned method, the step of, Is a steady state feature stream Dimension number of (a); Is based on a sample The first obtained in the process of calculating the steady-state global features Maintaining importance weights of steady-state features; Is the first The dimensional steady-state characteristics are based on target weights preset by the physical characteristics of the equipment; (3) Confidence calibration loss The expression of (2) is: ; In the above-mentioned method, the step of, The confidence level is divided into boxes; Is the first The number of samples in each bin; Is the first Average raw confidence of samples in individual bins; Is the first The actual recognition accuracy of the samples in the individual bins.

Description

Intelligent identification linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep-shallow dual-path network Technical Field The invention relates to the technical field of power system protection and operation and maintenance, in particular to an intelligent identification and self-adaptive protection method for power distribution equipment of 35kV and below. Background In the field of power system protection, conventional relay protection methods, such as overcurrent, overvoltage, overload, low-cycle load shedding, low-voltage load shedding, and the like, are generally implemented by relying on preset static fixed values and a throw-out and throw-back plate. The method takes the basic electric quantity such as current, voltage and the like collected on site as a criterion, and is set once when the equipment is put into operation, and then the equipment runs for a long time. However, conventional static modes fail to adequately account for dynamic changes in the plant's own state parameters (e.g., insulation resistance, dielectric loss angle) and system operating conditions (e.g., upper and lower loop parameters, load levels) during long-term operation. Therefore, the method is difficult to adapt to long-term operation scenes such as equipment aging, load fluctuation and the like, so that the fault identification accuracy rate is low for the small-scale equipment with special parameters such as reactors or unobvious operation characteristics. To improve the fault recognition accuracy, several improvements have been proposed by those skilled in the art. One idea is to use a fault diagnosis method based on a convolutional neural network, and extract features through a single-path convolutional operation to identify. Although the accuracy is improved to a certain extent, the network structure is difficult to simultaneously meet the dual requirements of the quick action of fault removal and the accuracy of fault discrimination, and about 15-20% of faults can not be effectively identified according to actual application feedback. Another concept is to attempt to introduce a feature contribution degree allocation mechanism with fixed weight to determine faults, aiming at strengthening the effect of key features. However, the fixed weight distribution mode has poor adaptability to dynamic operation scenes such as severe load fluctuation caused by starting and stopping of a motor and the like, and power grid harmonic interference and the like. Moreover, the method generally only depends on the fluctuation degree of a single feature to correct the confidence degree of the model output, so that the 'over-confidence' misjudgment is easy to occur due to the instantaneous mutation of the feature, or the 'under-sensitivity' misjudgment is easy to occur due to the slow change of the steady-state feature, and the overall misjudgment rate is high. In addition, the fixed value in the existing protection strategy is mostly static preset, and the fixed value cannot be linked and dynamically adjusted with the real-time identification result and the health state of the equipment, so that the protection system cannot adaptively follow the aging and load change of the equipment, and the reliability of long-term operation of the protection system is limited. Disclosure of Invention The invention provides an intelligent identification linkage protection method for equipment of 35kV and below based on multi-source adaptive fusion and deep-shallow dual-path network, which aims to 1 effectively improve fault identification accuracy in power system protection, and particularly improve identification capability of small-scale equipment such as reactors. 2. The self-adaptive capacity of the protection method to dynamic operation scenes such as load fluctuation, harmonic interference and the like is improved, and the problems of 'overstattoo' misjudgment and 'undersensitive' missed judgment caused by fixed weight distribution and single characteristic correction are avoided. 3. The threshold parameters can be dynamically adjusted according to the real-time identification result and the health state of the equipment so as to adapt to the long-term operation scene of equipment aging and load change, and the overall reliability of the protection system is enhanced. The technical scheme of the invention is as follows: According to the intelligent identification linkage protection method of the equipment of 35kV and below based on the multi-source adaptive fusion and deep-shallow dual-path network, the steps S1 to S5 are circularly executed according to a preset sampling frequency, the comprehensive characteristics of the target equipment at the current moment are respectively obtained in each execution period, and the comprehensive characteristics of a plurality of continuous execution periods form a time sequence characteristic sequence for subsequent time sequence identification and protection j