A power grid power load state intelligent identification method

Abstract

The application discloses a kind of power grid power load state intelligent identification method, it is related to smart grid monitoring technical field.The application includes the following steps: edge perception terminal synchronous acquisition high-frequency electrical quantity and non-electrical quantity;Three-layer acceleration DTW algorithm is used to calculate the similarity of real-time data and standard feature library;Through dynamic programming backtracking path, realize time sequence accurate matching, combined with random forest classifier and three-dimensional feature space mahalanobis distance, construct composite evaluation index;Based on LSTM neural network, predict health index.The application realizes the synchronous acquisition of high-frequency electrical quantity and multi-modal non-electrical quantity by edge perception terminal, uses three-layer acceleration dynamic time warping algorithm to construct composite similarity index, combined with random forest classifier and three-dimensional feature space mahalanobis distance, realizes accurate identification of load state, improves power load monitoring efficiency and the accuracy of load identification result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent power grid monitoring, and particularly relates to an intelligent power grid power load state identification method. BACKGROUND

[0002] With the rapid development of China's economy, society and technology, intelligent power grid has become the future development direction of the power system, and efficient monitoring of power system load is an important part of realizing the intelligentization of power grid. At present, the power monitoring means is to install independent monitoring devices on each load, and the running state of each load is determined through individual monitoring, which consumes a lot of manpower, material resources and financial resources.

[0003] Firstly, there are data quality problems, diversity and complexity problems, efficiency and real-time problems in technology, specifically:

[0004] Data quality problem: the accuracy and reliability of the intelligent power grid load identification method are affected by the data quality. For example, the error of the power measurement equipment, the incompleteness or delay of data collection, etc. may cause the load identification result to be inaccurate. Therefore, improving the data quality and processing noise, missing value, etc. is an important technical challenge.

[0005] Diversity and complexity problem: the power grid load contains various types of loads, such as industrial, commercial, residential, etc., and has the complexity of space-time change. Therefore, when designing the intelligent power grid load identification method, algorithms and models that adapt to different load types and dynamic changes need to be considered.

[0006] Efficiency and real-time problem: intelligent power grid load identification needs to process a large amount of data, and has high requirements on computing resources and algorithm efficiency. At the same time, real-time is also a key factor, especially for power system scheduling and control decision-making.

[0007] Therefore, how to improve the calculation efficiency and real-time of the identification method is a technical problem to be solved by the present application. SUMMARY

[0008] The purpose of the present application is to provide an intelligent power grid power load state identification method, which realizes the Beidou timing synchronous collection of high-frequency electrical quantities and multi-modal non-electrical quantities through edge perception terminals, constructs a composite similarity index by using a three-layer accelerated dynamic time warping algorithm, realizes accurate load state identification by combining a random forest classifier and a three-dimensional feature space Mahalanobis distance, and solves the problems of time-consuming and laborious power load monitoring and inaccurate load identification results.

[0009] To solve the above technical problems, the present application is realized by the following technical scheme:

[0010] The application is an intelligent power load state recognition method for power grid, comprising the following steps:

[0011] Step S1: Real-time collection of high-frequency electrical quantities (harmonic resolution 0.1%) and non-electrical quantities (temperature ±0.5℃, vibration frequency 0-500Hz) by edge perception terminals (sampling rate ≥20kHz) deployed at power grid input nodes, wherein the high-frequency electrical quantities include voltage / current waveform, power parameter, harmonic spectrum and transient characteristic quantity; the non-electrical quantities include temperature parameter, humidity parameter, gas concentration data, vibration frequency and displacement and deformation data; the collected data is standardized according to time sequence;

[0012] Step S2: Under the best condition of power load state, the edge perception terminal collects the standard characteristic time sequence of power load, and compares the standard characteristic time sequence with the real-time collected time sequence characteristics, and calculates the distance between the edge perception terminal collected data and the standard data to judge the similarity between the two;

[0013] Step S3: In the laboratory environment, simulate various typical load conditions (including normal, overload, short circuit, arc fault, insulation deterioration and other 20 standard states), collect at least 100 complete data cycles for each standard state, construct a standard characteristic time sequence library, use hierarchical clustering algorithm to cluster and analyze the characteristic sequences of the same type of state, extract the typical characteristic mode of each type of state, and establish a three-dimensional characteristic space model for each type of load state; the three-dimensional characteristic space model includes electrical characteristic axis, mechanical characteristic axis and environmental characteristic axis;

[0014] Step S4: Input the distance value obtained in step S2 into the pre-trained random forest classifier, and combine the Mahalanobis distance calculation of the three-dimensional characteristic space to construct a composite similarity evaluation index;

[0015] Step S5: Based on the LSTM neural network, a time sequence prediction model is constructed, the characteristic sequence of 6 consecutive sampling periods is input into the model, the load state evolution trend is predicted, and a state health index calculation model is established;

[0016] Step S6: A power grid topology relationship model is constructed through a graph neural network, spatial correlation analysis is performed on the state data of adjacent nodes, when 3 or more nodes in a region are abnormal at the same time, a system-level fault warning is triggered;

[0017] Step S7: An online learning mechanism is established, the confirmed diagnosis result is fed back to the feature library, the DTW distance threshold and the classifier parameters are automatically optimized every quarter; when a new type of load is connected, the feature learning mode is automatically started, and knowledge sharing between different regional power grids is realized through transfer learning.

[0018] As a preferred technical solution, in the step S1, the specific process of standardizing the collected data according to the time sequence is as follows:

[0019] According to the collected data, a multi-dimensional time sequence of high-frequency electrical quantities and non-electrical quantities is constructed;

[0020] At the moment of sensor ADC sampling, the Beidou time is written by FPGA, and after embedding the 64-bit timestamp in the information packet header and performing data packet packaging, all time sequences are synchronized by using Beidou timing, and linear interpolation processing is performed on the data segment with missing timestamp;

[0021] The original sampling frequency of each channel is identified and uniformly adjusted to the system reference sampling rate;

[0022] A three-layer acceleration strategy is adopted to realize full-resolution fine alignment of data; the first layer completes coarse alignment of one-tenth of the sampling sequence; the second layer completes regional positioning of two-tenths of the sampling sequence; and the third layer completes full-resolution alignment;

[0023] The feature parameters of each sampling are combined into a sampling time feature sequence in chronological order.

[0024] As a preferred technical solution, in the step S2, the similarity calculation process between the data collected by the edge perception terminal and the standard data is as follows:

[0025] Step S21: Collect the time sequences of the edge perception terminal data and the standard data to be compared under the same time sequence;

[0026] Step S22: Create a two-dimensional matrix, and each element in the matrix represents the distance between the data points and the standard data points under the same time sequence;

[0027] Step S23: Use the dynamic programming method to fill the accumulated distance matrix from the top left corner of the distance matrix; from the top left corner (the starting frame) to the bottom right corner (the ending frame) of the distance matrix, calculate the minimum accumulated distance path of the two time sequences (the action to be identified and the template action), facilitate the search for the optimal alignment path, backtrack the minimum path from the bottom right corner to determine the best alignment method, and convert the exhaustive problem with exponential complexity into a polynomial time through dynamic programming, which significantly improves the calculation efficiency, and at the same time, through the recursive strategy of accumulated minimum distance, avoids local noise interference (such as light change or background interference), and ensures global optimal matching;

[0028] Step S24: For each element in the accumulated distance matrix, calculate its value as the minimum accumulated distance of the three adjacent elements above, left and right, to find the best alignment path between the two time sequences; ​

[0029] Step S25: After filling the cumulative distance matrix, start backtracking from the bottom right corner, moving upwards and to the left along the path with the smallest cumulative distance until returning to the top left corner. This path represents the best alignment between the two time series.

[0030] Step S26: Calculate the similarity. The value in the lower right corner of the cumulative distance matrix is ​​the DTW distance between the two time series.

[0031] As a preferred technical solution, in step S21, the collected data points and standard data points are described in chronological order as a feature time series, where each series represents a data sample collected by the edge sensing terminal. The specific formula is as follows:

[0032] ;

[0033] ;

[0034] In the formula, and They represent the first The state component to be judged and the first... Characteristic time series of standard load state data and The length of the sequence is represented by the Euclidean distance formula:

[0035] ;

[0036] Calculated and The distance between each component is obtained as a result. The matrix, i.e.:

[0037] ;

[0038] In the formula, Indicates the state component to be judged. The k-th dimension feature of the frame, The first component of the standard load state represents the standard load state component. The k-th dimension feature of the frame, Indicates the total number of feature dimensions. Indicates the first of two sequences Frame and the Frame differences.

[0039] As a preferred technical solution, the specific process of tracing the backtracking path in step S25 is as follows:

[0040] Starting from the bottom right corner (end point) of the cumulative distance matrix, corresponding to the last frame of the two feature time series;

[0041] The accumulated distance values of the three adjacent cells of the current position, i.e. the left, the upper and the left upper cells, are compared, and the minimum value is selected to determine the moving direction; when moving to the left, it indicates that the current frame of the standard feature time sequence needs to be matched with the next frame of the test sequence (the standard feature time axis is stretched); when moving upwards, it indicates that the current frame of the sampling time feature sequence needs to be matched with the next frame of the standard feature time sequence (the sampling feature time axis is compressed); when moving to the left upper, it indicates that the current frames of the standard feature time sequence and the sampling time feature sequence are directly aligned;

[0042] The coordinates of each step of movement are stored in the path set to form a reverse sequence from the end point to the start point;

[0043] When the path is traced back to the left upper corner of the matrix (the start point), the process is stopped, and the reverse output path is the optimal alignment sequence; the dynamic programming ensures that the traced back path is the only solution with the minimum global accumulated distance, avoiding the local optimal trap, and the traced back path directly shows the frame level correspondence relationship of the sequence, facilitating the analysis of the time sequence difference of the data.

[0044] As a preferred technical solution, in the step S3, the standard feature time sequence library is constructed according to the following procedure:

[0045] In the laboratory environment, a programmable load simulation platform is established, which integrates an electrical parameter generator (0.1 Hz-10 MHz frequency band), a mechanical vibration table (0-1 kHz adjustable) and an environmental simulation cabin (temperature -40℃-150℃, humidity 10%-95% controllable). The nanosecond level synchronous acquisition of electrical quantities (sampling rate≥1 MHz), mechanical quantities (sampling rate≥50 kHz) and environmental quantities (sampling rate≥1 kHz) is realized by using an optical fiber synchronous network; at the same time, a simulation protocol of 21 types of standard working conditions is formulated, each type of working condition contains 5 typical intensity levels and 3 duration modes;

[0046] The original data is enhanced, and dynamic feature extraction is performed, and the improved Procrustes analysis method is used for time and space alignment optimization; the original data enhancement expands the sample diversity by adding Gaussian noise (SNR≥30 dB), time stretching (±20%), random missing (≤5%) and other methods;

[0047] A multi-scale clustering architecture is created, which respectively performs fast pre-clustering based on DTW distance for reducing the calculation amount, fine clustering combined with Mahalanobis distance for eliminating the dimensional influence between features, and visual verification by applying topological preserving dimension reduction (t-SNE), automatically determines the clustering center by density peak detection, extracts each type of state to build a three-dimensional feature space model; the extraction of each type of state includes core feature mode (appearance probability>80%), edge feature mode (20%<probability≤80%) and abnormal feature mode (probability≤20%).

[0048] As a preferred technical solution, in the step S5, the time series prediction model comprises three layers of spatio-temporal feature extraction layers;The first layer adopts a bidirectional LSTM to extract local time sequence features;The second layer combines a key feature strengthening module of an attention mechanism;The third layer introduces a causal convolutional network to capture long-period dependencies;The time series prediction model is dynamically inputted with one or more of a plurality of dimensional feature sequences of 6 consecutive periods, a device historical health index curve, a similar device degradation mode library and real-time meteorological data.

[0049] As a preferred technical solution, in the step S6, the graph neural network structure comprises node feature and edge weight calculation;The node feature comprises fusion of electrical quantity, mechanical quantity and environmental quantity to construct a 128-dimensional feature vector;The variable weight adopts a hybrid weight model of line impedance, physical distance and power transmission amount;An automatic updating mechanism is also provided, which adjusts the topological connection strength according to the power grid operation state every 15 minutes;The spatio-temporal graph convolution network comprises a spatial dimension for capturing the implicit relationship between nodes through a graph attention mechanism (GAT) and a time dimension for processing minute-level data fluctuations by integrating a TCN time convolution module;The system fault early warning mechanism is as follows: when a single node is abnormal (primary warning), a local diagnosis program is triggered;When 2 adjacent nodes are abnormal (intermediate warning), a regional scanning mode is started;When ≥3 nodes are abnormal (advanced warning), a system protection plan is activated.

[0050] As a preferred technical solution, in the step S7, the specific process of adaptive learning and optimization is as follows: an online learning mechanism with double-channel verification is established, the device inversion data (confidence ≥0.85) and the expert annotation results are fused, and a sliding window algorithm is used to dynamically update the feature library, so as to ensure the timeliness and accuracy of knowledge iteration;Secondly, a seasonal parameter self-optimization module is designed, the DTW dynamic regularization threshold (±15% seasonal fluctuation) is automatically calibrated based on kernel density estimation, and the classifier hyperparameter (50 generation population size) is evolved combined with genetic algorithm, so as to realize periodic self-tuning of algorithm parameters;For the new load access scene, a small sample transfer learning framework is developed, the feature similarity matching (Top-3 candidate) and GAN data enhancement (sample expansion) are used to complete the safety adaptation within 7 days of observation period;Finally, a federal knowledge sharing network is constructed, the differential privacy technology is used to realize cross-regional feature parameter exchange (80% shared weight of electrical features), and a collaborative evolution mode of "distributed training-centralized aggregation" is formed.

[0051] The application has the following beneficial effects:

[0052] The application realizes Beidou timing synchronous collection of high-frequency electrical quantities and multi-modal non-electrical quantities through an edge-aware terminal, adopts a three-layer accelerated dynamic time warping algorithm to construct a composite similarity index, and realizes accurate identification of load states by combining a random forest classifier and a three-dimensional feature space Mahalanobis distance, thereby improving the efficiency of power load monitoring and the accuracy of load identification results.

[0053] The application accurately maps the frame correspondence relationship of two time series through a backtracking path, allows sequences of different lengths to realize similarity calculation through nonlinear alignment, ensures that the backtracking path is the only solution with the minimum global cumulative distance through dynamic programming, avoids local optimal traps, and intuitively displays the frame-level correspondence relationship of the sequences through the backtracking path, thereby facilitating the analysis of the time sequence difference of data.

[0054] The application realizes time-space alignment optimization through an improved Procrustes analysis method, accelerates sinc calculation through segmented polynomial approximation, realizes parallel interpolation operation in FPGA, reduces data transfer overhead through three-layer caching, provides time-aligned feature sequences for subsequent DTW algorithms, realizes accurate time correlation of mechanical vibration signals and electrical abnormalities, realizes early warning of faults, and improves state evaluation efficiency.

[0055] The application combines the time sequence alignment capability of DTW with the statistical characteristics of Mahalanobis distance, solves the problem that time sequence features and static features are difficult to be unified in traditional methods, and improves the fine-grained classification accuracy while maintaining real-time performance (<50ms response) through a hierarchical progressive classification strategy.

[0056] Of course, implementing any product of the application does not necessarily need to achieve all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description, and obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0058] Figure 1 A flowchart of an intelligent power grid load state identification method of the application;

[0059] Figure 2 A flowchart of standardization processing of collected data according to time sequence;

[0060] Figure 3 A flowchart of similarity calculation between data collected by an edge-aware terminal and standard data. DETAILED DESCRIPTION

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0063] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-3 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0064] Please see Figure 1 As shown, the present invention is an intelligent identification method for power grid load status, comprising the following steps:

[0065] Step S1: Real-time acquisition of high-frequency electrical quantities (harmonic resolution 0.1%) and non-electrical quantities (temperature ±0.5℃, vibration frequency 0-500Hz) is achieved through edge sensing terminals (sampling rate ≥20kHz) deployed at the power grid input nodes. High-frequency electrical quantities include voltage / current waveforms, power parameters, harmonic spectra, and transient characteristics; non-electrical quantities include temperature parameters, humidity parameters, gas concentration data, vibration frequency, and displacement and deformation data. The acquired data is then standardized according to time series.

[0066] The deployment location of the edge sensing terminal must meet the requirement of synchronous acquisition of high-frequency electrical and non-electrical quantities; specifically as follows:

[0067] The location of the edge-aware terminal deployment includes: power transmission key nodes, user-side key equipment, and new energy access points; wherein, the power transmission key nodes are, for example, an integrated electric field sensor and temperature vibration negative composite probe installed at the cross arm of a tower, which is used to detect the contamination state of insulators and the dancing characteristics of conductors, a micro sensor terminal built-in cable joint box, which realizes high-frequency capture (sampling rate ≥ 20 kHz) and synchronous detection of joint temperature (temperature ± 0.5℃); the user-side key equipment is, for example, a vibration sensor (0-500Hz frequency response) deployed on the surface of the heat sink of a distribution transformer body, which, in combination with a winding outlet end electrical quantity acquisition module, realizes early overload warning; a multi-physical quantity fusion terminal is installed at the busbar connection point in the low-voltage electrical cabinet, which synchronously collects harmonics and contact point temperature rise; the new energy access point is, for example, a wide frequency domain acquisition module (0.1Hz-10MHz) deployed at the output end of a photovoltaic inverter, which is used to capture harmonic distortion and power fluctuation characteristics; a vibration-temperature collaborative sensing unit is deployed inside the fan converter cabinet, which identifies mechanical wear and abnormal heat dissipation;

[0068] Step S2: Under the best condition of power load state, the edge-aware terminal collects the standard characteristic time series of the power load, and compares the standard characteristic time series with the real-time collected time series characteristics, calculates the distance between the edge-aware terminal collected data and the standard data, and judges the similarity between the two;

[0069] Step S3: In a laboratory environment, simulate various typical load conditions (including normal, overload, short circuit, arc fault, insulation deterioration, etc. 20 standard states), collect at least 100 complete data periods for each standard state, construct a standard characteristic time series library, use hierarchical clustering algorithm to cluster and analyze the characteristic sequences of the same type of state, extract the typical characteristic mode of each type of state, and establish a three-dimensional characteristic space model for each type of load state; the three-dimensional characteristic space model includes an electrical characteristic axis, a mechanical characteristic axis, and an environmental characteristic axis;

[0070] Step S4: input the distance value obtained in step S2 into the pre-trained random forest classifier, and construct a composite similarity evaluation index combined with the Mahalanobis distance calculation of the three-dimensional characteristic space;

[0071] Step S5: based on the LSTM neural network, a time series prediction model is constructed, the characteristic sequence of the continuous 6 sampling periods is input into the model, the load state evolution trend is predicted, and a state health index calculation model is established;

[0072] Step S6: a power grid topology relationship model is constructed through a graph neural network, spatial correlation analysis is performed on the state data of adjacent nodes, and when 3 or more nodes in a region are abnormal at the same time, a system-level fault warning is triggered;

[0073] Step S7: Establish an online learning mechanism to feed back the confirmed diagnosis results to the feature library, automatically optimize the DTW distance threshold and classifier parameters every quarter, and automatically start feature learning mode when new loads are connected to realize knowledge sharing between different regional power grids through transfer learning.

[0074] Referring to Figure 2 As shown in FIG. 1, in step S1, the specific process of standardizing the collected data according to time sequence is as follows:

[0075] According to the collected data, a multi-dimensional time sequence of high-frequency electrical quantities and non-electrical quantities is constructed.

[0076] At the moment of sensor ADC sampling, the Beidou time is written by FPGA, and after embedding 64-bit timestamp in the header of the information packet and performing data packet packaging, all time sequences are synchronized by Beidou timing, and the data segment with missing timestamp is processed by linear interpolation.

[0077] The original sampling frequency of each channel is identified and adjusted to the system reference sampling rate.

[0078] A three-layer acceleration strategy is adopted to realize full-resolution fine alignment of data; the first layer completes coarse alignment of one-tenth sampling sequence; the second layer completes regional positioning of one-half sampling sequence; and the third layer completes full-resolution alignment; the distance matrix is divided into GPU calculation blocks, and CUDA kernel function is used to accelerate cumulative distance calculation, and shared memory is used to reduce global memory access;

[0079] The feature parameters of each sampling are combined into a sampling time feature sequence in chronological order, which is used for matching of the DTW algorithm.

[0080] In specific implementation, edge perception terminals are deployed at key nodes of 500kV transmission lines, and the following sensors are configured:

[0081] ‌The electrical quantity collection is a wideband current transformer (sampling rate 20kHz, harmonic resolution 0.1%);

[0082] ‌The non-electrical quantity collection includes an infrared temperature measurement probe (±0.5℃), a micro-vibration sensor (0-500Hz), and an ultraviolet imager (detects corona discharge);

[0083] The standardization process is as follows: the collected capture voltage waveform distortion rate (THD), leakage current harmonics (3 / 5 / 7 times), synchronous recording of insulator surface temperature gradient and vibration main frequency offset are used to construct multi-dimensional time series; FPGA writes 64-bit timestamp (Beidou-3 timing accuracy ± 30ns) at the moment of ADC sampling, and Lagrange interpolation is used to reconstruct the 5ms data segment lost due to signal interference; unify the original sampling rate: electrical quantity 20kHz, vibration quantity 10kHz, temperature 1Hz, unify to the reference 20kHz through sinc interpolation, and calculate and optimize using piecewise polynomial approximation;

[0084] The three-layer acceleration strategy is as follows:

[0085] Hierarchy Processing content Technical implementation First layer Downsample to 2kHz coarse alignment Calculate coarse matching region of two sequences DTW distance Second layer Downsample to 10kHz region positioning Reduce search range within coarse alignment region Third layer Full resolution 20kHz fine alignment GPU parallel computation

[0086] A feature vector containing the following dimensions is generated: [timestamp, fundamental amplitude, 3rd harmonic content, temperature range, vibration energy 125-150Hz frequency band].

[0087] Please refer to Figure 3 As shown in

[0088] Step S21: Collect the time series of the edge perception terminal data and the standard data to be compared under the same time series;

[0089] Step S22: Create a two-dimensional matrix, and each element in the matrix represents the distance between the collected data points and the standard data points under the same time series;

[0090] Step S23: Use dynamic programming method to fill the cumulative distance matrix from the top left corner; from the top left corner (starting frame) to the bottom right corner (ending frame), calculate the minimum cumulative distance path of the two time series (the action to be recognized and the template action), which facilitates the search for the optimal alignment path, and the dynamic programming converts the exponential complexity of the exhaustive problem into a polynomial time, significantly improving the calculation efficiency. At the same time, through the recursive strategy of cumulative minimum distance, local noise interference (such as light change or background interference) is avoided, ensuring global optimal matching;

[0091] Step S24: For each element in the cumulative distance matrix, calculate its value as the minimum cumulative distance of its left upper, upper and left adjacent elements, to find the best alignment path between the two time series;

[0092] Step S25: After filling the cumulative distance matrix, start backtracking from the bottom right corner, moving upwards and to the left along the path with the smallest cumulative distance until returning to the top left corner. This path represents the best alignment between the two time series.

[0093] Step S26: Calculate the similarity. The value in the lower right corner of the cumulative distance matrix is ​​the DTW distance between the two time series.

[0094] In step S21, the collected data points and standard data points are described in chronological order as a feature time series, where each series represents a data sample collected by the edge sensing terminal. The specific formula is as follows:

[0095] ;

[0096] ;

[0097] In the formula, and They represent the first The state component to be judged and the first... Characteristic time series of standard load state data and The length of the sequence is represented by the Euclidean distance formula:

[0098] ;

[0099] Calculated and The distance between each component is obtained as a result. The matrix, i.e.:

[0100] ;

[0101] In the formula, Indicates the state component to be judged. The k-th dimension feature of the frame, The first component of the standard load state represents the standard load state component. The k-th dimension feature of the frame, Indicates the total number of feature dimensions. Indicates the first of two sequences Frame and the Frame differences.

[0102] In step S25, the specific process of backtracking the path is as follows:

[0103] Starting from the bottom right corner (end point) of the cumulative distance matrix, corresponding to the last frame of the two feature time series;

[0104] Compare the cumulative distance values ​​of the three adjacent cells to the left, above, and upper left of the current position, and move in the direction corresponding to the minimum value. When moving to the left, it means that the current frame of the standard feature time series needs to match the next frame of the test sequence (stretching the standard feature time axis). When moving upward, it means that the current frame of the sampled time feature sequence needs to match the next frame of the standard feature time series (compressing the sampled feature time axis). When moving to the upper left, it means that the current frames of the standard feature time series and the sampled time feature sequence are directly aligned.

[0105] Store the coordinates of each step of movement into the path set to form a reverse sequence from the end point to the beginning point;

[0106] The backtracking path stops when it reaches the top left corner (starting point) of the matrix, and the reverse output path is the best aligned sequence. Dynamic programming ensures that the backtracking path is the unique solution with the minimum global cumulative distance, avoiding the trap of local optima. At the same time, the backtracking path intuitively shows the frame-level correspondence of the sequence, which is convenient for analyzing the temporal differences of the data.

[0107] In practice, during the data preparation phase, 100 sets of electrical quantity (voltage waveform, current harmonics) and mechanical quantity (tank vibration frequency) data were collected by simulating the rated load condition of the transformer in the laboratory, and a standard characteristic sequence was constructed:

[0108] ;

[0109] An edge sensing terminal collects real-time operating data of the main transformer in a substation and generates a sequence to be tested.

[0110] ;

[0111] When constructing the distance matrix, an 80×100 Euclidean distance matrix will be calculated, where the elements are... express:

[0112] ;

[0113] In the formula, This represents the effective value of the voltage waveform. The content of the 5th harmonic current. The dominant frequency of the oil tank vibration. The effective value of the voltage waveform acquired in real time. The effective value of the voltage waveform in the j-th group of the standard feature library. This refers to the real-time acquisition of the 5th harmonic current content. For the first in the standard feature library The content of the fifth harmonic current; The main frequency of the fuel tank vibration is collected in real time. For the first in the standard feature library Group vibration main frequency, subscript (1-100) represents 100 groups of independent sampling periods;

[0114] Take , For example, initialize the cumulative matrix as follows: ;

[0115] The path backtracking is as follows:

[0116] Backtracking step Matrix coordinates Moving direction Alignment relationship explanation 1 （80,100） Starting point Sequence end alignment 2 （79,99） Top left Direct frame alignment 3 （78,98） Top left Direct frame alignment ... ... ... ... 42 （50,60） Up Test sequence 50th frame aligns with standard sequence 61st frame Final （0,0） No Cumulative DTW distance = 18.7

[0117] According to the above table, the normalized similarity is calculated as follows: Since the normal range should be greater than 0.85, the calculated normalized similarity 0.051 deviates from 0.85 quite seriously, and combined with the vibration frequency offset +12%, it is determined to be early warning of winding overheating; After using GPU acceleration, the calculation time is only 6.3ms (satisfying the real-time requirement of <50ms), and at the same time, through inversion test, the false alarm rate is reduced by 37% compared with the traditional correlation coefficient method.

[0118] In step S3, the standard feature time sequence library construction process is as follows:

[0119] In the laboratory environment, a programmable load simulation platform is established, which integrates an electrical parameter generator (0.1Hz-10MHz frequency band), a mechanical vibration table (0-1kHz adjustable) and an environmental simulation cabin (temperature-40℃-150℃, humidity 10%-95% controllable). The nanosecond-level synchronous acquisition of electrical quantities (sampling rate≥1MHz), mechanical quantities (sampling rate≥50kHz) and environmental quantities (sampling rate≥1kHz) is realized by using optical fiber synchronous network; At the same time, the simulation protocol of 21 types of standard working conditions is formulated, each type of working condition contains 5 kinds of typical intensity grades and 3 kinds of duration modes;

[0120] The original data is enhanced, and dynamic feature extraction is performed, and the improved Procrustes analysis method is used for time and space alignment optimization; The original data enhancement expands the sample diversity by adding Gaussian noise (SNR≥30dB), time stretching (±20%), random missing (≤5%) and other ways;

[0121] Specifically, the dynamic feature extraction includes electrical features, mechanical features and environmental features; The electrical features include extracting 37-dimensional features such as harmonic energy ratio, transient oscillation frequency and zero sequence component asymmetry; The mechanical features include calculating 15-dimensional features such as vibration main frequency offset, envelope distortion rate and resonance band energy distribution; The environmental features include deriving 8-dimensional features such as temperature rise rate, humidity-temperature coupling coefficient and gas diffusion gradient.

[0122] A multi-scale clustering architecture is created, respectively, a fast pre-clustering based on DTW distance is used to reduce the calculation amount, a fine clustering combined with Mahalanobis distance is used to eliminate the dimensional influence between features, and a visual verification of topological preserving dimensionality reduction (t-SNE) is applied, the clustering center is automatically determined by density peak detection, and the state of each class is extracted to build a three-dimensional feature space model; extraction of each state includes core feature mode (appearance probability > 80%), edge feature mode (20% < probability ≤ 80%) and abnormal feature mode (probability ≤ 20%).

[0123] Since traditional Procrustes analysis can only process data of the same dimension, the present innovation realizes the spatial mapping of cross-sampling rate features by introducing time-frequency joint resampling and dynamic weight distribution mechanism, and the specific time-space alignment optimization of the improved Procrustes analysis method is as follows:

[0124] Taking the time axis of electrical quantity as the reference coordinate system, the time anchor points of mechanical quantity and environmental quantity are established through the Beidou timestamp, and the reference coordinate system is established;

[0125] Dynamic resampling is performed, and sinc function interpolation is used to realize the conversion of non-uniform sampling to 1MHz, and the specific formula is as follows:

[0126] ;

[0127] ;

[0128] In the formula, is the resampled mechanical signal, is the original mechanical vibration signal, is the dynamic weight coefficient, is the sinc difference value kernel function, is the mechanical signal sampling time, is the mechanical signal sampling interval, is the number of neighborhood points participating in interpolation; is the resampled environmental signal, is the original environmental monitoring signal, is the environmental quantity credible weight, is the environmental signal sampling interval, is the difference window width.

[0129] For transient features that cannot be aligned (such as short circuit impact), dynamic time warping (DTW) is used for local fine tuning.

[0130] In step S4, the Mahalanobis distance calculation combined with the three-dimensional feature space is used to construct a composite similarity evaluation index. The Mahalanobis distance is calculated for the standardized feature vector, the DTW distance is normalized to a similarity value in the interval [0, 1] by a sigmoid function, and a composite similarity index is constructed. The composite similarity index is , a dynamic threshold mechanism is set, such as a normal state threshold = 0.85 and a warning state threshold = 0.7. At the same time, an online learning mechanism is set: the classifier parameters are automatically updated every 24 hours, an incremental learning mode is triggered when a new type of load is connected, and the model robustness is improved by enhancing through adversarial samples.

[0131] In specific implementation, the standardized feature vector is obtained from a 110kV substation main transformer edge perception terminal:

[0132] ;

[0133] In the formula, respectively represent the harmonic distortion rate, the 5th harmonic content, the vibration main frequency, and the temperature rise. The covariance matrix Σ of the electrical-mechanical-environmental three-dimensional feature space is calculated based on the standard feature library, and the calculation formula is as follows:

[0134] ;

[0135] In the formula, is the normal state mean vector [0.05, 0.03, 120Hz, 8℃]; the measured real-time current waveform and the standard waveform DTW distance is 28.6; and the conversion through the sigmoid function is:

[0136] ;

[0137] In the formula, 25 is the dynamic threshold reference value, and 5 is the slope adjustment parameter.

[0138] The composite index is constructed, and the weight distribution is (Mahalanobis distance), (DTW):

[0139] ;

[0140] The dynamic threshold is compared: the normal threshold is 0.85 (down to 0.82 in summer), and the warning threshold is 0.7 (current 0.42 < 0.7). Therefore, according to the above embodiment, it is concluded that the three-level warning of insulation deterioration is triggered (combined with a vibration main frequency offset of +20%).

[0141] In step S5, the time series prediction model comprises three layers of spatio-temporal feature extraction layers; wherein the first layer adopts a bidirectional LSTM to extract local time series features; the second layer combines a key feature strengthening module of an attention mechanism; and the third layer introduces a causal convolution network to capture long-period dependencies; one or more of a continuous 6-period multi-dimensional feature sequence, a device historical health index curve, a similar device degradation mode library, and real-time meteorological data are dynamically input into the time series prediction model; a composite monitoring and evaluation model is constructed, and the specific formula is as follows: ; wherein, α+β+γ=1, α, β, γ are weights and the weights are automatically adjusted with the service life of the device; is a dynamic time warping distance, representing the time series difference between the current state and the standard state, represents a critical DTW distance threshold.

[0142] In specific implementation, the multi-dimensional data of a 220kV substation #3 main transformer for a continuous 6-period (15 minutes per period) is input, including: electrical quantity sequence , mechanical quantity sequence , and environmental quantity sequence .

[0143] The three layers of spatio-temporal feature extraction layers comprise:

[0144] The first layer is a bidirectional LSTM layer, with a structure of 2 layers of 128-unit LSTM connected bidirectionally; the output captures local time series features (such as harmonic distortion rate mutation points) to generate a 256-dimensional feature vector;

[0145] The second layer is an attention mechanism layer, which calculates the attention weights of electrical, mechanical, and environmental features, respectively , , and ; the vibration main frequency 142Hz (more than baseline +18%) obtains the highest weight 0.71;

[0146] The third layer is a causal convolution layer, with a structure of 3 layers of dilated convolution (dilation coefficients 1 / 3 / 5), 64 3×3 convolution kernels, and an output: identifying the acceleration trend of temperature rise rate (slope change +40% from the 4th period);

[0147] The composite evaluation model is calculated, and the dynamic weight adjustment is performed, α=0.55, β=0.30, γ=0.15; the difference between the current state and the standard state is , and the critical threshold is ; therefore, the health index is: ;

[0148] Known health index threshold: normal range > 0.75, early warning threshold 0.6; while the current health index is 0.53 < 0.6; trigger insulation aging secondary warning, and propose optimization scheme such as: shorten the monitoring period to 5 minutes, start oil chromatography online monitoring, arrange partial discharge test within 7 days, etc.

[0149] In step S6, the graph neural network structure includes node feature and edge weight calculation; wherein the node feature includes fusion of electrical quantity, mechanical quantity and environmental quantity to construct a 128-dimensional feature vector; the variable weight adopts a hybrid weight model of line impedance, physical distance and power transmission amount; at the same time, an automatic updating mechanism is set, and the topological connection strength is adjusted according to the power grid operation state every 15 minutes; the space-time graph convolution network includes a spatial dimension for capturing the implicit relationship between nodes through a graph attention mechanism (GAT) and a time dimension for processing minute-level data fluctuations by integrating a TCN time convolution module; the system fault warning mechanism is as follows: when a single node is abnormal (primary warning), a local diagnosis program is triggered; when 2 adjacent nodes are abnormal (intermediate warning), a regional scanning mode is started; when ≥3 nodes are abnormal (advanced warning), a system protection plan is activated; when the abnormal nodes form a closed loop, it is determined as power grid oscillation (response time < 200ms); when the abnormality propagates along the power transmission direction, it is determined as a cascading failure (prediction accuracy is improved); when the abnormal nodes are randomly distributed, combined with meteorological data, external interference is judged;

[0150] In specific implementation, real-time monitoring data of 3 adjacent substations (nodes A / B / C) in a certain regional power grid are acquired; electrical quantity: voltage harmonic distortion rate (A: 5.2%, B: 4.8%, C: 7.5%); mechanical quantity: main transformer vibration main frequency offset (A: +8Hz, B: +12Hz, C: +15Hz); environmental quantity: temperature rise rate (A: 0.8℃ / h, B: 1.2℃ / h, C: 2.5℃ / h); feature engineering is as follows: electrical feature axis adopts normalized harmonic energy ratio + zero sequence unbalance (32 dimensions); mechanical feature axis is vibration frequency offset + envelope distortion rate (48 dimensions); environmental feature axis is temperature rise gradient + humidity coupling coefficient (48 dimensions); 128-dimensional feature vector (32+48+48) is generated for each node;

[0151] Then, taking node A-B as an example, the hybrid weight model is:

[0152] ;

[0153] In the formula, the line impedance Z=0.15Ω, the physical distance d=5.2km, and the power transmission amount P=18MW;

[0154] Among them, the spatial dimension (GAT mechanism): the attention weight of node A to B / C is 0.63, 0.37 respectively; The implicit influence of node C vibration anomaly on A / B is captured (correlation coefficient 0.78); Time dimension (TCN module): analyze the minute level data fluctuation, identify the node C temperature rise acceleration trend (slope mutation +40%);

[0155] The system fault early warning mechanism is:

[0156] ‌Primary warning‌ (single point anomaly of node C), trigger local diagnosis: oil chromatographic analysis shows that the acetylene content exceeds the threshold value (3.2ppm);

[0157] ‌Intermediate warning‌ (adjacent anomaly of node B / C), start regional scanning: find that the 110kV line insulator is deteriorated (leakage current +25%);

[0158] ‌Advanced warning‌ (node A / B / C anomaly at the same time), activate protection plan: remove node C load (priority level III), enable standby line to transfer power, and issue unmanned aerial vehicle inspection instructions.

[0159] In step S7, the specific process of adaptive learning and optimization is as follows: an online learning mechanism of double-channel verification is established, the device inversion data (confidence ≥0.85) and expert annotation results are fused, a sliding window algorithm is used to dynamically update the feature library, and the timeliness and accuracy of knowledge iteration are ensured; Secondly, a seasonal awareness parameter self-optimization module is designed, based on kernel density estimation, the DTW dynamic regularization threshold (±15% seasonal fluctuation) is automatically calibrated, and the genetic algorithm is used to evolve the classifier hyperparameters (50 generation population size), realizing the periodic self-tuning of algorithm parameters; For the new load access scene, a small sample transfer learning framework is developed, through feature similarity matching (Top-3 candidate) and GAN data enhancement (sample expansion), the safety adaptation is completed within 7 days of observation period; Finally, a federal knowledge sharing network is constructed, differential privacy technology is used to realize cross-regional feature parameter exchange (electrical characteristics 80% shared weight), forming a "distributed training-centralized aggregation" co-evolution mode.

[0160] In specific implementation, when a new energy storage power station is added in a provincial power grid, the system synchronously receives device inversion data including PCS converter harmonic characteristics (confidence 0.87) and expert annotation including manually annotated "peak clipping and valley filling mode" operation label; The feature library is updated as follows: the sliding window size is to retain the latest 30 days of data; The dynamic weight is the new data weight coefficient 0.7 and the historical data 0.3; The abnormal filtering is to automatically remove samples with confidence less than 0.85;

[0161] In the seasonal awareness parameter optimization, taking the summer electricity peak as an example, the DTW threshold is adjusted as follows:

[0162] ;

[0163] Benchmark threshold is 0.85, actual adjustment: from 0.85 to 0.93 (July measured value).

[0164] It is worth noting that the above system embodiments, including each unit is only according to the logical division of function, but not limited to the above division, as long as the corresponding function can be realized; in addition, the specific name of each functional unit is only to facilitate mutual differentiation, and is not used to limit the protection scope of the present application.

[0165] In addition, those skilled in the art can understand that all or part of the steps of the above-mentioned embodiments can be completed by programs instructing relevant hardware, and the corresponding programs can be stored in a computer readable storage medium.

[0166] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and do not limit the application to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present application. The present application selects and describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can well understand and utilize the present application. The present application is limited by the claims and their entire scope and equivalents.

Claims

1. A method for intelligent identification of power grid load status, characterized in that, Includes the following steps: Step S1: Collect high-frequency electrical and non-electrical quantities in real time through edge sensing terminals deployed at the power grid input nodes. The high-frequency electrical quantities include voltage / current waveforms, power parameters, harmonic spectra, and transient characteristics. The non-electrical quantities include temperature parameters, humidity parameters, gas concentration data, vibration frequency, and displacement and deformation data. Standardize the collected data according to the time series. Step S2: Create a standard feature time series of power load collected by the edge sensing terminal under the optimal power load condition, and compare the standard feature time series with the real-time collected time series features. The similarity between the two is judged by calculating the distance between the data collected by the edge sensing terminal and the standard data by using the Dynamic Time Warping (DTW) algorithm. Step S3: In a laboratory environment, simulate various typical load conditions, collect multiple sets of complete data cycles for each standard state, construct a standard feature time series library, use hierarchical clustering algorithm to perform cluster analysis on the feature sequences of the same type of state, extract the typical feature patterns of each type of state, and establish a three-dimensional feature space model for each type of load state. Step S4: Input the distance value obtained in step S2 into the pre-trained random forest classifier, and combine it with Mahalanobis distance calculation in the three-dimensional feature space to construct a composite similarity evaluation index. Step S5: Construct a time series prediction model based on LSTM neural network, input the feature sequence of 6 consecutive sampling periods into the model, complete the prediction of the load state evolution trend, and establish a state health index calculation model; Step S6: Construct a power grid topology model using a graph neural network, perform spatial correlation analysis on the state data of adjacent nodes, and trigger a system-level fault warning when three or more nodes in a region are simultaneously abnormal. Step S7: Establish an online learning mechanism to feed back the confirmed diagnostic results to the feature library and automatically optimize the DTW distance threshold and classifier parameters every quarter; when a new load is connected, the feature learning mode is automatically started to achieve knowledge sharing between different regional power grids through transfer learning.

2. The intelligent identification method for power grid load status according to claim 1, characterized in that, In step S1, the specific process for standardizing the collected data according to the time series is as follows: Construct multi-dimensional time series of high-frequency electrical and non-electrical quantities based on the collected data; At the moment of sampling by the sensor ADC, the BeiDou time is written by the FPGA. After embedding a 64-bit timestamp in the header of the information packet and encapsulating the data packet, all time series are synchronized by BeiDou time synchronization, and linear interpolation is performed on the data segments with missing timestamps. Identify the original sampling frequency of each channel and adjust it uniformly to the system reference sampling frequency; A three-layer acceleration strategy is adopted to achieve full-resolution fine alignment of data; The three-layer acceleration strategy includes: the first layer completes coarse alignment of the one-tenth downsampled sequence; the second layer completes region localization of the one-half downsampled sequence; and the third layer completes full-resolution alignment. The feature parameters of each sample are combined into a sampling time feature sequence in chronological order.

3. The intelligent identification method for power grid load status according to claim 1, characterized in that, In step S2, the similarity calculation process between the data collected by the edge sensing terminal and the standard data is as follows: Step S21: Collect time series of edge sensing terminal data and standard data that need to be compared under the same time series. Step S22: Create a two-dimensional matrix, where each element... This represents the distance between the collected data points and the standard data points within the same time series. Step S23: Using the DTW algorithm, start from the top left corner of the cumulative distance matrix and gradually fill the cumulative distance matrix; Step S24: For each element in the cumulative distance matrix Calculate its value as the current distance plus the minimum cumulative distance among its three adjacent elements to the left, top, and left, and find the best alignment path between the two time series. Step S25: After filling the cumulative distance matrix, start backtracking from the bottom right corner, moving upwards and to the left along the path with the smallest cumulative distance until returning to the top left corner. This path represents the best alignment between the two time series. Step S26: Calculate the similarity. The value in the lower right corner of the cumulative distance matrix is ​​the DTW distance between the two time series.

4. The intelligent identification method for power grid load status according to claim 3, characterized in that, In step S21, the collected data points and standard data points are described in chronological order as a feature time series, where each series represents a data sample collected by the edge sensing terminal. The specific formula is as follows: ； ； In the formula, and They represent the first The state component to be judged and the first... Characteristic time series of standard load state data and The length of the sequence is represented by the Euclidean distance formula: ； Calculated and The distance between each component is obtained as a result. The matrix, i.e.: ； In the formula, Indicates the state component to be judged. The k-th dimension feature of the frame, The first component of the standard load state represents the standard load state component. The k-th dimension feature of the frame, Indicates the total number of feature dimensions. Indicates the first of two sequences Frame and the Frame differences.

5. The intelligent identification method for power grid load status according to claim 3, characterized in that, In step S25, the specific process of tracing the path back is as follows: Starting from the bottom right corner of the cumulative distance matrix, it corresponds to the last frame of the two feature time series; Compare the cumulative distance values ​​of the three adjacent cells to the left, above, and upper left of the current position, and move in the direction corresponding to the minimum value. When moving to the left, it means that the current frame of the standard feature time series needs to match the next frame of the sampled time feature series. When moving upward, it means that the current frame of the sampled time feature series needs to match the next frame of the standard feature time series. When moving to the upper left, it means that the current frame of the standard feature time series and the current frame of the sampled time feature series are directly aligned. Store the coordinates of each step of movement into the path set to form a reverse sequence from the end point to the beginning point; The path stops when it backtracks to the top left corner of the matrix, and the reversed output path is the best aligned path.

6. The intelligent identification method for power grid load status according to claim 1, characterized in that, In step S3, the construction process for building the standard feature time series library is as follows: In a laboratory environment, a programmable load simulation platform was established, integrating an electrical parameter generator, a mechanical vibration table, and an environmental simulation chamber. A fiber optic synchronous network was used to achieve nanosecond-level synchronous acquisition of electrical, mechanical, and environmental quantities. Simultaneously, simulation protocols for 21 standard working conditions were developed, with each working condition including 5 typical intensity levels and 3 duration modes. The original data is augmented and dynamic features are extracted. Spatiotemporal alignment optimization is performed using an improved Procrustes analysis method. A multi-scale clustering architecture is created, which performs fast pre-clustering based on DTW distance, fine clustering combined with Mahalanobis distance, and visual verification using topology-preserving dimensionality reduction. Cluster centers are automatically determined by density peak detection, and each class state is extracted for 3D feature space modeling.

7. The intelligent identification method for power grid load status according to claim 1, characterized in that, In step S5, the time series prediction model includes three spatiotemporal feature extraction layers; the first layer uses bidirectional LSTM to extract local temporal features; the second layer combines a key feature enhancement module with an attention mechanism; the third layer introduces a causal convolutional network to capture long-term dependencies; and the time series prediction model is dynamically input with one or more of the following: a multidimensional feature sequence of six consecutive periods, the historical health index curve of the equipment, a library of degradation patterns of similar equipment, and real-time meteorological data.

8. The intelligent identification method for power grid load status according to claim 1, characterized in that, In step S6, the graph neural network structure includes node features and edge weight calculation; the node features include a 128-dimensional feature vector constructed by fusing electrical quantities, mechanical quantities, and environmental quantities; the edge weight adopts a hybrid weight model of line impedance, physical distance, and power transmission; at the same time, an automatic update mechanism is set to adjust the topology connection strength every 15 minutes according to the power grid operation status.

9. The intelligent identification method for power grid load status according to claim 1, characterized in that, In step S7, the specific process of adaptive learning and optimization is as follows: A dual-channel verification online learning mechanism was established, which integrates device inversion data and expert annotation results, and dynamically updates the feature library using a sliding window algorithm; Design a parameter self-optimization module for seasonal perception, automatically calibrate the DTW distance threshold based on kernel density estimation, and combine it with the hyperparameters of the genetic algorithm to evolve the classifier, so as to achieve periodic self-tuning of the algorithm parameters; For novel load access scenarios, a few-shot transfer learning framework was developed, which completed safe adaptation within a 7-day observation period through feature similarity matching and GAN data augmentation. Construct a federated knowledge-sharing network, employ differential privacy technology to achieve cross-regional feature parameter exchange, and complete the co-evolutionary model.

Images