Power grid digital security risk dynamic early warning system based on deep learning algorithm

Abstract

The present application relates to the technical field of power grid safety risk early warning, and relates to a power grid digital safety risk dynamic early warning system based on a deep learning algorithm, which is modularly designed through a voltage atlas acquisition module, a voltage deviation calculation module, a voltage risk quantification module, and a power grid safety early warning module, and realizes efficient monitoring, analysis and management of power grid voltage fluctuations. The construction of this system not only improves the scientificity and timeliness of power grid safety management, but also lays an important technical foundation for the intelligentization and modernization of power grids, and has indispensable importance for ensuring the safety, stability and sustainable operation of power grids.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power grid safety risk early warning, and relates to a power grid digital safety risk dynamic early warning system based on a deep learning algorithm. BACKGROUND

[0002] With large-scale access of distributed new energy, especially photovoltaic power generation, the power flow stability of the power grid becomes more and more complex. Especially in the low-voltage distribution network area, due to the high intermittency and high uncertainty of photovoltaic power generation, when sunny days and cloudy days alternate, the local voltage of some areas will appear short-time "climbing" or "falling" phenomenon, which affects the operation of surrounding electrical equipment.

[0003] In the prior art, for the power grid safety risk management after large-scale access of distributed new energy, especially photovoltaic power generation, there are still prominent defects and technical drawbacks. First, from the monitoring and sensing point of view, at present, the low-voltage distribution network generally lacks sufficient intensive and high-sensitivity real-time monitoring devices, and the existing voltage acquisition equipment mostly operates at a low time resolution, which cannot accurately capture the short-time voltage climbing or falling caused by the rapid fluctuation of photovoltaic power generation in the millisecond or second range, thereby causing a delay between the occurrence of the anomaly and the detection response, and leading to the difficulty in timely reflecting the operation state of the power grid. Secondly, in terms of data processing and feature extraction, the traditional method still mainly uses simple statistical indicators such as mean and variance, and for the disturbance mode of photovoltaic power generation with high frequency, small amplitude and superimposed randomness, the traditional indicators cannot effectively extract the fluctuation precursor features, nor can they identify abnormal trajectories from multi-dimensional joint features, resulting in low risk discrimination accuracy. SUMMARY

[0004] In view of the problems existing in the prior art, the application provides a power grid digital safety risk dynamic early warning system based on a deep learning algorithm, which is used to solve the above technical problems.

[0005] In order to achieve the above purpose and other purposes, the technical scheme adopted by the application is as follows:

[0006] The application provides a power grid digital safety risk dynamic early warning system based on a deep learning algorithm, which comprises:

[0007] The voltage atlas acquisition module: the voltage sensor continuously collects voltage signals of the power grid nodes to obtain original voltage signal data; and the original voltage signal data is subjected to voltage micro-disturbance trajectory atlas modeling to obtain micro-disturbance trajectory atlas data;

[0008] The voltage deviation calculation module: based on the micro-disturbance trajectory atlas data, normal trajectory feature space data is extracted to obtain normal trajectory feature space data; and real-time trajectory deviation degree is calculated according to the normal trajectory feature space data to generate trajectory deviation degree;

[0009] The voltage risk quantification module: based on the trajectory deviation degree, the abnormal precursor identification data of voltage fluctuation evolution risk quantification is obtained; according to the abnormal precursor identification data, the voltage fluctuation evolution risk quantification data is obtained;

[0010] The power grid safety warning module: based on the fluctuation evolution risk quantification data, it is compared with the preset multi-level risk threshold in real time, and different levels of risk warning signals are dynamically generated according to the comparison result; the risk warning signal is directly pushed to the power grid dispatching center warning terminal to drive it to realize visual alarm and execute corresponding risk disposal plan.

[0011] As described above, the power grid digital safety risk dynamic early warning system based on deep learning algorithm provided by the application has at least the following beneficial effects:

[0012] The power grid digital safety risk dynamic early warning system based on deep learning algorithm provided by the application, through the combination of the voltage atlas acquisition module, the voltage deviation calculation module, the voltage risk quantification module and the power grid safety warning module, not only improves the monitoring and management efficiency of the power grid in the technical aspect, but also provides necessary guarantee for ensuring the safe operation of the power grid. In the current power environment with frequent perturbations and increasing uncertainty, especially with the large-scale access of distributed new energy, the stability and reliability of the power grid become increasingly complex and important. Through this modular system architecture, all-around perception, analysis and response of voltage signals can be realized, thereby showing significant benefits and necessity in many aspects.

[0013] Firstly, the voltage atlas acquisition module can realize real-time and accurate acquisition of the voltage state of the power grid through continuous voltage signal acquisition of the power grid node by the voltage sensor. This process not only improves the real-time of monitoring, but also generates visual atlas data of voltage fluctuation through voltage micro-perturbation trajectory atlas modeling, so that the running state of the power grid is clear and visible, laying a foundation for subsequent analysis. This real-time monitoring and data visualization can help management personnel quickly understand the voltage conditions of each node of the power grid, discover problems in time and take measures, effectively reducing the potential risks caused by delayed response.

[0014] Secondly, the voltage deviation calculation module can analyze the deviation degree of the normal range of voltage fluctuation and the actual situation through the extraction of normal trajectory feature space and real-time trajectory deviation degree calculation. This process helps to further identify abnormal behaviors of voltage fluctuation and can issue an alarm in time when a slight voltage change occurs. With the access of new energy, the voltage of the power grid changes frequently and has burstiness. Early identification of these abnormal phenomena can help operators take necessary control measures in advance, thereby reducing the influence of voltage fluctuation on power grid equipment and user equipment and ensuring the safe operation of the power grid.

[0015] In the voltage risk quantification module, the evolution risk of voltage fluctuation is quantitatively analyzed through identification of abnormal precursors, providing a scientific basis for evaluating the stability of the power grid. Compared with the traditional experience judgment method, the risk quantification method based on data driving can improve the accuracy and reliability of risk assessment, so that the power grid dispatching can make scientific decisions based on quantitative data. This data-based decision will greatly reduce the uncertainty and blindness of the power grid in operation, and improve the flexibility and response capability of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0017] Figure 1 The schematic diagram of the connection of the modules of the system of the present application. DETAILED DESCRIPTION

[0018] The above description of the present application is only an example and explanation of the concept of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, as long as they do not deviate from the concept of the present application or exceed the scope defined by the present claims, and they should belong to the protection scope of the present application. EMBODIMENT

[0019] Please refer to Figure 1 As shown in the figure, the power grid digital security risk dynamic early warning system based on deep learning algorithm includes a voltage map acquisition module, a voltage deviation calculation module, a voltage risk quantification module and a power grid safety warning module.

[0020] The above modules are connected by wired and / or wireless connection to realize data transmission between the modules.

[0021] Voltage map acquisition module: continuous voltage signal acquisition of power grid nodes is carried out by voltage sensor to obtain original voltage signal data; voltage micro-disturbance trajectory map modeling is carried out on the original voltage signal data to obtain micro-disturbance trajectory map data.

[0022] The operation logic of the voltage map acquisition module is as follows:

[0023] Synchronously sample three-phase voltages of a power grid node through a voltage sensor to obtain original voltage signal time series data at a preset sampling frequency; perform sliding window difference denoising processing on the original voltage signal time series data to obtain denoised differential voltage fluctuation data; perform multi-scale time-frequency feature decomposition on the differential voltage fluctuation data to generate a joint feature data set containing time-domain waveform features and frequency-domain energy distribution features;

[0024] Based on the joint feature data set, construct a voltage fluctuation baseline parameter matrix through normal operating condition historical data, the baseline parameter matrix containing an amplitude variation threshold and a frequency coupling coefficient;

[0025] Perform hour-by-hour comparison between the real-time collected joint feature data set and the baseline parameter matrix to identify voltage micro-disturbance segment data exceeding the amplitude variation threshold;

[0026] Perform time-frequency feature trajectory parameterization processing on the voltage micro-disturbance segment data to extract three core feature parameters of each disturbance segment, namely, a starting phase, a disturbance duration cycle and a frequency band energy transition rate;

[0027] Model the core feature parameters of each disturbance segment according to a time sequence to generate micro-disturbance trajectory map data.

[0028] In specific embodiments, the complete data processing link of the voltage atlas acquisition module is implemented as follows: the original data is obtained by sampling the three-phase voltage of the power grid with GPS clock synchronization through a high-precision voltage sensor, the original data is processed by sliding window difference denoising, a 128-sample-point window length and a 75% overlap rate are selected, a sequence difference operation AV=V[t]-V[t-32] is performed to suppress common-mode noise, and a 32-sample-point step corresponds to a 20 ms time resolution. Based on the improved Morlet wavelet, multi-scale time-frequency decomposition is carried out, and time-domain feature parameters are extracted in the divided 1-50 times fundamental frequency band: the rising slope k1=(V1-V2) / At1, where V1 is the extreme voltage, V2 is the baseline voltage, and At1 is the rising time; the fluctuation peak density p1=N1 / T2, where N1 is the number of threshold fluctuations, and T2 is the 80 ms window length; the frequency domain features are combined to generate a feature set by calculating the energy intensity E1=√(∑|C_i|²) / N of each frequency band, C_i is the wavelet coefficient, and N is the total number of sampling points. Based on six months of historical data, a baseline parameter matrix is constructed, including an amplitude threshold d1= m1+3s1, m1 is the mean value, s1 is the standard deviation, a frequency coupling coefficient b1=E1 / E2, E1 is the high-frequency energy, and E2 is the low-frequency energy. In real-time monitoring, when the time-domain amplitude exceeds d1 and b1 deviates from the baseline value by more than ±15%, the perturbation segment is marked, and the core parameters are extracted: the starting phase f1 is determined by the minimum voltage difference detection, the disturbance duration T1=t_end-t_start is accurate to 0.5 cycles, and the energy transition rate g1 is calculated by cubic exponential smoothing. Finally, a three-dimensional feature space is constructed based on (f1, T1, g1), and a two-dimensional dynamic trajectory atlas is generated by cubic spline interpolation, where the horizontal axis maps the phase shift, the vertical axis represents the energy transition intensity, and the trajectory curvature K=|d²y / dx²| / (1+(dy / dx)²)^(3 / 2) quantifies the nonlinear characteristics of the disturbance evolution.

[0029] It should be noted that the real-time collected joint feature data set is compared with the baseline parameter matrix segment by segment, and the voltage micro-perturbation segment data exceeding the amplitude change threshold is identified; the specific operation steps are as follows:

[0030] The real-time collected joint feature data set is segmented and arranged in time sequence to obtain joint feature segmented data segment by segment;

[0031] The joint feature segmented data segment by segment is matched with the baseline parameter matrix to generate baseline comparison data;

[0032] The amplitude change of each period is calculated based on the baseline comparison data, and the amplitude change data is obtained; the amplitude change data is compared with the preset amplitude change threshold segment by segment to obtain comparison result data;

[0033] According to the comparison result data, part with amplitude change exceeding the threshold is screened out, and is identified as voltage micro-disturbance segment data.

[0034] In the embodiment of the present application, when the real-time collected joint feature data set is compared with the baseline parameter matrix in each time period, the joint feature data set of the current monitoring period is first segmented into continuous time window units according to the original data collection time stamp, and the time period division of each window strictly corresponds to the historical data time period of the baseline parameter matrix. For each real-time time window unit, its time domain amplitude feature sequence and frequency domain coupling coefficient sequence are extracted respectively, and are matched and mapped with the preset parameters of the corresponding period in the baseline matrix. Specifically, the sliding average of the real-time fluctuation peak density is compared with the baseline amplitude threshold value (composed of the statistical mean value of the historical data plus three times the standard deviation) horizontally, and the deviation percentage of the real-time frequency domain coupling coefficient from the baseline value is calculated. In this process, a double verification mechanism is introduced: for any window unit, when the number of continuous fluctuations of the time domain amplitude exceeds 150% of the baseline statistical peak value or the single fluctuation amplitude exceeds the upper threshold, a primary abnormality flag is triggered; further combined with the real-time offset of the high frequency and low frequency energy ratio in the window, a composite condition verification is performed, wherein a difference of ± 20% from the baseline value is determined as a significant deviation. After comparing all window units, the abnormality flag window is clustered using a space-time correlation analysis method, and when three or more adjacent window units trigger the abnormality flag and the frequency domain offset direction is consistent, it is determined that there is an effective voltage micro-disturbance segment in the time sequence. Finally, the start and end time stamps of the disturbance segment are accurately located through a sliding overlap verification mechanism, and the joint feature data set of all window units in the disturbance segment is packaged as voltage micro-disturbance segment data output. The sliding overlap verification mechanism is to trace back two windows and extend one window backward.

[0035] It should be noted that the starting phase, disturbance duration period, and frequency band energy transition rate of each disturbance segment are extracted as three core feature parameters, and the specific operation steps are as follows:

[0036] Perform time-frequency decomposition processing on the voltage micro-disturbance segment data to obtain time-frequency distribution data of the disturbance segment;

[0037] Identify the starting point position of the disturbance signal based on the time-frequency distribution data, calculate and extract the starting phase of the disturbance segment to obtain starting phase data;

[0038] Combine the starting phase data with the time-frequency distribution data, analyze the duration range of the disturbance signal on the time axis, extract the duration period of the disturbance segment to obtain disturbance duration period data;

[0039] According to the disturbance duration period data, the energy evolution of the time-frequency distribution data is evaluated, the energy transfer proportion of the disturbance signal between different frequency bands is calculated, and frequency band energy transition rate data is obtained;

[0040] The starting phase data, the disturbance duration period data and the frequency band energy transition rate data are uniformly output as the core characteristic parameters of the voltage micro-disturbance section.

[0041] After the voltage micro-disturbance section data is acquired in the embodiment of the application, the specific implementation of the characteristic parameter extraction operation is as follows: first, the voltage micro-disturbance section is subjected to time-frequency decomposition by improved short-time Fourier transform, an analysis window of 100 ms and an overlap rate of 75% are set, and the time-domain waveform is converted into time-frequency matrix data with time-frequency-energy three-dimensional characteristics. Based on the matrix, a double-threshold detection method is used to identify the disturbance starting point, the time position of the time-frequency energy sudden increase point is scanned, and when the high-frequency energy amplitude exceeds 200% of the baseline level and the low-frequency energy attenuation exceeds 30% in the continuous three analysis windows, the starting point is determined, and the voltage waveform of the first five fundamental periods before the starting point is intercepted for phase detection. The phase calculation uses the zero-crossing comparison method after band-pass filtering: the 50 Hz notch filter is used to eliminate the fundamental component, the remaining harmonic components are normalized, the zero-crossing point of the first negative peak is taken as the disturbance starting phase reference point, and the resolution is accurate to ±5° phase. In the duration period calculation stage, the time-frequency matrix is expanded along the time axis to form an energy attenuation curve, and the adaptive sliding window technology is used to mark the time cut-off point of the energy regression baseline: when the main frequency band energy is stable within ±15% of the baseline mean value in the continuous three windows, the disturbance termination is determined, and the time difference between the starting point and the termination point is the disturbance duration period, and the measurement resolution is set to 0.5 fundamental periods. When evaluating the frequency band energy transition rate, the time-frequency matrix is divided into six characteristic frequency bands at an interval of 100 Hz, the energy integral values of each frequency band in the disturbance duration period are calculated, and the energy proportion change rate is calculated based on the main disturbance frequency band. The specific method is to take the ratio of the energy proportion in the first 1 / 3 period to the energy proportion in the last 1 / 3 period, and if the ratio is >1, it indicates that the high-frequency energy is enhanced, and if the ratio is <1, it indicates that the energy is transferred to low frequency.

[0042] The voltage deviation calculation module: based on the micro-disturbance trajectory map data, the normal trajectory characteristic space data is obtained; the real-time trajectory deviation degree is calculated according to the normal trajectory characteristic space data, and the trajectory deviation degree is generated;

[0043] In one possible design, the operation logic of the voltage deviation calculation module is as follows:

[0044] The micro-disturbance trajectory map data is acquired, wherein the micro-disturbance trajectory map data includes three core characteristic parameters of the starting phase, the disturbance duration period and the frequency band energy transition rate of each disturbance section, the data is uniformly formatted and checked for integrity, and the characteristic parameter pre-processing data is obtained;

[0045] Based on the characteristic parameter pre-processing data, the starting phase distribution of different disturbance sections is aggregated and analyzed, and the disturbance starting phase distribution data is obtained.

[0046] The disturbance initial phase distribution data and the disturbance duration period data are jointly modeled, the time characteristics of the disturbance section under normal operation state are extracted, and normal trajectory time characteristic data are generated;

[0047] The normal trajectory time characteristic data and the frequency band energy transition rate data are fused, a normal trajectory characteristic space is constructed, and normal trajectory characteristic space data are obtained;

[0048] The real-time collected voltage trajectory characteristic parameters are correspondingly matched with the normal trajectory characteristic space data, the trajectory deviation value is calculated, and the trajectory deviation degree is obtained.

[0049] In the embodiment of the application, when the voltage deviation calculation module is operated, firstly, the input micro-disturbance trajectory atlas data are standardized and pretreated, the initial phase of each disturbance section is uniformly converted into a normalized phase angle based on the power frequency cycle, the range of 0-360° is mapped into the unit value of 0-1, the disturbance duration period is converted into the integer format of milliseconds, the frequency band energy transition rate is processed by percentage, and the sliding window verification method is used to interpolate and fill the missing data, so that the integrity and time sequence alignment of the characteristic parameters are ensured. Subsequently, initial phase aggregation analysis is performed: in the continuous monitoring period, the disturbance section occurrence frequency in each power frequency phase interval is counted, a phase-frequency distribution histogram is generated, and the abnormal aggregation area exceeding three standard deviations of the historical same period is marked. The phase distribution data and the disturbance duration period are jointly modeled by the kernel density estimation method, the conditional probability relationship between them is calculated, the conditional probability relationship is represented as the probability of a specific phase angle appearing corresponding to the duration of the disturbance, and the duration bandwidth of the 80% confidence interval under the normal operation state is extracted as the normal trajectory time characteristic. Then, the time characteristic and the energy transition rate are mapped in a three-dimensional space: in the time sequence coordinate system, the X-axis corresponds to the phase distribution mean value, the Y-axis represents the upper and lower limits of the duration bandwidth, and the Z-axis represents the positive and negative offset amplitude of the energy transition rate. The manifold learning algorithm is used to reduce the dimension of the historical normal data to generate a reference characteristic space. In the real-time data processing stage, the dynamically collected disturbance characteristic parameters are projected into the space after standardization, and the improved Hausdorff distance algorithm is used to calculate the minimum deviation distance between the real-time trajectory point and the boundary of the normal characteristic cloud. When the real-time phase offset exceeds 30% of the width of the baseline space projection area or the energy transition direction is opposite to the reference trend, a secondary deviation alarm is triggered. Finally, the deviation distance is converted into a deviation degree index of 0-100 through normalization, and the time sequence moving average filter is combined to eliminate short-term fluctuation interference, and the trajectory deviation degree curve is output.

[0050] In one possible design, the acquisition logic of the normal trajectory time characteristic data is as follows:

[0051] The disturbance starting phase distribution data is time-sequentially arranged to obtain starting phase sequence data; the starting phase sequence data is correspondingly matched with the disturbance duration period data to obtain phase-period corresponding data;

[0052] The time coverage range of the disturbance section is calculated based on the phase-period corresponding data to obtain disturbance time coverage data;

[0053] The disturbance time coverage data is trend-fitted to extract the typical time characteristic mode of the disturbance section under the normal operation state, and initial time characteristic data is generated; the initial time characteristic data is jointly corrected with the phase-period corresponding data to obtain corrected time characteristic data; and the corrected time characteristic data is output as normal trajectory time characteristic data.

[0054] In the acquisition operation of the normal trajectory time characteristic data, the starting phase of the historical disturbance section is first segmented and arranged in time order, the phase data is resampled at equal intervals with a 15-minute granularity window, the linear interpolation method is used to compensate the phase values of the missing time slots, and a continuous and complete starting phase time sequence is formed. Then, a mapping relationship table is established between the phase sequence and the corresponding disturbance duration period through a timestamp accurate matching mechanism, and a sliding difference window technology is used to identify abnormal associated points: when the length mutation rate of the two adjacent period data exceeds 50%, a verification mark is automatically inserted. Based on the associated data table, the time coverage range of each disturbance section is calculated, and the specific method is to take the backward extension of the starting phase corresponding time point to form a time coverage window, and to perform fusion processing on the overlapping windows to generate a continuous coverage time interval. Multi-scale trend fitting is performed on the coverage interval, a cubic spline curve is used to fit the daily cycle fluctuation trend at the hour level dimension, and an improved empirical mode decomposition is used to extract the intrinsic time modal component at the minute level dimension. The modal whose main component energy proportion exceeds 70% is reconstructed as an initial time characteristic curve. The curve is bidirectionally corrected with the original phase-period data: when forward verification is performed, the matching degree of the actual phase distribution and the envelope area of the characteristic curve is calculated, and the curve curvature of the section deviating more than twice the standard deviation is adjusted; when reverse verification is performed, the Fourier phase compensation algorithm is used to eliminate the time sequence deviation error, and finally the corrected time characteristic data is generated.

[0055] The voltage risk quantification module: based on the trajectory deviation degree, abnormal precursor identification of voltage fluctuation is performed to obtain abnormal precursor identification data; and voltage fluctuation evolution risk quantification is performed according to the abnormal precursor identification data to obtain fluctuation evolution risk quantification data;

[0056] In one possible design, the operation logic of the voltage risk quantification module is as follows:

[0057] The trajectory deviation degree data is obtained, and the trajectory deviation degree data is analyzed in each period to obtain trajectory deviation degree time sequence data;

[0058] extracting fluctuation anomaly features based on the trajectory deviation degree time series data, identifying abnormal precursor signals appearing in the voltage fluctuation process, and obtaining abnormal precursor identification data;

[0059] jointly analyzing the abnormal precursor identification data and the trajectory deviation degree time series data, calculating the occurrence frequency and the sustained intensity of the abnormal precursor signals, and obtaining abnormal precursor quantification data;

[0060] based on the abnormal precursor quantification data, performing evolution trend modeling to predict the change direction and amplitude of the voltage fluctuation in a future short time window, and obtaining voltage fluctuation evolution trend data;

[0061] comparing the voltage fluctuation evolution trend data with the preset risk evaluation standard, quantifying the risk level, and obtaining fluctuation evolution risk quantification data.

[0062] The implementation process of the voltage risk quantification module in the embodiment of the application is as follows: first, the input trajectory deviation degree data is divided into sliding windows with a time granularity of 5 minutes, the original time series data is denoised by using a double exponential smoothing method, and a standardized deviation degree curve with trend representation capability is formed. Based on the curve, abnormal precursor signal detection is performed: the absolute value of the deviation degree change rate between adjacent windows is calculated, when the change rates of three consecutive windows exceed the 95% confidence interval of the same period in history and the cumulative growth amplitude reaches 200% of the baseline level, it is determined that an effective abnormal precursor signal is detected. For the identified precursor signal cluster, the density peak clustering algorithm is used to aggregate discrete signals with similar spatial distribution characteristics into abnormal events, the trapezoidal integration method is used to calculate the area of each event under the deviation degree curve as the sustained intensity index, and the triggering frequency in each ten minutes is calculated according to the time interval of the event occurrence. In the evolution trend modeling stage, a deep prediction model based on a time convolution network is constructed, the high-density sampling deviation degree sequence in the past two hours at the current time is taken as the input, multi-scale time series features are extracted through multiple layers of hollow convolution, the deviation degree extreme value and change direction angle in the future 15-minute window are predicted, and the direction angle is divided into eight risk evolution quadrants with an interval of 45°. Finally, the predicted deviation degree extreme value is dynamically matched with the grid safe operation standard, when the extreme value exceeds the first warning threshold, a high risk level is generated, when the threshold value is in the range of 80-150% of the baseline value, a medium risk value is calculated according to the fault historical probability of the quadrant where the change direction angle is located, and when the baseline value is less than 80%, a low risk level is output.

[0063] In one possible design, the abnormal precursor identification data acquisition logic is as follows:

[0064] performing trend analysis on the trajectory deviation degree time series data, and obtaining trajectory deviation degree trend data;

[0065] Based on the trajectory deviation degree trend data, the rising section and the falling section of the time sequence fluctuation are calculated for fluctuation amplitude, and trajectory fluctuation amplitude data is obtained;

[0066] According to the trajectory fluctuation amplitude data, the local extreme points are identified, and the deviation degree of the extreme points is calculated, and the extreme point deviation data is obtained;

[0067] The extreme point deviation data and the trajectory deviation degree trend data are compared, and the abnormal fluctuation feature information is extracted, and the abnormal fluctuation feature data is generated;

[0068] Based on the abnormal fluctuation feature data, the abnormal precursor signals appearing in the voltage fluctuation process are identified, and the abnormal precursor identification data is obtained.

[0069] In the implementation of the acquisition process of the abnormal precursor identification data of the embodiment of the application, first, the trajectory deviation degree time sequence data is subjected to three-order Savitzky-Golay filtering processing to smooth the noise, while the mutation characteristics of the time sequence curve are retained. Based on the filtered data, a variable window length sliding analysis algorithm is used for trend analysis: in the rising trend section, the analysis window is adaptively shortened to 5 seconds, the first derivative of each data point in the window is calculated, and the mean value is taken as the rising trend strength; in the stable trend section, the window is expanded to 60 seconds, and the least square fitting line slope is used to quantify the trend direction stability. Then, the fluctuation amplitude is quantitatively evaluated, the cumulative deviation growth from the starting point to the peak of each rising section is calculated, and the maximum drawdown rate from the peak to the bottom of each falling section is calculated, and the fluctuation intensity is represented by the defined amplitude drawdown ratio, wherein the amplitude drawdown ratio is represented by the absolute value of the rising cumulative amount / the falling drawdown amount. In the local extreme point identification link, an improved double-threshold peak detection method is used: when the deviation degree values of three consecutive data points exceed twice the standard deviation of the previous window mean value and the slope signs of adjacent points are reversed, the candidate extreme points are marked, and further verified by the consistency of the trend direction of the five sampling points before and after the extreme points to exclude false peak interference. For the verified extreme points, the offset amount (Euclidean distance in three-dimensional space) of the extreme points from the reference position of the same phase section in the same period is calculated, and the offset risk index is calculated by combining the amplitude drawdown ratio of the trend section. Based on the joint mapping analysis of the offset index and the trend stability parameter, the abnormal fluctuation feature matrix is defined: when the offset risk index exceeds the threshold value and the trend stability parameter of the corresponding period is lower than the lower limit of the normal interval, the abnormal fluctuation event is determined. Finally, through the time sequence matching mode mining technology, the aggregation occurrence rule of similar feature events in three continuous fluctuation periods is detected, and when the event density reaches the preset alarm threshold, the abnormal precursor signal output is triggered.

[0070] In one possible design, the acquisition logic of the abnormal precursor quantitative data is as follows:

[0071] The abnormal precursor identification data is subjected to time sequence marking processing to obtain abnormal precursor time sequence marking data;The abnormal precursor time sequence marking data is compared with the trajectory deviation degree time sequence data to obtain abnormal precursor joint comparison data;Based on the abnormal precursor joint comparison data, the number of times of occurrence of the abnormal precursor signal within a certain time window is counted to obtain abnormal precursor occurrence frequency data;

[0072] The abnormal precursor occurrence frequency data and the trajectory deviation degree time sequence data are combined to calculate the duration of the abnormal precursor signal in time to obtain abnormal precursor duration intensity data;The abnormal precursor occurrence frequency data and the abnormal precursor duration intensity data are subjected to comprehensive processing to generate abnormal precursor quantification data.

[0073] In the process of acquiring the abnormal precursor quantification data, first, the abnormal precursor identification data is subjected to precise time stamp marking: using a GPS synchronous clock to synchronize each precursor signal trigger time to a precision of microseconds, and dividing the time slots on the time axis at an interval of 0.1 seconds to establish a time sequence marking index table. Then, the marked data and the trajectory deviation degree time sequence data are subjected to three-dimensional matching comparison through a sliding window correlation algorithm: in the time dimension, a correlation window with a forward extension of 50 ms and a backward coverage of 200 ms is set, and when there are deviation degree jumps and precursor signal trigger events in the window at the same time, it is recorded as an effective correlation event, and multi-dimensional joint comparison data containing time offset, deviation amplitude ratio and signal type are generated. Based on this data, the precursor signal occurrence rate index is calculated within a 15-minute statistical cycle, and the window is dynamically divided in specific implementation: the first 5 minutes are divided into 15-second granularity windows to count the average number of occurrences, and the last 10 minutes are subjected to inhibitory counting through an exponential decay weighting algorithm. The duration intensity quantification adopts an integral method with time decay, taking the deviation degree amplitude corresponding to each precursor signal as the reference value, multiplying its duration (the time length between the signal start and the adjacent deviation degree inflection point), and then performing time decay compensation according to the exponential function (the weight decays by 15% for every 1 minute from the current time). Finally, after normalizing the occurrence rate and intensity indicators, the comprehensive risk index is calculated through principal component analysis: setting the occurrence frequency weight coefficient as 0.4 and the duration intensity as 0.6, linearly combining the normalized indicators to generate a precursor quantification value on a 0-100 scale.

[0074] In one possible design, the voltage fluctuation evolution trend data acquisition logic is as follows:

[0075] The abnormal precursor quantification data is subjected to time sequence arrangement to obtain abnormal precursor time sequence data;Based on the abnormal precursor time sequence data, the fluctuation directionality analysis is performed to identify the rising trend or the falling trend of the voltage fluctuation to obtain fluctuation direction data;

[0076] The fluctuation direction data and the abnormal precursor quantization data are combined, the amplitude change range in a short time window is calculated, and fluctuation amplitude data are obtained;Trend fitting is performed according to the fluctuation direction data and the fluctuation amplitude data, and evolution trend fitting data are generated;

[0077] The evolution trend fitting data are corrected to remove abnormal interference points, and voltage fluctuation evolution trend data are obtained.

[0078] In the process of acquiring voltage fluctuation evolution trend data, first, the abnormal precursor quantization data is time axis reorganization and integrity repair: the discrete quantization value is resampled into an equidistant sequence with 100 milliseconds as the reference granularity, the data missing time slot caused by communication delay is compensated by cubic spline interpolation, and continuous and complete abnormal precursor time series data are generated. Based on this time series data, the fluctuation direction is analyzed, and the trend baseline is calculated in each 30-second analysis window by using the weighted moving average algorithm. When the difference between the latest quantization value and the baseline value exceeds 2.5 times the standard deviation of the window for 3 consecutive sampling points, the direction determination is triggered, and a direction persistence verification mechanism is introduced at the same time: if the duration of the upward trend is less than 40% of the window width, it is downgraded to noise disturbance. For the trend segment that passes the verification, the variable scale sliding extreme value detection method is further used to calculate the amplitude range: in the upward trend segment, the detection window is narrowed to 5 seconds to extract the peak value, and in the downward trend segment, it is expanded to 15 seconds to capture the bottom. The absolute difference between adjacent extreme points defines the transient amplitude. Based on the coordinated analysis of direction and amplitude data, a two-dimensional trend feature matrix is constructed: the row vector marks the positive and negative attributes of the trend direction, and the column vector records the gradient distribution of the amplitude change. The long short-term memory neural network is used to deeply model this matrix, and the feature sequence of the previous 1 hour is input to predict the fluctuation evolution trajectory of the next 5 minutes. During network training, a bidirectional time attention mechanism is added to enhance the key feature capture ability. The prediction result is filtered and corrected by a dynamic threshold: when the residual error between the predicted value and the actual monitoring value at a certain time exceeds 150% of the maximum residual error of the same period, Kalman filtering is triggered for real-time correction, and a smoothed trend curve is generated by combining the weighted average of the previous three prediction points. Finally, the spatial density clustering is used to remove isolated abnormal points, and the data segment that meets the trend direction for more than three consecutive sampling points is retained to form a voltage fluctuation evolution trend data set that eliminates sharp peak interference.

[0079] The power grid safety warning module: based on the fluctuation evolution risk quantization data, it is compared with the preset multi-level risk threshold in real time, and different levels of risk warning signals are dynamically generated according to the comparison result;The risk warning signal is directly pushed to the power grid dispatching center warning terminal to drive it to perform visual alarm and execute the corresponding risk disposal plan.

[0080] In one possible design, the operation logic of the power grid safety warning module is:

[0081] Acquire volatility evolution risk quantification data and organize it in chronological order to obtain risk quantification time series data; compare the risk quantification time series data with preset multi-level risk thresholds level by level to obtain risk threshold comparison data;

[0082] Based on risk threshold comparison data, risk level ranges are identified, and corresponding risk warning signals are dynamically generated to obtain risk warning signal data. The risk warning signal data is then directly pushed to the warning terminal of the power grid dispatch center to drive it to perform visual alarm processing to obtain visual alarm data.

[0083] Based on the visualized alarm data, a response plan matching the risk level is executed to obtain risk response execution data, which is then output by the system.

[0084] The implementation process of the power grid safety early warning module in this embodiment of the invention is as follows: First, the real-time input fluctuation evolution risk quantification data is processed with millisecond-level timestamp alignment, and then converted into a standardized risk index with a granularity of 30 seconds using a sliding window integral algorithm. At the same time, a time axis mapping relationship with the historical risk database is established. In the risk threshold comparison stage, a multi-track dynamic threshold system is set up—a basic threshold (historical mean + 2 times the standard deviation), an early warning threshold, and an emergency threshold (90% of the real-time load capacity). A three-level nested verification mechanism is used to match risk levels: when the risk index is within the range of the basic threshold and the early warning threshold for three consecutive minutes, a blue early warning is triggered; when the early warning threshold is exceeded, it is converted to an orange level two alarm; if the emergency threshold is touched twice within five minutes, it is upgraded to a red highest alarm.

[0085] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0086] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0088] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power grid digital security risk dynamic early warning system based on a deep learning algorithm, characterized in that, The system comprises the following modules: The voltage atlas acquisition module: through the voltage sensor, the continuous voltage signal of the power grid node is collected to obtain the original voltage signal data; the original voltage signal data is modeled by voltage micro-disturbance trajectory atlas to obtain micro-disturbance trajectory atlas data; The voltage deviation calculation module: based on the micro-disturbance trajectory atlas data, the normal trajectory feature space data is obtained by extracting the normal trajectory feature space; the trajectory deviation degree is calculated according to the normal trajectory feature space data to generate the trajectory deviation degree; The operation logic of the voltage deviation calculation module is as follows: Obtain the micro-disturbance trajectory atlas data, wherein the micro-disturbance trajectory atlas data includes three core feature parameters of the starting phase, the disturbance duration period and the frequency band energy transition rate of each disturbance segment, and the data is uniformly formatted and checked for integrity to obtain feature parameter preprocessed data, wherein the frequency band energy transition rate data is obtained by calculating the energy transfer rate of the disturbance signal between different frequency bands; Based on the feature parameter preprocessed data, the starting phase distribution of different disturbance segments is aggregated and analyzed to obtain disturbance starting phase distribution data; The disturbance starting phase distribution data and the disturbance duration period data are jointly modeled to extract the time characteristics of the disturbance segment under normal operation to generate normal trajectory time characteristic data; The normal trajectory time characteristic data and the frequency band energy transition rate data are fused to construct a normal trajectory feature space to obtain normal trajectory feature space data; The real-time collected voltage trajectory feature parameters are matched with the normal trajectory feature space data to calculate the trajectory deviation value to obtain the trajectory deviation degree; The voltage risk quantification module: based on the trajectory deviation degree, the voltage fluctuation abnormal precursor identification data is obtained; According to the abnormal precursor identification data, the voltage fluctuation evolution risk quantification data is obtained; The power grid safety early warning module: based on the fluctuation evolution risk quantification data, it is compared with the preset multi-level risk threshold in real time, and different levels of risk warning signals are dynamically generated according to the comparison result; the risk warning signal is directly pushed to the power grid dispatching center early warning terminal to drive it to perform visual alarm and execute the corresponding risk disposal plan.

2. The power grid digitization security risk dynamic early warning system based on a deep learning algorithm according to claim 1, characterized in that, The operation logic of the voltage atlas acquisition module is as follows: The three-phase voltage of the power grid node is synchronously sampled by the voltage sensor to obtain the original voltage signal time series data at a preset sampling frequency; the original voltage signal time series data is processed by sliding window difference denoising to obtain denoised difference voltage fluctuation data; the difference voltage fluctuation data is decomposed by multi-scale time-frequency feature to generate a joint feature data set containing time domain waveform features and frequency domain energy distribution features; Based on the joint feature data set, a voltage fluctuation baseline parameter matrix is constructed by normal working condition historical data, and the baseline parameter matrix includes amplitude variation threshold and frequency coupling coefficient; The real-time collected joint feature data set is compared with the baseline parameter matrix in each period to identify the voltage micro-disturbance segment data exceeding the amplitude variation threshold; Perform time-frequency feature trajectory parameterization processing on the voltage micro-disturbance segment data, and extract three core feature parameters of each disturbance segment, including starting phase, disturbance duration period, and frequency band energy transition rate; Model the core feature parameters of each disturbance segment in time sequence to generate micro-disturbance trajectory atlas data. 3.The power grid digitization security risk dynamic early warning system based on deep learning algorithm of claim 2, characterized in that, Compare the real-time collected joint feature data set with the baseline parameter matrix in each time segment to identify voltage micro-disturbance segment data that exceeds the amplitude change threshold; the specific operation steps are: Obtain the real-time collected joint feature data set, and segment and organize it in time sequence to obtain joint feature segmented data in each time segment; Correspondingly match the joint feature segmented data in each time segment with the baseline parameter matrix to generate baseline comparison data; Calculate the amplitude change of each time segment based on the baseline comparison data, and obtain amplitude change data; Compare the amplitude change data with the preset amplitude change threshold in each time segment to obtain comparison result data; According to the comparison result data, filter out the part whose amplitude change exceeds the threshold, and identify it as voltage micro-disturbance segment data.

4. The power grid digitization security risk dynamic early warning system based on a deep learning algorithm according to claim 2, characterized in that, Extract three core feature parameters of each disturbance segment, including starting phase, disturbance duration period, and frequency band energy transition rate; The specific operation steps are: Perform time-frequency decomposition processing on the voltage micro-disturbance segment data to obtain time-frequency distribution data of the disturbance segment; Based on the time-frequency distribution data, identify the starting point position of the disturbance signal, calculate and extract the starting phase of the disturbance segment to obtain starting phase data; Combine the starting phase data with the time-frequency distribution data to analyze the duration range of the disturbance signal on the time axis, extract the duration period of the disturbance segment, and obtain disturbance duration period data; According to the disturbance duration period data, perform energy evolution evaluation on the time-frequency distribution data, calculate the energy transfer proportion of the disturbance signal between different frequency bands, and obtain frequency band energy transition rate data; Uniformly output the starting phase data, disturbance duration period data, and frequency band energy transition rate data as the core feature parameters of the voltage micro-disturbance segment.

5. The deep learning algorithm-based power grid digitization security risk dynamic early warning system according to claim 1, characterized in that, The logic for obtaining normal trajectory time feature data is: Arrange the disturbance starting phase distribution data in time sequence to obtain starting phase sequence data; correspondingly match the starting phase sequence data with the disturbance duration period data to obtain phase-period corresponding data; Based on the phase-period corresponding data, calculate the time coverage of the disturbance segment to obtain disturbance time coverage data; Trend fitting is performed on the disturbance time coverage data to extract the typical time feature mode of the disturbance segment under normal operating conditions to generate initial time feature data; Jointly correct the initial time feature data with the phase-period corresponding data to obtain corrected time feature data; output the corrected time feature data as normal trajectory time feature data.

6. The deep learning algorithm-based power grid digitization security risk dynamic early warning system according to claim 1, characterized in that, The operation logic of the voltage risk quantification module is: Obtain trajectory deviation degree data, and analyze the trajectory deviation degree data in each time segment to obtain trajectory deviation degree time series data; Based on the trajectory deviation degree time series data, extract fluctuation abnormal feature to identify abnormal precursor signals in the voltage fluctuation process to obtain abnormal precursor identification data; The abnormal precursor identification data is combined with the trajectory deviation degree time sequence data to analyze, the occurrence frequency and the continuous intensity of the abnormal precursor signal are calculated, and abnormal precursor quantitative data is obtained; Based on the abnormal precursor quantitative data, the evolution trend modeling is carried out, the change direction and amplitude of the voltage fluctuation in the future short time window are predicted, and voltage fluctuation evolution trend data is obtained; The voltage fluctuation evolution trend data is compared with the preset risk evaluation standard, the risk level is quantified, and fluctuation evolution risk quantitative data is obtained.

7. The deep learning algorithm-based power grid digitization security risk dynamic early warning system according to claim 6, characterized in that, The logic for obtaining abnormal precursor identification data is as follows: The trajectory deviation degree time sequence data is analyzed to obtain trajectory deviation degree trend data; Based on the trajectory deviation degree trend data, the fluctuation amplitude of the rising section and the falling section of the time sequence fluctuation is calculated to obtain trajectory fluctuation amplitude data; According to the trajectory fluctuation amplitude data, the local extreme points are identified, and the deviation degree of the extreme points is calculated to obtain extreme point deviation data; The extreme point deviation data is combined with the trajectory deviation degree trend data to extract abnormal fluctuation feature information, and abnormal fluctuation feature data is generated; Based on the abnormal fluctuation feature data, the abnormal precursor signal appearing in the voltage fluctuation process is identified, and the abnormal precursor identification data is obtained. 8.The deep learning algorithm-based power grid digitization security risk dynamic early warning system according to claim 6, characterized in that, The logic for obtaining abnormal precursor quantitative data is as follows: The abnormal precursor identification data is time sequence marked to obtain abnormal precursor time sequence marked data; the abnormal precursor time sequence marked data is compared with the trajectory deviation degree time sequence data to obtain abnormal precursor combined comparison data; based on the abnormal precursor combined comparison data, the number of times of the abnormal precursor signal appearing in a certain time window is counted to obtain abnormal precursor occurrence frequency data; Combined with the abnormal precursor occurrence frequency data and the trajectory deviation degree time sequence data, the duration of the abnormal precursor signal in time is calculated to obtain abnormal precursor continuous intensity data; the abnormal precursor occurrence frequency data and the abnormal precursor continuous intensity data are comprehensively processed to generate abnormal precursor quantitative data. 9.The power grid digitization security risk dynamic early warning system based on deep learning algorithm of claim 6, wherein, The logic for obtaining voltage fluctuation evolution trend data is as follows: The abnormal precursor quantitative data is time sequence arranged to obtain abnormal precursor time sequence data; Based on the abnormal precursor time sequence data, fluctuation directionality analysis is performed to identify the rising trend or the falling trend of the voltage fluctuation, and fluctuation direction data is obtained; The fluctuation direction data is combined with the abnormal precursor quantitative data to calculate the amplitude variation range in the short time window, and fluctuation amplitude data is obtained; According to the fluctuation direction data and the fluctuation amplitude data, trend fitting is performed to generate evolution trend fitting data; The evolution trend fitting data is corrected to remove abnormal interference points, and voltage fluctuation evolution trend data is obtained.

10. The deep learning algorithm-based power grid digitization security risk dynamic early warning system according to claim 1, characterized in that, The operation logic of the power grid safety early warning module is as follows: The fluctuation evolution risk quantitative data is obtained, and it is arranged in time sequence to obtain risk quantitative time sequence data; the risk quantitative time sequence data is compared with the preset multi-level risk threshold value step by step to obtain risk threshold comparison data; Based on the risk threshold comparison data, the risk level interval is identified, and the corresponding level risk early warning signal is dynamically generated to obtain risk early warning signal data; The risk early warning signal data is directly pushed to an early warning terminal of a power grid dispatching center, driving the early warning terminal to perform visual alarm processing, and obtaining visual alarm data; Based on the visual alarm data, a disposal plan matched with the risk level is executed, risk disposal execution data is obtained, and the execution data is taken as system output.

Images