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

Abstract

The invention 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. And efficient monitoring, analysis and management of the voltage fluctuation of the power grid are realized. The construction of the system not only improves the scientificity and timeliness of power grid safety management, but also lays an important technical foundation for intelligentization and modernization of the power grid, and has indispensable importance for guaranteeing the safety, stability and sustainable operation of the power grid.

Description

Technical Field

[0001] This invention belongs to the field of power grid security risk early warning technology, and relates to a dynamic early warning system for digital security risks of power grids based on deep learning algorithms. Background Technology

[0002] With the large-scale integration of distributed renewable energy sources, especially photovoltaic power generation, the power flow stability of the power grid has become increasingly complex. Especially in low-voltage distribution network areas, due to the high intermittency and uncertainty of photovoltaic power generation, short-term local voltage "climbing" or "dropping" phenomena may occur in some areas when sunny and cloudy days alternate, affecting the operation of surrounding electrical equipment.

[0003] In existing technologies, there are still significant shortcomings and technical drawbacks in grid security risk management for distributed renewable energy, especially photovoltaic (PV) power, after large-scale grid integration. Firstly, from a monitoring and sensing perspective, low-voltage distribution networks generally lack sufficiently dense and highly sensitive real-time monitoring devices. Most existing voltage acquisition equipment operates with low time resolution, failing to accurately capture short-term voltage rises or falls caused by rapid fluctuations in PV power generation within milliseconds or seconds. This results in a delay between anomaly occurrence and detection response, making it difficult to reflect the grid's operational status in a timely manner. Secondly, in terms of data processing and feature extraction, traditional methods still rely primarily on simple statistical indicators such as mean and variance. For the high-frequency, small-amplitude, and superimposed random disturbance patterns of PV power generation, traditional indicators cannot effectively extract precursor features of fluctuations, nor can they identify abnormal trajectories from multi-dimensional joint features, resulting in low accuracy in risk assessment. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention provides a dynamic early warning system for digital safety risks of power grids based on deep learning algorithms, which is used to solve the above-mentioned technical problems.

[0005] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows: This invention provides a dynamic early warning system for digital safety risks in power grids based on deep learning algorithms. The system includes: Voltage spectrum acquisition module: continuously acquires voltage signals from grid nodes using voltage sensors to obtain raw voltage signal data; models voltage micro-perturbation trajectory data from the raw voltage signal data to obtain micro-perturbation trajectory data; Voltage deviation calculation module: Extracts normal trajectory feature space from micro-perturbation trajectory map data to obtain normal trajectory feature space data; calculates real-time trajectory deviation based on normal trajectory feature space data to generate trajectory deviation. Voltage risk quantification module: Based on trajectory deviation, it identifies abnormal precursors of voltage fluctuations and obtains abnormal precursor identification data; based on the abnormal precursor identification data, it quantifies the risk of voltage fluctuation evolution and obtains fluctuation evolution risk quantification data. Power grid safety early warning module: Based on the fluctuation evolution risk quantification data, it compares the data with preset multi-level risk thresholds in real time, and dynamically generates risk warning signals of different levels according to the comparison results; the risk warning signals are directly pushed to the early warning terminal of the power grid dispatch center, driving it to perform visual alarm and execute the corresponding risk handling plan.

[0006] As described above, the dynamic early warning system for digital power grid security risks based on deep learning algorithms provided by this invention has at least the following beneficial effects: The present invention provides a dynamic early warning system for digital power grid security risks based on deep learning algorithms. By combining a voltage spectrum acquisition module, a voltage deviation calculation module, a voltage risk quantification module, and a power grid security early warning module, it not only improves the efficiency of power grid monitoring and management at the technical level but also provides necessary guarantees for ensuring the safe operation of the power grid. In the current power environment characterized by frequent micro-disturbances and increasing uncertainty, especially with the large-scale integration of distributed renewable energy sources, the stability and reliability of the power grid have become increasingly complex and important. This modular system architecture enables comprehensive perception, analysis, and response to voltage signals, demonstrating significant benefits and necessity in many aspects.

[0007] First, the voltage spectrum acquisition module continuously collects voltage signals from grid nodes using voltage sensors, enabling real-time and accurate acquisition of the grid's voltage status. This process not only improves the real-time performance of monitoring but also generates visualized voltage fluctuation data through voltage micro-disturbance trajectory modeling, making the grid's operating status clearly visible and laying the foundation for subsequent analysis. This real-time monitoring and data visualization helps managers quickly understand the voltage status of each grid node, promptly identify problems, and take measures, effectively reducing potential risks caused by delayed response.

[0008] Secondly, the voltage deviation calculation module, through the extraction of normal trajectory feature space and real-time trajectory deviation calculation, can analyze the degree of deviation between the normal range of voltage fluctuations and the actual situation. This process helps to further identify abnormal voltage fluctuation behavior and can issue timely alarms when even minor voltage changes occur. With the integration of new energy sources, the voltage fluctuations in the power grid are frequent and sudden. Early identification of these anomalies can help operators take necessary control measures in advance, thereby reducing the impact of voltage fluctuations on power grid equipment and user equipment, and ensuring the safe operation of the power grid.

[0009] In the voltage risk quantification module, the identification of anomaly precursors and the quantitative analysis of the evolutionary risk of voltage fluctuations provide a scientific basis for assessing grid stability. Compared with traditional experience-based judgment methods, data-driven risk quantification methods can improve the accuracy and reliability of risk assessment, enabling grid dispatching to make scientific decisions based on quantitative data. This data-driven decision-making will significantly reduce the uncertainty and blind spots in grid operation, improving the grid's flexibility and responsiveness. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0012] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention. Example

[0013] Please see Figure 1 As shown, a dynamic early warning system for digital security risks in power grids based on deep learning algorithms is presented. The system includes a voltage spectrum acquisition module, a voltage deviation calculation module, a voltage risk quantification module, and a power grid security early warning module. The various modules mentioned above are connected via wired and / or wireless means to enable data transmission between them; Voltage spectrum acquisition module: continuously acquires voltage signals from grid nodes using voltage sensors to obtain raw voltage signal data; models voltage micro-perturbation trajectory data from the raw voltage signal data to obtain micro-perturbation trajectory data; The operation logic of the voltage spectrum acquisition module is as follows: The three-phase voltage of the power grid node is synchronously sampled by a voltage sensor to obtain the original voltage signal time series data at a preset sampling frequency; the original voltage signal time series data is subjected to sliding window differential noise reduction processing to obtain the noise-reduced differential voltage fluctuation data; the differential voltage fluctuation data is decomposed into multi-scale time-frequency features to generate a joint feature dataset containing time-domain waveform features and frequency-domain energy distribution features. Based on the joint feature dataset, a voltage fluctuation baseline parameter matrix is ​​constructed using historical data under normal operating conditions. The baseline parameter matrix includes an amplitude change threshold and a frequency coupling coefficient. The real-time acquired joint feature dataset is compared with the baseline parameter matrix on a time-by-time basis to identify voltage micro-perturbation segments that exceed the amplitude change threshold. The voltage micro-perturbation segment data is subjected to time-frequency characteristic trajectory parameterization processing to extract three core characteristic parameters for each perturbation segment: the initial phase, the duration of the perturbation, and the frequency band energy transition rate. The core characteristic parameters of each disturbance segment are dynamically modeled according to the time series to generate micro-disturbance trajectory map data.

[0014] In a specific embodiment, the complete data processing link of the voltage spectrum acquisition module is implemented as follows: the original data is obtained by GPS clock synchronization sampling of the three-phase voltage of the power grid through a high-precision voltage sensor. The original data is processed by sliding window differential noise reduction. A window length of 128 sampling points and an overlap rate of 75% are selected. Sequence differential operation ΔV=V[t]-V[t-32] is performed to suppress common-mode noise. The step size of 32 sampling points corresponds to a time resolution of 20ms. Multi-scale time-frequency decomposition was performed based on an improved Morlet wavelet, extracting time-domain feature parameters within the divided 1-50 times fundamental frequency bands: rising slope k1=(V1-V2) / Δt1, where V1 is the extreme voltage, V2 is the baseline voltage, and Δt1 is the rise time; peak fluctuation density ρ1=N1 / T2, where N1 is the number of fluctuations exceeding the threshold, and T2 is the 80ms window duration; frequency domain features were generated by calculating the energy intensity of each frequency band E1=√(∑|C_i|²) / N to generate a joint feature set, where C_i is the wavelet coefficient and N is the total number of sampling points. A baseline parameter matrix was constructed based on six months of historical data, including the amplitude threshold δ1=μ1+3σ1, where μ1 is the mean and σ1 is the standard deviation, and the frequency coupling coefficient β1=E1 / E2, where E1 is the high-frequency energy and E2 is the low-frequency energy. In real-time monitoring, when the time domain amplitude exceeds δ1 and β1 deviates from the baseline value by more than ±15%, a micro-perturbation segment is marked, and core parameters are extracted: the initial phase φ1 is determined by minimum voltage difference detection, the perturbation duration T1 = t_end - t_start is accurate to 0.5 cycles, and the energy transition rate γ1 is calculated using cubic exponential smoothing. Finally, a three-dimensional feature space is constructed based on (φ1, T1, γ1), and a two-dimensional dynamic trajectory map is generated by cubic spline interpolation, where the horizontal axis maps the phase offset, the vertical axis represents the energy transition intensity, and the trajectory curvature κ = |d²y / dx²| / (1+(dy / dx)²)^(3 / 2) quantifies the nonlinear characteristics of the perturbation evolution.

[0015] It should be added that the real-time acquired joint feature dataset is compared with the baseline parameter matrix time-by-time to identify voltage micro-perturbation segments that exceed the amplitude change threshold; the specific operation steps are as follows: Obtain the real-time collected joint feature dataset and organize it into segments according to time order to obtain segmented joint feature data for each time period; The joint feature segmentation data for each time period is matched with the baseline parameter matrix to generate baseline comparison data; The amplitude change is calculated for each time period based on the baseline comparison data, and the amplitude change data is obtained. The amplitude change data is then compared with the preset amplitude change threshold for each time period to obtain the comparison result data. Based on the comparison results, the portion of amplitude change exceeding the threshold was selected and identified as voltage micro-perturbation segment data.

[0016] In this embodiment of the invention, when performing time-by-time comparison between the real-time acquired joint feature dataset and the baseline parameter matrix, the joint feature dataset of the current monitoring period is first divided into continuous time window units according to the original data acquisition timestamp. The time period division of each window strictly corresponds to the historical data time periods 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 and matched with the preset parameters of the corresponding time period in the baseline matrix. Specifically, the moving average of the real-time fluctuation peak density is compared horizontally with the baseline amplitude threshold (composed of the statistical mean of historical data plus three times the standard deviation), and the percentage deviation of the real-time frequency-domain coupling coefficient from the baseline value is calculated. A dual verification mechanism is introduced in this process: for any window unit, when the number of consecutive fluctuations in the time-domain amplitude exceeds 150% of the baseline statistical peak or the amplitude of a single fluctuation exceeds the upper limit of the threshold, a primary anomaly marker is triggered; further, a composite condition verification is performed by combining the real-time offset of the high-frequency to low-frequency energy ratio within the window, where a difference of ±20% from the baseline value is judged as a significant deviation. After comparing all window units, spatiotemporal correlation analysis is used to cluster the anomaly-marked windows. When three or more adjacent window units trigger anomaly marking and their frequency domain offset directions are consistent, the time series is determined to have an effective voltage micro-perturbation segment. Finally, a sliding overlap verification mechanism is used to accurately locate the start and end timestamps of the perturbation segment, and the joint feature dataset of all window units in between is packaged into the voltage micro-perturbation segment data output. The sliding overlap verification mechanism involves tracing back two windows and extending one window backward.

[0017] It should be added that three core feature parameters are extracted for each perturbation segment: the initial phase, the duration of the perturbation, and the frequency band energy transition rate. The specific operation steps are as follows: Time-frequency decomposition processing is performed on the voltage micro-perturbation segment data to obtain the time-frequency distribution data of the perturbation segment; The starting point of the disturbance signal is identified based on the time-frequency distribution data, and the starting phase of the disturbance segment is calculated and extracted to obtain the starting phase data. By combining the initial phase data with the time-frequency distribution data, the duration of the disturbance signal on the time axis is analyzed, the duration period of the disturbance segment is extracted, and the disturbance duration period data is obtained. Based on the disturbance duration period data, the energy evolution of the time-frequency distribution data is evaluated, and the energy transfer ratio of the disturbance signal between different frequency bands is calculated to obtain the frequency band energy transition rate data. The initial phase data, disturbance duration period data, and frequency band energy transition rate data are output in a unified manner as the core characteristic parameters of the voltage micro-disturbance segment.

[0018] In this embodiment of the invention, after acquiring the voltage micro-perturbation segment data, the specific implementation of the feature parameter extraction operation is as follows: First, the voltage micro-perturbation segment is subjected to improved short-time Fourier transform for time-frequency decomposition. A 100ms analysis window and a 75% overlap rate are set to convert the time-domain waveform into time-frequency matrix data with three-dimensional time-frequency-energy characteristics. Based on this matrix, a dual-threshold detection method is used to identify the disturbance initiation point. By scanning the time position of the time-frequency energy surge point, when the high-frequency energy increase exceeds 200% of the baseline level and the low-frequency energy attenuation exceeds 30% within three consecutive analysis windows, it is determined as the initiation point. The voltage waveform of the five fundamental cycles before the initiation point is then extracted for phase detection. The phase calculation adopts the zero-crossing comparison method after bandpass filtering: the fundamental component is eliminated by a 50Hz notch filter, the remaining harmonic components are normalized, and the zero-crossing point that first reaches the negative peak value is taken as the disturbance initiation phase reference point, accurate to ±5° phase resolution. During the duration calculation phase, the time-frequency matrix is ​​expanded along the time axis to form an energy decay curve. An adaptive sliding window technique is used to mark the time cutoff point for energy return to the baseline: the disturbance is considered terminated when the energy in the main frequency band remains stable within ±15% of the baseline mean for three consecutive windows. The time difference between the start and end points is the disturbance duration. The measurement resolution is set to 0.5 fundamental cycles. For frequency band energy transition rate evaluation, the time-frequency matrix is ​​divided into six characteristic frequency bands at 100Hz intervals. The energy integral value of each frequency band within the disturbance duration is calculated. The rate of change of its energy proportion is calculated based on the main disturbance frequency band. Specifically, the ratio of the energy proportion in the first 1 / 3 of the cycle to the energy proportion in the second 1 / 3 of the cycle is taken. If the ratio > 1, it indicates high-frequency energy enhancement; if < 1, it indicates a shift to lower frequencies.

[0019] Voltage deviation calculation module: Extracts normal trajectory feature space from micro-perturbation trajectory map data to obtain normal trajectory feature space data; calculates real-time trajectory deviation based on normal trajectory feature space data to generate trajectory deviation. In one possible design, the operating logic of the voltage deviation calculation module is as follows: Acquire micro-perturbation trajectory map data, wherein the micro-perturbation trajectory map data includes three core feature parameters for each perturbation segment: the initial phase, the duration of the perturbation, and the frequency band energy transition rate. Perform format uniformity and integrity checks on the data to obtain feature parameter preprocessing data. Aggregate analysis of the initial phase distribution of different disturbance segments is performed based on the preprocessed data of characteristic parameters to obtain the initial phase distribution data of disturbance; By jointly modeling the initial phase distribution data of the disturbance and the duration period data of the disturbance, the temporal characteristics of the disturbance segment under normal operating conditions are extracted, and normal trajectory temporal characteristic data are generated. Normal trajectory time feature data and frequency band energy transition rate data are fused to construct a normal trajectory feature space, thus obtaining 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 and obtain the trajectory deviation degree.

[0020] In this embodiment of the invention, when the voltage deviation calculation module is operated, the input micro-disturbance trajectory map data is first standardized and preprocessed. The starting phase of each disturbance segment is uniformly converted into a normalized phase angle based on the power frequency period, and the 0-360° range is mapped to a 0-1 per-unit value. The disturbance duration period is converted into a millisecond-level integer format, and the frequency band energy transition rate is processed as a percentage. At the same time, the sliding window verification method is used to interpolate and fill missing data to ensure the integrity of feature parameters and time sequence alignment. Then, the starting phase aggregation analysis is performed: within the continuous monitoring period, the frequency of disturbance segments in each power frequency phase interval is counted, and a phase-frequency distribution histogram is generated. Abnormal clusters exceeding three times the standard deviation of the historical period are marked. The phase distribution data and the disturbance duration period are jointly modeled using the kernel density estimation method, and the conditional probability relationship between the two is calculated. The conditional probability relationship represents the probability of a disturbance of a corresponding duration occurring at a specific phase angle. The duration bandwidth of the 80% confidence interval under normal operating conditions is extracted as the normal trajectory time feature. Then, a three-dimensional spatial mapping is performed between temporal features and energy transition rates: in the temporal coordinate system, the X-axis corresponds to the mean of the phase distribution, the Y-axis represents the upper and lower limits of the duration bandwidth, and the Z-axis characterizes the positive and negative offset magnitude of the energy transition rate. A manifold learning algorithm is used to reduce the dimensionality of historical normal data to generate a baseline feature space. In the real-time data processing stage, the dynamically acquired perturbation feature parameters are standardized and projected into this space. A modified Hausdorff distance algorithm is used to calculate the minimum deviation distance between the real-time trajectory point and the boundary of the normal feature cloud: when the real-time phase offset exceeds 30% of the width of the baseline spatial projection area or the energy transition direction is opposite to the baseline trend, a secondary deviation alarm is triggered. Finally, the deviation distance is converted into a deviation index of 0-100 through normalization processing, and combined with temporal moving average filtering to eliminate short-term fluctuation interference, outputting the trajectory deviation curve.

[0021] In one possible design, the logic for obtaining normal trajectory time feature data is as follows: The initial phase distribution data of the disturbance is sorted in time order to obtain the initial phase sequence data; the initial phase sequence data is matched with the disturbance duration period data to obtain the phase-period correspondence data. The time coverage range of the disturbance segment is calculated based on the phase-period correspondence data to obtain the disturbance time coverage data; Trend fitting is performed on the disturbance time coverage data to extract typical time feature patterns of the disturbance segment under normal operating conditions, generating initial time feature data; the initial time feature data is jointly corrected with phase-period corresponding data to obtain corrected time feature data; the corrected time feature data is output as normal trajectory time feature data.

[0022] In this embodiment of the invention, when acquiring the normal trajectory time feature data, the initial phase of the historical disturbance segment is first segmented and organized according to the chronological order of occurrence. Phase data is resampled at equal intervals using a 15-minute granularity window, and linear interpolation is used to compensate for missing time slots, forming a continuous and complete initial phase time series. Subsequently, a mapping table is established between the phase series and the corresponding disturbance duration period using a timestamp precise matching mechanism. Anomaly correlation points are identified using a sliding difference window technique: when the duration abrupt change rate of two adjacent periods exceeds 50%, a check mark is automatically inserted. Based on the correlation data table, the time coverage range of each disturbance segment is calculated line by line. Specifically, a time coverage window is formed by extending backward from the time point corresponding to the initial phase, and overlapping windows are fused to generate a continuous coverage time interval. Multi-scale trend fitting is performed on the coverage interval. At the hourly level, cubic spline curves are used to fit the daily periodic fluctuation trend. At the minute level, improved empirical mode decomposition is used to extract inherent time mode components, and modes with a principal component energy ratio exceeding 70% are reconstructed as initial time feature curves. The curve is bidirectionally corrected with the original phase-period data: during forward verification, the matching degree between the actual phase distribution and the envelope region of the characteristic curve is calculated, and the curve curvature is adjusted for segments that deviate by more than twice the standard deviation; during reverse verification, the timing offset error is eliminated based on the Fourier phase compensation algorithm, and finally the corrected time characteristic data is generated.

[0023] Voltage risk quantification module: Based on trajectory deviation, it identifies abnormal precursors of voltage fluctuations and obtains abnormal precursor identification data; based on the abnormal precursor identification data, it quantifies the risk of voltage fluctuation evolution and obtains fluctuation evolution risk quantification data. In one possible design, the operating logic of the voltage risk quantification module is as follows: The trajectory deviation data is obtained, and the trajectory deviation data is analyzed time-by-time to obtain the trajectory deviation time series data. Based on the time series data of trajectory deviation, the abnormal fluctuation features are extracted, and the abnormal precursor signals that appear during the voltage fluctuation process are identified to obtain abnormal precursor identification data. By jointly analyzing the abnormal precursor identification data and the trajectory deviation time series data, the occurrence frequency and duration of the abnormal precursor signals are calculated to obtain the abnormal precursor quantitative data. Based on the quantitative data of anomaly precursors, we can model the evolution trend and predict the direction and amplitude of voltage fluctuations in the future short time window to obtain voltage fluctuation evolution trend data. By comparing the voltage fluctuation evolution trend data with the preset risk assessment standards, the risk level is quantified, and the fluctuation evolution risk quantification data is obtained.

[0024] The implementation process of the voltage risk quantification module in this embodiment of the invention is as follows: First, the input trajectory deviation data is divided into sliding windows with a 5-minute time granularity. A double exponential smoothing method is used to denoise the original time-series data, forming a standardized deviation curve with trend representation capabilities. Based on this curve, abnormal precursor signal detection is implemented: the absolute value of the deviation change rate between adjacent windows is calculated. When the change rate of three consecutive windows exceeds the 95% confidence interval of the same period in history and the cumulative increase reaches 200% of the baseline level, it is determined to be a valid abnormal precursor signal. For the identified precursor signal clusters, discrete signals with similar spatial distribution characteristics are aggregated into abnormal events using a density peak clustering algorithm. Simultaneously, the trapezoidal integral method is used to calculate the area under the deviation curve for each event as a persistence intensity index, and the trigger frequency within every ten minutes is calculated in conjunction with the event occurrence time interval. In the evolutionary trend modeling stage, a deep prediction model based on a temporal convolutional network is constructed. Using a high-density sampled deviation sequence from the previous two hours as input, multi-scale temporal features are extracted through multi-layer dilated convolution to predict the extreme deviation values ​​and change direction angles within a 15-minute window. The direction angles are divided into eight risk evolution quadrants at 45° intervals. Finally, the predicted extreme deviation values ​​are dynamically matched with power grid safety operation standards. When the extreme value exceeds the first-level warning threshold, a high-risk level is generated. Within 80-150% of the threshold and benchmark value, a medium-risk value is calculated based on the historical fault probability of the quadrant where the change direction angle is located. When the value is below 80% of the benchmark value, a low-risk level is output.

[0025] In one possible design, the logic for acquiring abnormal precursor identification data is as follows: By performing trend analysis on the time series data of trajectory deviation, the trend data of trajectory deviation change is obtained; Based on the trajectory deviation change trend data, the fluctuation amplitude of the rising and falling segments of the time series fluctuation is calculated to obtain the trajectory fluctuation amplitude data. Identify local extreme points based on trajectory fluctuation amplitude data, calculate the degree of offset of extreme points, and obtain extreme point offset data; By jointly comparing the extreme point offset data with the trajectory deviation change trend data, abnormal fluctuation feature information is extracted and abnormal fluctuation feature data is generated. Based on the abnormal fluctuation characteristic data, abnormal precursor signals that appear during voltage fluctuations are identified, and abnormal precursor identification data is obtained.

[0026] In the specific implementation of the abnormal precursor identification data acquisition process in this embodiment of the invention, firstly, a third-order Savitzky-Golay filter is performed on the trajectory deviation time series data to smooth noise while preserving the abrupt change characteristics of the time series curve. Based on the filtered data, a variable window length sliding analysis algorithm is used for trend analysis: during the upward trend segment, the analysis window is adaptively shortened to the 5-second level, the first derivative of each data point within the window is calculated, and the mean is taken as the upward trend strength; during the stable trend segment, the window is extended to the 60-second level, and the slope of the least squares fitted line is used to quantify the stability of the trend direction. Subsequently, the fluctuation amplitude is quantitatively evaluated. For each upward segment, the cumulative deviation growth from the starting point to the peak is calculated, and for each downward segment, the maximum retracement rate from the peak to the trough is calculated. The magnitude of fluctuation is characterized by the defined increase-retracement ratio, where the increase-retracement ratio is represented by the absolute value of the cumulative increase / decline retracement. In the local extreme point identification stage, an improved dual-threshold peak detection method is adopted: when the deviation value of three consecutive data points exceeds twice the standard deviation of the mean of the preceding window and the slope sign of adjacent points reverses, they are marked as candidate extreme points. False peak interference is further eliminated by verifying the consistency of the trend direction of the five preceding and following sampling points. For verified extreme points, the offset (Euclidean distance in three-dimensional space) between their reference position and the same phase segment in the same historical period is calculated, and the offset risk index is calculated by weighting the increase-decrease ratio of the trend segment. Based on the joint mapping analysis of the offset index and the trend stability parameter, an abnormal fluctuation feature matrix is ​​defined: when the offset risk index exceeds the threshold and the trend stability parameter of the corresponding period is lower than the lower limit of the normal range, it is determined to be an abnormal fluctuation event. Finally, through time series matching pattern mining technology, the clustering pattern of similar feature events is detected within three consecutive fluctuation cycles. When the event density reaches a preset alarm threshold, an abnormal precursor signal is triggered.

[0027] In one possible design, the logic for obtaining quantitative data on anomaly precursors is as follows: The abnormal precursor identification data is processed by time-series marking to obtain abnormal precursor time-series marked data; the abnormal precursor time-series marked data is compared with the trajectory deviation time-series data to obtain abnormal precursor joint comparison data; based on the abnormal precursor joint comparison data, the number of times the abnormal precursor signal occurs within a certain time window is counted to obtain abnormal precursor occurrence frequency data. By combining the frequency data of abnormal precursors and the time series data of trajectory deviation, the duration of the abnormal precursor signal in time is calculated to obtain the duration intensity data of the abnormal precursor; the frequency data of abnormal precursors and the duration intensity data of abnormal precursors are then processed to generate quantitative data of abnormal precursors.

[0028] In the acquisition of abnormal precursor quantitative data, this invention first performs precise timestamp marking on the abnormal precursor identification data: a GPS synchronous clock is used to synchronize the trigger time of each precursor signal to microsecond-level accuracy, and time slots are divided on the time axis at 0.1-second intervals to establish a time-series marking index table. Then, the marked data and trajectory deviation time-series data are matched and compared in three dimensions using a sliding window association algorithm: an association window is set with a forward extension of 50ms and a backward coverage of 200ms in the time dimension. When both a deviation jump and a precursor signal trigger event exist simultaneously within the window, it is recorded as a valid association event, generating multi-dimensional joint comparison data including time offset, deviation increase ratio, and signal type. Based on this data, the precursor signal occurrence rate index is calculated within a 15-minute statistical period. In specific implementation, the window is dynamically segmented: the average number of occurrences is counted in 15-second granular windows for the first 5 minutes, and the high-frequency burst events are suppressed and counted using an exponential decay weighted algorithm for the next 10 minutes. The sustained intensity quantification employs an integral method with time decay. The deviation increase corresponding to each precursor signal is used as the baseline value, multiplied by its duration (the time between the signal's initiation and the adjacent deviation inflection point), and then time decay compensation is applied using an exponential function (the weight decreases by 15% for every minute remaining from the current time). Finally, after normalizing the occurrence rate and intensity indicators, a comprehensive risk index is calculated using principal component analysis: setting the occurrence frequency weight coefficient to 0.4 and the sustained intensity to 0.6, a linear combination of the normalized indicators generates precursor quantification values ​​on a 0-100 scale.

[0029] In a possible design, the logic for obtaining voltage fluctuation evolution trend data is as follows: The quantitative data of anomaly precursors are processed into time series data to obtain anomaly precursor time series data; based on the anomaly precursor time series data, fluctuation directionality analysis is performed to identify the upward or downward trend of voltage fluctuations and obtain fluctuation direction data. By combining fluctuation direction data with anomaly precursor quantification data, the amplitude variation range within a short time window is calculated to obtain fluctuation amplitude data; trend fitting is performed based on fluctuation direction data and fluctuation amplitude data to generate evolution trend fitting data; The evolution trend fitting data is corrected to remove abnormal interference points, and the voltage fluctuation evolution trend data is obtained.

[0030] In the implementation of this invention, during the acquisition of voltage fluctuation evolution trend data, the time axis of the anomaly precursor quantization data is first reshaped and its integrity is restored: the discrete quantization values ​​are resampled into equally spaced sequences with a granularity of 100 milliseconds, and cubic spline interpolation is used to compensate for data gaps caused by communication delays, generating continuous and complete anomaly precursor time series data. Based on this time series data, fluctuation direction analysis is performed. Within each 30-second analysis window, a trend baseline is calculated using a 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 three consecutive sampling points, a direction determination is triggered. At the same time, a direction persistence verification mechanism is introduced—if the duration of the upward trend is less than 40% of the window width, it is downgraded to noise disturbance. For the trend segments that pass the verification, the amplitude range is further calculated using a variable-scale sliding extreme value detection method: the detection window is narrowed to 5 seconds to extract the peak value in the upward trend segment, and the downward trend segment is expanded to 15 seconds to capture the trough. The transient amplitude is defined by the absolute difference between adjacent extreme points. Based on the collaborative analysis of direction and amplitude data, a two-dimensional trend feature matrix is ​​constructed: row vectors mark the positive and negative attributes of the trend direction, and column vectors record the gradient distribution of amplitude changes. A long short-term memory neural network is used for deep modeling of this matrix. Inputting the feature sequence from the previous hour, the system predicts the fluctuation trajectory for the next 5 minutes. A bidirectional temporal attention mechanism is incorporated during network training to enhance the ability to capture key features. The prediction results are corrected by dynamic threshold filtering: when the residual between the predicted value and the actual monitored value at a certain moment exceeds 150% of the maximum residual for the same period in history, a Kalman filter is triggered for real-time correction. Simultaneously, a smoothed trend curve is generated by combining the weighted average of the previous three prediction points. Finally, isolated outliers are removed through spatial density clustering, retaining data segments with three or more consecutive sampling points that conform to the trend direction, forming a voltage fluctuation evolution trend dataset that eliminates spike interference.

[0031] Power grid safety early warning module: Based on the fluctuation evolution risk quantification data, it compares the data with preset multi-level risk thresholds in real time, and dynamically generates risk warning signals of different levels according to the comparison results; the risk warning signals are directly pushed to the early warning terminal of the power grid dispatch center, driving it to perform visual alarm and execute the corresponding risk handling plan.

[0032] In one possible design, the operating logic of the power grid safety early warning module is as follows: 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; 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. 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.

[0033] 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. Simultaneously, a timeline 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 standard deviations), an early warning threshold, and an emergency threshold (90% of real-time load capacity). A three-level nested verification mechanism is used for risk level matching: a blue early warning is triggered when the risk index remains within the range of the basic threshold and the early warning threshold for three consecutive minutes; a second-level orange alert is triggered when the early warning threshold is exceeded; and a red highest-level alert is triggered if the emergency threshold is reached twice within five minutes.

[0034] 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.

[0035] 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.

[0036] 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.

[0037] 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 dynamic early warning system for digital power grid security risks based on deep learning algorithms, characterized in that, The system includes the following modules: Voltage spectrum acquisition module: continuously acquires voltage signals from grid nodes using voltage sensors to obtain raw voltage signal data; models voltage micro-perturbation trajectory data from the raw voltage signal data to obtain micro-perturbation trajectory data; Voltage deviation calculation module: Extracts normal trajectory feature space from micro-perturbation trajectory map data to obtain normal trajectory feature space data; calculates real-time trajectory deviation based on normal trajectory feature space data to generate trajectory deviation. Voltage risk quantification module: Based on trajectory deviation, it identifies abnormal voltage fluctuation precursors and obtains abnormal precursor identification data; Voltage fluctuation evolution risk is quantified based on abnormal precursor identification data to obtain fluctuation evolution risk quantification data; Power grid safety early warning module: Based on the fluctuation evolution risk quantification data, it compares the data with preset multi-level risk thresholds in real time, and dynamically generates risk warning signals of different levels according to the comparison results; the risk warning signals are directly pushed to the early warning terminal of the power grid dispatch center, driving it to perform visual alarm and execute the corresponding risk handling plan.

2. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 1, characterized in that, The operation logic of the voltage spectrum acquisition module is as follows: The three-phase voltage of the power grid node is synchronously sampled by a voltage sensor to obtain the original voltage signal time series data at a preset sampling frequency; the original voltage signal time series data is subjected to sliding window differential noise reduction processing to obtain the noise-reduced differential voltage fluctuation data; the differential voltage fluctuation data is decomposed into multi-scale time-frequency features to generate a joint feature dataset containing time-domain waveform features and frequency-domain energy distribution features. Based on the joint feature dataset, a voltage fluctuation baseline parameter matrix is ​​constructed using historical data under normal operating conditions. The baseline parameter matrix includes an amplitude change threshold and a frequency coupling coefficient. The real-time acquired joint feature dataset is compared with the baseline parameter matrix on a time-by-time basis to identify voltage micro-perturbation segments that exceed the amplitude change threshold. The voltage micro-perturbation segment data is subjected to time-frequency characteristic trajectory parameterization processing to extract three core characteristic parameters for each perturbation segment: the initial phase, the duration of the perturbation, and the frequency band energy transition rate. The core characteristic parameters of each disturbance segment are dynamically modeled according to the time series to generate micro-disturbance trajectory map data.

3. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 2, characterized in that, The real-time acquired joint feature dataset is compared with the baseline parameter matrix time-by-time to identify voltage micro-perturbation segments that exceed the amplitude change threshold; the specific operation steps are as follows: Obtain the real-time collected joint feature dataset and organize it into segments according to time order to obtain segmented joint feature data for each time period; The joint feature segmentation data for each time period is matched with the baseline parameter matrix to generate baseline comparison data; The amplitude change was calculated for each time period based on baseline comparison data, and the amplitude change data was obtained. The amplitude change data is compared with the preset amplitude change threshold in time intervals to obtain the comparison result data; Based on the comparison results, the portion of amplitude change exceeding the threshold was selected and identified as voltage micro-perturbation segment data.

4. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 2, characterized in that, Three core feature parameters were extracted for each perturbation segment: the initial phase, the duration of the perturbation, and the frequency band energy transition rate. The specific operating steps are as follows: Time-frequency decomposition processing is performed on the voltage micro-perturbation segment data to obtain the time-frequency distribution data of the perturbation segment; The starting point of the disturbance signal is identified based on the time-frequency distribution data, and the starting phase of the disturbance segment is calculated and extracted to obtain the starting phase data. By combining the initial phase data with the time-frequency distribution data, the duration of the disturbance signal on the time axis is analyzed, the duration period of the disturbance segment is extracted, and the disturbance duration period data is obtained. Based on the disturbance duration period data, the energy evolution of the time-frequency distribution data is evaluated, and the energy transfer ratio of the disturbance signal between different frequency bands is calculated to obtain the frequency band energy transition rate data. The initial phase data, disturbance duration period data, and frequency band energy transition rate data are output in a unified manner as the core characteristic parameters of the voltage micro-disturbance segment.

5. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 1, characterized in that, The operating logic of the voltage deviation calculation module is as follows: Acquire micro-perturbation trajectory map data, wherein the micro-perturbation trajectory map data includes three core feature parameters for each perturbation segment: the initial phase, the duration of the perturbation, and the frequency band energy transition rate. Perform format uniformity and integrity checks on the data to obtain feature parameter preprocessing data. Aggregate analysis of the initial phase distribution of different disturbance segments is performed based on the preprocessed data of characteristic parameters to obtain the initial phase distribution data of disturbance; By jointly modeling the initial phase distribution data of the disturbance and the duration period data of the disturbance, the temporal characteristics of the disturbance segment under normal operating conditions are extracted, and normal trajectory temporal characteristic data are generated. Normal trajectory time feature data and frequency band energy transition rate data are fused to construct a normal trajectory feature space, thus obtaining 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 and obtain the trajectory deviation degree.

6. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 5, characterized in that, The logic for obtaining normal trajectory time feature data is as follows: The initial phase distribution data of the disturbance is sorted in time order to obtain the initial phase sequence data; the initial phase sequence data is matched with the disturbance duration period data to obtain the phase-period correspondence data. The time coverage range of the disturbance segment is calculated based on the phase-period correspondence data to obtain the disturbance time coverage data; Trend fitting is performed on the disturbance time coverage data to extract typical time feature patterns of the disturbance segment under normal operating conditions, and initial time feature data is generated. The initial time feature data and the phase-period corresponding data are jointly corrected to obtain the corrected time feature data; the corrected time feature data is output as the normal trajectory time feature data.

7. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 1, characterized in that, The operating logic of the voltage risk quantification module is as follows: The trajectory deviation data is obtained, and the trajectory deviation data is analyzed time-by-time to obtain the trajectory deviation time series data. Based on the time series data of trajectory deviation, the abnormal fluctuation features are extracted, and the abnormal precursor signals that appear during the voltage fluctuation process are identified to obtain abnormal precursor identification data. By jointly analyzing the abnormal precursor identification data and the trajectory deviation time series data, the occurrence frequency and duration of the abnormal precursor signals are calculated to obtain the abnormal precursor quantitative data. Based on the quantitative data of anomaly precursors, we can model the evolution trend and predict the direction and amplitude of voltage fluctuations in the future short time window to obtain voltage fluctuation evolution trend data. By comparing the voltage fluctuation evolution trend data with the preset risk assessment standards, the risk level is quantified, and the fluctuation evolution risk quantification data is obtained.

8. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 7, characterized in that, The logic for obtaining abnormal precursor identification data is as follows: By performing trend analysis on the time series data of trajectory deviation, the trend data of trajectory deviation change is obtained; Based on the trajectory deviation change trend data, the fluctuation amplitude of the rising and falling segments of the time series fluctuation is calculated to obtain the trajectory fluctuation amplitude data. Identify local extreme points based on trajectory fluctuation amplitude data, calculate the degree of offset of extreme points, and obtain extreme point offset data; By jointly comparing the extreme point offset data with the trajectory deviation change trend data, abnormal fluctuation feature information is extracted and abnormal fluctuation feature data is generated. Based on the abnormal fluctuation characteristic data, abnormal precursor signals that appear during voltage fluctuations are identified, and abnormal precursor identification data is obtained.

9. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 7, characterized in that, The logic for obtaining quantitative data on anomaly precursors is as follows: The abnormal precursor identification data is processed by time-series marking to obtain abnormal precursor time-series marked data; the abnormal precursor time-series marked data is compared with the trajectory deviation time-series data to obtain abnormal precursor joint comparison data; based on the abnormal precursor joint comparison data, the number of times the abnormal precursor signal occurs within a certain time window is counted to obtain abnormal precursor occurrence frequency data. By combining the frequency data of abnormal precursors and the time series data of trajectory deviation, the duration of the abnormal precursor signal in time is calculated to obtain the duration intensity data of the abnormal precursor; the frequency data of abnormal precursors and the duration intensity data of abnormal precursors are then processed to generate quantitative data of abnormal precursors.

10. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 7, characterized in that, The logic for obtaining voltage fluctuation evolution trend data is as follows: The quantitative data of anomaly precursors are processed into time series data to obtain anomaly precursor time series data; Fluctuation directionality analysis is performed based on anomaly precursor time series data to identify the upward or downward trend of voltage fluctuations and obtain fluctuation direction data. By combining the fluctuation direction data with the quantitative data of anomaly precursors, the amplitude change range within a short time window is calculated to obtain the fluctuation amplitude data. Trend fitting is performed based on fluctuation direction data and fluctuation amplitude data to generate evolution trend fitting data; The evolution trend fitting data is corrected to remove abnormal interference points, and the voltage fluctuation evolution trend data is obtained.

11. The dynamic early warning system for digital power grid security risks based on deep learning algorithms according to claim 1, characterized in that, The operating logic of the power grid safety early warning module is as follows: 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; 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. Risk warning signal data is directly pushed to the warning terminal of the power grid dispatch center to drive it to perform visual alarm processing and obtain visual alarm data; 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.

Images