Power distribution network disturbance identification method and system based on multiple time scales

Abstract

The invention provides a power distribution network disturbance identification method and system based on multiple time scales, and the method comprises the following steps: S1, achieving the standardized processing and noise characteristic modeling of measurement data through building a mathematical relation among node voltage, a phase angle and a measurement quantity, and reflecting a system topological structure and parameter sensitivity through a Jacobian matrix; s2, firstly, based on the step S1, capturing dynamic feature changes in different time ranges; extracting features through feature engineering, optimizing a feature set by adopting a recursive feature elimination method, and improving the accuracy and stability of the model; s3, after the classification framework preliminarily confirms that a traditional event belongs to a large class, training and verifying feature data extracted in the previous step by adopting an XGBoost algorithm, realizing high-precision identification of multiple classes of disturbance events through hyper-parameter optimization and cross validation, and outputting an event class and a confidence result; according to the technical scheme, intelligent analysis of the operation state of the power distribution network and accurate identification of abnormal events can be realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and status analysis technology for power systems, and in particular to a method and system for identifying disturbances in distribution networks based on multiple time scales. Background Technology

[0002] In modern power systems, with the large-scale integration of distributed energy resources and power electronic devices, the system structure and operating characteristics are undergoing profound changes. Distribution networks are gradually evolving from traditional passive power supply networks into active networks characterized by high dynamism, high uncertainty, and multi-energy coupling. During system operation, various disturbances and abnormal events (including voltage sags, current surges, frequency fluctuations, islanding effects, communication delays, and data synchronization failures) exhibit more complex spatiotemporal characteristics and nonlinear propagation properties. This complexity makes the detection and identification of abnormal events in power systems one of the key issues in the safe operation of the power grid. Failure to detect and identify these abnormal events in a timely and accurate manner may lead to malfunctions of protection devices, delays in dispatching decisions, and consequently trigger chain reactions, threatening the safety and stability of the entire power system.

[0003] In recent years, with the development and deployment of synchronous phasor measurement unit (PMU) technology, the wide-area measurement capabilities of power systems have been significantly improved. PMUs can synchronously acquire key operating parameters such as voltage phasors, current phasors, frequency, and phase angle with millisecond-level resolution, providing high-precision data support for dynamic monitoring, disturbance analysis, and state estimation of power systems. Wide-area measurement systems (WAMS) based on PMUs have become an important foundation for power systems to achieve state awareness, anomaly detection, and online diagnosis. However, in actual distribution network environments, PMU data is often affected by communication links, equipment noise, synchronization deviations, and hardware faults, leading to quality problems such as noise pollution, sample frame loss, abnormal drift, or timing mismatch. This poor data can be misjudged as abnormal events during disturbance detection or can mask the true disturbance characteristics, thereby reducing the accuracy and robustness of detection algorithms.

[0004] Traditional methods for detecting abnormal events in power systems mainly fall into three categories: threshold analysis, signal processing, and machine learning. Threshold-based detection methods set thresholds for variables such as voltage, current, and frequency; when a measured value exceeds the threshold range, it is considered an abnormal event. These methods are simple to implement and computationally inexpensive, but the thresholds are usually set manually based on experience, making them difficult to adapt to complex operating conditions and prone to false positives due to noise. Signal processing-based detection methods can uncover the time-frequency characteristics of disturbance signals, but their performance remains limited for events with diverse disturbance types, short durations, or weak amplitudes. In recent years, machine learning and deep learning technologies have been introduced into the field of power system anomaly detection, enabling disturbance event identification through training classification models. While these methods offer high detection accuracy, they rely on a large number of labeled samples, facing challenges in practical engineering applications such as scarce labels, variable data distribution, and insufficient generalization ability.

[0005] In distribution network scenarios, the diversity of disturbance signals and the lack of data labels make it difficult to generalize supervised learning methods. Meanwhile, the complex coupling characteristics of power grid operation states mean that the response patterns of different disturbance events may overlap in the time and frequency domains, making it difficult for traditional classification algorithms to effectively distinguish them. Furthermore, the limited deployment density of PMUs and complex communication conditions in distribution networks exacerbate data quality issues. Traditional detection methods based on fixed models or single features experience a significant drop in accuracy and stability when faced with real-world environmental noise and uncertain disturbances.

[0006] To address the aforementioned issues, unsupervised learning methods are introduced into the field of power system disturbance detection. By using a data-driven approach to mine the latent structural features of disturbance samples, event identification can be achieved under label-less conditions. Through feature self-learning and cluster analysis, unsupervised methods can not only identify known types of disturbances but also automatically discover and classify unknown or mixed disturbances, thus overcoming the shortcomings of traditional methods in terms of sample dependence and threshold setting. Combining high spatiotemporal resolution measurement data provided by the Power Management Unit (PMU), a strategy integrating feature extraction, dimensionality reduction analysis, and clustering algorithms is expected to achieve efficient detection and accurate identification of power system anomalies without prior knowledge.

[0007] However, unsupervised learning-based power system anomaly detection is still in the research stage, and suffers from problems such as insufficient model interpretability, reliance on human experience for feature selection, and room for improvement in real-time performance. How to achieve stable, accurate, and real-time anomaly detection and identification while considering PMU data quality fluctuations and the complex dynamic environment of the distribution network has become an important research direction and engineering requirement in the field of intelligent power system monitoring. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a method and system for identifying power distribution network disturbances based on multiple time scales, in order to solve the problems of existing technologies such as reliance on a large number of manually labeled samples, insufficient ability to identify unknown disturbances, and decreased detection accuracy when data quality issues exist. This invention innovatively achieves pre-identification and reclassification of abnormal events in power systems by integrating signal processing and unsupervised learning techniques. It can accurately identify multiple disturbance types under label-less conditions and maintain high sensitivity and robustness to weak disturbance signals in complex environments.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a method for identifying distribution network disturbances based on multiple time scales, comprising the following steps:

[0010] Step S1: By establishing the mathematical relationship between node voltage, phase angle and measured quantity, the measurement data is standardized and noise characteristics are modeled. The Jacobian matrix is ​​used to reflect the system topology and parameter sensitivity. On this basis, the weighted least squares method is used to estimate the state of the noisy measurement data collected by SCADA, and redundant and error information is removed to obtain the estimated value that best reflects the true operating state of the system.

[0011] Step S2: First, based on step S1, capture the dynamic feature changes within different time ranges; then, extract time-domain and frequency-domain features through feature engineering, and optimize the feature set using a recursive feature elimination method to improve the accuracy and stability of the model; on this basis, classify and identify typical disturbance events such as tripping, load shedding, short circuits, and circuit breakers by utilizing the frequency and phase angle variation patterns under multiple time scales. At the same time, frequency or voltage anomalies that do not conform to the traditional pattern are marked as non-traditional events, providing a foundation for subsequent refined anomaly identification.

[0012] Step S3: After initially identifying the major categories of traditional events within the classification framework, the XGBoost algorithm is used to train and validate the feature data extracted in the previous step. High-precision identification of multiple types of disturbance events is achieved through hyperparameter optimization and cross-validation, and the event category and confidence results are output. For unknown events output by the model, hierarchical clustering is used to achieve automatic grouping and refined classification, thereby identifying new or undefined events such as frequency oscillations, islanding, and power quality disturbances.

[0013] In a preferred embodiment, step S1 includes the following steps:

[0014] Step S11, State estimation: The real-time state of the power network is characterized by integrating multi-source inputs; for network segments that need to meet physical constraints but cannot be directly measured, pseudo-measured values ​​based on load forecasting and generation plans are used for estimation; the number and distribution of measurement points in the network directly determine the observability of the system; when the redundancy is sufficient, the state estimation technique allows temporary deletion of observation data without affecting the expected value, thereby effectively reducing the impact of data distortion.

[0015] The state estimator generates a reliable system state estimate by eliminating redundant information and measurement errors. When the monitoring range of the power system is limited, a suitable measurement matrix is ​​constructed by selecting several rows from the complete observation matrix. It is assumed that the dimension of all measurement matrices satisfies m>n. As a core tool for efficient and rapid power flow distribution, the state predictor needs to deduce state variables from the inaccurate measurement data obtained from the SCADA system.

[0016] (1)

[0017] in, This represents the weighted least squares estimate of the state variables; It is the weight matrix, which is a diagonal matrix. The reciprocal of this matrix represents the noise variance matrix. It is a Jacobian matrix, and T denotes the transpose.

[0018] In a preferred embodiment, step S1 further includes the following steps:

[0019] Step S12, perform state estimation using weighted least squares:

[0020] By calculating the variance of the measurement error, the weighting factor is obtained. ;

[0021] The purpose of WLS is as follows:

[0022] (2)

[0023] Where N represents the total number of measurements involved in the estimation, and j represents the measurement number; The variance represents the measurement error. Represents an m-dimensional measurement vector. This represents an m-dimensional estimation vector.

[0024] In a preferred embodiment, step S1 further includes the following steps:

[0025] Step S13, Bad Data Detection:

[0026] A bad data detection analyzer is embedded in the state estimation to identify and remove outlier or erroneous measurements; definition: (3)

[0027] in, This represents the difference between the measured data and the estimated data. This represents the weighted least squares estimate of the state variable vector x;

[0028] Estimated data

[0029] (4)

[0030] Residues written as

[0031] (5)

[0032] in, It is the identity matrix. Here is the gain matrix. This is the weight matrix. It is a Jacobian matrix;

[0033] As evaluation indicators and chi-square tests are used as detection tools, their objective function is defined as follows:

[0034] (6)

[0035] in, Indicates the number of measurements, with the weighting factor set to [value]. ;

[0036] if , Let be the objective function. Let α be the critical value of the chi-square distribution with k degrees of freedom and a significance level of α. When the log-likelihood value of k degrees of freedom is satisfied, the state estimate is considered accurate and reliable; otherwise, it indicates the presence of bad data in the system.

[0037] In a preferred embodiment, step S2 includes the following steps:

[0038] Step S21, Data Preparation:

[0039] The data is divided into time windows based on interference events in historical fault logs; normal classes include N50, N40, N30, N20 and N10, Pre, Post; segmented analysis allows detection of warning signals that occur at different times; multi-scale analysis is performed using 30-second, 60-second and 180-second window sizes for each segment to capture short-term and long-term dynamics;

[0040] Step S22, Feature Engineering:

[0041] A comprehensive set of features is extracted to capture the complex dynamics of the power system; these features include time-domain, frequency-domain, and system dynamics-specific metrics; in the testing methodology, recursive feature elimination is used as the feature selection method; and the model is simplified by reducing irrelevant features.

[0042] In a preferred embodiment, step S2 further includes the following steps:

[0043] Step S23, Pre-classification:

[0044] The dataset is divided into training and testing sets in a 90:10 ratio;

[0045] Anomalies lasting 10 seconds to 1 minute and those lasting less than 1 second are distinguished by their frequency behavior at different time scales. Then, based on the characteristics of a frequency change slope greater than 0 for tripping events and a frequency change slope less than 0 for load shedding events, tripping and load shedding are differentiated. The distinction between short-circuit and tripping events depends on the duration of their voltage phase angle anomalies. Specifically, tripping events are caused by immediate line disconnection after the event, allowing for faster fault resolution and restoration of operation, with an impact on the voltage phase angle of less than 20 seconds, exhibiting anomalies in short-time characteristics. Short-circuit events have a longer impact on the voltage phase angle, showing anomalies not only in ultra-short-time characteristics but also in short-time characteristics. Frequency and phase angle anomalies that do not conform to traditional event patterns are recorded as non-traditional or unknown events and will be re-identified in subsequent steps. Furthermore, if anomalies in voltage amplitude occur even when no anomalies are observed in the transmission network, it is determined that the disturbance originates from the local distribution network.

[0046] In a preferred embodiment, step S3 includes the following steps:

[0047] Step S31, XGBoost construction:

[0048] The feature data extracted from S2 is used as the feature input of the samples, and training and validation sets are constructed in combination with existing labels. During data preparation, normalization and denoising are performed, and class imbalance is addressed. XGBoost is used during the training phase, and the key parameters for parameter tuning include tree depth, number of leaves, learning rate, feature subsampling ratio, and regularization term. Cross-validation is used to select the optimal hyperparameters, and multi-class classification performance is evaluated using recall and confusion matrix metrics. After training, the predicted probabilities are calibrated for subsequent threshold selection and uncertainty estimation. The output results are the major class label and corresponding probability and confidence score for each event, which are used to guide the next step of subdivision and manual review.

[0049] In a preferred embodiment, step S3 further includes the following steps:

[0050] Step S32, analyze intra-class differences:

[0051] After initially classifying the major categories, in-depth feature analysis is performed on the samples within each major category to discover intra-category differences: including statistical analysis, time series morphological analysis, and time-varying curve characteristics of discriminator scores in each time window;

[0052] In a preferred embodiment, step S3 further includes the following steps:

[0053] Step S33, Identification of Unknown Events:

[0054] In XGBoost, each iteration finds the leaf with the largest split gain from all current leaves and splits it, enabling the model to complete class identification in a very short time, thus outperforming other decision tree methods. Hierarchical clustering is flexible, automatically adding groups through splitting, and clustering events with similar characteristics into the same subclass, which is convenient for subsequent research, analysis and labeling.

[0055] The present invention also provides a distribution network disturbance identification system based on multiple time scales, including a processor, a memory and a bus, wherein the memory stores machine-readable instructions executed by the processor;

[0056] When the system is running, the processor and the memory communicate via a bus, and the machine-readable instructions are executed by the processor as described in the multi-time-scale power distribution network disturbance identification method.

[0057] Compared with existing technologies, this invention has the following advantages: Based on a multi-timescale distribution network disturbance identification method and system, this invention first establishes a mathematical model of the power system, uses weighted least squares for state estimation, and utilizes the Jacobian matrix to reflect the system's topological characteristics and parameter sensitivity, thereby obtaining a realistic and reliable system state. Secondly, the data is divided into different time windows according to disturbance events, extracting time-domain, frequency-domain, and system dynamic features, and employing a recursive feature elimination algorithm to optimize the feature set. By analyzing the differences in frequency and phase angle variations at different time scales, preliminary classification of typical disturbance events such as generator tripping, load shedding, short circuits, and outages is achieved, while also identifying abnormal voltage disturbances and non-traditional events. Finally, the XGBoost algorithm, combined with cross-validation and parameter optimization, is used to improve the model's generalization ability, outputting event categories and confidence scores. Attached Figure Description

[0058] Figure 1 This is a flowchart of the residual measurement technology described in step S1 of the present invention;

[0059] Figure 2 This is a schematic diagram of the fault detection and residual analysis process described in step S1 of the present invention;

[0060] Figure 3 This is a schematic diagram of the disturbance identification process described in step S2 of the present invention;

[0061] Figure 4 This is a diagram of the XGBoost algorithm architecture described in step S3 of the present invention. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0063] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0064] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0065] A method for identifying disturbances in distribution networks based on multiple time scales, referencing Figure 1-4 This includes the following steps:

[0066] Step S1: By establishing the mathematical relationship between node voltage, phase angle, and measured quantities, the measurement data is standardized and noise characteristics are modeled. The Jacobian matrix is ​​used to reflect the system topology and parameter sensitivity. Based on this, the weighted least squares method is used to perform state estimation on the noisy measurement data acquired by SCADA, eliminating redundancy and error information to obtain the estimate that best reflects the true operating state of the system.

[0067] Step S2: First, based on Step S1, capture the dynamic feature changes within different time ranges; then, extract time-domain and frequency-domain features through feature engineering, and optimize the feature set using a recursive feature elimination method to improve the model's accuracy and stability. On this basis, classify and identify typical disturbance events such as generator tripping, load shedding, short circuits, and circuit breaker tripping using the frequency and phase angle variation patterns across multiple time scales. Simultaneously, frequency or voltage anomalies that do not conform to traditional patterns are marked as non-traditional events, providing a foundation for subsequent refined anomaly identification.

[0068] Step S3: After initially identifying the major categories of traditional events within the classification framework, the XGBoost algorithm is used to train and validate the feature data extracted in the previous step. Through hyperparameter optimization and cross-validation, high-precision identification of multiple types of disturbance events is achieved, and the event category and confidence score results are output. For unknown events output by the model, hierarchical clustering is used to achieve automatic grouping and refined classification, thereby identifying novel or undefined events such as frequency oscillations, islanding events, and power quality disturbances, effectively improving the comprehensiveness and intelligence of abnormal event identification.

[0069] Furthermore, step S1 specifically includes the following steps:

[0070] In power system anomaly identification research, system modeling plays a core role in the initial data processing. By establishing the mathematical relationship between node voltage, phase angle, and measured quantities, the standardization of measured data and noise characteristic modeling are achieved, providing a calculable residual benchmark for subsequent anomaly detection. This stage not only defines the mapping relationship between state variables and measured variables but also reflects the system topology and parameter sensitivity through the Jacobian matrix, thus laying the mathematical foundation for the location and identification of anomalies.

[0071] Step S11: State estimation:

[0072] State predictors are information processing methods that characterize the real-time state of a power network by integrating multiple input sources. System parameter estimation, mathematical models, test data, and prior knowledge from different inputs and outputs are typical examples of such data sources. For network segments that need to meet physical constraints but cannot be directly measured, estimations are typically made using pseudo-measurements based on load forecasts and generation plans. The number and distribution of measurement points in the network directly determine the observability of the system. Observable islands refer to interconnected areas in the network where the flow of all branches is known. When redundancy is sufficient, state estimation techniques can allow temporary deletion of observation data without affecting expected values, thereby effectively reducing the impact of data distortion. This state estimation method can efficiently address errors, noise, and data inconsistencies caused by equipment failures or network connectivity problems.

[0073] The state estimator generates a reliable system state estimate by eliminating redundant information and measurement errors. Therefore, its output accurately reflects the actual state of the system. When the monitoring range of the power system is limited (some buses and lines are not covered by meters), a suitable measurement matrix can be constructed by selecting several rows from the complete observation matrix. This study assumes that the dimension of all measurement matrices satisfies m>n. As a core tool for efficient and rapid power flow distribution, the state predictor needs to deduce state variables as accurately as possible from the inaccurate measurement data obtained from the SCADA system.

[0074] (1)

[0075] in, This represents the weighted least squares estimate of the state variables. It is the weight matrix, which is a diagonal matrix. The reciprocal of this matrix represents the noise variance matrix. It is a Jacobian matrix.

[0076] Step S12: State estimation using weighted least squares:

[0077] The core objective of least squares theory is to solve data fitting problems. Its variant, the weighted least squares rule, is based on the following assumptions: the off-diagonal elements of the residual correlation matrix are all zero, and the error terms are independent. This method also requires that the error terms follow a normal distribution, with a covariance matrix of... The mean vector is zero.

[0078] The weighting factor is a parameter added to optimize the fitting process. Because the accuracy of measuring instruments varies, their error values ​​also differ. Instruments with smaller error values ​​have larger weighting factors, and vice versa. The weighting factor can be obtained by calculating the variance of the measurement error. .

[0079] The purpose of WLS is as follows:

[0080] (2)

[0081] Where N represents the total number of measurements involved in the estimation, and j represents the measurement number; The variance represents the measurement error. Represents an m-dimensional measurement vector. This represents an m-dimensional estimation vector.

[0082] Step S13: Bad Data Detection

[0083] The best estimators not only provide state estimates but also corresponding covariance or error matrices. The covariance matrix simultaneously provides measurement reliability and error prediction. Ideally, the difference between the estimated error and the actual error should remain constant. If this difference increases significantly, it indicates that the system may be encountering complex or abnormal operating conditions. Analyzing these deviations is the core objective of differential data detection, aiming to ensure that the state estimate perfectly matches the system's actual operating behavior and physical structure. By detecting erroneous data, the system can identify, understand, and reduce measurement errors.

[0084] Bad data detection analyzers are typically embedded in state estimation to identify and remove outlier or erroneous measurements. Definition: (3)

[0085] in, This represents the difference between the measured data and the estimated data. This represents the weighted least squares estimate of the state variable vector x.

[0086] Estimated data

[0087] (4)

[0088] Residues can also be written as

[0089] (5)

[0090] in, It is the identity matrix. Here is the gain matrix. This is the weight matrix. It is a Jacobian matrix.

[0091] The core objective of defective data detection is to identify, locate, and eliminate measurement errors in the system. In power systems, two detection methods can be used to identify abnormal data: evaluation metrics and maximum normalized residuals (GNR). Evaluation metrics and the chi-square test can be used as detection tools, and their objective function is defined as follows:

[0092] (6)

[0093] in, Indicates the number of measurements, with the weighting factor set to [value]. .

[0094] if , Let be the objective function. Let α be the critical value of the chi-square distribution with k degrees of freedom and a significance level of α. When the log-likelihood value of k degrees of freedom is satisfied, the state estimate can be considered accurate and reliable; otherwise, it indicates that there is bad data in the system.

[0095] Furthermore, step S2 specifically includes the following steps:

[0096] Step S21: Data Preparation

[0097] Data is divided into time windows based on interference events in historical fault logs. Normal categories include N50, N40, N30, N20, and N10 (representing 50, 40, 30, 20, and 10 minutes before, respectively, before, after, and after the interference), Pre (immediately before the interference), and Post (immediately after the interference). Segmented analysis allows for the detection of early warning signals that may appear at different times. Multi-scale analysis uses 30-second, 60-second, and 180-second window sizes for each segment to capture short-term and long-term dynamics.

[0098] Step S22: Feature Engineering:

[0099] A comprehensive set of features is extracted to capture the complex dynamics of power systems. These features include time-domain, frequency-domain, and system dynamics-specific metrics such as covariance and detrended fluctuation analysis. They are used to capture time-based relationships and long-term patterns. In the tested approach, recursive feature elimination is used as the feature selection method. Reducing irrelevant features simplifies the model, making it more efficient and reliable. Furthermore, the visualization of the top 20 features selected by recursive feature elimination is significantly clearer.

[0100] Step S23: Pre-classification:

[0101] The dataset is divided into training and test sets in a 90:10 ratio, which ensures robust training on 90% of the data and validation on the remaining 10%.

[0102] The basic principle of classification is to match the anomalous characteristics of disturbances with the typical features of each major category of disturbances at different time scales. Specifically, by observing the frequency behavior at different time scales (short-term), anomalies lasting 10 seconds to 1 minute (corresponding to generator tripping and load shedding events) and anomalies lasting less than 1 second (including short-circuit and tripping events) can be distinguished. Then, by observing that the frequency change slope of generator tripping events is greater than 0, while the frequency change slope of load shedding events is less than 0, generator tripping and load shedding can be identified. The distinction between short-circuit and tripping events depends on the different durations of their voltage phase angle anomalies. Specifically, tripping events are usually caused by the immediate disconnection of the line after the event occurs, allowing for rapid fault resolution and restoration of operation. The impact on the voltage phase angle is less than 20 seconds, exhibiting anomalies in the short-time characteristics. Short-circuit events have a longer impact on the voltage phase angle, showing anomalies not only in the ultra-short-time characteristics but also in the short-time characteristics. This allows for the classification of four traditional disturbance events that affect the transmission network. Frequency and phase angle anomalies that do not conform to traditional event patterns are recorded as non-traditional or unknown events and will be re-identified in subsequent steps. Additionally, if voltage amplitude anomalies occur even when no abnormalities are observed in the transmission network, it is determined that the disturbance originates from the local distribution network.

[0103] Furthermore, step S3 specifically includes the following steps:

[0104] Step S31: XGBoost Construction:

[0105] The feature data extracted from S2 is used as the feature input for the samples, and training and validation sets are constructed by combining them with existing labels. Data preparation includes normalization and denoising, as well as handling class imbalance. XGBoost is used during the training phase, with key parameter tuning including tree depth / number of leaves, learning rate, feature subsampling ratio, and regularization term. Cross-validation is used to select the optimal hyperparameters, and multi-class classification performance is evaluated using metrics such as recall and confusion matrix. After training, the predicted probabilities are calibrated for subsequent threshold selection and uncertainty estimation. The output results are the major class label for each event (e.g., short circuit, trip, generator disconnection, load shedding, power quality in the distribution network, unknown) and the corresponding probability / confidence, used to guide further subdivision and manual verification.

[0106] Step S32: Analyze within-class differences:

[0107] After initially classifying the samples into broad categories, in-depth feature analysis is performed on the samples within each category to identify intra-category differences. This includes statistical analysis (mean, variance, skewness, kurtosis), temporal morphology analysis (amplitude of abrupt changes, duration, oscillation frequency), and time-varying curve characteristics of discriminator scores across different time windows (e.g., peak value, peak duration, number of persistent anomalous windows). Based on these features, a similarity metric is designed / selected: if the discriminator score vector is the primary metric, Euclidean distance or cosine similarity can be used.

[0108] Step S33: Identification of Unknown Events

[0109] XGBoost, in each iteration, selects the leaf with the highest splitting gain from all current leaves for splitting, enabling the model to complete large-class identification in a very short time, thus outperforming other decision tree methods. Hierarchical clustering, on the other hand, offers flexibility, automatically adding groups through splitting, clustering events with similar characteristics into the same subclass for easier subsequent research, analysis, and labeling. For example, a large class of unknown events might contain subclasses such as frequency oscillations and islands, and the oscillation event subclass might contain oscillation events of varying durations and severity. Similarly, power quality disturbances might include specific categories such as voltage surges, voltage drops, voltage swells, voltage dips, and voltage interruptions. Another advantage of this classification method is that if false detections occur, similar false detection events will be clustered into the same subclass, resulting in higher efficiency for whole-class processing.

Claims

1. A multi-time scale based power distribution network disturbance identification method, characterized in that, Comprising the following steps: Step S1: Through the establishment of mathematical relationship between node voltage, phase angle and measurement, the standardization of measurement data and the modeling of noise characteristics are realized, the mapping relationship between state variables and measurement variables is defined, and the system topology and parameter sensitivity are reflected through Jacobian matrix; Step S2: Through feature engineering, time domain and frequency domain features are extracted, and the feature set is optimized; On this basis, the frequency and phase angle change law under multi-time scale is used to classify and identify typical disturbance events such as generator tripping, load shedding, short circuit and tripping, and the frequency or voltage anomalies that do not conform to the traditional mode are marked as non-traditional events; Step S3: After the classification framework preliminarily confirms the category of traditional events, XGBoost algorithm is used to train and verify the feature data, and the event category and confidence result are output; For unknown events output by the model, hierarchical clustering method is used to realize automatic grouping and detailed classification.

2. The multi-time scale based power distribution network disturbance identification method of claim 1, wherein, Step S1 includes the following steps: Step S11, state estimation: the real-time state of the power network is represented by integrating multiple source inputs; For network segments that need to meet physical constraints but cannot be directly measured, pseudo-measurement values based on load prediction and power generation plan are used for estimation; The number and distribution of measurement points in the network directly determine the observability of the system; When the redundancy is sufficient, the state estimation technique allows temporary deletion of observation data without affecting the expected value, thereby effectively reducing the impact of data distortion; The state estimator eliminates redundant information and eliminates measurement errors to generate a reliable system state estimate; When the monitoring range of the power system is limited, a suitable measurement matrix is constructed by selecting several rows from the complete observation matrix; It is assumed that the dimension of all measurement matrices satisfies m>n; As a core tool for efficient and fast allocation of power flow, the state predictor needs to calculate the state variables from the inaccurate measurement data obtained from the SCADA system; （1） wherein denotes a weighted least squares estimate of the state variable; is a weight matrix, is a diagonal matrix, the inverse of which represents the noise variance matrix; is a Jacobian matrix, T denotes the transpose.

3. The multi-time scale based power distribution network disturbance identification method of claim 2, wherein, Step S1 further includes the following steps: Step S12, state estimation by weighted least squares: By calculating the variance of the measurement error, the weight factor is obtained as ; The purpose of WLS is as follows （2） where N represents the total number of measurements participating in the estimation, and j represents the measurement number; σ2represents the variance of the measurement error, represents the m-dimensional measurement vector, represents the m-dimensional estimation vector.

4. The multi-time scale based power distribution network disturbance identification method of claim 3, wherein, Step S1 further includes the following steps: Step S13, bad data detection: Bad data detection analyzers are embedded in the state estimation for identifying and rejecting abnormal or erroneous measurements; defined as: (3) wherein represents the difference between the measured data and the estimated data; represents the weighted least squares estimate of the state variable vector x; Estimated data （4） The residue is written as （5） wherein is the identity matrix, is the gain matrix, is the weight matrix, is the Jacobian matrix; The evaluation index and chi-square test are used as detection tools, and the objective function is defined as （6） wherein represents the number of measurements, and the weighting factor is set to ; If , is the objective function, is the critical value of the chi-square distribution with k degrees of freedom and significance level α; when the log-likelihood value of k degrees of freedom satisfies, the state estimation value is considered accurate and reliable; otherwise, it indicates that there is bad data in the system.

5. The multi-time scale based power distribution system disturbance identification method of claim 1, wherein, Step S2 includes the following steps: Step S21, data preparation: The data is divided into time windows according to the disturbance events in the historical fault log; Normal class includes N50, N40, N30, N20 and N10, Pre, Post; Segment analysis allows detection of early warning signals appearing at different times; Multi-scale analysis uses 30s, 60s and 180s window sizes for each segment to capture short-term and long-term dynamics; Step S22, feature engineering: A complete set of features is extracted to capture the complex dynamics of the power system; These features include time domain, frequency domain and system dynamics specific metrics; In the tested method, recursive feature elimination is used as the feature selection method; The model is simplified by reducing irrelevant features.

6. The multi-time scale based power distribution network disturbance identification method of claim 5, wherein, Step S2 further includes the following steps: Step S23, pre-classification: The data set is divided into training set and test set according to the ratio of 90:10; The abnormality with a duration of 10s~1min and the abnormality less than 1s are distinguished by the performance of frequency at different time scales; then, the tripping and load shedding are distinguished by the characteristics that the slope of frequency change of tripping is greater than 0 and the slope of frequency change of load shedding is less than 0; the short circuit and the trip-out event are distinguished according to the different durations of voltage phase angle abnormality, specifically, the trip-out event is caused by the immediate removal of the line after the event occurs, the fault can be removed and the operation is recovered quickly, and the influence on the voltage phase angle is less than 20s, which is abnormal in the short-time feature; the influence of the short circuit event on the voltage phase angle lasts for a long time, and the abnormality is detected in the short-time feature in addition to the abnormality in the ultra-short-time feature; the frequency and phase angle abnormality that does not conform to the traditional event rule is recorded as a non-traditional event or an unknown event, which will be re-identified in the subsequent steps; in addition, if the voltage amplitude is abnormal when no abnormality occurs in the power transmission network, it is judged that the disturbance originates from the local distribution network.

7. The multi-time scale based power distribution system disturbance identification method of claim 1, wherein, Step S3 comprises the following steps: Step S31, XGBoost construction: The feature data extracted in S2 is input as the feature of the sample, and the existing label is used to build the training set and the verification set; normalization and denoising are performed during data preparation, and class imbalance is processed; XGBoost is used in the training stage, and the parameters include the depth of the tree, the number of leaves, the learning rate, the feature subsampling ratio and the regularization term; the optimal hyperparameters are selected by cross-validation, and the recall rate and the confusion matrix are used to evaluate the multi-class classification performance; after the training is completed, the prediction probability is calibrated for subsequent threshold selection and uncertainty estimation; the output result is the large class label and the corresponding probability and confidence of each event, which is used to guide the next step of subdivision and manual review.

8. The multi-time scale based power distribution system disturbance identification method of claim 1, wherein, Step S3 specifically comprises the following steps: Step S32, analysis of intra-class differences: After the large class is preliminarily divided, the samples in each large class are analyzed in depth to find the intra-class differences: including statistical analysis, time series pattern analysis and the time-varying curve characteristics of the discriminator score in each time window.

9. The multi-time scale based power distribution system disturbance identification method of claim 1, wherein, Step S33, unknown event identification: XGBoost finds the leaf with the maximum splitting gain from all the current leaves in each loop to split, so that the model can complete the large class identification in a very short time, thereby being superior to other decision tree methods; the hierarchical clustering method has flexibility, and events with similar characteristics are clustered in the same small class to facilitate subsequent research and analysis and labeling.

10. A power distribution network disturbance identification system based on multiple time scales, comprising a processor, a memory and a bus, wherein the memory stores machine readable instructions executed by the processor; When the system is running, the processor communicates with the memory through the bus, and the machine readable instructions are executed by the processor to perform the power distribution network disturbance identification method based on multiple time scales as claimed in any one of claims 1 to 9.

Images