Training Method for Semi-Real-Time Prediction Models Based on Data Distribution Drift Hierarchical Triggering

Abstract

This application relates to the field of model training technology, specifically to a semi-real-time prediction model training method based on hierarchical triggering of data distribution drift. The method includes: extracting feature values ​​from transmission current signals to train a fault prediction model for real-time prediction of fault feature values ​​at all upcoming data collection times within a preset time period; calculating distribution drift evaluation values ​​and persistence evaluation values ​​to assess the significance of data distribution drift persistence within a preset adjacent time period before the current data collection time; statistically analyzing the quartiles of historical persistence evaluation values ​​to perform semi-real-time update training of the fault prediction model using hierarchical triggering; and using the real-time updated fault prediction model to predict transmission line faults. This application aims to improve the real-time accuracy of fault prediction models in predicting transmission line faults, thereby more effectively ensuring the safety and reliability of transmission lines.

Description

Technical Field

[0001] This application relates to the field of model training technology, specifically to a semi-real-time prediction model training method based on data distribution drift-triggered hierarchical triggering. Background Technology

[0002] With the transformation of the energy structure and the continuous development of smart grid construction, transmission lines, as the core carrier of power transmission in the power system, directly affect the safety and reliability of power supply due to changes in their transmission status. Traditional transmission fault detection methods often only identify faults after they occur, resulting in severe response lag and low detection efficiency. Currently, with the continuous development of artificial intelligence and machine learning technologies, to address the problems of response lag and low detection efficiency in traditional transmission fault detection methods, predictive neural network models are commonly used to perform deep learning on multiple feature parameters of transmission lines. These models then predict faults in transmission lines, improving the timeliness and efficiency of fault detection.

[0003] Currently, predictive neural network models mainly include RNN (Recurrent Neural Networks) and LSTM (Long Short-Term Memory Neural Networks). Existing technologies generally use predictive neural network models to predict transmission line faults. This requires deep learning of multiple feature parameters of the transmission line to train the predictive neural network model, aiming to improve the accuracy of transmission line fault detection. However, because power distribution data on transmission lines is prone to data distribution drift—for example, transmission faults caused by aging insulators often lead to significant changes in power distribution data and transmission characteristic data over time—existing technologies cannot accurately use a hierarchical triggering method for semi-real-time updating and training of the fault prediction model. This results in a continuous decline in the accuracy of the fault prediction model in predicting transmission line faults, failing to guarantee the safety and reliability of transmission lines. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a semi-real-time prediction model training method based on data distribution drift hierarchical triggering, thereby resolving the existing problems.

[0005] The semi-real-time prediction model training method based on data distribution drift-triggered hierarchical triggering in this application adopts the following technical solution:

[0006] One embodiment of this application provides a method for training a semi-real-time prediction model based on data distribution drift-triggered hierarchical triggering, the method comprising the following steps:

[0007] Step 1: Real-time acquisition of transmission current signals on the transmission line using Rogowski coil sensors, and extraction of time-series data of the first feature value, second feature value, and fault feature value at the acquisition time from the transmission current signals to train a fault prediction model for real-time prediction of fault feature values ​​at all acquisition times within a preset future time period.

[0008] Step 2: Construct two-dimensional fault feature points at the acquisition time using the first and second feature values. Based on the proportion of historical two-dimensional fault feature points in the truncated neighborhood dataset of each adjacent acquisition time and the Euclidean distance between each adjacent acquisition time's two-dimensional fault feature point and all two-dimensional fault feature points in its truncated neighborhood dataset, calculate the distribution drift evaluation value for each adjacent acquisition time. This value is used to evaluate the significance of the distribution drift of power distribution data and transmission fault data at adjacent acquisition times.

[0009] Step 3: Identify each neighboring abnormal moment based on the distribution drift assessment value; calculate the persistence assessment value of the current acquisition moment based on the time proportion of neighboring abnormal moments and the average time interval proportion of adjacent neighboring abnormal moments, which is used to assess the significance of the persistence of data distribution drift in the preset neighboring time period before the current acquisition moment.

[0010] Step 4: Calculate the quartiles of historical persistence assessment values ​​to train the fault prediction model in semi-real-time using a graded triggering method, and use the real-time updated fault prediction model to predict transmission line faults.

[0011] Preferably, the training method of the fault prediction model is as follows: the time series data of the first feature value and the time series data of the second feature value of all collection times within a preset historical period are used as the two input features for training the prediction neural network model, and the time series data of the fault feature value is used as the output label for training the prediction neural network model. The prediction neural network model is trained to obtain the fault prediction model.

[0012] Preferably, the first feature value is the standard deviation of the local current sequence at the acquisition time, the second feature value is the information entropy of the local current sequence at the acquisition time, and the fault feature value is the sum of the abnormality of all IMF components after removing the IMF component with the smallest abnormality after performing mode decomposition on the local current sequence at the acquisition time.

[0013] Preferably, the local current sequence at the acquisition time consists of the current signal within a preset time period prior to the acquisition time in the transmission current signal.

[0014] Preferably, the degree of anomaly is measured by the permutation entropy of the IMF components.

[0015] Preferably, the distribution drift evaluation value is negatively correlated with the proportion of historical two-dimensional fault feature points in the truncated neighborhood dataset at each adjacent acquisition time, and positively correlated with the Euclidean distance between the two-dimensional fault feature points at each adjacent acquisition time and all two-dimensional fault feature points in the truncated neighborhood dataset.

[0016] Preferably, the method for obtaining the truncated neighborhood dataset for each adjacent acquisition time is as follows: DPC density peak clustering is performed on the two-dimensional fault feature points of all historical acquisition times of the most recently trained fault prediction model and the two-dimensional fault feature points of all adjacent acquisition times of the current acquisition time, and the truncation distance in the preset algorithm is used to identify the truncated neighborhood dataset of the two-dimensional fault feature points of adjacent acquisition times.

[0017] Preferably, the persistence assessment value is positively correlated with the time proportion and negatively correlated with the time interval proportion.

[0018] Preferably, the method for identifying the nearest abnormal moment is as follows: the maximum inter-class variance algorithm is used to segment the distribution drift evaluation values ​​of all the nearest acquisition moments of the current acquisition moment, and the nearest acquisition moments whose distribution drift evaluation values ​​are greater than or equal to the segmentation threshold are identified as the nearest abnormal moments of the current acquisition moment.

[0019] Preferably, the step of using a tiered triggering method to perform semi-real-time update training of the fault prediction model includes:

[0020] The quartiles of the persistence evaluation values ​​of all historical acquisition times before the current acquisition time are calculated, including the lower quartile, median, and upper quartile. The lower quartile, median, and upper quartile are multiplied by a preset adjustment coefficient, which are used as the first judgment threshold, the second judgment threshold, and the third judgment threshold, respectively.

[0021] If the persistence evaluation value at the current collection time is less than the first judgment threshold, the fault prediction model will not be updated and trained in semi-real-time.

[0022] If the persistence evaluation value at the current acquisition time is greater than or equal to the first judgment threshold and less than the second judgment threshold, the fault prediction model is fine-tuned.

[0023] If the persistence evaluation value at the current acquisition time is greater than or equal to the second judgment threshold and less than the third judgment threshold, the fault prediction model is incrementally learned.

[0024] If the persistence evaluation value at the current acquisition time is greater than or equal to the third judgment threshold, the fault prediction model will be fully retrained.

[0025] This application has at least the following beneficial effects:

[0026] (1) The present invention fully considers the fluctuation and disorder of the transmission current on the transmission line, constructs two-dimensional fault feature points for each adjacent acquisition time of the current acquisition time, and accurately evaluates the significance of data distribution drift in the power distribution data and transmission fault data of the adjacent acquisition time through the local distribution characteristics of the two-dimensional fault feature points of each adjacent acquisition time, thereby more accurately reflecting the adaptability of the fault prediction model trained based on historical power distribution data.

[0027] (2) The present invention identifies the neighboring abnormal time when the data distribution drift phenomenon is more prominent by using the distribution drift evaluation value of all neighboring acquisition time, and evaluates the significance of the persistence of data distribution drift in the preset neighboring time period before the current acquisition time by the distribution status of neighboring abnormal time, so as to more accurately reflect the severity of data distribution drift of power distribution data on the transmission line at the current acquisition time.

[0028] (3) By using the quartiles of the continuous evaluation values ​​of all historical acquisition times before the current acquisition time, the present invention can more accurately use a hierarchical triggering method to perform semi-real-time update training on the fault prediction model, enhance the adaptability of the fault prediction model, and improve the real-time accuracy of the fault prediction model in predicting power transmission faults, thereby more effectively ensuring the safety and reliability of power transmission lines. Attached Figure Description

[0029] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart of the semi-real-time prediction model training method based on data distribution drift-triggered hierarchical triggering provided in this application. Detailed Implementation

[0031] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive objective, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the semi-real-time prediction model training method based on data distribution drift hierarchical triggering proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0032] Unless otherwise defined, 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.

[0033] The following, in conjunction with the accompanying drawings, details the specific scheme of the semi-real-time prediction model training method based on data distribution drift hierarchical triggering provided in this application.

[0034] One embodiment of this application provides a semi-real-time prediction model training method based on data distribution drift hierarchical triggering.

[0035] Specifically, the following training method for a semi-real-time prediction model based on data distribution drift and hierarchical triggering is provided. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0036] Step 1: Real-time acquisition of transmission current signals on the transmission line using Rogowski coil sensors, and extraction of time-series data of the first feature value, second feature value, and fault feature value at the acquisition time from the transmission current signals to train a fault prediction model for real-time prediction of fault feature values ​​at all acquisition times within a preset future time period.

[0037] To improve the accuracy of predictive neural network models in predicting power transmission faults, it is necessary to accurately adopt a hierarchical triggering method to perform semi-real-time update training on the predictive neural network models, thereby more effectively ensuring the safety and reliability of power transmission lines.

[0038] First, Rogowski coil sensors are installed on the transmission line to collect the transmission current signal in real time. The transmission current signal within a preset time period before each acquisition moment is arranged in chronological order and recorded as a local current sequence for each acquisition moment. As an example, this embodiment uses a raw sampling rate of 10kHz for the transmission current signal to meet the requirements for capturing low-frequency fault features. A sliding window mechanism is used to extract features from the local current sequence at a frequency of 10Hz (sliding step size of 0.1s). To balance the computational complexity of EEMD mode decomposition, the local current sequence is downsampled before being input into the mode decomposition algorithm, after anti-aliasing filtering and high-frequency information statistical mapping, for example, downsampling to 200Hz–500Hz. Simultaneously, a controlled-size historical point set is constructed from historical two-dimensional fault feature points using sliding window aggregation or representative sampling methods for subsequent DPC clustering analysis. The preset time period is set to 1s.

[0039] Furthermore, to predict transmission line faults using a predictive neural network model, the local current sequence at each acquisition time is used as the input to the mode decomposition algorithm. The mode decomposition algorithm can be either Empirical Mode Decomposition (EMD) or Ensemble Empirical Mode Decomposition (EEMD). In this embodiment, EEMD is used to obtain the various Intrinsic Mode Functions (IMFs) in different frequency bands, and the anomaly degree of each IMF component is calculated. The anomaly degree can be measured by permutation entropy or sample entropy. This embodiment uses permutation entropy to measure the anomaly degree. The IMF component corresponding to the minimum anomaly degree represents a current characteristic component with strong regularity and high autocorrelation, while other IMF components represent fault characteristic components caused by transmission line faults. EEMD is a well-known technique and will not be elaborated further.

[0040] Therefore, the IMF component with the lowest anomaly level among all IMF components is removed, and the sum of the anomaly levels of the remaining IMF components is recorded as the fault characteristic value at each acquisition time, reflecting the salience of fault characteristics on the transmission line. Simultaneously, the standard deviation and information entropy of the local current sequence at each acquisition time are calculated and used as the first and second characteristic values ​​at each acquisition time, respectively, reflecting the volatility and disorder of the transmission current signal, and are used to train the fault prediction model for transmission line faults.

[0041] Furthermore, the time-series data of the first feature value, the second feature value, and the fault feature value for all collection moments within a preset historical period are obtained through the above method. The time-series data of the first and second feature values ​​are used as the two input features for training the predictive neural network model, and the time-series data of the fault feature value is used as the output label for training the predictive neural network model. The predictive neural network model can be an RNN or an LSTM (Long Short-Term Memory) neural network. This embodiment uses an LSTM, and the model training uses the Adam optimizer with a learning rate of 0.001 and 1000 iterations. After training, a fault prediction model is obtained, which is used to predict the fault feature values ​​for all collection moments within a future preset period in real time. The LSTM is a well-known technology and will not be described further. In this embodiment, the historical period is set to all collection moments within one week prior to the current collection moment.

[0042] Step 2: Construct two-dimensional fault feature points at the acquisition time using the first and second feature values. Based on the proportion of historical two-dimensional fault feature points in the truncated neighborhood dataset of each adjacent acquisition time and the Euclidean distance between each adjacent acquisition time's two-dimensional fault feature point and all two-dimensional fault feature points in its truncated neighborhood dataset, calculate the distribution drift evaluation value for each adjacent acquisition time. This value is used to evaluate the significance of the distribution drift of power distribution data and transmission fault data at adjacent acquisition times.

[0043] In general, due to the easy drift of power distribution data on transmission lines, the power distribution data and its transmission characteristics data often change significantly over time. In order to more accurately predict transmission line faults, it is necessary to accurately use a hierarchical triggering method to perform semi-real-time update training on the fault prediction model to avoid affecting the prediction accuracy of transmission line faults.

[0044] In order to use a tiered triggering method to train the fault prediction model in semi-real-time, it is necessary to accurately identify the data distribution drift of power data characteristics on the transmission line.

[0045] Accordingly, this application denotes the two-dimensional data points composed of the first and second feature values ​​at each historical acquisition moment during the most recent fault prediction model training as the two-dimensional fault feature points at each historical acquisition moment, reflecting the fluctuation and disorder characteristics of the transmission current during the historical model training process; wherein, before using the first and second feature values ​​at each historical acquisition moment to form two-dimensional data points, in order to eliminate the dimensional influence between the first and second feature values, the first and second feature values ​​are first subjected to Z-score standardization processing.

[0046] Simultaneously, each acquisition time within a preset adjacent time period before the current acquisition time is recorded as an adjacent acquisition time. The two-dimensional data point composed of the first and second feature values ​​of each adjacent acquisition time is recorded as a two-dimensional fault feature point of each adjacent acquisition time, reflecting the fluctuation and disorder characteristics of the transmission current in the recent time period of the current acquisition time. In this embodiment, the preset adjacent time period is 1 hour.

[0047] Furthermore, the two-dimensional fault feature points from all historical acquisition times during the most recent fault prediction model training, as well as the two-dimensional fault feature points from all neighboring acquisition times at the current acquisition time, are used as inputs to the DPC density peak clustering algorithm. In this embodiment, the truncation distance is set to the average number of neighbors of each two-dimensional fault feature point, which is 1%-2% of the total number of two-dimensional fault feature points. The DPC density peak clustering algorithm is used to obtain the truncated neighborhood dataset of each two-dimensional fault feature point at each neighboring acquisition time, reflecting the local distribution characteristics of the two-dimensional fault feature points in the neighboring acquisition times. Here, the number of neighbors refers to the number of truncated neighborhood data points for each two-dimensional fault feature point.

[0048] It should be noted that when there is an overlap between the historical acquisition time and the neighboring acquisition time of the current acquisition time, the overlapping time is analyzed as the neighboring acquisition time during clustering.

[0049] Generally, the greater the Euclidean distance between a two-dimensional fault feature point at a nearby acquisition time and all two-dimensional fault feature points in its truncated neighborhood dataset, and the fewer the number of two-dimensional fault feature points at historical acquisition times in the truncated neighborhood dataset, the more likely the power distribution data and transmission fault data at nearby acquisition times will experience data distribution drift, thus affecting the prediction accuracy of transmission line faults.

[0050] Therefore, in order to evaluate the data distribution drift phenomenon of power distribution data and transmission fault data at adjacent acquisition times, for the truncated neighborhood dataset of two-dimensional fault feature points at each adjacent acquisition time, the proportion of the number of two-dimensional fault feature points at historical acquisition times to the total number of two-dimensional fault feature points in the truncated neighborhood dataset is counted. This proportion is denoted as the percentage of the number of historical two-dimensional fault feature points in the truncated neighborhood dataset at each adjacent acquisition time. The smaller the percentage, the smaller the proportion of historical two-dimensional fault feature points in the truncated neighborhood dataset corresponding to the adjacent acquisition time. In this case, the power distribution data and transmission fault data are more likely to experience data distribution drift.

[0051] Based on the above analysis, this application calculates the distribution drift evaluation value of the adjacent acquisition time based on the proportion of the number of historical two-dimensional fault feature points in the truncated neighborhood dataset at each adjacent acquisition time and the Euclidean distance between the two-dimensional fault feature point at each adjacent acquisition time and all two-dimensional fault feature points in its truncated neighborhood dataset. This value is used to evaluate the significance of the distribution drift of the power distribution data and transmission fault data at adjacent acquisition times.

[0052] Preferably, in this application, the distribution drift evaluation value is negatively correlated with the proportion of the number of historical two-dimensional fault feature points in the truncated neighborhood dataset at each adjacent acquisition time, and positively correlated with the Euclidean distance between the two-dimensional fault feature point at each adjacent acquisition time and all two-dimensional fault feature points in its truncated neighborhood dataset.

[0053] It is understandable that a positive correlation means that the dependent variable increases as the independent variable increases, and decreases as the independent variable decreases; a negative correlation means that the dependent variable decreases as the independent variable increases, and increases as the independent variable decreases. This is determined by the actual application, and this application does not impose any special restrictions.

[0054] Specifically, in this embodiment, the distribution drift evaluation value for each adjacent acquisition time is calculated as follows:

[0055]

[0056] In the formula, The mean of the Euclidean distances between a two-dimensional fault feature point at each adjacent acquisition time and all two-dimensional fault feature points in its truncated neighborhood dataset is given. The percentage of historical two-dimensional fault feature points in the truncated neighborhood dataset for each adjacent acquisition time. The first parameter is preset and its function is to avoid the denominator being 0. The value range is 0.01-0.1, and in this embodiment, the value is 0.05.

[0057] It should be noted that existing methods for measuring data distribution drift generally use data interpolation or Euclidean distance. This application uses a feature measurement method based on Euclidean distance, combined with the proportion of historical two-dimensional fault feature points within the truncated neighborhood dataset, to evaluate the data distribution drift of power distribution data and transmission fault data at adjacent acquisition times. Furthermore, the formula for calculating the distribution drift evaluation value uses an inverse proportional relationship to assess the data distribution drift. Specifically, the proportion of historical two-dimensional fault feature points within the truncated neighborhood dataset is inversely proportional to the distribution drift evaluation value, while the mean of the Euclidean distance is directly proportional to the distribution drift evaluation value.

[0058] The distribution drift evaluation value reflects the evaluation result of the significance of data distribution drift in the power distribution data and transmission fault data at the adjacent acquisition time. The larger the distribution drift evaluation value, the higher the significance of data distribution drift in the power distribution data and transmission fault data on the transmission line at this time. At this time, the adaptability of the fault prediction model trained based on historical power distribution data is poor and it cannot accurately predict transmission line faults. Therefore, it is necessary to adopt a hierarchical triggering method to perform semi-real-time update training of the fault prediction model more accurately.

[0059] Step 3: Identify each neighboring abnormal moment based on the distribution drift assessment value; calculate the persistence assessment value of the current acquisition moment based on the time proportion of neighboring abnormal moments and the average time interval proportion of adjacent neighboring abnormal moments, which is used to assess the significance of the persistence of data distribution drift in the preset neighboring time period before the current acquisition moment.

[0060] Furthermore, to more accurately employ a tiered triggering method for semi-real-time update training of the fault prediction model, the distribution drift evaluation values ​​of all neighboring acquisition times at the current acquisition time are used as input to the Otsu's inter-class variance algorithm. The Otsu's algorithm is used to obtain a segmentation threshold, and each neighboring acquisition time whose distribution drift evaluation value is greater than or equal to the segmentation threshold is recorded as a neighboring anomalous time at the current acquisition time. These neighboring anomalous times exacerbate the data distribution drift phenomenon in the power distribution data on transmission lines, necessitating precise tiered triggering for semi-real-time update training of the fault prediction model.

[0061] Additionally, it should be noted that if the calculated segmentation threshold is less than the preset minimum safety threshold, it is determined that there are no abnormal moments in the current time period. The minimum safety threshold can be determined by the implementer using the average of segmentation thresholds calculated from historical periods without abnormal moments; details will not be elaborated further.

[0062] Generally, if the number of adjacent abnormal moments in the preset adjacent time period before the current acquisition time is larger and the time interval between each two adjacent abnormal moments is smaller, it indicates that the data distribution drift in the preset adjacent time period before the current acquisition time is more persistent. At this time, the adaptability of the fault prediction model trained based on historical power distribution data is worse, which can easily lead to a continuous decline in the accuracy of the fault prediction model in predicting power transmission faults.

[0063] Therefore, the ratio of the number of adjacent abnormal times in the preset adjacent time period before the current acquisition time to the total number of all adjacent acquisition times is recorded as the time proportion of data distribution drift in the preset adjacent time period before the current acquisition time. The larger the time proportion, the higher the persistence of data distribution drift in the preset adjacent time period before the current acquisition time, which is likely to cause the accuracy of the fault prediction model in predicting power transmission faults to continuously decrease.

[0064] Simultaneously, the time length of the preset adjacent time period before the current collection time is calculated, and the average time interval between all adjacent abnormal times within the preset adjacent time period before the current collection time is calculated. The ratio of the average time interval to the total time interval is recorded as the proportion of time interval of data distribution drift within the preset adjacent time period before the current collection time.

[0065] Specifically, if the number of adjacent abnormal times in the preset adjacent time period before the current acquisition time is less than 2, then the time interval ratio of data distribution drift in the preset adjacent time period before the current acquisition time is assigned a value of 1. The time interval ratio reflects the persistence of data distribution drift in the preset adjacent time period before the current acquisition time. The smaller the time interval ratio, the higher the persistence of data distribution drift in the preset adjacent time period before the current acquisition time, and the easier it is to affect the accuracy of the fault prediction model in predicting power transmission faults.

[0066] Based on the above analysis, this application calculates the persistence assessment value of the current acquisition time based on the time proportion of adjacent abnormal moments and the average time interval proportion of adjacent abnormal moments, which is used to assess the significance of the persistence of data distribution drift in the preset adjacent time period before the current acquisition time.

[0067] Preferably, in this application, the persistence assessment value is positively correlated with the time proportion and negatively correlated with the time interval proportion.

[0068] Specifically, in this embodiment, the persistence evaluation value at the current acquisition moment is calculated as follows:

[0069]

[0070] In the formula, The preset percentage of time for data distribution drift within the adjacent time period before the current collection time. The percentage of time intervals in the data distribution drift within the adjacent time period prior to the current collection time is preset. The second parameter is preset and its function is to avoid the denominator being 0. Its value range is 0.01-0.1, and in this embodiment, the value is 0.05.

[0071] It should be noted that for measurement methods of feature parameters with positive and negative proportional relationships, the measurement can be performed in the form of fractions or function mappings. As shown in the calculation method above, the time proportion is directly proportional to the persistence evaluation value, and the time interval proportion is inversely proportional to the persistence evaluation value. Therefore, the persistence evaluation value is measured by taking the time proportion as the numerator and the time interval proportion as the denominator.

[0072] The persistence evaluation value reflects the significance of the persistence of data distribution drift in the preset adjacent time period before the current acquisition time. The larger the persistence evaluation value, the more severe the data distribution drift of the power distribution data on the transmission line at the current acquisition time, resulting in poor accuracy of transmission line fault detection. Therefore, it is necessary to adopt a more accurate hierarchical triggering method to perform semi-real-time update training on the fault prediction model to improve the real-time accuracy of the fault prediction model in predicting transmission line faults.

[0073] Step 4: Calculate the quartiles of historical persistence assessment values ​​to train the fault prediction model in semi-real-time using a graded triggering method, and use the real-time updated fault prediction model to predict transmission line faults.

[0074] Furthermore, this application employs a tiered triggering method to perform semi-real-time update training on the fault prediction model. It measures the persistence evaluation value at the time of real-time collection of power distribution data on the transmission line, and statistically analyzes the quartiles of the persistence evaluation values ​​at all historical collection times prior to the current collection time, including the lower quartile, median, and upper quartile. The lower quartile, median, and upper quartile are each multiplied by a preset adjustment coefficient, with the coefficient ranging from 1.5 to 2.5; in this embodiment, it is set to 2, which serves as the first, second, and third judgment thresholds, respectively.

[0075] The continuous evaluation value at the current acquisition moment is compared with the judgment threshold, and the semi-real-time update training of the fault prediction model is triggered in stages according to the comparison result:

[0076] If the persistence evaluation value at the current collection time is less than the first judgment threshold, it means that there is no data distribution drift in the power distribution data on the transmission line at this time. At this time, there is no need to perform semi-real-time update training on the fault prediction model. The fault prediction model after the last training can continue to use the fault prediction model after the last training to predict the fault feature values ​​of all collection times in the future preset period in real time.

[0077] If the persistence evaluation value at the current collection time is greater than or equal to the first judgment threshold and less than the second judgment threshold, it means that the power distribution data on the transmission line has slightly drifted. At this time, the fault prediction model after the last training is selected to be fine-tuned. The power distribution data and transmission fault data of the most recent 12 hours are extracted from the real-time data stream as the incremental training set to fine-tune the fault prediction model after the last training. Then, the fault prediction model after the model fine-tuning is used to predict the fault feature values ​​of all collection times in the next second in real time.

[0078] If the persistence evaluation value at the current collection time is greater than or equal to the second judgment threshold and less than the third judgment threshold, it means that the power distribution data on the transmission line has experienced moderate drift. In this case, incremental learning is selected for the fault prediction model after the last training. During incremental learning, online stochastic gradient descent or online support vector machine can be used to semi-real-time update and train the fault prediction model after the last training. Then, the fault prediction model after incremental learning is used to predict the fault feature values ​​of all collection times in the next second in real time.

[0079] If the persistence evaluation value at the current collection time is greater than or equal to the third judgment threshold, it means that the power distribution data on the transmission line has drifted significantly. At this time, the fault prediction model after the last training is selected to be fully retrained. Then, the fault prediction model after full retraining is used to predict the fault feature values ​​of all collection times within the preset time period in real time.

[0080] It is important to note that during most of the stable operation of transmission lines, the persistence assessment values ​​are generally extremely low and fluctuate little, resulting in a low frequency of triggering full retraining during transmission line fault prediction. Furthermore, during the period of triggering semi-real-time model update training, the old model continues to be used to predict transmission faults until the new model is trained and deployed, at which point the model is switched to predict transmission faults.

[0081] Among them, model fine-tuning, incremental learning, or full retraining of predictive neural networks are all well-known techniques, and the specific processes will not be elaborated here.

[0082] Therefore, a semi-real-time update and training method is adopted for the fault prediction model using a hierarchical triggering approach. The updated fault prediction model then predicts fault feature values ​​for all data collection moments within a preset future time period. If any fault feature value within this period exceeds a fault threshold, a line fault is considered imminent. This type of fault is caused by insulator aging and is a low-frequency fault. Professional electricians must promptly repair this fault to ensure the safety and reliability of the transmission line. The fault threshold can be obtained by segmenting the fault feature values ​​monitored during historical transmission processes using the Otsu's inter-class variance algorithm, a well-known technique whose specific process will not be elaborated upon here.

[0083] Thus, the invention of a semi-real-time prediction model training method based on data distribution drift hierarchical triggering was completed.

[0084] The above technical features constitute the preferred embodiment of this application, which has strong adaptability and the best implementation effect. Unnecessary technical features can be added or removed according to actual needs to meet the needs of different situations.

Claims

1. A method for semi-real-time predictive model training based on hierarchical triggering of data distribution shift, characterized in that, The method includes the following steps: Step 1: Real-time acquisition of transmission current signals on the transmission line using Rogowski coil sensors, and extraction of time-series data of the first feature value, second feature value, and fault feature value at the acquisition time from the transmission current signals to train a fault prediction model for real-time prediction of fault feature values ​​at all acquisition times within a preset future time period. Step 2: Construct two-dimensional fault feature points at the acquisition time using the first and second feature values. Based on the proportion of historical two-dimensional fault feature points in the truncated neighborhood dataset of each adjacent acquisition time and the Euclidean distance between each adjacent acquisition time's two-dimensional fault feature point and all two-dimensional fault feature points in its truncated neighborhood dataset, calculate the distribution drift evaluation value for each adjacent acquisition time. This value is used to evaluate the significance of the distribution drift of power distribution data and transmission fault data at adjacent acquisition times. Step 3: Identify each neighboring abnormal moment based on the distribution drift assessment value; calculate the persistence assessment value of the current acquisition moment based on the time proportion of neighboring abnormal moments and the average time interval proportion of adjacent neighboring abnormal moments, which is used to assess the significance of the persistence of data distribution drift in the preset neighboring time period before the current acquisition moment. Step 4: Calculate the quartiles of historical persistence assessment values ​​to train the fault prediction model in semi-real-time using a graded triggering method, and use the real-time updated fault prediction model to predict transmission line faults. The distribution drift evaluation value is negatively correlated with the proportion of the number of historical two-dimensional fault feature points in the truncated neighborhood dataset at each adjacent acquisition time, and positively correlated with the Euclidean distance between the two-dimensional fault feature point at each adjacent acquisition time and all two-dimensional fault feature points in its truncated neighborhood dataset. The method for obtaining the truncated neighborhood dataset for each adjacent acquisition time is as follows: DPC density peak clustering is performed on the two-dimensional fault feature points of all historical acquisition times of the most recently trained fault prediction model and the two-dimensional fault feature points of all adjacent acquisition times of the current acquisition time. The truncation distance in the preset algorithm is used to identify the truncated neighborhood dataset of the two-dimensional fault feature points of adjacent acquisition times. The persistence assessment value is positively correlated with the time proportion and negatively correlated with the time interval proportion; the method for identifying adjacent abnormal moments is as follows: The Otsu's inter-class variance algorithm is used to segment the distribution drift evaluation values ​​of all neighboring acquisition times at the current acquisition time. Neighboring acquisition times with distribution drift evaluation values ​​greater than or equal to the segmentation threshold are identified as neighboring abnormal times at the current acquisition time.

2. The semi-real-time predictive model training method based on data distribution shift hierarchical trigger of claim 1, wherein, The training method for the fault prediction model is as follows: the time series data of the first feature value and the time series data of the second feature value of all collection times within a preset historical period are used as the two input features for training the prediction neural network model, and the time series data of the fault feature value is used as the output label for training the prediction neural network model. The prediction neural network model is then trained to obtain the fault prediction model.

3. The semi-real-time prediction model training method based on data distribution drift hierarchical triggering as described in claim 2, characterized in that, The first feature value is the standard deviation of the local current sequence at the acquisition time, the second feature value is the information entropy of the local current sequence at the acquisition time, and the fault feature value is the sum of the abnormality of all IMF components after removing the IMF component with the smallest abnormality after performing mode decomposition on the local current sequence at the acquisition time.

4. The semi-real-time prediction model training method based on data distribution drift hierarchical triggering as described in claim 3, characterized in that, The local current sequence at the acquisition time consists of the current signal within a preset time period prior to the acquisition time in the transmission current signal.

5. The semi-real-time prediction model training method based on data distribution drift hierarchical triggering as described in claim 3, characterized in that, The degree of anomaly is measured by the permutation entropy of the IMF components.

6. The semi-real-time prediction model training method based on data distribution drift hierarchical triggering as described in claim 1, characterized in that, The method of using a tiered triggering approach to perform semi-real-time update training of the fault prediction model includes: The quartiles of the persistence evaluation values ​​of all historical acquisition times before the current acquisition time are calculated, including the lower quartile, median, and upper quartile. The lower quartile, median, and upper quartile are multiplied by a preset adjustment coefficient, which are used as the first judgment threshold, the second judgment threshold, and the third judgment threshold, respectively. If the persistence evaluation value at the current collection time is less than the first judgment threshold, the fault prediction model will not be updated and trained in semi-real-time. If the persistence evaluation value at the current acquisition time is greater than or equal to the first judgment threshold and less than the second judgment threshold, the fault prediction model is fine-tuned. If the persistence evaluation value at the current acquisition time is greater than or equal to the second judgment threshold and less than the third judgment threshold, the fault prediction model is incrementally learned. If the persistence evaluation value at the current acquisition time is greater than or equal to the third judgment threshold, the fault prediction model will be fully retrained.

Images