An ultrahigh voltage direct current transmission line audible noise invalid data cleaning method

Abstract

A method for cleaning invalid data of audible noise from ultra-high voltage direct current (UHVDC) transmission lines includes the following steps: Collected audible noise A-weighted sound level data is divided into multiple frequency bands in a certain octave band format; the validity of the A-weighted sound level data is determined using an AM-LSTM-LightGBM model, and invalid data is directly removed; missing values ​​in the A-weighted sound level data after deleting invalid data are filled using mean interpolation; and L5, L50, and L95 values ​​are calculated for the datasets without invalid data removal, the datasets after invalid data removal based on the AM-LSTM-LightGBM model, and the datasets with invalid data removed and filled using mean interpolation. Compared with traditional methods, this method achieves higher invalid data classification accuracy without considering meteorological factors; and because it considers the filling of missing values, it can provide a more complete dataset for subsequent scientific research.

Description

Technical Field

[0001] This invention relates to the field of audible noise monitoring technology for power transmission lines, specifically to a method for cleaning invalid data of audible noise from ultra-high voltage direct current transmission lines. Background Technology

[0002] Audible noise constrains the selection of conductors, optimization design, engineering investment, and environmental protection of transmission lines. For overhead lines with voltage levels above 500kV, audible noise surpasses radio interference as the controlling factor in the electromagnetic environment assessment of transmission lines. Accurately measuring audible noise data of high-voltage transmission lines is of great significance for studying the generation mechanism and prevention measures of audible noise in ultra-high-voltage transmission lines, and is crucial for line design and environmental protection.

[0003] Audible noise testing is typically conducted outdoors, and there are usually numerous external noise sources around the test lines for power transmission lines. Based on past experience, continuously recorded audible noise measurements show significant fluctuations due to the random nature of weather conditions and corona discharge. However, when the instrument is functioning correctly and there is no external interference, these fluctuations are relatively small in the short term. Invalid measurements caused by external noise sources often deviate from the normal fluctuation range. For small amounts of audible noise data, manual methods are usually used to remove outliers. Currently, however, the use of sound level meters for automatic continuous acquisition of audible noise data from power transmission lines results in a large volume of data, including many invalid data points, making manual data cleaning time-consuming and labor-intensive.

[0004] In recent years, domestic scholars have studied the relationship between the A-weighted sound level of corona audible noise and its various spectral components in the positive and negative conductors of ultra-high voltage direct current (UHVDC) transmission lines through corona cage experiments, discovering some correlations between the A-weighted sound level and its spectral components. This provides a new approach for cleaning invalid audible noise data; however, the application of this technology in cleaning invalid audible noise data has not yet been reported.

[0005] Noise from different sources in the external environment may be concentrated in certain frequency bands. Since the equivalent A-weighted sound level is a weighted sum of the energy of each frequency within the audible noise frequency range, the equivalent A-weighted sound level may not change significantly when there is external environmental noise, but it will produce a large spike in the corresponding 1 / 3 octave band component. Traditional methods for solving the problem of invalid audible noise data detection in UHVDC transmission lines do not consider the data characteristics of each frequency band component of the noise, resulting in poor detection accuracy and frequent missed detections and false detections. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for cleaning invalid data with audible noise from ultra-high voltage direct current transmission lines. Compared with traditional methods, this method achieves higher accuracy in classifying invalid data without considering meteorological factors. Furthermore, by taking into account the imputation of missing values, it can provide a more complete dataset for subsequent scientific research.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for cleaning invalid data from audible noise in ultra-high voltage direct current transmission lines includes the following steps:

[0009] Step 1: Divide the collected audible noise A-weighted sound level data into multiple frequency bands in a certain octave band format, use the AM-LSTM-LightGBM model to determine the validity of the A-weighted sound level data, and directly remove invalid data;

[0010] Step 2: Fill in the missing values ​​in the A-weighted sound level data after deleting invalid data in Step 1 using the mean imputation method. The formula is:

[0011]

[0012] Where, x t+i This represents the fill value to be inserted at time t+i; x i This represents the missing i-th A-weighted sound level data; i represents the i-th missing data; x t This represents the valid data at the time preceding consecutive missing data; x t+n+1 This represents the valid data at the next time step after consecutive missing data; n represents the number of consecutive missing data.

[0013] Step 3: Calculate the L5, L50, and L95 values ​​for the datasets without invalid data removal, the datasets after invalid data removal based on the AM-LSTM-LightGBM model in Step 1, and the datasets after invalid data deletion and filling in the complete datasets using mean interpolation in Step 2.

[0014] By analyzing the changes in L5, L50, and L95 values, it can be determined whether the invalid data cleaning method used in this invention will affect the subsequent analysis of audible noise data of transmission lines.

[0015] In step 1, the audible noise data is the audible noise data of the UHVDC transmission line. The audible noise components of 33 frequency bands from 12.5Hz to 20000Hz are divided in the form of 1 / 3 octave band.

[0016] The method of this invention uses the equivalent A-weighted sound level of audible noise and its components in each frequency band as input to mine the temporal characteristics of each frequency band component, which can improve the accuracy of invalid data detection.

[0017] In step 1, the AM-LSTM-LightGBM model includes an LSTM model, a LightGBM model, and an AM mechanism; the LSTM model is used to extract the feature vector of a 33-dimensional audible noise sequence, and then the AM mechanism and the LightGBM model are combined to remove invalid data.

[0018] The AM-LSTM-LightGBM model workflow includes the following steps:

[0019] ①. Data preprocessing: Normalize the input multidimensional noisy time series separately, and divide the training set and test set in the same proportion;

[0020] ②. LSTM Feature Extraction: Extract time features from the normalized multidimensional noisy time series using an LSTM neural network.

[0021] ③. Add attention mechanism AM: Use attention mechanism to selectively learn the temporal features extracted by LSTM, selectively extract some features from the output vector at each time step and then perform weighted summation to obtain a more representative feature vector γ.

[0022] ④. Invalid data classification: The feature vector containing audible noise temporal information extracted by AM-LSTM is input into the LightGBM layer for training to separate invalid data.

[0023] In step 1, the time window length parameter of the LSTM neural network model is set to 16.

[0024] The impact of moving window length on model detection performance was tested by setting 10 different moving window lengths: 4, 6, 8, 10, 12, 14, 16, 18, 20, and 22. The results showed that the accuracy and recall were lowest when the moving window length was 4, with recall being the metric for evaluating the model's performance on invalid data detection. This may be because the moving window length is too short, preventing the model from learning effective temporal feature representations. As the moving window length increased, both accuracy and recall improved significantly. The model achieved the highest accuracy and recall with a moving window length of 16, but further increases in moving window length did not significantly improve these metrics and even resulted in a slight decrease. This phenomenon may be due to redundant input information caused by an excessively long moving window, leading to a decline in the model's learning ability. Therefore, the method of this invention uses 16 as the optimal moving window length setting.

[0025] In step 2, the mean imputation method uses the preceding and following valid data of the missing data to calculate, which can reduce the estimation bias that may be caused by directly deleting invalid data and construct a complete dataset.

[0026] In step 3, the L5, L50, and L95 values ​​refer to the values ​​that rank in the top 5%, top 50%, and top 95% of the data when arranged from largest to smallest. These values ​​are statistical values ​​used to study the distribution characteristics of the data. By comparing the changes in statistical values ​​before and after data cleaning, the reliability of the data removal method can be analyzed. Generally, if the statistical values ​​do not change significantly before and after invalid data cleaning, meaning the distribution characteristics remain essentially unchanged, it indicates that the invalid data cleaning method is effective.

[0027] This invention discloses a method for cleaning invalid data from audible noise in ultra-high voltage direct current transmission lines, with the following technical advantages:

[0028] 1) In step 1 of this invention, the noise generated by different external noise sources may be concentrated in certain frequency bands. Using the equivalent A-weighted sound level of audible noise and its 1 / 3 octave band components as input, the temporal characteristics of each frequency band component can be mined, which can improve the accuracy of invalid data detection.

[0029] 2) In step 2 of this invention, directly deleting invalid data from the original data would damage the integrity and continuity of the data, negatively impacting subsequent data analysis. To reduce the potential bias in statistical values ​​and estimates caused by directly deleting invalid data, and to construct a complete dataset as much as possible, mean imputation is used to fill in the invalid data. Its advantage lies in its robustness; it does not disrupt the overall data distribution and preserves the historical patterns of the data as much as possible.

[0030] 3) In step 3 of this invention, the statistical values—L5, L50, and L95—reflect the distribution characteristics of the data. The reliability of the data removal method is analyzed by comparing the changes in these statistical values ​​before and after data cleaning. If the statistical values ​​do not change significantly before and after invalid data cleaning, meaning the distribution characteristics remain essentially unchanged, it proves that the invalid data cleaning method is effective. Attached Figure Description

[0031] Figure 1 This is a flowchart of the implementation process of the present invention.

[0032] Figure 2 Flowchart for determining invalid data in the AM-LSTM-LightGBM model. Detailed Implementation

[0033] A method for cleaning invalid audible noise data in ultra-high voltage direct current (UHVDC) transmission lines is disclosed. First, the advantages of LSTM neural networks in processing multidimensional and nonlinear time-series data are utilized, using it as a feature extraction layer. Second, an attention mechanism is introduced as an attention layer to address the potential performance degradation of the feature extraction layer as the dimension and length of the time series increase. Then, the extracted features are classified using LightGBM to detect invalid data. Finally, invalid data is deleted and imputed using mean imputation, resulting in complete audible noise data after invalid data cleaning. This invention utilizes the temporal characteristics of audible noise data, comprehensively considering the 1 / 3 octave band components of 33 audible noises as part of the LSTM model's feature extraction data, selecting an appropriate moving window length, and finally imputing the deleted invalid data using mean imputation.

[0034] A method for cleaning invalid audible noise data from ultra-high voltage direct current (UHVDC) transmission lines is proposed. This method utilizes the frequency band components of audible noise data and employs AM and LSTM-LightGBM algorithms to identify invalid data. Based on an LSTM neural network, features are extracted from the frequency band data at each time stamp. A feature dimension attention mechanism is introduced to adaptively allocate weights to characterize the expressive power of key feature information. LightGBM is then used to classify the extracted features and detect invalid data. Invalid data is deleted and then filled in using mean interpolation. The L5, L50, and L95 values ​​are calculated for the datasets without invalid data removal, the datasets with invalid data removed using the AM and LSTM-LightGBM methods, and the datasets with invalid data removed and filled in using mean interpolation. Analysis of the changes in L5, L50, and L95 values ​​confirms that the invalid data cleaning method does not adversely affect subsequent audible noise data analysis.

[0035] Specifically, the following steps are included:

[0036] Step 1: Divide the collected audible noise data into 33 frequency bands in the form of 1 / 3 octave bands, use the AM-LSTM-LightGBM model to determine the validity of the data, and directly remove invalid data;

[0037] Step 2: Impute missing values ​​in the dataset after Step 1 using mean imputation.

[0038]

[0039] Where n represents the number of consecutive missing data points, x t x represents the valid data at the time preceding the consecutive missing data. t+n+1 x represents the valid data at the next time step after consecutive missing data.t+i This represents the fill value to be inserted at time t+i;

[0040] Step 3: For the dataset without invalid data removal, the dataset after invalid data removal based on the AM-LSTM-LightGBM model, and the dataset after deleting invalid data and filling in the complete dataset using mean interpolation, calculate the L5, L50, and L95 values ​​respectively. By analyzing the changes in the L5, L50, and L95 values, determine whether the invalid data cleaning method will have an adverse effect on the subsequent audible noise data analysis.

[0041] The AM-LSTM-LightGBM model includes an LSTM model, a LightGBM model, and an AM mechanism;

[0042] Long Short-Term Memory (LSTM) is a deep learning algorithm for processing time series data. LSTM units use a three-gate structure to update and utilize historical information, enabling prediction and feature extraction of time series data, particularly effective for high-dimensional and long-length time series. This study utilizes LSTM to extract feature vectors from 33-dimensional audible noise sequences, and then combines AM mechanisms and the LightGBM classification model to remove invalid data.

[0043] LSTM is described in existing technical literature [1], [2], and [3]:

[0044] Literature [1]: BENGIO Y. Learning long-term dependencies with gradientdescent is difficult [J]. IEEE Transactions on Neural Networks, 2002, 5(2): 157-166.

[0045] Reference [2]: Liu Yunpeng, Xu Ziqiang, et al. Prediction method of dissolved gas concentration in transformer oil based on empirical mode decomposition and long short-term memory neural network [J]. Proceedings of the CSEE, 2019, 39(13): 3998-4005

[0046] Reference [3]: Zhu Lingjian, Xun Zihan, Wang Yuxin, et al. Short-term power load forecasting based on CNN-Bi LSTM [J]. Power System Technology, 2021, 45(11):4532-4538.

[0047] Attention mechanism (AM) is a mechanism that enhances the contribution of important features when processing high-dimensional and long sequence data, thereby improving model training efficiency and convergence speed. AM is described in existing literature [4], [5], and [6]:

[0048] Literature [4]: ​​MNIH V, HEESS N, GRAVES A. Recurrent models of visual attention[C]. Advances in Neural Information Processing Systems, 2014: 2204-2212.

[0049] Reference [5]: Zhu Lingjian, Xun Zihan, Wang Yuxin, et al. Short-term power load forecasting based on CNN-Bi LSTM [J]. Power System Technology, 2021, 45(11):4532-4538.

[0050] Reference [6]: Gao Guangshang. A review of research on attention mechanisms in deep learning recommendation models [J]. Computer Engineering and Applications, 2022, 58(9):9-16.

[0051] Lightweight Gradient Boosting Machines (LightGBM) are a common method for solving classification prediction problems. Gradient boosting decision trees are generated through multiple iterations of the training data. In each iteration, gradient information is used to refit a new tree and add it to the previous iteration's tree. In function space, this process can be viewed as a continuously iterative linear combination.

[0052]

[0053] Where χ is the function space of the iterative tree; f q (x i ) represents the predicted value of the i-th sample in the q-th tree. The LightGBM classification strategy refers to existing technology literature [7], and literature [8] describes it as follows:

[0054] Reference [7]: Zhu Lingjian, Xun Zihan, Wang Yuxin, et al. Short-term power load forecasting based on CNN-Bi LSTM [J]. Power System Technology, 2021, 45(11):4532-4538.

[0055] Reference [8]: Gu Chumei, Cao Jianjun, et al. Individual identification of LightGBM radiation source based on ant colony parameter optimization [J]. Computer Engineering and Science, 2023, 45(1): 86-93.

[0056] The invalid data detection model based on AM-LSTM-LightGBM is designed and implemented by combining the above three intelligent algorithms, involving a total of four processes:

[0057] ①. Data preprocessing: Normalize the input multidimensional noisy time series separately, and divide the training set and test set in the same proportion;

[0058] ②. LSTM Feature Extraction: Extract time features from the normalized multidimensional noisy time series using an LSTM neural network.

[0059] ③. Add attention mechanism AM: Use attention mechanism to selectively learn the temporal features extracted by LSTM, selectively extract some features from the output vector at each time step and then perform weighted summation to obtain a more representative feature vector γ.

[0060] ④. Invalid data classification: The feature vector containing audible noise temporal information extracted by AM-LSTM is input into the LightGBM layer for training to separate invalid data.

[0061] Training data: Take the first 30% of the noise components in the dataset as training data, and then take the first 16 groups of labeled multidimensional time series as the first training data. Use the learned feature vectors and the 17th group of noise data vectors for classification training to determine whether the 17th group of data is abnormal. Move the window one unit at a time to complete the training process.

[0062] In step 1, the audible noise data is the audible noise data of the ±800kV UHVDC transmission line, which is decomposed into 33 frequency band components in 1 / 3 octave band form: 12.5Hz, 16Hz, 20Hz, 25Hz, 31.5Hz, 40Hz, 50Hz, 63Hz, 80Hz, 100Hz, 125Hz, 160Hz, 200Hz, 250Hz, 315Hz, 400Hz, 500Hz, 630Hz, 800Hz, 1000Hz, 1250Hz, 1600Hz, 2000Hz, 2500Hz, 3150Hz, 4000Hz, 5000Hz, 6300Hz, 8000Hz, 10000Hz, 12500Hz, 16000Hz, and 20000Hz.

[0063] In step 1, each frequency band component corresponding to each timestamp is used as a set of inputs, and the moving window length of the LSTM feature extraction layer is set to 16. Each step is one unit, and the invalidity of the data is determined sequentially.

[0064] In step 2, if there is only one missing value, the missing value is filled with the average of the data before and after the missing value. If there are consecutive missing data, the missing value is filled with the mean imputation method.

[0065]

[0066] Where n represents the number of consecutive missing data points, x t x represents the valid data at the time preceding the consecutive missing data. t+n+1x represents the valid data at the next time step after consecutive missing data. t+i This represents the fill value to be inserted at time t+i;

[0067] In step 3, L5, L50, and L95 refer to the values ​​that are in the top 5%, top 50%, and top 95% positions when the data is arranged from largest to smallest.

[0068] Verification Example:

[0069] To further analyze a method for cleaning invalid data of audible noise from an ultra-high voltage direct current (UHVDC) transmission line, invalid data cleaning was performed on the noise data of a ±800kV DC transmission line collected between 0:00 and 6:00 AM on a clear night on November 26 of a certain year.

[0070] The audible noise components of transmission lines in 33 frequency bands from 12.5Hz to 20000Hz at a timestamp are used as a set of inputs. The invalid data detection model of AM-LSTM-LightGBM is used to remove invalid data step by step. Then, the missing data after removal is filled in with mean interpolation to construct a complete dataset and complete the invalid data cleaning process.

[0071] The L5, L50, and L95 values ​​were calculated for the dataset without invalid data removal, the dataset after invalid data removal using the AM-LSTM-LightGBM invalid data detection model, and the dataset after invalid data deletion and mean interpolation. By analyzing the changes in the L5, L50, and L95 values, it was determined whether the invalid data cleaning method used in this invention would affect the subsequent analysis of audible noise data of transmission lines. The results showed that the invalid data cleaning method of this invention does not interfere with the subsequent analysis of audible noise data of transmission lines.

[0072] Table 1 shows the maximum, minimum, L5, L50, and L95 values ​​of the dataset after the invalid data was removed using the AM-LSTM-LightGBM invalid data detection model and after missing data was filled in using the mean imputation method.

[0073] Table 1 Validation of Data Validity

[0074]

[0075] By comparing the L5, L50, and L95 values, it can be found that after cleaning invalid data with audible noise using the method of the present invention, the L50 and L95 values ​​remain basically unchanged, with only the maximum value and L5 value decreasing. This does not affect the reliability of the data and provides convenience for subsequent data analysis.

Claims

1. A method for cleaning invalid data from audible noise in ultra-high voltage direct current transmission lines, characterized in that... Includes the following steps: Step 1: Divide the collected audible noise A-weighted sound level data into multiple frequency bands in a certain octave band format, use the AM-LSTM-LightGBM model to determine the validity of the A-weighted sound level data, and directly remove invalid data; Step 2: Fill in the missing values ​​in the A-weighted sound level data after deleting invalid data in Step 1 using the mean imputation method. The formula is: （1）； in, Indicates the insertion point The fill value at time; Indicates the missing first A-weighted sound level data; Indicates the first One missing data point; This represents the valid data at the moment preceding the consecutive missing data. This represents valid data at the time point following a series of missing data. Indicates the number of consecutive missing data; Step 3: Calculate the L5, L50, and L95 values ​​for the datasets without invalid data removal, the datasets after invalid data removal based on the AM-LSTM-LightGBM model in Step 1, and the datasets after deleting invalid data and filling in the complete datasets using mean interpolation in Step 2. In step 1, the AM-LSTM-LightGBM model includes an LSTM model, a LightGBM model, and an AM mechanism. The LSTM model is used to extract the feature vector of a 33-dimensional audible noise sequence, and then the AM mechanism and LightGBM model are combined to remove invalid data. The AM-LSTM-LightGBM model workflow includes the following steps: ① Data preprocessing: Normalize the input multidimensional noisy time series and divide it into training and test sets in the same proportion; ②LSTM Feature Extraction: Extract time features from the normalized multidimensional noisy time series using an LSTM neural network; ③ Add attention mechanism AM: Use attention mechanism to selectively learn the temporal features extracted by LSTM, selectively extract some features from the output vector at each time step and then perform weighted summation to obtain a more representative feature vector γ; ④ Invalid data classification: Input the feature vector containing audible noise temporal information extracted by AM-LSTM into the LightGBM layer for training to separate invalid data.

2. The method for cleaning invalid data of audible noise in ultra-high voltage direct current transmission lines according to claim 1, characterized in that: By analyzing the changes in L5, L50, and L95 values, subsequent analysis of audible noise data of transmission lines can be conducted.

3. The method for cleaning invalid data of audible noise in ultra-high voltage direct current transmission lines according to claim 1, characterized in that: In step 1, the audible noise data is the audible noise data of the UHVDC transmission line. The audible noise components of 33 frequency bands from 12.5Hz to 20000Hz are divided in the form of 1 / 3 octave band.

4. The method for cleaning invalid data of audible noise in ultra-high voltage direct current transmission lines according to claim 1, characterized in that: In step 1, the time window length parameter of the LSTM neural network model is set to 16.

5. The method for cleaning invalid data of audible noise in ultra-high voltage direct current transmission lines according to claim 1, characterized in that: In step 2, the mean imputation method uses the preceding and following valid data of the missing data to calculate, which can reduce the estimation bias that may be caused by directly deleting invalid data and construct a complete dataset.

6. The method for cleaning invalid data of audible noise in ultra-high voltage direct current transmission lines according to claim 1, characterized in that: In step 3, L5, L50, and L95 refer to the values ​​that are in the top 5%, top 50%, and top 95% positions when the data is arranged from largest to smallest.

Images