Power measurement data dynamic filling method based on CNN-BiLSTM

Abstract

The application provides a kind of power measurement data dynamic filling method based on CNN-BiLSTM, based on historical contemporaneous similar data, using convolutional neural network, bidirectional long short-term memory network realizes the dynamic missing filling of data on the intelligent fusion terminal of transformer area.The application considers the time sequence characteristics, historical characteristics and the correlation between data and other influencing factors of power measurement data, and constructs a data dynamic prediction model based on CNN-BiLSTM network, wherein a convolutional neural network CNN is used to extract local data features, a bidirectional long short-term memory network is used to learn time sequence characteristics, correlation and historical characteristics of measurement data, and finally based on a dynamic fusion strategy, the missing data is dynamically filled by combining historical characteristics and prediction results, and the filling effect is evaluated in different scenarios.The application method is simple, easy to implement, has strong generalization ability, high accuracy, and can adapt to high missing rate and complex data scenarios.

Description

Technical Field

[0001] This invention relates to the field of electric power measurement data filling technology, specifically a dynamic filling method for electric power measurement data based on CNN-BiLSTM. Background Technology

[0002] With the development of power distribution networks and the investment in new energy power generation, the scale of the power grid is constantly expanding, and various power electronic devices are being put into the grid. The operating environment of the power system is becoming increasingly complex, posing greater challenges to ensuring its safety. In recent years, with the continuous development of big data technology and its application in power grid security protection, combined with the increasing amount of measurement data from the power system, the power system can predict its operating status through data-driven methods, thereby preventing potential dangers and responding promptly. At the same time, with the rapid development of power system measurement technology and the continuous decline in measurement costs, measurement data is showing a rapid growth trend, gradually acquiring the characteristics of big data. However, during the data acquisition and transmission process, communication problems caused by network fluctuations and signal delays lead to varying degrees of data loss. Since the operating status of the power system is obtained through calculations of a series of electrical measurement data, the completeness and accuracy of the measurement data affect the analysis of the power system's operating status. Therefore, accurate data filling is particularly important for better determining the operating status of the power system and preventing potential dangers.

[0003] Although power measurement data imputation has been studied to some extent, several key issues and challenges remain. First, existing research has not comprehensively considered the key factors influencing power data imputation, such as the temporal and historical characteristics of power data, and the correlations between data points. These factors all affect the imputation results to varying degrees, and the vast scope of historical power data means that current research has not fully utilized this large amount of historical data from the same period. Second, most existing studies have not effectively considered the combined effect of these factors on the imputation results, resulting in less than ideal imputation outcomes. Finally, many existing studies rely solely on complex models for data imputation, making it difficult to efficiently handle data inputs of different types and missing rates. Therefore, developing a power data imputation method that comprehensively considers key influencing factors and is simple and easy to implement is urgently needed.

[0004] Terminology Explanation:

[0005] Historical data from the same period in the past: data from the same or similar time point in the past.

[0006] Power measurement data: Power measurement data are numerical information about the operating status of power systems obtained through various measuring devices and technologies. They include basic electrical quantity data such as voltage, power, and current, power quality data such as frequency and harmonics, and other types of data.

[0007] Time series characteristics: The various properties and patterns contained in time series data, used to describe the trend, periodicity, seasonality, volatility and correlation of data changes over time. Summary of the Invention

[0008] To address the aforementioned problems, the present invention aims to provide a dynamic imputation method for electrical measurement data based on CNN-BiLSTM. This method considers factors such as the temporal characteristics, historical features, and correlations between data points. A missing data prediction model based on a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (BiLSTM) network is constructed. The CNN is used to extract local features of the data, and the BiLSTM network is used to learn the temporal series features, correlations, and historical features of the measurement data. Finally, based on a dynamic fusion strategy, missing data is dynamically imputed by combining historical features and prediction results, and the imputation effect is evaluated. The method of this invention is simple, easy to implement, has strong generalization ability, high accuracy, and can adapt to scenarios with high missing rates and complex data.

[0009] The technical solution adopted in this invention is as follows:

[0010] A dynamic padding method for power measurement data based on CNN-BiLSTM includes the following steps:

[0011] Preprocessing of raw electrical force measurement data is performed based on time-tolerant alignment, dynamic interpolation, and construction of time features;

[0012] Historical features are constructed based on historical data from the same period in the preprocessed data;

[0013] Based on the input format required by the data dynamic prediction model, the input data is standardized and sequenced. The input data includes data that is related to the data to be filled, lagged data, historical data, and time characteristic data.

[0014] A dynamic data prediction model based on CNN-BiLSTM network is established to dynamically predict missing data.

[0015] Dynamic imputation of missing data based on prediction results and historical features: calculate the standard deviation of prediction results and historical features to determine the weights; and then use the weighted fusion of prediction results and historical features to obtain the final imputed value.

[0016] Furthermore, it also includes:

[0017] Post-processing and effect evaluation: The final filled values ​​were smoothed using a sliding window mean; the mean square error (MSE) and root mean square error (RMSE) were used to evaluate the filling effect.

[0018] Furthermore, the preprocessing of the original electrical force measurement data based on time-tolerant alignment, dynamic interpolation, and the construction of time features includes:

[0019] Time-Tolerant Alignment: Generate theoretical time series, generate theoretical time points within a specified time period at fixed time intervals; match the theoretical time series and the actual time series with the nearest one through the difference margin, and incorporate the data corresponding to the actual time into the theoretical time data frame, and filter the data columns to form an aligned dataset;

[0020] Dynamic interpolation: The original data is segmented and processed. The standard deviation within the sliding window is calculated. The interpolation method is selected based on the standard deviation threshold: cubic spline interpolation is used when the fluctuation is large and the number of valid data points is ≥4; otherwise, linear interpolation is used. If the fluctuation is small, the linear interpolation result is retained.

[0021] Constructing time features: Time is converted into trigonometric function form, and the sine and cosine components corresponding to hours, minutes, and seconds are calculated through conversion formulas to characterize the periodicity of time.

[0022] Furthermore, the formula for calculating the standard deviation is shown in equation (1):

[0023] (1)

[0024] Where n is the number of data points. It is the actual value of the i-th data point. It is the mean of the data within a fixed-length window.

[0025] Furthermore, the conversion formulas are shown in equations (2) and (3):

[0026] (2)

[0027] (3)

[0028] Where 's' is the timestamp converted to seconds. It is the number of seconds in a day, and the formula for calculating s is shown in equation (4):

[0029] (4)

[0030] Hour, minute, and second correspond to the hour, minute, and second of the timestamp, respectively.

[0031] Furthermore, the construction of historical features based on historical data from the same period in the preprocessed data includes: reading the time point of the target day, searching for A-phase power data at each of the K_STEPS time steps before and after the current time point to obtain historical data prev_values ​​and next_values; if historical data at the same time point exists, filtering the historical data by double-condition verification to find the mean of the historical data as the historical feature; otherwise, calculating the mean of prev_values ​​and next_values ​​as the historical feature.

[0032] Furthermore, the standardization and sequence construction of the input data based on the input format required by the dynamic data prediction model includes:

[0033] Step 3.1: Data Standardization

[0034] The QuantileTransformer method is used to transform the input data into data that follows a uniform distribution, thereby scaling data with different features to the same scale;

[0035] Step 3.2: Constructing the sequence

[0036] Feature sequences x and corresponding target values ​​y are extracted from standardized time series data using a sliding window approach. The resulting feature sequence x is a three-dimensional array with shape (n_samples, N_STEPS, n_features), where n_samples is the number of samples, N_STEPS is the number of time steps in each feature sequence, and n_features is the number of features. The target value y is a one-dimensional array with shape (n_samples), containing the target value corresponding to each feature sequence.

[0037] Furthermore, the establishment of a dynamic data prediction model based on a CNN-BiLSTM network, and the dynamic prediction of missing data based on the dynamic data prediction model, includes:

[0038] Step 4.1: Construct a neural network model containing one Conv1D convolutional layer, two BiLSTM layers, two LeakyReLU activation layers, one Dropout layer, and one fully connected layer (Dense). Use the Adam optimizer and mean squared error loss function, and set an early stopping mechanism.

[0039] Step 4.2: Dynamic Prediction: Initialize the prediction buffer, generate a dynamic sequence seq containing feature data of N_STEPS time steps point by point. If the prediction result of the previous time point exists, update it to the A-phase power lag feature position of the last time step of seq. Use the data dynamic prediction model to predict seq to obtain the prediction result pred of the current time point. If there are missing values ​​in the target column of the current time point, store pred in the corresponding buffer and update the corresponding value in the target column.

[0040] Furthermore, the weight calculation formula is shown in equation (5):

[0041] (5)

[0042] Where hist_std and pre_std are the standard deviations of historical features and prediction results, respectively, and weight represents the weight of the prediction result in the fusion;

[0043] The weighted fusion formula is shown in equation (6):

[0044] (6)

[0045] Where final is the final filled value, pre is the prediction result, and hist is the historical feature.

[0046] Furthermore, the formula for calculating the mean squared error (MSE) is shown in equation (7):

[0047] (7)

[0048] Where n is the number of observations. It is the i-th true value. This is the corresponding prediction result;

[0049] The formula for calculating the root mean square error (RMSE) is shown in equation (8):

[0050] (8).

[0051] The beneficial effects of this invention are:

[0052] 1) This invention proposes a dynamic filling method for power measurement data based on CNN-BiLSTM. It addresses the problems of existing methods, such as insufficient consideration of factors, inadequate utilization of historical data, and over-reliance on models. It comprehensively considers the temporal characteristics, historical characteristics, and correlations between data in power data, and achieves accurate filling of missing measurement data.

[0053] 2) This invention constructs a dynamic data prediction model based on CNN-BiLSTM network. By dynamically fusing the prediction results with historical features, it fully incorporates the application of historical data from the same period, solving the problem that existing methods rely too much on model prediction results and do not make full use of historical features when filling power data, thus realizing dynamic data filling.

[0054] 3) Based on real-time available electrical force measurement data, this invention can adapt to complex situations such as high missing rates and multiple types of inputs, providing important support for solving the problem of large amounts of missing measurement data. Attached Figure Description

[0055] Figure 1 This is a flowchart of a method for dynamically filling in electrical measurement data based on CNN-BiLSTM, according to an embodiment of the present invention.

[0056] Figure 2 A flowchart of time-tolerant alignment provided for embodiments of the present invention;

[0057] Figure 3 A flowchart of dynamic interpolation provided for embodiments of the present invention;

[0058] Figure 4 A flowchart illustrating the construction history features provided in this embodiment of the invention;

[0059] Figure 5 A diagram illustrating the construction method of a sliding window provided in an embodiment of the present invention;

[0060] Figure 6 A detailed structural diagram of the model provided in the embodiments of the present invention;

[0061] Figure 7 A flowchart of dynamic prediction provided for embodiments of the present invention;

[0062] Figure 8 A dynamic fusion flowchart provided for embodiments of the present invention;

[0063] Figure 9 A diagram illustrating the power data filling effect provided in an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] like Figure 1As shown, this embodiment of the invention provides a method for dynamically filling in electrical measurement data based on CNN-BiLSTM, including the following steps:

[0066] Step 1: Data Preprocessing

[0067] Data preprocessing consists of three parts: time-tolerant alignment, dynamic interpolation, and construction of time features.

[0068] 1) Time-tolerant alignment

[0069] The process of time fault tolerance alignment is as follows Figure 2 As shown, when generating the theoretical time series, theoretical time points are generated at 2-second intervals within a specified time period based on the obtained base date. Then, the theoretical time series is placed into an aligned data frame. Next, fault-tolerant matching is performed, and the theoretical time points and actual time points are matched as close as possible by setting an appropriate difference margin. Then, the data corresponding to the actual time is merged into the theoretical time data frame. Finally, the data columns are filtered for subsequent analysis and processing.

[0070] 2) Dynamic interpolation

[0071] The idea behind dynamic interpolation is to use cubic spline interpolation and linear interpolation respectively to handle different fluctuations in the data. The implementation steps of dynamic interpolation are as follows: Figure 3 As shown, the original data is first linearly interpolated, then the window size is set and the standard deviation of each time period is calculated sequentially. Next, based on the set threshold, it is determined whether the rolling standard deviation is greater than the threshold. If it is greater than the threshold, this continuous time period is selected for processing of segments with large fluctuations; otherwise, segments with small fluctuations are processed. When processing segments with large fluctuations, it is first determined whether the effective data points meet the requirements of cubic spline interpolation. Cubic spline interpolation requires no less than 4 effective data points. If the requirements are met, cubic spline interpolation is used; otherwise, the linear interpolation result is used. When processing segments with small fluctuations, the linear interpolation result is retained. The formula for calculating the standard deviation is shown in equation (1):

[0072] (1)

[0073] Where n is the number of data points. It is the actual value of the i-th data point. It is the mean of the data within a fixed-length window.

[0074] 3) Constructing time features

[0075] By converting time into trigonometric function form, the model can better learn the periodic characteristics of time series data. The conversion formulas are shown in equations (2) and (3) respectively:

[0076] (2)

[0077] (3)

[0078] Where 's' is the timestamp converted to seconds. It is the number of seconds in a day. The formula for calculating s is shown in equation (4):

[0079] (4)

[0080] Hour, minute, and second correspond to the hour, minute, and second of the timestamp, respectively.

[0081] Step 2: Constructing Historical Features

[0082] The steps to construct historical features are as follows: Figure 4 As shown, first, the time point of the target day is read. Then, at the current time point, the A-phase power data is searched for K_STEPS time steps forward and backward, and recorded as prev_values ​​and next_values. Next, A-phase power data at the same time point in the past is searched. If it exists, it is compared with the mean of prev_values ​​and next_values ​​for a dual-condition verification. By judging whether the threshold requirement is met, the mean of historical data with similar values ​​is selected as historical features. If the A-phase power data at the same time point in the past does not exist, the mean of prev_values ​​and next_values ​​is calculated as the historical features.

[0083] Step 3: Data Standardization and Sequence Construction

[0084] Step 3.1: Data Standardization

[0085] The QuantileTransformer method is used to transform the input data into data that follows a uniform distribution, thereby scaling data with different features to the same scale and reducing the impact of outliers in the data.

[0086] Step 3.2: Constructing the sequence

[0087] After standardization, the data needs to be converted into a format suitable for model input, such as... Figure 5As shown, a sliding window approach is used to extract feature sequences x and corresponding target values ​​y from standardized time series data. The resulting feature sequence x is a three-dimensional array of shape (n_samples, N_STEPS, n_features), where n_samples is the number of samples, N_STEPS is the number of time steps in each feature sequence, and n_features is the number of features. The target value y is a one-dimensional array of shape (n_samples), containing the target value corresponding to each feature sequence.

[0088] Step 4: Model Building and Dynamic Prediction

[0089] Step 4.1: Model Building

[0090] The specific structure of the model is as follows Figure 6 As shown, the model consists of one Conv1D convolutional layer, two BiLSTM layers, two LeakyReLU activation layers, one Dropout layer, and one fully connected Dense layer. The Adam optimizer was chosen, with a learning rate of 0.001 and a loss function of 'mse'. An early stopping mechanism was also implemented, stopping training prematurely when the model's loss function has almost stopped decreasing.

[0091] Step 4.2: Dynamic Prediction

[0092] The process of dynamic prediction is as follows Figure 7 As shown, in the initialization phase, a prediction result buffer of the same length as the target data is initialized, and the target data is standardized using the previously fitted standardized parameters. Next, prediction is performed point-by-point starting from the N_STEPS time point. Each time, a dynamic sequence seq containing feature data from N_STEPS time steps is constructed. If a prediction result exists for the previous time point, it is updated to the A-phase power lag feature position of the last time step of the seq. Then, the model is used to predict the seq to obtain the prediction result pred for the current time point. If there are missing values ​​in the target column at the current time point, pred is stored in the corresponding buffer, and the corresponding value in the target column is updated.

[0093] Step 5: Dynamic Fusion

[0094] Dynamic fusion process such as Figure 8 As shown, dynamic fusion combines the model prediction results and historical features in different proportions to fully utilize the information from both. The weight calculation formula is shown in equation (5):

[0095] (5)

[0096] Where hist_std and pre_std are the standard deviations of historical features and prediction results, respectively, and weight represents the weight of the prediction result in the fusion.

[0097] The weighted fusion formula is shown in equation (6):

[0098] (6)

[0099] Where final is the final filled value, pre is the prediction result, and hist is the historical data.

[0100] Step 6: Post-processing and effect evaluation

[0101] Step 6.1: Post-processing

[0102] After the prediction results are dynamically fused with historical features, the resulting data will be averaged using a sliding window to calculate the average value of data from adjacent time points.

[0103] Step 6.2: Effectiveness Evaluation

[0104] The mean squared error (MSE) and root mean squared error (RMSE) were used to evaluate the performance. The formula for calculating the mean squared error (MSE) is shown in equation (7):

[0105] (7)

[0106] Where n is the number of observations. It is the i-th true value. This is the corresponding prediction result.

[0107] The formula for calculating the root mean square error (RMSE) is shown in equation (8):

[0108] (8)

[0109] In this embodiment, the measurement data acquisition cycle is 2 seconds, the input features are 6, and the measurement data from 19:00:00 to 21:00:00 over 6 days are selected as input to fill in the missing data of phase A power on day 6. The curve of the filled result and the original data are plotted. The data filling is performed using LSTM and BiLSTM dynamic prediction models as well as CNN-BiLSTM model without dynamic filling. The accuracy and efficiency of the proposed model in filling the data are verified by comparison.

[0110] Figure 9The image shown is a diagram illustrating the filling effect provided by an embodiment of the present invention. The blue data points represent the original data, and the orange curve represents the data change after filling. It can be seen that the shapes of the two are well-matched with minimal error, indicating that the present invention can accurately fill in missing measurement data.

[0111] Table 1 compares the results of four models with different missing rates under random missing values.

[0112]

[0113] As shown in Table 1, the imputation errors of all four models increase with the increase of the missing data rate, indicating that the imputation performance of the models decreases as the number of missing data increases. The reason for this is that the more missing data there is, the fewer features the model learns from the data, thus reducing the imputation effect. When comparing the imputation effects of the four models at the same missing data rate, it was found that the CNN-BiLSTM dynamic prediction model consistently had the smallest imputation error regardless of the missing data rate. Specifically, when the missing data rate was no greater than 20%, the imputation errors of the LSTM, BiLSTM, and CNN-BiLSTM dynamic prediction models were similar, even identical. However, when the missing data rate was greater than 20%, the imputation error of the CNN-BiLSTM dynamic prediction model was significantly lower than that of the LSTM and BiLSTM dynamic prediction models. At a missing data rate of 40%, the imputation error of the CNN-BiLSTM dynamic prediction model decreased by 11.0% and 6.9% compared to the LSTM and BiLSTM models, respectively. Furthermore, comparing CNN-BiLSTM networks with and without dynamic imputation, it was found that under different missing rates, the error of the model with dynamic imputation was lower than that of the model without dynamic imputation, and this difference was greater as the missing rate increased. Analysis suggests that with a high missing rate, the model learns fewer data features, and relying solely on the model's predictions to imput data results in poor performance. Combining historical features with prediction results for dynamic imputation effectively addresses the problem of over-reliance on model predictions, thereby improving the final imputation quality.

[0114] Table 2. Comparison of four models with different deletion types at a 30% deletion rate.

[0115]

[0116] As shown in Table 2, the CNN-BiLSTM dynamic prediction model performs better overall compared to the other three models under different missing types. Specifically, the proposed model has the lowest error rate in both random and combined missing cases, but its performance is inferior to the LSTM and BiLSTM dynamic prediction models in cases of continuous missing data. Furthermore, compared to the other three, the CNN-BiLSTM model without dynamic imputation has a significantly larger imputation error.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for dynamically filling in electrical power measurement data based on CNN-BiLSTM, characterized in that, Includes the following steps: Preprocessing of raw electrical force measurement data is performed based on time-tolerant alignment, dynamic interpolation, and construction of time features; Historical features are constructed based on historical data from the same period in the preprocessed data; Based on the input format required by the data dynamic prediction model, the input data is standardized and sequenced. The input data includes data that is related to the data to be filled, lagged data, historical data, and time characteristic data. A dynamic data prediction model based on a CNN-BiLSTM network is established to dynamically predict missing data. Dynamic imputation of missing data based on prediction results and historical features: calculate the standard deviation of prediction results and historical features, and determine the weights; The final filling value is obtained based on the weighted fusion prediction results and historical features; The preprocessing of the raw electrical force measurement data based on time-tolerant alignment, dynamic interpolation, and the construction of time features includes: Time-Tolerant Alignment: Generate theoretical time series, generate theoretical time points within a specified time period at fixed time intervals; match the theoretical time series and the actual time series with the nearest one through the difference margin, and incorporate the data corresponding to the actual time into the theoretical time data frame, and filter the data columns to form an aligned dataset; Dynamic interpolation: The original data is segmented and processed. The standard deviation within the sliding window is calculated. The interpolation method is selected based on the standard deviation threshold: cubic spline interpolation is used when the fluctuation is large and the number of valid data points is ≥4; otherwise, linear interpolation is used. If the fluctuation is small, the linear interpolation result is retained. Constructing time characteristics: Time is converted into trigonometric function form, and the sine and cosine components corresponding to hours, minutes, and seconds are calculated through conversion formulas to characterize the periodicity of time.

2. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, Also includes: Post-processing and effect evaluation: The final filled values ​​were smoothed using a sliding window mean, and the filling effect was evaluated using mean square error (MSE) and root mean square error (RMSE).

3. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, The formula for calculating the standard deviation is shown in equation (1): (1)； Where n is the number of data points. It is the actual value of the i-th data point. It is the mean of the data within a fixed-length window.

4. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, The conversion formulas are shown in equations (2) and (3): (2)； (3)； Where 's' is the timestamp converted to seconds. It is the number of seconds in a day, and the formula for calculating s is shown in equation (4): (4)； Hour, minute, and second correspond to the hour, minute, and second of the timestamp, respectively.

5. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, The construction of historical features based on historical data from the same period in the preprocessed data includes: reading the time point of the target day, searching for A-phase power data at each of the K_STEPS time steps before and after the current time point to obtain historical data prev_values ​​and next_values; if historical data at the same time point exists, filtering the historical data by double-condition verification to find the mean of the historical data as the historical feature; otherwise, calculating the mean of prev_values ​​and next_values ​​as the historical feature.

6. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, The process of standardizing and constructing sequences from input data based on the input format required by the dynamic data prediction model includes: Step 3.1: Data Standardization The QuantileTransformer method is used to transform the input data into data that follows a uniform distribution, thereby scaling data with different features to the same scale; Step 3.2: Constructing the sequence Feature sequences x and corresponding target values ​​y are extracted from standardized time series data using a sliding window approach. The resulting feature sequence x is a three-dimensional array with shape (n_samples, N_STEPS, n_features), where n_samples is the number of samples, N_STEPS is the number of time steps in each feature sequence, and n_features is the number of features. The target value y is a one-dimensional array with shape (n_samples), containing the target value corresponding to each feature sequence.

7. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, The establishment of a dynamic data prediction model based on a CNN-BiLSTM network, and the dynamic prediction of missing data based on the dynamic data prediction model, includes: Step 4.1: Construct a neural network model containing one Conv1D convolutional layer, two BiLSTM layers, two LeakyReLU activation layers, one Dropout layer, and one fully connected layer (Dense). Use the Adam optimizer and mean squared error loss function, and set an early stopping mechanism. Step 4.2: Dynamic Prediction: Initialize the prediction buffer, generate a dynamic sequence seq containing feature data of N_STEPS time steps point by point. If the prediction result of the previous time point exists, update it to the A-phase power lag feature position of the last time step of seq. Use the data dynamic prediction model to predict seq to obtain the prediction result pred of the current time point. If there are missing values ​​in the target column of the current time point, store pred in the corresponding buffer and update the corresponding value in the target column.

8. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, The weight calculation formula is shown in equation (5): (5)； Where hist_std and pre_std are the standard deviations of historical features and prediction results, respectively, and weight represents the weight of the prediction result in the fusion; The weighted fusion formula is shown in equation (6): (6)； Where final is the final filled value, pre is the prediction result, and hist is the historical feature.

9. The method for dynamically filling in electrical measurement data based on CNN-BiLSTM according to claim 1, characterized in that, The formula for calculating the mean squared error (MSE) is shown in equation (7): (7)； Where n is the number of observations. It is the i-th true value. This is the corresponding prediction result; The formula for calculating the root mean square error (RMSE) is shown in equation (8): (8)。

Images