PMU harmonic data missing filling method and device based on multi-measuring-point data correlation

Abstract

The invention provides a PMU harmonic data missing filling method and device based on multi-measuring-point data correlation, and aims to provide a higher-precision solution for a power grid harmonic monitoring system. When the harmonic data of the target station are missing, PMU measurement data of the target station and the adjacent monitoring stations in the same time period are extracted, and the PMU measurement data are divided into a known data set and a missing data set according to the integrity of the harmonic data; and performing strong correlation screening on the known data set: calculating type-by-type Pearson's correlation coefficients of harmonic data of the target station and PMU data of the target station and PMU data of other stations, removing data types lower than a threshold value, and retaining strong correlation data. Strong correlation data is used as an input feature, a target harmonic real value is used as an output label, a CNN-LSTM model (CNN extracts spatial features and LSTM captures time dependence) is adopted for training, and parameters are optimized through mean square errors until the error of a verification set reaches the standard. And finally, predicting a harmonic value of the missing data set by using the model, and finishing data restoration by taking a prediction result as a filling value.

Description

Technical Field

[0001] The present invention relates to the technical field of filling missing harmonic data of PMU. Background Art

[0002] Power grid companies are actively promoting the construction of power grid harmonic monitoring and analysis systems, and have completed the installation of harmonic monitoring equipment of synchronized phasor measurement units (PMUs) in some substations of 220 kV and above and major power plants. With the expansion of the scale of the power system and the rapid development of big data technology, PMUs will provide more important technical guarantees for the safe and stable operation of the power grid. However, in the actual operation of the power grid, problems such as communication jams and hardware failures occur from time to time, resulting in quality problems such as different degrees of missing of PMU measured harmonic data at a single or adjacent multiple sites. The missing data will directly affect the accuracy of key links such as harmonic state estimation, harmonic source location and responsibility division of the power system.

[0003] In the early stage, limited by the insufficient coverage rate of PMU harmonic monitoring equipment, the method for filling missing harmonic data mainly relied on the time series characteristics of a single measurement point itself for completion, and the spatio-temporal correlation characteristics between adjacent sites were not fully explored. This one-dimensional filling strategy led to insufficient mining of potential data information, directly affecting the accuracy and reliability of the filling results. With the full coverage of the PMU monitoring network, the collection and analysis of massive multi-source harmonic data have become possible, which lays a foundation for constructing a collaborative filling method based on the spatio-temporal correlation characteristics of multi-measurement point data. By integrating the dynamic coupling relationship of harmonic measurement data of adjacent sites, the new filling technology can effectively improve data integrity, which is of great significance for ensuring the accuracy of power grid harmonic monitoring and supporting the safe and stable operation of the power system. Summary of the Invention

[0004] The present invention proposes a method and device for filling missing PMU harmonic data based on the correlation of multi-measurement point data, aiming to provide a higher-precision and more effective solution for the problem of filling missing PMU harmonic data in the power grid harmonic monitoring and analysis system, so as to ensure the accuracy and integrity of PMU harmonic data.

[0005] First aspect, a method for filling missing PMU harmonic data, comprising: when there is missing harmonic data at a target site, extracting PMU measurement data of the target site and one or more adjacent monitoring sites within the same sampling period, and performing dataset division: classifying the PMU measurement data with complete harmonic data as a known dataset, and dividing the PMU measurement data with missing harmonic data into a missing dataset to be filled; performing strong correlation screening on the known dataset: calculating the per-type Pearson correlation coefficient between the harmonic data of the target site and other types of PMU measurement data of the site itself, and at the same time calculating the per-type Pearson correlation coefficient between the harmonic data of the target site and the PMU measurement data of adjacent sites. If the correlation coefficient between the harmonic data of the target site and a certain data type is lower than the threshold, then eliminate this data type; if the correlation coefficient exceeds or is equal to the threshold, then retain this data type, and finally construct a strong correlation known dataset that only contains strongly correlated data types; using the data types strongly correlated with the harmonic data of the target site in the strong correlation known dataset as input features, and using the true value of the harmonic data of the target site as the output label to train a filling model; using the trained filling model to predict the missing harmonic data of the target site in the missing dataset, and using the prediction result as the filling value to complete the filling of the missing part of the PMU harmonic data.

[0006] Second aspect, a device for filling missing PMU harmonic data, comprising: a dataset construction module configured to: when there is missing harmonic data at a target site, extract PMU measurement data of the target site and one or more adjacent monitoring sites within the same sampling period, and perform dataset division: classifying the PMU measurement data with complete harmonic data as a known dataset, and dividing the PMU measurement data with missing harmonic data into a missing dataset to be filled; a strong correlation screening module configured to perform strong correlation screening on the known dataset: calculating the per-type Pearson correlation coefficient between the harmonic data of the target site and other types of PMU measurement data of the site itself, and at the same time calculating the per-type Pearson correlation coefficient between the harmonic data of the target site and the PMU measurement data of adjacent sites. If the correlation coefficient between the harmonic data of the target site and a certain data type is lower than the threshold, then eliminate this data type; if the correlation coefficient exceeds or is equal to the threshold, then retain this data type, and finally construct a strong correlation known dataset that only contains strongly correlated data types; a training module configured to use the data types strongly correlated with the harmonic data of the target site in the strong correlation known dataset as input features, and use the true value of the harmonic data of the target site as the output label to train a filling model; a data filling module configured to use the trained filling model to predict the missing harmonic data of the target site in the missing dataset, and use the prediction result as the filling value to complete the filling of the missing part of the PMU harmonic data.

[0007] In a third aspect, a computer includes: a processor; a memory including one or more computer program modules; wherein, the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules include instructions for implementing the PMU harmonic data missing filling method described above.

[0008] In a fourth aspect, a non-transitory computer-readable instruction is provided for storing, and when the non-transitory computer-readable instruction is executed by a computer, it can implement the PMU harmonic data missing filling method described above. Description of the Drawings

[0009] Figure 1 is a schematic flow chart of a PMU harmonic data missing filling method based on multi-measurement point data correlation according to an embodiment of the present invention.

[0010] Figure 2 is a structural diagram of a filling model according to an embodiment of the present invention.

[0011] Figure 3 is a power grid topological structure according to an embodiment of the present invention.

[0012] Figure 4 is a heat map of the correlation between the fifth harmonic voltage amplitude of the BH station and the data types of other stations according to an embodiment of the present invention. Detailed Embodiments

[0013] Figure 1 Illustrates a PMU harmonic data missing filling method based on multi-measurement point data correlation. The method is introduced in detail below.

[0014] Step 1, PMU harmonic data missing detection and sample division.

[0015] Each PMU monitoring node collects a number of (such as 3600) data samples (i.e., PMU measurement data) at a fixed period (such as every hour). The PMU measurement data includes the fundamental wave voltage amplitude V1, the fundamental wave current amplitude I1, the active power P, and the reactive power Q, and the harmonic voltage amplitude V h and the harmonic current amplitude I h , where h represents the number of a certain harmonic.

[0016] First, perform harmonic component missingness detection on the PMU measurement data of each monitoring site: If the harmonic component data of the target site is missing, calculate its harmonic data missing rate (the number of missing harmonic samples / the total number of harmonic samples × 100%); Subsequently, extract the PMU measurement data of the target site and one or more adjacent monitoring sites within the same sampling period, and perform dataset partitioning: Classify the PMU measurement data with complete harmonic component data as the known dataset, while the PMU measurement data with missing harmonic component data is divided into the missing dataset to be filled. That is, the known dataset includes the PMU measurement data (including harmonic data) of the target site and the PMU measurement data of the sites adjacent to the target site. The missing dataset includes the PMU measurement data of the target site (missing harmonic data) and the PMU measurement data of the sites adjacent to the target site.

[0017] Step 2, perform multi - measurement - point data correlation analysis to construct a strongly correlated known dataset.

[0018] When constructing a strongly correlated known dataset, calculate the per - type Pearson correlation coefficient r between the harmonic data of the target site and other PMU measurement data types of the site itself (such as fundamental wave voltage amplitude, fundamental wave current amplitude, active power, reactive power, etc.); calculate the per - type Pearson correlation coefficient r between the harmonic data of the target site and the PMU measurement data types of adjacent sites (including fundamental wave voltage amplitude, fundamental wave current amplitude, active power, reactive power, harmonic voltage amplitude, harmonic current amplitude, etc.). The calculation formula is as follows:

[0019]

[0020] In the formula: m is the number of samples; X i and Y i are the measured values of two data types respectively; and are the means of the two data types respectively.

[0021] Set the correlation coefficient r threshold (such as r ≥ 0.6). Those skilled in the art can adjust the threshold according to the power grid topology structure or harmonic propagation characteristics. If the correlation coefficient between the harmonic data of the target site and a certain data type (whether it is other PMU data of the target site itself or the PMU data of adjacent sites) is lower than the threshold: Eliminate the uncorrelated data type (such as the fundamental wave current amplitude of the target site or the reactive power of adjacent sites), and retain the harmonic data field of the target site to avoid the interference of low - correlation data on the filling model. If the correlation coefficient exceeds the threshold: Retain the association relationship between the target harmonic data and this data type, because it has spatio - temporal coupling characteristics and can provide high - value association information for subsequent filling.

[0022] By this method, only other data types that are strongly correlated with the target harmonic data are retained (such as the fundamental voltage amplitude of the target site itself or the harmonic current amplitude of adjacent sites), thereby improving the calculation efficiency and accuracy of the filling model.

[0023] The present invention considers the correlation between the measurement sites with missing harmonic data and the data of multiple adjacent measurement sites, uses the Pearson correlation method to analyze the internal relationships and laws between the data of different measurement sites, screens out the data types with strong correlation with the harmonic data in the measurement sites, improves the utilization value of the data, and then obtains more effective data samples, laying a solid foundation for the accuracy of the subsequent model.

[0024] Step 3, train the filling model.

[0025] The data filling model adopts a spatio-temporal feature fusion architecture (CNN-LSTM model) in which a convolutional neural network (CNN) and a long short-term memory neural network (LSTM) are connected in series.

[0026] In terms of input design, the multi-dimensional data strongly correlated with the target harmonic data screened out from the strongly correlated known data set is used as the model input, including the PMU measurement data types of the target site itself (such as fundamental voltage / current amplitude, active / reactive power) and the strongly correlated PMU data fields of adjacent sites (such as harmonic voltage / current amplitude). The input data is organized in the form of a time series to form a three-dimensional tensor (number of samples × time step × feature dimension), where the feature dimension covers all the screened strongly correlated data fields.

[0027] In terms of the model architecture, first, one-dimensional convolution operations are used to extract the spatial correlation of the input data in the feature dimension (such as local patterns between different PMU data fields), and a pooling layer is used to reduce the data dimension; then the extracted spatial features are input into the LSTM network to capture the dynamic dependence relationships in the time dimension (such as the fluctuation rules of harmonic data over time), and finally the time series prediction results are output. The training objective is to use the true value of the target harmonic data as the supervision signal, and by minimizing the mean square error (MSE) loss function between the predicted value and the true value, the backpropagation algorithm is used to optimize the model parameters. The convergence condition is set as: when the prediction error (such as root mean square error RMSE) of the model on the validation set does not decrease for several consecutive rounds of iteration, or is lower than a preset threshold (such as RMSE < 0.1PU), stop training and save the optimal model parameters to complete the model construction.

[0028] This design ensures that the model can effectively learn the mapping relationship between harmonic data and other strongly correlated data through joint modeling of spatial and temporal features, and at the same time guarantees the generalization performance of the model through a dynamic error monitoring mechanism.

[0029] Specifically, the CNN-LSTM model includes an input layer, a CNN convolutional layer, a CNN pooling layer, an LSTM layer, a fully connected layer, and an output layer. The specific operation of the CNN-LSTM modeling process is as follows:

[0030] (1) Input layer

[0031] The input layer is responsible for receiving a strongly correlated known data set, screening out multi-dimensional data fields that are strongly correlated with the target harmonic data, and organizing these data into a matrix form and inputting them into the CNN convolutional layer.

[0032] (2) CNN convolutional layer

[0033] The data matrix from the input layer is input into the CNN convolutional layer, and a convolutional kernel of i×j is used to perform a convolutional operation on the data matrix to extract local features of the input data. After the operation, a new data matrix is obtained, where the parameters i and j are both positive integers. The CNN convolution operation formula is as follows:

[0034]

[0035] In the formula, C is the data matrix output by the CNN convolutional layer; X is the input data matrix; W cnn is the weight matrix; b cnn is the bias; is the convolution operation; f is the ReLU activation function.

[0036] (3) CNN pooling layer

[0037] After receiving the data matrix output by the CNN convolutional layer, the CNN pooling layer performs a pooling operation using a pooling kernel of a×b, and then simplifies these data features to reduce the dimension, so as to reduce the computational complexity and avoid overfitting, where the parameters a and b are both positive integers. The CNN pooling layer operation formula is as follows:

[0038] P = maxpooling(C)

[0039] In the formula, P is the data matrix output by the CNN pooling layer; maxpooling(·) is the max pooling function.

[0040] (4) LSTM layer

[0041] After flattening the data matrix output by the CNN pooling layer, it is input into the LSTM layer. The LSTM layer can handle long-term dependencies in sequence data, thereby predicting the harmonic data matrix. The LSTM layer includes three gated units: a forget gate, an input gate, and an output gate, as well as a cell state. The prediction process of the LSTM long short-term memory neural network for the harmonic data matrix is as follows:

[0042] 1) The update formula of the LSTM forget gate is as follows:

[0043] f t = σ(W f ·[h t-1 , x t + b f )

[0044] In the formula, f t is the output value of the forget gate; σ is the sigmoid activation function; W f is the weight of the forget gate; h t-1 is the cell output at the previous moment; x t is the cell input at time t; b f is the bias parameter of the forget gate.

[0045] 2) The update formula of the LSTM input gate is as follows:

[0046] i t = σ(W i ·[h t-1 , x t + b i )

[0047] g t = tanh(W c ·[h t-1 , x t + b c )

[0048] In the formula, i t is the output value of the input gate; W i is the weight of the input gate; b i is the bias parameter of the input gate; tanh is the activation function; g t is the cell state of the current input; W c is the weight corresponding to the cell state of the current input; b c is the bias parameter corresponding to the current cell state.

[0049] 3) The formula for updating the cell state using the LSTM forget gate and input gate is as follows:

[0050] c t = f t · c t-1 + i t · g t

[0051] In the formula, c t is the updated cell state value; c t-1 is the cell state value at the previous moment.

[0052] 4) The update formula for the LSTM output gate is as follows:

[0053] o t = σ(W0 · [h t-1 , x t + b0)

[0054] h t = o t · tanh(c t )

[0055] In the formula: o t is the output value of the previous step; h t is the output value of the output gate at the current moment; W0 is the weight of the output gate; b0 is the bias parameter of the output gate.

[0056] (5) Fully connected layer and output layer

[0057] Send the harmonic data matrix output by the LSTM output gate to the fully connected layer, use the fully connected layer to map the harmonic data matrix to a space with the same dimension as the original harmonic data, and finally output the predicted value of the harmonic data through the output layer.

[0058] Compare the predicted value of the harmonic data output by the CNN-LSTM model with the true value of the harmonic data in the known dataset, and calculate the prediction error using the mean absolute percentage error (MAPE). Then, by continuously adjusting the parameters of the CNN convolutional layer, CNN pooling layer, and LSTM layer, when the prediction error of the model is less than the threshold, the model establishment / training is completed.

[0059] The present invention utilizes the advantages of CNN in feature extraction and LSTM in data prediction, and thus constructs a CNN-LSTM data missing filling model based on the correlation of multi-measurement point data, which has high accuracy and improves the integrity and reliability of PMU harmonic data in the power grid harmonic monitoring system.

[0060] Step 4, fill the missing dataset.

[0061] In the missing sample dataset, by using the data fields strongly correlated with the target harmonic data as the model input, use the established filling model for prediction, and use the prediction result as the filling value, so as to effectively fill the missing part of the PMU harmonic data.

[0062] The present invention proposes an embodiment of a PMU harmonic data missing filling device. The device includes a dataset construction module, a strong correlation screening module, a training module, and a data filling module.

[0063] The dataset construction module is configured to: when there is missing harmonic data at the target site, extract the PMU measurement data of the target site and one or more adjacent monitoring sites within the same sampling period, and perform dataset division: classify the PMU measurement data with complete harmonic data as the known dataset, and divide the PMU measurement data with missing harmonic data into the missing dataset to be filled.

[0064] The strong correlation screening module is configured to perform strong correlation screening on the known dataset: calculate the per-type Pearson correlation coefficient between the harmonic data of the target site and other PMU measurement data types of the site itself, and at the same time calculate the per-type Pearson correlation coefficient between the harmonic data of the target site and the PMU measurement data types of adjacent sites. If the correlation coefficient between the harmonic data of the target site and a certain data type is lower than the threshold, then eliminate this data type; if the correlation coefficient exceeds or is equal to the threshold, then retain this data type, and finally construct a strong correlation known dataset that only contains strongly correlated data types.

[0065] The training module is configured to use the data types strongly correlated with the harmonic data of the target site in the strong correlation known dataset as input features, and use the true values of the harmonic data of the target site as output labels to train the filling model.

[0066] The data filling module is configured to use the trained filling model to predict the missing harmonic data of the target site in the missing dataset, and use the prediction result as the filling value to complete the filling of the missing part of the PMU harmonic data.

[0067] Only the components of the device are outlined here. For the specific implementation methods, please refer to the description of the embodiments of the PMU harmonic data missing filling method associated therewith.

[0068] The following provides an application case: use the PMU measurement data of BH Station, a 500 kV substation, and 4 sites connected to it. Among them, AS Station and QY Station are 500 kV substations, DQ Station is a power plant, and XR Station is a converter station. Figure 3 The main wiring diagram of the regional system is shown.

[0069] Taking the case where the PMU has five missing harmonic voltage amplitude data during the period from 1 pm to 2 pm on a certain day on Bilibili (BH) station as an example, the PMU measurement data of BH station and the adjacent 4 measurement stations during this period are extracted. Each station in the dataset collects 3600 samples, and the types of PMU measurement data for each station include active power P, reactive power Q, fundamental voltage amplitude V1, fundamental current amplitude I1, fifth harmonic voltage amplitude V5, and fifth harmonic current amplitude I5. The selected original data does not contain missing values, and the missing values required in the algorithm are randomly deleted manually. The samples with missing fifth harmonic voltage amplitude data are set as the missing dataset, and the remaining samples are set as the known dataset.

[0070] In the known dataset, the Pearson correlation analysis method is used to calculate the correlation coefficients between the fifth harmonic voltage amplitude of BH station and the data types of other stations. The calculation results are presented in a heatmap, as Figure 4 shown. In order to retain the important correlation information in the input data, according to the Pearson correlation evaluation criteria, 16 data types with strong correlation (the absolute value of the correlation coefficient is greater than 0.6) are retained.

[0071] The known dataset selected by the multi-measurement point data correlation method is imported into the CNN-LSTM model. The other data types strongly correlated with the harmonic data are set as the inputs of the model, and the harmonic data is set as the output of the model. The CNN-LSTM model is used to learn the mapping relationship between the harmonic data and other data types. During the model establishment process, the model parameters are continuously optimized to make the prediction results output by the CNN-LSTM model closer to the true values. When the prediction error of the model is less than 0.03, the model establishment is completed.

[0072] The parameters of the CNN-LSTM model are as follows: the size of the CNN convolution kernel is 15×16, and the activation function is ReLu; the size of the CNN pooling kernel is 2×1, and the pooling method is max pooling; the number of LSTM layers is 1, and the number of neurons in each layer is 50.

[0073] In the missing sample set, the data types strongly correlated with the harmonic data are used as the inputs of the model. The data missing filling model based on multi-measurement point data correlation is used to predict the PMU harmonic missing data in the missing sample set, and the prediction results are used as the missing data filling values.

[0074] To verify the accuracy of the present invention, a traditional CNN-LSTM data filling model based on the correlation of single-measurement point data is used for comparison. Two error metrics, the mean absolute error (MAE) and the mean absolute percentage error (MAPE), are adopted as the criteria for evaluating the effectiveness of the model. Table 1 shows the comparison of the filling effects between the method of the present invention and the traditional filling method. It can be seen that the two error metrics of the model proposed in the present invention are smaller, indicating that the data filling method based on the correlation of multi-measurement point data has a higher filling accuracy.

[0075] Table 1 Comparison of the filling effects between the method of the present invention and the traditional filling method

[0076]

[0077]

[0078] The present invention also provides an embodiment of a computer. The computer includes a processor and a memory. The memory is used to store non-transitory computer-readable instructions (such as one or more computer program modules). The processor is used to run the non-transitory computer-readable instructions, and when the non-transitory computer-readable instructions are run by the processor, one or more steps in the above-mentioned PMU measurement harmonic data missing filling method can be executed. The memory and the processor can be interconnected through a bus system and / or other forms of connection mechanisms.

[0079] For example, the processor can be a central processing unit (CPU), a graphics processing unit (GPU), or other forms of processing units with data processing capabilities and / or program execution capabilities. For example, the central processing unit (CPU) can be of the X86 or ARM architecture, etc. The processor can be a general-purpose processor or a special-purpose processor, and can control other components in the computer to execute the desired functions.

[0080] For example, the memory can include any combination of one or more computer program products. The computer program products can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory can include, for example, random access memory (RAM) and / or cache memory, etc. Non-volatile memory can include, for example, read-only memory (ROM), hard disk, erasable programmable read-only memory (EPROM), compact disc read-only memory (CD-ROM), USB memory, flash memory, etc. One or more computer program modules can be stored on the computer-readable storage media, and the processor can run one or more computer program modules to implement various functions of the computer.

[0081] The present invention also provides a computer-readable storage medium for storing non-transitory computer-readable instructions, which can implement one or more steps in the above PMU measurement harmonic data missing filling method when executed by a computer. When the PMU measurement harmonic data missing filling method provided by the embodiments of the present invention is implemented in the form of software and sold or used as an independent product, it can be stored in a computer-readable storage medium. For the relevant description of the storage medium, reference can be made to the corresponding description of the memory in the computer system above, which will not be elaborated here.

Claims

1. A method for filling missing PMU harmonic data, characterized in that, Including: When there are missing harmonic data at the target site, extract the PMU measurement data of the target site and one or more adjacent monitoring sites within the same sampling period, and perform dataset partitioning: classify the PMU measurement data with complete harmonic data as the known dataset, and the PMU measurement data with missing harmonic data as the missing dataset to be filled; Perform strong correlation screening on the known dataset: calculate the per-type Pearson correlation coefficient between the harmonic data of the target site and other PMU measurement data types of the site itself, and at the same time calculate the per-type Pearson correlation coefficient between the harmonic data of the target site and the PMU measurement data types of adjacent sites. If the correlation coefficient between the harmonic data of the target site and a certain data type is lower than the threshold, then eliminate this data type; If the correlation coefficient exceeds or is equal to the threshold, then retain this data type, and finally construct a strong correlation known dataset that only contains strongly correlated data types; Use the data types strongly correlated with the harmonic data of the target site in the strong correlation known dataset as input features, and the true value of the harmonic data of the target site as the output label to train the filling model; Use the trained filling model to predict the missing harmonic data of the target site in the missing dataset, and use the prediction result as the filling value to complete the filling of the missing part of the PMU harmonic data.

2. The PMU harmonic data missing filling method according to claim 1, wherein The threshold of the correlation coefficient is greater than or equal to 0.

6.

3. The PMU harmonic data missing filling method according to claim 1, characterized in that The filling model adopts a cascaded architecture of CNN and LSTM.

4. The PMU harmonic data missing filling method according to claim 3, wherein The CNN extracts the spatial correlation of the input data in the feature dimension through one-dimensional convolution operations, and reduces the data dimension through the pooling layer; the LSTM layer receives the spatial features output by the CNN layer, captures the dynamic dependencies in the time dimension, and finally outputs the prediction result of the harmonic data of the target site.

5. A PMU harmonic data missing filling device, characterized in that Including: A dataset construction module, configured to: when there are missing harmonic data at the target site, extract the PMU measurement data of the target site and one or more adjacent monitoring sites within the same sampling period, and perform dataset partitioning: classify the PMU measurement data with complete harmonic data as the known dataset, and the PMU measurement data with missing harmonic data as the missing dataset to be filled; A strong correlation screening module, configured to perform strong correlation screening on the known dataset: calculate the per-type Pearson correlation coefficient between the harmonic data of the target site and other PMU measurement data types of the site itself, and at the same time calculate the per-type Pearson correlation coefficient between the harmonic data of the target site and the PMU measurement data types of adjacent sites. If the correlation coefficient between the harmonic data of the target site and a certain data type is lower than the threshold, then eliminate this data type; If the correlation coefficient exceeds or is equal to the threshold, then retain this data type, and finally construct a strong correlation known dataset that only contains strongly correlated data types; A training module, configured to use the data types strongly correlated with the harmonic data of the target site in the strong correlation known dataset as input features, and the true value of the harmonic data of the target site as the output label to train the filling model; A data filling module, configured to use a trained filling model to predict the missing harmonic data of the target site in the missing dataset, and use the prediction result as the filling value to complete the filling of the missing part of the PMU harmonic data.

6. The PMU harmonic data missing filling device according to claim 5, wherein The threshold of the correlation coefficient is greater than or equal to 0.

6.

7. The PMU harmonic data missing filling device according to claim 5, characterized in that, The filling model adopts a cascaded architecture of CNN and LSTM.

8. The PMU harmonic data missing filling device according to claim 5, wherein The CNN extracts the spatial correlation of the input data in the feature dimension through one-dimensional convolution operations, and reduces the data dimension through a pooling layer; the LSTM layer receives the spatial features output by the CNN layer, captures the dynamic dependence in the time dimension, and finally outputs the prediction result of the harmonic data of the target site.

9. A computer, characterized in that, Comprising: A processor; A memory, including one or more computer program modules; Wherein, the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules include instructions for implementing the PMU harmonic data missing filling method according to any one of claims 1-4.

10. A non-transitory computer-readable instruction storage, characterized in that, When the non-transitory computer-readable instructions are executed by a computer, the PMU harmonic data missing filling method according to any one of claims 1-4 can be implemented.

Images