Method and system for repairing abnormal data of power distribution network based on deep learning

Abstract

The invention provides a deep learning-based power distribution network abnormal data restoration method and system. The method comprises the steps of obtaining operation data of each load node in a target power distribution network; performing second sorting on all the time sequence data according to the number of the missing values to obtain a second sorting result, and constructing a label pair according to the second sorting result; inputting the label pairs into an improved NeuralProphet model for training and prediction, and obtaining prediction data of corresponding moments through a trend module, a season module, a festival module, a future regression module, a trend regression module and an autoregression module in the model; and correcting the error data based on a pre-trained LSTM model to obtain a prediction error, performing calculation according to the prediction error and the prediction data to obtain target data of each moment under each type, and filling missing data in each type of time sequence data according to the target data. The accuracy of the abnormal data restoration result of the power distribution network can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data repair, and in particular to a method and system for repairing abnormal data in a distribution network based on deep learning. Background Art

[0002] The theoretical line loss of the distribution network is an important comprehensive technical and economic index of power enterprises. Accurate theoretical line loss calculation requires accurate and comprehensive user power consumption data. Due to the influence of the accuracy of data acquisition equipment, interference in the data transmission process, and other factors, data anomalies or missing phenomena will inevitably occur in the source data, making it difficult to support the development of theoretical line loss calculation work.

[0003] In the prior art, mainly the machine learning method is adopted, a neural network is constructed, and a large amount of data is used for training to automatically learn the complex patterns in the data, and then the data is predicted and repaired.

[0004] However, although the machine learning method for repairing abnormal data in theoretical line loss calculation has the characteristics of strong adaptability, high flexibility, and self-learning, this method also has certain limitations, specifically as follows: First, the model itself has applicable conditions and limitations. There is a large amount of uncertain information in the theoretical line loss calculation data. There are some differences in reflecting data information among various single prediction models, and each has its own applicable conditions and limitations. For example, the traditional NeuralProphet model will inevitably lose some data information when predicting data sequences.

[0005] Second, when there are many random non-linear correlation factors, such as the production plans of large users, weather, temperature, humidity, electricity price adjustment, etc., the random factors have a significant impact on data prediction. The machine learning model cannot capture the non-linear correlation behind the random factors, resulting in low accuracy of the abnormal data repair results. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for repairing abnormal data in a distribution network based on deep learning, aiming to solve the problem of low accuracy existing in the traditional distribution network data repair method.

[0007] In the first aspect, the present invention provides a method for repairing abnormal data in a distribution network based on deep learning, and the method includes: Obtain the operation data of each load node in the target distribution network, where the operation data at least includes voltage, current, active power, reactive power, active power consumption, and reactive power consumption, and perform a first sorting on each type of operation data in chronological order to obtain time series data corresponding to each type of operation data; Obtain the number of missing values for each time series data, perform a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and construct label pairs according to the second sorting result; Input the label pairs into an improved NeuralProphet model for training and prediction, and respectively pass through the trend module, seasonal module, holiday module, future regression module, trend regression module, and autoregressive module in the model to obtain the predicted data at the corresponding moment; Obtain error data based on the real data and predicted data at the same moment, correct the error data based on a pre-trained LSTM model to obtain the prediction error, calculate the target data at each moment for each type according to the prediction error and the predicted data, and fill in the missing data in each time series data according to the target data.

[0008] Further, the steps of obtaining the number of missing values for each time series data, performing a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and constructing label pairs according to the second sorting result include: Assume the second sorting result is: ; Among them, , , , , , respectively represent the number of missing values of the first, second, third, fourth, fifth, and sixth time series data; Define the time series data corresponding to , , , , , as , , , , , , and pre-fill the missing values in each time series data according to the preset value to obtain the pre-filled time series data , , , , , ; Using the i-th pre-filled time series data as the label, except Construct label pairs using all pre-filled time series data except the above as features to predict the missing data of the i-th time series data.

[0009] Furthermore, the step of inputting the label pairs into the improved NeuralProphet model for training and prediction, and obtaining the predicted data at the corresponding moment through the trend module, seasonal module, holiday module, future regression module, trend regression module, and autoregressive module in the model respectively includes: Calculate the predicted data according to the following formula: ; Where, represents the predicted data at the t-th moment, represents the data at the t-th moment predicted by the trend module, represents the data at the t-th moment predicted by the seasonal module, represents the data at the t-th moment predicted by the holiday module, represents the data at the t-th moment predicted by the future regression module, represents the data at the t-th moment predicted by the trend regression module, represents the data at the t-th moment predicted by the autoregressive module.

[0010] Furthermore, obtain the data predicted by the trend module according to the following formula: ; Where, represents the growth rate, represents the time interval between any two adjacent data in the time series data, represents the data at the -th moment predicted by the trend module, the -th moment and the t-th moment are two adjacent moments; Calculate the data predicted by the seasonal module according to the following formula: ; Where, represents the cosine term coefficient, represents the sine term coefficient, represents the number of Fourier series, j is the summation index, with a value range of 1 to k, representing the j-th term in the series, and p represents the period length of the seasonal pattern; Calculate the data predicted by the holiday module according to the following formula: ; Where, represents the effect of a specific holiday or event at the t-th moment, e represents that each event is a binary variable, ; The data predicted by the future regression module is calculated according to the following formula: ; Wherein, represents the effect of a specific future regression quantity at the t-th moment, represents the selection of the addition or multiplication mode; The data predicted by the trend regression module is calculated according to the following formula: ; Wherein, , respectively represent the data at the -th and -th moments predicted by the trend regression module, represents the autoregressive feedforward neural network for time series, , , respectively represent the -th, -th, and -th historical covariates; The data predicted by the autoregressive module is calculated according to the following formula: ; Wherein, , , respectively represent the data at the t-th, -th, and -th moments predicted by the autoregressive module, , , respectively represent the -th, -th, and -th historical covariates.

[0011] Furthermore, the step of obtaining error data based on the real data and the predicted data at the same moment includes: The error data is calculated according to the following formula: ; Wherein, represents the error data at the t-th moment, represents the real data at the t-th moment.

[0012] Furthermore, the step of correcting the error data based on the pre-trained LSTM model to obtain the prediction error includes: The prediction error is obtained according to the following formula: ; Among them, represents the prediction error at the t-th moment, represents the activation function, represents the forgetting gate weight matrix, 、 represent the state information at the t-th and (t - 1)-th moments, represents the bias term, represents the output control quantity at the t-th moment, represents the hyperbolic tangent function, represents the candidate memory cell state, temporarily storing new information, represents the output gate weight matrix, represents the candidate memory cell weight matrix, represents the output gate bias, represents the candidate memory cell bias, and represent the memory cell state at the t-th moment and the memory cell state at the (t - 1)-th moment respectively, represents the forgetting gate that determines how much of the previous memory to retain, represents the input gate that determines how much new information to update.

[0013] Furthermore, the step of calculating the target data at each moment for each type according to the prediction error and prediction data, and filling the missing data in each time series data according to the target data includes: Calculating the target data according to the following formula: ; Among them, represents the target data at the t-th moment; Obtaining the data positions of each missing data in the i-th type of time series data according to the pre-filling result, and obtaining the corresponding moment serial numbers according to the data positions, so as to replace the pre-filled preset values with the corresponding target data according to the moment serial numbers.

[0014] In a second aspect, the present invention provides a distribution network abnormal data repair system based on deep learning, and the system includes: An operation data acquisition module, configured to acquire operation data of each load node in the target distribution network, where the operation data at least includes voltage, current, active power, reactive power, active power consumption, and reactive power consumption, and perform a first sorting on each type of operation data in chronological order to obtain time series data corresponding to each type of operation data; A sorting module, configured to obtain the number of missing values for each type of time series data, and perform a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and construct a label pair according to the second sorting result; A prediction module, configured to input the label pair into an improved NeuralProphet model for training and prediction, and respectively pass through a trend module, a season module, a festival module, a future regression module, a trend regression module, and an autoregressive module in the model to obtain predicted data at corresponding moments; A data filling module, configured to obtain error data according to the real data and the predicted data at the same moment, correct the error data based on a pre-trained LSTM model to obtain a prediction error, calculate the target data at each moment for each type according to the prediction error and the predicted data, and fill in the missing data in each type of time series data according to the target data.

[0015] In a third aspect, the present invention provides a storage medium storing one or more programs, which when executed by a processor implement the above-mentioned method for repairing abnormal data in a distribution network based on deep learning.

[0016] In a fourth aspect, the present invention provides an electronic device, where the electronic device includes a memory and a processor, and: The memory is used to store a computer program; The processor is configured to implement the above-mentioned method for repairing abnormal data in a distribution network based on deep learning when executing the computer program stored on the memory.

[0017] Compared with the prior art, the embodiments of the present invention have the following advantages: 1. By combining two models, NeuralProphet and LSTM, a high-precision abnormal data repair technical framework that combines decomposing a complex model and supplementing a random non-linear part is proposed. First, the NeuralProphet algorithm model is used to train data such as n-dimensional time series voltage, active power, and reactive power for theoretical line loss calculation to obtain predicted values and calculate a residual sequence. Then, an LSTM neural network model is established for the random non-linear part of the residual data for prediction correction. Finally, the repaired value is reconstructed.

[0018] 2. By designing an improved NeuralProphet-LSTM model, it can be used to process irregular data and effectively solve problems such as large errors in repair results and unstable prediction effects caused by abnormal data fluctuations. It can still make effective predictions in the face of external events such as holidays, meteorological anomalies, and electricity price adjustments. Compared with traditional neural networks such as the SVM model, NeuralProphet, and LSTM, the NeuralProphet-LSTM combined model constructed by the residual correction method has improved data repair accuracy and can more fully extract the complex information of the theoretical line loss calculation data of the distribution network. This model can not only decompose the composition components through the built-in additive model to intuitively reproduce the fluctuation characteristics of different time scales in the data, but also efficiently fit the abnormal fluctuations in the data and make effective predictions by itself, with more accurate predictions than traditional single neural network models. Description of the Drawings

[0019] Figure 1 It is a flowchart of a method for repairing abnormal data of a distribution network based on deep learning proposed in an embodiment of the present invention; Figure 2 It is a schematic structural diagram of a system for repairing abnormal data of a distribution network based on deep learning proposed in an embodiment of the present invention.

[0020] The following specific embodiments will further illustrate the present invention in conjunction with the above drawings. Specific Embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein shall have the ordinary meaning as understood by those of ordinary skill in the art in the field to which the present invention belongs. The words such as "including" used herein mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects.

[0022] As Figure 1 shown, an embodiment of the present invention provides a method for repairing abnormal data of a distribution network based on deep learning. The method includes steps S101 to S104, where: Step S101: Obtain the operation data of each load node in the target distribution network. The operation data includes at least voltage, current, active power, reactive power, active power consumption, and reactive power consumption, and perform a first sorting on each type of operation data in chronological order to obtain time series data corresponding to each type of operation data; It should be noted that the first sorting refers to sorting in chronological order, that is, each time series data includes data at multiple moments, and for any two adjacent moment data, the moment with the earlier sorting is earlier.

[0023] Step S102: Obtain the number of missing values of each time series data, and perform a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and construct a label pair according to the second sorting result; It should be noted that in this step, before obtaining the number of missing values of each time series data, it is also necessary to identify the positions of all missing data in each time series data. Specifically: perform abnormal data identification based on limit learning and multi-dimensional data outlier identification on the operation data such as voltage, active power, reactive power, and power consumption of each load node on a typical day for power flow calculation, and call function to return a for the position of all outliers, realizing the identification of the missing data position.

[0024] Furthermore, after identifying the positions of each missing data, the number of missing values of each time series data can be accurately obtained.

[0025] Specifically, assume the second sorting result is: ; Among them, , , , , , respectively represent the number of missing values of the first, second, third, fourth, fifth, and sixth types of time series data; Define the time series data corresponding to , , , , , as , , , , , respectively. Pre-fill the missing values in each time series data according to the preset value to obtain the pre-filled time series data , , , , , ; Using the pre-filled time series data of the i-th type as the label, and using all the pre-filled time series data except as features to construct a label pair for predicting the missing data of the i-th type of time series data.

[0026] By way of example and not limitation, to reduce the impact of missing data on data repair, the operation of filling null values is first performed, that is, pre-filling. Assuming the corresponding data is , the representation is as follows:

[0027] where represents the normal value, represents , that is, there is data missing or abnormal at this moment. The current data at time T is pre-filled with zero values to obtain , and the representation is as follows:

[0028] Respectively assume , , , , the corresponding data are respectively , , , , , and similarly, the pre-filling of , , , , data can be performed, which will not be elaborated here.

[0029] Taking , , , , 's data as features, 's data as the label, to construct a label pair, specifically as follows: .

[0030] Step S103: Input the tag pairs into the improved NeuralProphet model for training and prediction. Through the trend module, seasonal module, holiday module, future regression module, trend regression module, and autoregressive module in the model respectively, the predicted data at the corresponding time is obtained; It should be noted that when is predicted, in this step, the improved NeuralProphet model is used to for training and prediction. Since , , , , the missing abnormal data has been pre-filled, the data of and the normal data in can be directly used. Through fitting the trend module, seasonal module, holiday module, future regression module, trend regression module, and autoregressive module for prediction. The trend module considers parameters such as trend change points, detects the moment of trend change in the time series, breaks through the traditional linear trend assumption, enables the model to more flexibly adapt to the dynamic changes of the data, and has a more accurate grasp of the long-term trend; the seasonal module considers parameters such as cycle length and seasonal patterns, breaks through the fixed seasonal pattern, and can more flexibly and accurately capture the seasonal laws of the data; the holiday module models the impact of specific holidays or special events on the time series data, makes up for the deficiency of the traditional model in considering special events, and makes the prediction results closer to the actual situation; the future regression module incorporates external future known information into the prediction model, considers parameters such as external regression factors and their coefficients, innovatively expands the input dimension of the model, and makes full use of the value of external information; the trend regression module combines trend information and regression analysis, considers parameters such as trend regression factors and their weights, can better explain the reasons for trend changes, and improves the fitting ability and prediction accuracy of the trend; the autoregressive module utilizes the autocorrelation of the time series data, considers parameters such as the lag order of the data, and proposes a method to capture the inherent law of the time series itself, realizing the full utilization of the value of the data itself. The multi-module fitting prediction synthesizes various characteristics and laws such as long-term trend, periodic change, external information, special events, and autocorrelation. Each module complements and cooperates with each other, enabling the prediction model to adapt to various complex time series data, and can greatly improve the accuracy and reliability of the prediction. After model fitting and parameter tuning, the corresponding predicted data

[0031] is obtained. Specifically, the predicted data is calculated according to the following formula: ; where represents the predicted data at the t-th moment, represents the data at the t-th moment predicted by the trend module, Data at the t-th moment predicted by the season module Data at the t-th moment predicted by the festival module Data at the t-th moment predicted by the future regression module Data at the t-th moment predicted by the trend regression module Data at the t-th moment predicted by the autoregressive module

[0032] In addition, the data predicted by the trend module is obtained according to the following formula: ; where represents the growth rate represents the time interval between any two adjacent data in the time series data represents the data at the -th moment predicted by the trend module. The -th moment and the t-th moment are two adjacent moments; The data predicted by the season module is calculated according to the following formula: ; where represents the cosine term coefficient represents the sine term coefficient represents the number of Fourier series. j is the summation index, ranging from 1 to k, representing the j-th term in the series, and p represents the period length of the seasonal pattern; The data predicted by the festival module is calculated according to the following formula: ; where represents the effect of a specific festival or event at the t-th moment. e represents that each event is a binary variable ; The data predicted by the future regression module is calculated according to the following formula: ; where represents the effect of a specific future regressor at the t-th moment represents the selection of additive or multiplicative mode; The data predicted by the trend regression module is calculated according to the following formula: ; where and respectively represent the data at the -th and -th moments predicted by the trend regression module Denote the autoregressive feedforward neural network for time series, which takes \(p\) historical covariates as input and, after passing through the network, obtains the output at \(h\) steps, which helps to improve the fitting speed and accuracy. , , respectively denote the -th, -th, -th historical covariates; The data predicted by the autoregressive module is calculated according to the following formula: ; where , , respectively denote the data at the \(t\)-th, -th, -th time points predicted by the autoregressive module. , , respectively denote the -th, -th, -th historical covariates. The trend regressor module has the same function as the autoregressive module, which takes \(p\) historical covariates as input and, after passing through the network, obtains the output at \(h\) steps. The difference is that the last observed value of the covariates is used as the input of the module.

[0033] Step S104: Obtain the error data according to the real data and the predicted data at the same time point, and correct the error data based on the pre-trained LSTM model to obtain the prediction error. Calculate the target data at each time point for each type according to the prediction error and the predicted data, and fill in the missing data in each time series data according to the target data.

[0034] In this step, first, calculate the error data according to the following formula: ; where denotes the error data at time \(t\), denotes the real data at time \(t\).

[0035] In addition, in addition to the data relationships that the NeuralProphet model can capture, the 24-point data of a single load node also contains a random non-linear part, such as meteorological events, electricity price adjustment events, etc. The long short-term memory network (LSTM) model is used to characterize the non-linear relationship therein. Use the LSTM model to predict the error and optimize the parameters, correct the original error value, and obtain a new prediction error , and the calculation process of LSTM is as follows: The prediction error is calculated according to the following formula: ; where represents the prediction error at the t-th moment, represents the activation function, represents the forget gate weight matrix, 、 represent the state information at the t-th and (t - 1)-th moments, represents the bias term, represents the output control quantity at the t-th moment, represents the hyperbolic tangent function, represents the candidate memory cell state, temporarily storing new information, represents the output gate weight matrix, represents the candidate memory cell weight matrix, represents the output gate bias, represents the candidate memory cell bias, and represent the memory cell state at the t-th moment and the memory cell state at the (t - 1)-th moment respectively, represents the forget gate that determines how much of the previous memory to retain, represents the input gate that determines how much new information to update.

[0036] By inputting the state information at the previous moment and the feature information at the current moment , connecting the two, multiplying by the forget gate weight matrix , adding the bias term , and passing it into the sigmoid activation function. This value determines whether the neuron state value at the previous moment is to be remembered or forgotten. Also has a forgetting function and represents the value to be updated. Through the activation function, a new candidate state vector is created. and will determine which information is important and needs to be retained. Finally, the information to be forgotten is added to the new part to be remembered, which is the neuron state to be input to the next node .

[0037] In addition, the prediction result of the NeuralProphet model and the error prediction result are combined, and finally the prediction result of the NeuralProphet-LSTM combined model is obtained, that is .

[0038] Repeat the above steps to obtain the model predicted voltage data at time t-1 and the model predicted voltage data at time t-2 and so on, and complete the filling of all missing and abnormal dynamic 24-point voltage data for theoretical line loss calculation by analogy.

[0039] Similarly, data filling can be performed for current, active power, reactive power, active energy, reactive energy, etc. Finally, the repaired results of abnormal data are output in a data format that can be recognized by the distribution network power flow calculation algorithm.

[0040] In summary, the embodiments of the present invention have the following advantages: 1. By combining two models, NeuralProphet and LSTM, a high-precision abnormal data repair technology framework that combines the decomposition of complex models and the supplementation of random non-linear parts is proposed. First, the NeuralProphet algorithm model is used to train data such as n-dimensional time series voltage, active power, and reactive power for theoretical line loss calculation to obtain predicted values and calculate the residual sequence. Then, an LSTM neural network model is established for the residual data of the random non-linear part for prediction correction, and finally, the repaired value is reconstructed.

[0041] 2. By designing an improved NeuralProphet-LSTM model, it can be used to process irregular data, and can effectively solve problems such as large errors in repair results and unstable prediction effects caused by abnormal data fluctuations. It can still make effective predictions in the face of external event impacts such as holidays, meteorological anomalies, and electricity price adjustments. Compared with traditional neural networks such as the SVM model, NeuralProphet, and LSTM, the NeuralProphet-LSTM combined model constructed by the residual correction method has improved data repair accuracy and can better extract the complex information of the distribution network theoretical line loss calculation data. This model can not only decompose the component composition through the built-in additive model to intuitively reproduce the fluctuation characteristics of different time scales in the data, but also can efficiently fit the abnormal fluctuations in the data and make effective predictions by itself, and the prediction is more accurate than that of traditional single neural network models.

[0042] As Figure 2 shown, an embodiment of the present invention also provides a distribution network abnormal data repair system based on deep learning. The system includes: An operation data acquisition module 10, configured to acquire operation data of each load node in the target distribution network. The operation data at least includes voltage, current, active power, reactive power, active energy, and reactive energy, and perform a first sorting on each type of operation data in chronological order to obtain time series data corresponding to each type of operation data; A sorting module 20, configured to obtain the number of missing values of each type of time series data, and perform a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and construct label pairs according to the second sorting result; A prediction module 30, configured to input the label pairs into an improved NeuralProphet model for training and prediction, and respectively pass through a trend module, a season module, a festival module, a future regression module, a trend regression module, and an autoregressive module in the model to obtain predicted data at corresponding moments; A data filling module 40, configured to obtain error data according to the real data and the predicted data at the same moment, correct the error data based on a pre-trained LSTM model to obtain a prediction error, calculate target data at each moment under each type according to the prediction error and the predicted data, and fill in the missing data in each type of time series data according to the target data.

[0043] On the other hand, the present invention also proposes a storage medium, on which one or more programs are stored, and when the program is executed by a processor, the above-mentioned method for repairing abnormal data in a distribution network based on deep learning is implemented.

[0044] On the other hand, the present invention also proposes an electronic device, including a memory and a processor, where the memory is used to store a computer program, and the processor is used to execute the computer program stored on the memory to implement the above-mentioned method for repairing abnormal data in a distribution network based on deep learning.

[0045] Those skilled in the art can understand that the logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0046] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing it in a suitable manner if necessary, and then storing it in a computer memory.

[0047] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, the multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and the like.

[0048] Although the embodiments of the present invention have been described in detail above, it is obvious to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as described in the claims. Moreover, the present invention described herein can have other embodiments and can be implemented or realized in various ways.

Claims

1. A method for repairing abnormal data in a distribution network based on deep learning, characterized in that, The method includes: Obtain the operation data of each load node in the target distribution network. The operation data includes at least voltage, current, active power, reactive power, active energy, and reactive energy, and perform a first sorting on each type of operation data in chronological order to obtain time series data corresponding to each type of operation data; Obtain the number of missing values of each time series data, perform a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and construct label pairs according to the second sorting result; Input the label pairs into an improved NeuralProphet model for training and prediction, and pass through the trend module, seasonal module, festival module, future regression module, trend regression module, and autoregressive module in the model respectively to obtain the prediction data at the corresponding moment; Obtain error data based on the real data and prediction data at the same moment, correct the error data based on a pre-trained LSTM model to obtain a prediction error, calculate the target data at each moment for each type according to the prediction error and prediction data, and fill in the missing data in each time series data according to the target data.

2. The method for repairing abnormal data of a distribution network based on deep learning according to claim 1, wherein The step of obtaining the number of missing values of each time series data, performing a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and constructing label pairs according to the second sorting result includes: Assume the second sorting result is: ; Among them, , , , , , respectively represent the number of missing values of the first, second, third, fourth, fifth, and sixth time series data; Definition and , , , , , The corresponding time series data are respectively , , , , , . The missing values in each time series data are respectively pre-filled according to the preset value to obtain the pre-filled time series data , , , , , ; Using the pre-filled time series data of the i-th type as the label, and using all the pre-filled time series data except as the features to construct label pairs, and predicting the missing data of the i-th type of time series data.

3. The method for repairing abnormal data of a distribution network based on deep learning according to claim 1, wherein, The step of inputting the label pairs into an improved NeuralProphet model for training and prediction, and passing through the trend module, seasonal module, festival module, future regression module, trend regression module, and autoregressive module in the model respectively to obtain the prediction data at the corresponding moment includes: Calculate the prediction data according to the following formula: ； Among them, represents the predicted data at the t-th moment, represents the data at the t-th moment predicted by the trend module, represents the data at the t-th moment predicted by the seasonal module, represents the data at the t-th moment predicted by the festival module, represents the data at the t-th moment predicted by the future regression module, represents the data at the t-th moment predicted by the trend regression module, represents the data at the t-th moment predicted by the autoregressive module.

4. The method for repairing abnormal data of a distribution network based on deep learning according to claim 3, wherein, Obtain the data predicted by the trend module according to the following formula: ； Among them, represents the growth rate, represents the time interval between any two adjacent data in the time series data, represents the data at the th moment predicted by the trend module. The th moment and the tth moment are two adjacent moments; Calculate the data predicted by the seasonal module according to the following formula: ； Among them, represents the cosine term coefficient, represents the sine term coefficient, represents the number of Fourier series, j is the summation index, with a value range from 1 to k, representing the j-th term in the series, and p represents the period length of the seasonal pattern; Calculate the data predicted by the festival module according to the following formula: ； Among them, represents the effect of a specific festival or event at the t-th moment, where e indicates that each event is a binary variable, ; Calculate the data predicted by the future regression module according to the following formula: ； Among them, represents the effect of a specific future regressor at the t-th moment, represents the selection of the additive or multiplicative mode; Calculate the data predicted by the trend regression module according to the following formula: ； Among them, and respectively represent the data at the -th and -th moments predicted by the trend regression module. represents the autoregressive feedforward neural network for time series. and and respectively represent the -th, -th, and -th historical covariates. Calculate the data predicted by the autoregressive module according to the following formula: ； Among them, , , respectively represent the data at the t-th, -th, and -th moments predicted by the autoregressive module. , , respectively represent the -th, -th, and -th historical covariates.

5. The method for repairing abnormal data of a distribution network based on deep learning according to claim 3, characterized in that The step of obtaining error data based on the real data and prediction data at the same moment includes: Calculate the error data according to the following formula: ； Among them, represents the error data at time t, represents the true data at time t.

6. The method for repairing abnormal data of a distribution network based on deep learning according to claim 5, wherein, The step of correcting the error data based on a pre-trained LSTM model to obtain a prediction error includes: Obtain the prediction error according to the following formula: ； Among them, represents the prediction error at the t-th moment, represents the activation function, represents the forgetting gate weight matrix, 、 represent the state information at the t-th and (t - 1)-th moments, represents the bias term, represents the output control quantity at the t-th moment, represents the hyperbolic tangent function, represents the candidate memory cell state, temporarily storing new information, represents the output gate weight matrix, represents the candidate memory cell weight matrix, represents the output gate bias, represents the candidate memory cell bias, and represent the memory cell state at the t-th moment and the memory cell state at the (t - 1)-th moment respectively, represents the forgetting gate that determines how much of the previous memory to retain, represents the input gate that determines how much new information to update.

7. The method for repairing abnormal data of a distribution network based on deep learning according to claim 6, wherein The step of calculating the target data at each moment for each type according to the prediction error and prediction data, and filling in the missing data in each time series data according to the target data includes: Calculate the target data according to the following formula: ； Among them, represents the target data at the t-th moment; Obtain the data position of each missing data in the i-th time series data according to the pre-filling result, and obtain the corresponding moment serial number according to the data position, so as to replace the pre-filled preset value with the corresponding target data according to the moment serial number.

8. A deep learning-based abnormal data repair system for a distribution network, characterized in that, The system includes: An operating data acquisition module, configured to acquire the operating data of each load node in a target distribution network, where the operating data at least includes voltage, current, active power, reactive power, active power consumption, and reactive power consumption, and perform a first sorting on each type of operating data in chronological order to obtain time series data corresponding to each type of operating data; A sorting module, configured to obtain the number of missing values of each time series data, and perform a second sorting on all the time series data according to the number of missing values to obtain a second sorting result, and construct a label pair according to the second sorting result; A prediction module, configured to input the label pair into an improved NeuralProphet model for training and prediction, and respectively pass through a trend module, a season module, a festival module, a future regression module, a trend regression module, and an autoregressive module in the model to obtain prediction data at corresponding moments; A data filling module, configured to obtain error data according to the real data and the prediction data at the same moment, correct the error data based on a pre-trained LSTM model to obtain a prediction error, calculate the target data at each moment under each type according to the prediction error and the prediction data, and fill in the missing data in each time series data according to the target data.

9. A storage medium, characterized in that, The storage medium stores one or more programs, and when the program is executed by a processor, it implements the deep learning-based distribution network abnormal data repair method according to any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: The memory is used for storing a computer program; When the processor is used to execute the computer program stored on the memory, it implements the deep learning-based distribution network abnormal data repair method according to any one of claims 1-7.

Images