Ammeter fault prediction method, system and equipment based on deep learning

Abstract

The invention relates to an electricity meter fault prediction method, system and device based on deep learning. The electric meter fault prediction method based on deep learning comprises the following steps: periodically collecting various data of an electric meter; preprocessing the collected data, and then performing clustering processing to obtain a plurality of clusters; the data in the multiple clusters are input into the deep learning model, each different cluster corresponds to different gate structure weights of the deep learning model, and finally a result value with time sequence features corresponding to the data is obtained. According to the invention, the data collected by the electricity meter is input into the deep learning network for fault prediction, features and modes can be automatically learned from the data, and prediction and classification are carried out. In the field of electric meter fault prediction, deep learning can learn the power utilization habit of a user and the operation mode of an electric meter by analyzing historical data of the electric meter so as to predict possible faults of the electric meter.

Description

Technical Field

[0001] The present invention relates to the technical field of electric meter fault prediction, and in particular to an electric meter fault prediction method, system, device and storage medium based on deep learning. Background Art

[0002] Electricity meters are important devices for measuring electric energy, and their normal operation is crucial to the stability of the power system and the safety of electricity use by users. However, during long-term use, electric meters may have various faults, such as inaccurate electricity measurement, abnormal current, and operating status word errors. These faults not only affect the user's electricity experience, but may also lead to instability and safety hazards in the power system.

[0003] Traditional meter fault detection methods mainly rely on manual inspections and regular maintenance, which is inefficient, costly, and unable to detect faults in a timely manner. With the development of smart grids, meters are becoming more and more intelligent, and data collection and transmission technologies have also been widely used. Therefore, using the data collected by meters to predict faults has become a new solution. Summary of the invention

[0004] Based on this, it is necessary to provide an electric meter fault prediction method, system, device and storage medium based on deep learning to address the problem that the existing methods for detecting electric meter faults are inefficient and untimely.

[0005] A method for predicting electric meter faults based on deep learning, comprising the following steps:

[0006] Periodically collect various data of electric meters;

[0007] After preprocessing, the collected data are clustered to obtain multiple clusters. The data in the multiple clusters are input into the deep learning model. Each different cluster corresponds to a different gate structure weight of the deep learning model through similarity, and finally a result value with time series characteristics corresponding to the data is obtained.

[0008] Compare the result value with the set threshold interval. If the result value is outside the threshold interval, the result value is defined as an abnormal point. Within a unit time, if an abnormal point appears, the counter is incremented by 1, and if a normal result value appears, the counter is decremented by 1. When the counter reaches the counting threshold, the meter is judged to be in a fault state, where the counting threshold is a positive integer. And, in all the collection cycles of the day, when the number of abnormal points exceeds 20% of the total number of result values, the meter is judged to be in a fault state.

[0009] The deep learning model is trained before use, and the training method includes the following steps:

[0010] The training set of electricity meter data is clustered to obtain N clusters, each of which contains i electricity meters;

[0011] Select the historical data of one of the cluster centers p to train the existing model to obtain the initial model m1;

[0012] Calculate the electric meter q in the cluster that has the highest similarity with the data feature of the center point p;

[0013] Input the historical data of meter q into the initial model m1 for training to obtain the next model m2; find meter a with the highest similarity to meter q data features, input the historical data of meter a into model m2 for training to obtain the next model m3; repeat the model training and search process until model m3 is obtained. i+1 ; N clusters get N models m i+1 ; N models m i+1 A deep learning model is constructed and the entire deep learning model training is completed.

[0014] As a preferred example, the various data of the electric meter include voltage, current, power consumption, and operating status.

[0015] As a preferred example, the preprocessing includes data cleaning processing, normalization processing and feature extraction processing.

[0016] As a preferred example, the clustering process adopts K-means algorithm; the number of clusters adopts elbow rule to select the optimal number of clusters.

[0017] As a preferred example, the method for setting the threshold interval includes the following steps:

[0018] Calculate the residual of the test set during model training. The calculation formula is:

[0019]

[0020] Where, E represents the daily residual of the test set; represents the daily electricity consumption forecast; Indicates the daily reading of the electric meter;

[0021] The K-Sigma algorithm is used to calculate the minimum K value that can satisfy the residual E predicted by the test set; after determining the K value, the threshold interval is calculated:

[0022] T u =T mean +(K×I range )

[0023] T L =T mean -(K×l range )

[0024] In the formula, is the upper limit of the threshold interval, T L is the lower limit of the threshold interval, T mean is the trimmed mean of the data, I range is the interquartile range of the data.

[0025] As a preferred example, in each cluster, the historical data used for model training belongs to set T, and the unused historical data belongs to set M; when set M is an empty set during model training, the entire deep learning model training is completed.

[0026] As a preferred example, when it is determined that the electric meter is in a fault state, an early warning mechanism is immediately triggered; the early warning mechanism includes sound reminders and sending SMS dispatch orders.

[0027] A meter fault prediction system based on deep learning, which uses the meter fault prediction method based on deep learning as described above; the meter fault prediction system comprises:

[0028] A data acquisition module, which is used to periodically collect various data of the electric meter;

[0029] The data preprocessing module is used to preprocess various types of collected data to standardize the data;

[0030] A training module, which is used to create and train deep learning models;

[0031] A fault prediction module, which is used to predict the fault condition of the electric meter based on the output results of the deep learning model;

[0032] The early warning and processing module is used to issue fault early warning and perform monitoring and processing according to the results of the electric meter fault prediction.

[0033] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the electric meter fault prediction method based on deep learning as described above are implemented.

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the electric meter fault prediction method based on deep learning as described above.

[0035] The beneficial effects of the present invention are:

[0036] 1. The present invention uses the data collected by the electric meter to input into the deep learning network for fault prediction, which can automatically learn features and patterns from the data, and make predictions and classifications. In the field of electric meter fault prediction, deep learning can analyze the historical data of the electric meter, learn the user's electricity usage habits and the operation mode of the electric meter, and thus predict the possible faults of the electric meter. Compared with the existing manual inspection method, the fault detection efficiency is greatly improved, and the fault is discovered in a timely manner.

[0037] 2. The present invention can monitor the operating status of the electric meter in real time, detect faults in time and issue early warnings, thereby improving the stability, safety and preventiveness of the power system. Moreover, by analyzing the user's electricity usage habits, it can provide users with more personalized electricity usage services and improve users' understanding and satisfaction with electricity usage.

[0038] 3. When there are a large number of users, it is very labor-intensive to establish the meter operation evaluation model one by one. The present invention uses transfer learning to transfer model parameters within the cluster to train the deep learning model, and migrates data between more similar meters, so that the prediction accuracy of the deep learning model is higher. In addition to accuracy, migrating the model between higher similarities will further reduce the deep learning time. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flowchart of an electric meter fault prediction method based on deep learning in an embodiment;

[0040] Figure 2 It is a training flow chart of deep learning model transfer learning in the embodiment;

[0041] Figure 3 This is a result diagram of the electric meter fault prediction for all normal data of electric power consumption in the embodiment;

[0042] Figure 4 This is a result diagram of the electric meter fault prediction after injecting fault data into the power consumption in the embodiment. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "or / and" used herein includes any and all combinations of one or more of the related listed items.

[0045] Please refer to Figure 1 , this embodiment provides an electric meter fault prediction method based on deep learning, which includes the following steps:

[0046] Periodically collect various data of the electric meter. Specifically, sensors can be used to collect and store electric meter data, and periodically collect the usage habits of the electric meter. The various types of data collected include important habitual periodic information such as the voltage, current, power consumption, and operating status of the electric meter. The collected data needs to be preprocessed. The preprocessing includes data cleaning, normalization, and feature extraction. Among them, data cleaning can be used to remove incomplete, inconsistent dirty data and noise that may exist in the data collected by the sensor. Dirty data and noise will affect the validity of the data used in the subsequent processing process. Normalization is used to standardize the data to the same scale, which helps to improve the prediction accuracy of the algorithm or model. The above processing can all use some existing algorithms or models. For example, data cleaning can use mean filling method, hot card filling method, data mining algorithm, etc. Normalization can use Min-Max standard normalization method for data sorting. Feature extraction can use normalization and statistical features (such as standard deviation). The main purpose of feature extraction is to extract key features that can reflect the operating status of the electric meter from the data collected by the electric meter. These key features will be passed as input to subsequent deep learning for fault prediction.

[0047] After preprocessing, various types of data are clustered to obtain multiple clusters. The data in multiple clusters are input into the deep learning model. Each different cluster corresponds to a different gate structure weight of the deep learning model through similarity, and finally obtains a result value with time series characteristics corresponding to the data. Among them, the clustering process can use the K-means algorithm and use the elbow rule to select the optimal number of clusters.

[0048] Compare the result value with the set threshold interval. If the result value is outside the threshold interval, the result value is defined as an abnormal point. Due to the volatility of power system operation, it is easy to produce false alarms based on a single abnormal point. Therefore, a method of using a counter is specially added to improve the data identification results. The use logic of the counter is: within a unit time, if an abnormal point appears, the counter is increased by 1, and if a normal result value appears, the counter is reduced by 1. When the counter reaches the counting threshold, the meter is judged to be in a fault state, where the counting threshold is a positive integer. In addition to statistical local anomalies, monitoring rules are set for the entire data to be evaluated for global monitoring: in all collection cycles of the day, when the number of abnormal points exceeds 20% of the total number of result values, the meter is judged to be in a fault state.

[0049] The above-mentioned deep learning model is trained before use, and the training of the deep learning model is one of the key points of the present invention. In this embodiment, the K-means algorithm in clustering is used as the basis for clustering to improve the similarity of user data patterns, and all values ​​in the acquired meter data are normalized to the same range. By using a recurrent neural network as the basis of the learning model, a gated recurrent unit model is used to simplify the long short-term memory network model, while maintaining functions similar to the long short-term memory network model and having a long-term memory function, the problem of gradient disappearance is solved. A chain transfer learning algorithm is used to simulate the habitual data of similar users using the meter, and a deep learning model is established for the user. Specifically, Figure 2 As shown, the method for deep learning model training includes the following steps:

[0050] Take the user's electricity consumption data as the object. Assuming that the electricity consumption curves of various users have some similarities, then their trained models also have similarities. If the initial prediction model is used as the starting point to build a model for the most similar meter, the training time of similar meter electricity consumption data will be greatly reduced.

[0051] The electricity consumption data of the electric meter is used as a training set, and its electricity consumption curve is clustered (this method can be used for the collected voltage, current and other data), and N clusters are obtained, each with i electric meters. Clustering is used to ensure that the electricity consumption patterns and electricity consumption behavior characteristics between each cluster are highly different. Among them, the clustering process can use the K-means algorithm and use the elbow rule to select the optimal number of clusters. After clustering, the electricity consumption characteristics within each cluster are highly similar, and the efficiency of transfer learning will also be improved.

[0052] The deep learning model architecture uses a recurrent neural network (RNN) model structure. The RNN contains internal states that can store information in sequential time steps, which makes it very suitable for time series prediction tasks. The gated recurrent unit (GRU) is a variant of the recurrent neural network (RNN) model commonly used in deep learning. It has good performance when processing sequence data, and compared with the traditional RNN model, the GRU can better capture long-term dependencies and solve the problem of gradient disappearance. The power consumption data of the center point p of one of the clusters is selected to train the GRU to obtain the initial model m1. Then the meter q in the cluster with the highest similarity in power consumption data features with the center point p is calculated, that is, the power consumption curves between the two are the most similar.

[0053] Input the electricity consumption data of meter q into the initial model m1 for training, and obtain the next model m2. Find the meter a with the highest similarity to the electricity consumption data of meter q, and input the electricity consumption data of meter a into model m2 for training, and obtain the next model m3. During the migration process, the structure and parameters of the source model remain unchanged. The weights of each gate structure only change under the training of the electricity consumption data of the migration target. Repeat the model training and search process until the model m3 is obtained. i+1 At the same time, in each cluster, the electricity consumption data used in model training belongs to set T, and the unused electricity consumption data belongs to set M. When set M is empty during model training, the entire deep learning model training is completed. Finally, N clusters get N models m i+1 ; N models m i+1 Constitute a complete deep learning model.

[0054] This method greatly reduces the training time of similar data through the strategy of model transfer learning. The above training method is proposed to optimize the route of transfer learning. In general, the determination of the migration direction of model parameters starts with similarity calculation. Calculate the similarity between each pair of data in the cluster. The focus of clustering is the similarity of data features, not the actual value of each meter data. Therefore, all values ​​in the meter data set are normalized to bring them to the same range. Min-Max normalization is performed on the meter data of all users, and the normalized data is used for similarity calculation. For i meters in the cluster, the result of the similarity calculation is a k×k upward sloping matrix. This similarity matrix is ​​denoted as S k×k The lower the relative distance value between two meters, the higher the similarity of the meters. The similarity calculation considers three different metrics, including Euclidean distance, cosine distance, and Manhattan distance.

[0055] In the process of model transfer learning, the sliding window is used to roll forward for the training set sequence. Take a piece of data as input, and then try to predict one or more values ​​in front of the window. Each predicted value is subtracted from the actual measured value to get the training loss. After a complete training set is selected, the feedback loss is used to optimize the weights and adjust the parameters by the optimizer. When the deep learning model reaches a fixed number of iterations or the mean square error (MSE) reaches the specified value, the model training ends and the model is established. The mean square error is calculated by squaring the difference between the predicted value and the actual measured value in each training set, and then taking the average of all the squared differences as the error measure. The smaller the mean square error value, the smaller the difference between the predicted value and the actual measured value, and the better the performance of the deep learning model.

[0056] In addition, for the setting of threshold intervals in the prediction of electric meter faults. In this embodiment, it is determined by the following method: taking the electricity consumption in the electric meter data as an example, the electric meter data collected by default has no error data, and 80% of these error-free electric meter data are divided into training sets and 20% into test sets. After the deep learning model is established with the training set, the electricity consumption is predicted for the test set. It is assumed that the electric meter is in good operating condition during the establishment of the deep learning model, so the residual of the corresponding test set is within the normal range, and its calculation formula is:

[0057]

[0058] Where, E represents the daily residual of the test set; represents the daily electricity consumption forecast; Represents the daily reading of the electric meter. The upper and lower limits of the error between the predicted residual of the test set and the actual residual, that is, the threshold interval, are determined by K-Sigma. The K value needs to be determined by the test set. The determination of the K value requires the minimum K value that can meet the residual E predicted by the test set. The setting of the threshold interval is the change of the time series characteristics before and after the comparison, and its calculation formula is:

[0059]

[0060] T L =T mean -(K×I range )

[0061] In the formula, is the upper limit of the threshold interval, T L is the lower limit of the threshold interval, T mean is the trimmed mean of the data, I range is the interquartile range of the data.

[0062] In another embodiment, the process is explained by taking the electricity consumption in the data collected by the electric meter as an example. In the process of collecting the electric meter data, the data is collected once at intervals of 15 minutes, and a total of 96 points are collected in 24 hours a day. The daily electricity consumption of the meter user is obtained by subtracting the electric energy indication. The electricity consumption data of 40 users for 400 consecutive days are selected as experimental sample data, and these data are thoroughly checked manually, all indicating that the corresponding electric meters are in normal operation. Chain migration is performed based on the deep learning model established in the above embodiment, and one of the users is selected to check whether the migrated model is in line with the user's electricity usage habits. After the deep learning model is chain-migrated to the user, the parameters and base values ​​of the deep learning model are adjusted at the same time, so that the deep learning model is more in line with the user's electricity usage habits and the operating status of the electric meter. Predictive analysis is performed on the electricity consumption collection, and the analysis results are as follows: Figure 3 As shown. Figure 3 It can be seen that the predicted residual of electricity consumption is always between the upper and lower limits of the residual threshold (that is, within the threshold interval), indicating that the clustered chain migration deep learning model can well predict the operating status of the user's meter after adjusting the parameters and base values. On the basis of the above, artificially modified faults are injected (that is, a certain offset is embedded in a random number of meter readings). According to the technical specifications of smart meters, the offset of an inaccurate meter refers to the offset of the reading (compared with the actual value), and the meter is judged to be abnormal if the relative error exceeds 2%. After fault injection, the deep learning model predicts the meter fault structure as follows Figure 4 As shown by Figure 4 It can be seen that the predicted residual of electricity consumption is outside the upper and lower limits of the residual threshold and corresponds to the injected fault data, indicating that the deep learning model can accurately predict the meter fault after the fault occurs.

[0063] In another embodiment, a meter fault prediction system based on deep learning is also proposed. It includes a data acquisition module, a data preprocessing module, a training module, a fault prediction module, an early warning and processing module, a module evaluation module and a data visualization module. Among them, the data acquisition module includes a sensor and a data transmission unit. The sensor is used to periodically collect various data of the meter, including important habitual periodic information such as the voltage, current, power consumption, and operating status of the meter. The data transmission unit is used to transmit the collected data completely to the data preprocessing module for data processing and normalization.

[0064] The data preprocessing module includes a data cleaning unit, a data standardization unit, and a data dimension reduction unit. The data cleaning unit is used to remove incomplete, inconsistent dirty data and noise that may exist in the data. Dirty data and noise will affect the clustering level of meter data and the model's perception and prediction level of time series characteristics, ensuring the validity of the data. The data standardization unit is used to standardize the data to the same scale. The data dimension reduction unit is used to reduce the dimension of the data, which is more effective for data processing and deep learning.

[0065] The training module is used to create and train deep learning models, which include clustering, chain migration algorithm, gated unit cycle model, and K-Sigma threshold algorithm. The above algorithms are used to create a user's periodic habitual electricity consumption data model, and the chain migration algorithm is used to migrate the model for users with similar electricity consumption data sets as the model basis for the user. Algorithms such as Euclidean distance and cosine distance are used to adjust the model parameters.

[0066] The fault prediction module is used to predict the results of the meter fault condition through the deep learning model output based on the collected real-time data, and determine whether the meter is faulty. The fault prediction module includes a counter and a fault classification unit. Due to the volatility of the operation of the power system, it is easy to produce false alarms based on a single abnormal point. In order to identify whether the meter is a real fault, it is necessary to exclude the data anomalies caused by the user's use of the meter mutation. The function of the counter is to judge it as abnormal only when the target continuously sends an abnormal signal within a certain period of time. The fault classification unit is used to classify the meter fault according to the judgment result of the counter. Through the fault prediction module, users can take preventive measures to avoid the occurrence of faults, reduce downtime and maintenance costs, and improve the reliability and stability of the meter. In addition, the unit can also help users optimize maintenance plans and achieve predictive maintenance, thereby improving the efficiency and performance of the entire system.

[0067] The early warning and processing module is used to issue fault warnings and perform monitoring and processing based on the results of the meter fault prediction. It can implement the output results of the monitoring fault prediction module. When the early warning and processing module detects that the prediction result is normal data, it will perform the following processing:

[0068] Data storage: Store normal data for subsequent analysis and reference. This data can be used to establish the normal operating mode of the equipment and provide a comparative basis for fault diagnosis.

[0069] Trend analysis: Perform trend analysis on normal data to understand how the operating status of the equipment changes over time. This helps to discover potential problem trends and take preventive measures in advance.

[0070] Compare with historical data: Compare the current normal data with historical data to determine whether the operating status of the meter has changed. If abnormal changes are found, further investigation and analysis may be required.

[0071] Performance update: Use normal data to evaluate the performance of the meter and more accurately reflect the actual operation of the equipment. Through continuous learning and optimization, the model can improve its ability to predict failures.

[0072] By processing the normal data in the prediction results, the early warning and processing modules can better understand the operating status of the equipment, discover potential problems in a timely manner, and provide support for equipment maintenance and management. This helps to improve the reliability and availability of equipment, reduce the risk of failure, and ensure the smooth progress of production. At the same time, the processing of normal data also provides a basis for the optimization and improvement of the fault prediction module, enabling it to continuously adapt to the changes and development of equipment.

[0073] When the early warning and processing module detects that the prediction result is abnormal data, it will perform the following processing:

[0074] Data storage: Store abnormal data for subsequent analysis and reference. These data can be used to store abnormal databases and provide reference for comparison of actual meter failures and usage time.

[0075] Abnormal analysis: abnormal data is analyzed to understand the changes in the operating status of the equipment over time. For some minor faults, it may automatically repair or adjust to avoid further expansion of the fault; for some serious faults, the early warning mechanism will be triggered immediately. The early warning mechanism can include various forms, such as sound, sending SMS dispatch, etc., so as to promptly notify relevant personnel to handle it.

[0076] Continuous monitoring: Record and track the troubleshooting process, and continue to focus on the operating status of the meter for subsequent analysis and evaluation.

[0077] Records and reports: Detailed fault reports can be generated, including information such as the time, cause, treatment measures and results of the fault, providing strong support for equipment maintenance and management to improve the effectiveness of fault prediction and treatment.

[0078] The early warning and processing module can effectively avoid the occurrence and expansion of faults through real-time monitoring and early warning, and timely processing measures, ensure the normal operation of the meter, and improve the reliability and stability of the system.

[0079] The module evaluation module is used to evaluate and optimize the performance of the deep learning model. The data visualization module is used to display the collected meter data and prediction results in the form of charts.

[0080] In other embodiments, a computer device is also provided. The computer device provided in this embodiment can take various forms, such as an intelligent terminal, a tablet computer, a laptop computer, a desktop computer, a rack server, a blade server, a tower server or a cabinet server (including an independent server, or a server cluster composed of multiple servers) that can execute a program. The computer device of this embodiment includes at least but is not limited to: a memory and a processor that can communicate with each other through a system bus. The memory stores a computer program. When the computer program is executed by the processor, the steps of the electric meter fault prediction method based on deep learning as described above are implemented.

[0081] In this embodiment, the memory (i.e., readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of a computer device, such as a hard disk or memory of the computer device.

[0082] In other embodiments, the memory may also be an external storage device of a computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc., equipped on the computer device. Of course, the memory may also include both an internal storage unit of the computer device and its external storage device. In this embodiment, the memory stores a computer program. When the computer program is executed by the processor, the steps of the electric meter fault prediction method based on deep learning as described above are implemented.

[0083] In some embodiments, the processor may be a central processing unit (CPU), a graphics processing unit (GPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor is generally used to control the overall operation of a computer device.

[0084] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting electric meter faults based on deep learning, characterized in that: It includes the following steps: Periodically collect various data of electric meters; After preprocessing, various types of collected data are clustered to obtain multiple clusters; The data in multiple clusters are input into the deep learning model. Each different cluster corresponds to a different gate structure weight of the deep learning model through similarity, and finally a result value with time series characteristics corresponding to the data is obtained; Compare the result value with the set threshold interval. If the result value is outside the threshold interval, the result value is defined as an abnormal point. Within a unit time, if an abnormal point appears, the counter is incremented by 1, and if a normal result value appears, the counter is decremented by 1. When the counter reaches the counting threshold, the meter is judged to be in a fault state, where the counting threshold is a positive integer. And, when the number of abnormal points exceeds 20% of the total number of result values ​​in all collection cycles of the day, the meter is judged to be in a fault state; The deep learning model is trained before use, and the training method includes the following steps: The training set of electricity meter data is clustered to obtain N clusters, each of which contains i electricity meters; Select the historical data of one of the cluster centers p to train the existing model to obtain the initial model m1; Calculate the electric meter q in the cluster with the highest data feature similarity to the center point p; Input the historical data of meter q into the initial model m1 for training to obtain the next model m2; find meter a with the highest similarity to meter q data features, input the historical data of meter a into model m2 for training to obtain the next model m3; repeat the model training and search process until model m3 is obtained. i+1 ; N clusters get N models m i+1 ; N models m i+1 A deep learning model is constructed and the entire deep learning model training is completed.

2. The electric meter fault prediction method based on deep learning according to claim 1 is characterized in that: The various data of the electric meter include voltage, current, power consumption, and operating status.

3. The electric meter fault prediction method based on deep learning according to claim 1 is characterized in that: The preprocessing includes data cleaning processing, normalization processing and feature extraction processing.

4. The electric meter fault prediction method based on deep learning according to claim 1 is characterized in that: The clustering process adopts K-means algorithm; the number of clusters adopts elbow rule to select the optimal number of clusters.

5. The electric meter fault prediction method based on deep learning according to claim 1 is characterized in that: The method for setting the threshold interval comprises the following steps: Calculate the residual of the test set during model training. The calculation formula is: Where, E represents the daily residual of the test set; represents the daily electricity consumption forecast; Indicates the daily reading of the electric meter; The K-Sigma algorithm is used to calculate the minimum K value that can satisfy the residual E predicted by the test set; after determining the K value, the threshold interval is calculated: T L =T mean -(K×I range ) In the formula, is the upper limit of the threshold interval, T L is the lower limit of the threshold interval, T mean is the trimmed mean of the data, I range is the interquartile range of the data.

6. The electric meter fault prediction method based on deep learning according to claim 1, characterized in that: In each cluster, the historical data used for model training belongs to set T, and the unused historical data belongs to set M. When set M is an empty set during model training, the entire deep learning model training is completed.

7. The electric meter fault prediction method based on deep learning according to claim 1, characterized in that: When the electric meter is judged to be in a fault state, an early warning mechanism is immediately triggered; the early warning mechanism includes sound reminders and sending SMS dispatch orders.

8. A meter fault prediction system based on deep learning, characterized in that: It uses the electric meter fault prediction method based on deep learning as described in any one of claims 1 to 7; The electric meter fault prediction system comprises: A data acquisition module, which is used to periodically collect various data of the electric meter; A data preprocessing module is used to preprocess various types of collected data to standardize the data; A training module, which is used to create and train deep learning models; A fault prediction module, which is used to predict the fault condition of the electric meter based on the output results of the deep learning model; The early warning and processing module is used to issue fault early warning and perform monitoring and processing according to the results of the electric meter fault prediction.

9. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the electric meter fault prediction method based on deep learning as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the electric meter fault prediction method based on deep learning as described in any one of claims 1 to 7 are implemented.

Images