Electric energy metering online monitoring method based on adaptive filtering and deep convolutional neural network

Abstract

The invention relates to the technical field of electric energy metering monitoring, in particular to an electric energy metering online monitoring method based on adaptive filtering and a deep convolutional neural network. According to the technical scheme, the method comprises the following steps: based on monitoring of correlation between electrical parameters and equipment characteristics, carrying out preliminary correlation analysis by using an intelligent algorithm according to electrical parameters, mutual inductor characteristics and electrical topological structure information collected by an electric energy metering device, and then carrying out global optimization algorithm processing. Noise interference in data is effectively removed through the adaptive filtering technology, the data quality is improved, a reliable data basis is provided for subsequent analysis and model training, the accuracy of a monitoring result is improved, high-precision description of errors of a measuring instrument is achieved, the precision of online monitoring of electric energy measurement is remarkably improved, and the method is suitable for popularization and application. Potential problems of the measuring instrument can be found in time, electric power transaction disputes and power grid operation risks caused by measuring errors are reduced, and stable operation and economic operation of an electric power system are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of electric energy metering and monitoring, and particularly to an on-line monitoring method for electric energy metering based on adaptive filtering and deep convolutional neural network. Background Art

[0002] In modern power systems, the accuracy and reliability of electric energy metering play a crucial role in power trading settlement, power grid operation management, and fair power consumption of users. With the rapid development of smart grids, the accuracy and reliability of electric energy metering are essential for the stable operation of power systems, the fairness and justice of power trading, and the rational power consumption of users. Traditional on-line monitoring methods for electric energy metering mainly rely on regular manual inspections and simple data records, and there are many limitations.

[0003] Currently, common on-line monitoring methods for electric energy metering include monitoring methods based on threshold judgment, monitoring methods based on simple statistical analysis, and some systems that introduce automation monitoring but have relatively single functions. The method based on threshold judgment issues an alarm when the monitored value exceeds the preset threshold range of parameters such as voltage, current, and power. The method based on simple statistical analysis mainly performs some basic statistical calculations on electric energy metering data, such as mean value, variance, etc., to judge whether the equipment operation status is normal. In addition, some existing automation monitoring systems can achieve automatic data collection and preliminary analysis, but there is still much room for improvement in intelligent diagnosis and data processing capabilities.

[0004] However, the above existing on-line monitoring methods for electric energy metering all have some technical problems that cannot be ignored: First, the monitoring real-time performance is poor. The threshold setting of the method based on threshold judgment mostly depends on experience or equipment standard parameters. In a complex and changeable power grid operation environment, such as frequent load fluctuations and intermittent interference, it is difficult to accurately adapt to the actual situation, resulting in frequent false alarms or missed alarms and unable to truly reflect the equipment operation status. Second, the fault mode recognition ability is insufficient. The method based on simple statistical analysis mainly relies on basic statistics such as mean value and variance to judge the equipment status. Facing complex fault modes, due to the lack of mining of deep-level correlation relationships in the data, it is difficult to accurately identify the fault type and cause, and the effective information provided for fault diagnosis is extremely limited. Third, the system adaptability is weak. Some existing preliminary automation monitoring systems usually do not have adaptive learning and optimization functions, cannot automatically adjust the monitoring strategy and model parameters according to the dynamic changes of the power grid operation status, and are difficult to meet the high requirements of smart grids for monitoring accuracy, dynamics, and intelligence, and cannot respond in time to new challenges brought by power grid changes.

[0005] Therefore, this application proposes an on-line monitoring method for electric energy metering based on adaptive filtering and deep convolutional neural network. Summary of the Invention

[0006] The object of the present invention is to address the problems of poor real-time monitoring and insufficient fault mode recognition ability in the power metering online monitoring method in the background technology, and to propose a power metering online monitoring method based on adaptive filtering and deep convolutional neural network.

[0007] The technical solution of the present invention: A power metering online monitoring method based on adaptive filtering and deep convolutional neural network includes the following steps:

[0008] Based on the monitoring of the association between electrical parameters and equipment characteristics, using the electrical parameters collected by the power metering device, the characteristics of the mutual inductor, and the electrical topology structure information, performing preliminary association analysis using intelligent algorithms, and then processing through a global optimization algorithm;

[0009] Construct a multi-source data deep learning model, using a deep convolutional neural network (CNN) or a recurrent neural network (RNN), extracting deep features from multi-source data such as power metering data, electrical topology data, and environmental data through training the model, and performing adaptive monitoring and analysis;

[0010] Perform adaptive filtering and data enhancement processing on the collected power metering data.

[0011] Optionally, the monitoring based on the association between electrical parameters and equipment characteristics specifically includes the following steps:

[0012] Feature extraction, extracting amplitude, phase, harmonic content, power factor fluctuation electrical parameter features from power metering data, extracting ratio deviation and phase angle error change rate features from the characteristics of the mutual inductor, and extracting node connection relationship and branch impedance features from the electrical topology structure;

[0013] Preliminary association and analysis, using an association analysis algorithm to perform preliminary association analysis on the extracted power metering data, the characteristics of the mutual inductor, and the electrical topology structure features;

[0014] Fine monitoring and optimization, using the particle swarm optimization (PSO) algorithm to optimize the results of the preliminary association analysis, and using the genetic algorithm to perform global optimization on all monitoring parameters.

[0015] Optionally, when extracting the amplitude feature from the power metering data, the mean and variance features are obtained through statistical analysis of the voltage and current amplitudes; when extracting the harmonic content feature, the harmonic components in the voltage and current are analyzed using the fast Fourier transform method.

[0016] Optionally, when extracting the ratio deviation feature from the characteristics of the mutual inductor, the ratio deviation is calculated according to the nominal ratio of the mutual inductor and the actually measured ratio; when extracting the phase angle error change rate feature, the change rate of the phase angle error of the mutual inductor over time is monitored.

[0017] Optionally, when extracting the node connection relationship features from the electrical topological structure, a graph theory algorithm is used to construct an electrical topology graph, and the connection degree of nodes and the adjacent node information features are extracted; when extracting the branch impedance features, the impedance values of each branch in the electrical topology are calculated.

[0018] Optionally, the correlation analysis algorithm is Pearson correlation coefficient or grey relational analysis, and the calculation formula of Pearson correlation coefficient is:

[0019]

[0020] where x i and y i are respectively the sample value sequences of the i-th current amplitude and the transformer ratio deviation, and are respectively the sample means of n sample current amplitudes and the transformer ratio deviation, and n is the number of samples.

[0021] Optionally, the construction of the multi-source data deep learning model specifically includes:

[0022] Network architecture design: For power metering data, CNN is used to extract the spatio-temporal features of electrical parameters; for electrical topology data, graph convolutional network (GCN) or similar methods are used to extract topological structure features; for environmental data, a fully connected layer is used to convert it into a feature vector and fuse it with other data features. The fusion methods include early fusion and late fusion, and attention mechanism or gated recurrent unit (GRU) is used for context learning and fusion of multi-source data information;

[0023] Training and optimization: Use the labeled data set containing power metering data, electrical topology data and environmental data for training, adopt semi-supervised learning or transfer learning methods, optimize the model through the loss function, and the training process adopts end-to-end learning method and uses data augmentation technology.

[0024] Optionally, the output calculation formula of the convolutional layer of the CNN is:

[0025]

[0026] where x i,j is the input image, k m,n is the convolutional kernel, m = 1, …, f h ; n = 1, …, f w , f h and f w are respectively the height and width of the convolutional kernel, b is the bias term, and after passing through the convolutional layer, the ReLU function is connected: y i′,j′ = max(0, y i′,j′ ).

[0027] Optionally, the formula for the attention mechanism to calculate the attention weights is:

[0028]

[0029] where e i = v T tanh(W q q + W h h i ), v is a learnable parameter vector that participates in the calculation of the intermediate quantity in the attention weight calculation, W q is a learnable weight matrix used to transform the query vector q, and W h is a learnable weight matrix used to transform the hidden state h corresponding to the i-th element in the input sequence; i denotes the transformation;

[0030] where e j = v T tanh(W q q + W h h j ), W h is a learnable weight matrix used to transform the hidden state h of the input sequence; j denotes the transformation; h j is the hidden state corresponding to the j-th element in the input sequence; when calculating the attention weight α i , e j is used to measure the degree of association between the j-th element in the input sequence and the query vector q. By performing exponentialization and normalization on all e j , the obtained attention weight α i can reflect the relative importance of each element in the input sequence for the current attention calculation, and then the context vector c is obtained by weighted summation:

[0031]

[0032] α i is the formula for the attention mechanism to calculate the attention weights.

[0033] Optionally, the adaptive filtering and data augmentation processing specifically includes:

[0034] Power metering data noise reduction, using Kalman filtering or wavelet transform denoising methods to perform noise reduction processing on power metering data;

[0035] Data augmentation, using data resampling or feature transformation methods to perform data augmentation processing on power metering data.

[0036] Compared with the prior art, the present application includes at least one of the following beneficial technical effects:

[0037] 1. The noise interference in the data is effectively removed through adaptive filtering technology, improving the data quality, providing a reliable data basis for subsequent analysis and model training, and thus enhancing the accuracy of the monitoring results.

[0038] 2. The deep convolutional neural network model can automatically extract the complex features of the data, and through the correlation learning of the electrical topology, transducer characteristics, and the random process of electrical parameters, it realizes the high-precision characterization of the errors of metering devices, significantly improving the accuracy of on-line monitoring of electric energy metering.

[0039] 3. The problem of real-time acquisition of errors without power interruption and without physical standard devices has been successfully solved, realizing the real-time on-line monitoring and fault warning of electric energy metering devices, timely discovering potential problems of metering devices, reducing power trading disputes and power grid operation risks caused by metering errors, and ensuring the stable and economic operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is the system architecture diagram of this embodiment;

[0041] Figure 2 is the data processing flow chart of this embodiment;

[0042] Figure 3 is the deep learning model architecture diagram of this embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] The technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present disclosure. The components of the embodiments of the present disclosure described and shown herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but merely represents selected embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[0044] I. Monitoring Method Based on the Association between Electrical Parameters and Equipment Characteristics

[0045] 1.1 Feature Extraction and Association

[0046] Feature extraction is a key step in electric energy metering monitoring. For electric energy metering data, common feature extraction methods include electrical parameter features such as amplitude, phase, harmonic content, and power factor fluctuation. For transducer characteristics, features such as ratio deviation and phase angle error change rate are extracted. For electrical topology structures, features such as node connection relationships and branch impedances are extracted.

[0047] Feature extraction of electric energy metering data:

[0048] Feature extraction of electrical parameter amplitude: By statistically analyzing the amplitudes of voltage and current, features such as their mean values and variances are obtained to judge the basic situation of electric energy metering.

[0049] Feature extraction of harmonic content: Using the Fast Fourier Transform (FFT) method to analyze the harmonic components in voltage and current, and determining the harmonic content features to detect the impact of harmonic interference in the power grid on metering.

[0050] Feature extraction of transducer characteristics:

[0051] Feature extraction of ratio deviation: According to the nominal ratio and the actually measured ratio of the transducer, the ratio deviation is calculated, which is an important index to measure the performance of the transducer.

[0052] Feature extraction of phase angle error change rate: Monitoring the change rate of the phase angle error of the transducer over time to detect potential faults of the transducer in a timely manner.

[0053] Feature extraction of electrical topology structure:

[0054] Feature extraction of node connection relationship: Using graph theory algorithms to construct an electrical topology graph, and extracting features such as the connection degree of nodes and adjacent node information to analyze the impact of the power transmission path on metering.

[0055] Feature extraction of branch impedance: Calculating the impedance values of each branch in the electrical topology, and combining with the electrical parameter features to judge the impact of line loss on metering.

[0056] 1.2 Preliminary correlation and analysis

[0057] Through the extracted electric energy metering data, transducer characteristics and electrical topology structure features, preliminary correlation analysis is carried out using correlation analysis algorithms (such as Pearson correlation coefficient, grey correlation analysis, etc.):

[0058] Pearson correlation coefficient: Used to calculate the linear correlation between electrical parameters and transducer characteristics, and determine the degree of association between the two. For example, analyzing the correlation between the current amplitude and the ratio deviation of the transducer to judge the impact of the transducer on current metering. Its calculation formula is:

[0059]

[0060] Where x i and y i are the sample value sequences of the i-th current amplitude and the ratio deviation of the transducer respectively, and are the sample means of n sample current amplitudes and the ratio deviation of the transducer respectively, and n is the number of samples.

[0061] Grey relational analysis: applicable to situations with less data volume and complex data relationships, can be used to analyze the correlation between changes in electrical topology structure and power metering errors, and find key influencing factors.

[0062] The result of the preliminary correlation analysis is to determine the preliminary correlation between various factors, providing a basis for subsequent monitoring and optimization.

[0063] 1.3 Fine Monitoring and Optimization

[0064] Optimization based on intelligent algorithms: Use the Particle Swarm Optimization (PSO) algorithm to optimize the results of the preliminary correlation analysis and reduce monitoring errors. The PSO algorithm finds the optimal monitoring model parameters through an iterative process to minimize the error between the monitoring results and the actual situation.

[0065] Steps of PSO:

[0066] Initialize the position and velocity of the particle swarm;

[0067] Calculate the fitness value (i.e., monitoring error) of each particle;

[0068] Update the velocity and position of the particle;

[0069] Judge whether the stop condition is satisfied. If not, repeat steps 2 and 3 until convergence.

[0070] Global optimization (such as genetic algorithm): Use the genetic algorithm to globally optimize all monitoring parameters, reduce error accumulation, and improve monitoring accuracy. The genetic algorithm simulates the biological evolution process, performs selection, crossover, and mutation operations on the parameters of the monitoring model, and searches for the optimal solution. Common genetic algorithm operations include the design and application of selection operators, crossover operators, and mutation operators.

[0071] II. Monitoring Method Based on Deep Learning Model of Multi-source Data

[0072] 2.1 Network Architecture Design

[0073] To achieve efficient monitoring and analysis of power metering data, a deep learning model can be used to learn the mapping relationship between multi-source data. A common architecture is a hybrid model based on Deep Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). The specific steps are as follows:

[0074] Input processing:

[0075] For power metering data: Use CNN to extract the spatio-temporal features of electrical parameters. For example, through the convolutional layer and pooling layer, process the waveform data of voltage and current to obtain their local and global features. Taking a simple convolutional layer as an example, assume the input image (the waveform of power metering data can be regarded as an image) is xi,j where \(i = 1,\ldots,h\); \(j = 1,\ldots,\omega\), where \(h\) and \(\omega\) are the height and width of the image respectively. The convolutional kernel is \(k\). m,n where \(m = 1,\ldots,f\). h and \(n = 1,\ldots,f\). w where \(f\). h and \(f\). w are the height and width of the convolutional kernel respectively. The output of the convolutional layer \(y\). i′,j′ is:

[0076]

[0077] where \(b\) is the bias term, \(i' = 1,\ldots,h - f+1\); \(j' = 1,\ldots,\omega - f + 1\). After passing through the convolutional layer, an activation function is usually connected, such as the ReLU function: \(y\). h = \(\max(0,y)\). w ) i′,j′ i′,j′

[0078] For electrical topology data: Convert the electrical topology diagram into matrix form, and use a graph convolutional network (GCN) or a similar method to extract topological structure features, such as node connectivity and hierarchical features.

[0079] For environmental data: Such as temperature and humidity, use a fully connected layer to convert it into a feature vector and fuse it with other data features.

[0080] Multi-source data fusion:

[0081] Feature fusion: Fuse the features of power metering data, electrical topology data, and environmental data. A multi-level fusion method can be adopted:

[0082] Early fusion: Perform fusion at the lower layer of the network, splice the features of different source data and input them, so that the model can learn the features of multiple data at the same time.

[0083] Late fusion: First extract the features of each source data through the network respectively, and then fuse the information at a higher level for comprehensive monitoring and analysis.

[0084] Fusion network design: Use an attention mechanism or a gated recurrent unit (GRU) for context learning and fusion of multi-source data information. Taking the attention mechanism as an example, let \(h\). i \((i = 1,\ldots,N)\) be the hidden state representation of the input sequence (such as the multi-source data feature sequence), and \(q\) be the query vector (which can be a vector related to the monitoring task). Calculate the attention weight \(\alpha\). i :

[0085]

[0086] ​​Among them, e i = v T tanh(W q q + W h h i ), where v is a learnable parameter vector that participates in the calculation of intermediate quantities in the calculation of attention weights, and W q is a learnable weight matrix used to transform the query vector q, and W h is a learnable weight matrix used to transform the hidden state h corresponding to the i-th element in the input sequence; i denotes the transformation;

[0087] Among them, e j = v T tanh(W q q + W h h j ), W h is a learnable weight matrix used to transform the hidden state h of the input sequence; j denotes the transformation; h j is the hidden state corresponding to the j-th element in the input sequence; when calculating the attention weight α i , e j is used to measure the degree of association between the j-th element in the input sequence and the query vector q. By performing exponentialization and normalization on all e j , the obtained attention weight α i can reflect the relative importance of each element in the input sequence for the current calculation of attention, and then the context vector c is obtained by weighted summation:

[0088]

[0089] α i is the formula for the attention mechanism to calculate the attention weight.

[0090] The attention mechanism can automatically focus on the data part that has a greater impact on the monitoring results, while the GRU can effectively handle the long-term dependencies in time series data and improve the model's monitoring ability for the dynamic changes of power metering.

[0091] 2.2 Training and Optimization

[0092] Use the labeled dataset containing power metering data, electrical topology data, and environmental data for training, and adopt semi-supervised learning or transfer learning methods to optimize the model through loss functions (such as mean square error, cross entropy).

[0093] During the training process, an end-to-end learning approach is adopted, directly from the original data to the monitoring results, reducing the error transmission in the intermediate links and not relying on traditional manual feature extraction methods. At the same time, data augmentation techniques are used, such as randomly sampling and adding noise to the power metering data, expanding the training data set, and improving the generalization ability of the model.

[0094] III. Adaptive Filtering and Data Augmentation Processing

[0095] 3.1 Denoising of Power Metering Data

[0096] Power metering data usually contains noise, especially in the case of strong power grid fluctuations or electromagnetic interference. To improve the data quality, the following denoising methods are usually adopted:

[0097] Kalman filter: By dynamically modeling and predicting electrical parameters, the weighted combination of measured values and predicted values is used to estimate the true value, effectively removing noise interference, especially suitable for processing power metering data with dynamic change characteristics. Its state equation is:

[0098] x k =F k x k-1 +B k u k +w k

[0099] where x k is the system state vector at time k, x k-1 is the state transition matrix, which is the system state vector at time k-1, B k is the control input matrix, u k is the control input vector, w k is the process noise vector, assuming it follows a Gaussian distribution with a mean of 0 and a covariance of Q k . The measurement equation is:

[0100] z k =H k x k +v k

[0101] where z k is the measurement vector at time k, H k is the measurement matrix, v k is the measurement noise vector, assuming it follows a Gaussian distribution with a mean of 0 and a covariance of R k .

[0102] Wavelet transform denoising: Using wavelet transform to decompose the electrical parameter signal into sub-bands of different frequencies, according to the distribution characteristics of noise in the high-frequency sub-bands, removing the noise components and retaining the main features of the signal, which has a good denoising effect on non-stationary signals.

[0103] 3.2 Data Augmentation

[0104] The sample size of power metering data may be limited. To improve the performance of the model, data augmentation processing is required. Common data augmentation methods include:

[0105] Data resampling: Randomly resample the power metering data to change the time interval or sampling frequency of the data, increasing the diversity of the data.

[0106] Feature transformation: Randomly transform the electrical parameter features, such as multiplying by a random coefficient or adding a random offset, to simulate different operating conditions and expand the dataset.

[0107] The present invention effectively removes the noise interference in the data through adaptive filtering technology, improves the data quality, provides a reliable data basis for subsequent analysis and model training, and thus enhances the accuracy of the monitoring results.

[0108] Among them, the deep convolutional neural network model can automatically extract the complex features of the data, and through the correlation learning of the electrical topology, transformer characteristics, and random process of electrical parameters, it realizes the high-precision characterization of the errors of metering devices, significantly improving the accuracy of online power metering monitoring. The present invention successfully solves the problem of real-time error acquisition without power interruption and without physical standard devices, realizes the real-time online monitoring and fault warning of power metering devices, timely discovers potential problems of metering devices, reduces power trading disputes and power grid operation risks caused by metering errors, and ensures the stable and economic operation of the power system.

[0109] Such as Figure 1 is the system architecture diagram, which clearly shows the connection relationships among the power metering device, sensor, data acquisition module, data preprocessing module, correlation analysis and monitoring module, deep learning model module, and result output and alarm module. The power metering device is connected to the data acquisition module through the sensor, and the data acquisition module transmits the collected data to the data preprocessing module for filtering, denoising, and enhancement processing. The processed data enters the correlation analysis and monitoring module and the deep learning model module respectively. The correlation analysis and monitoring module uses intelligent algorithms for preliminary correlation analysis and fine monitoring optimization, and feeds the results back to the deep learning model module. The deep learning model module extracts, fuses features and trains the model based on multi-source data, and finally outputs the monitoring results to the result output and alarm module. If the monitoring results exceed the preset threshold, the result output and alarm module will send an alarm signal.

[0110] Such as Figure 2It is a data processing flow chart, which details the entire process of power metering data from collection to the output of the final monitoring results. First is the data collection stage, including obtaining voltage, current, power, temperature, and humidity data from power metering devices and environmental sensors. Then it enters the data preprocessing stage, successively performing adaptive filtering (such as Kalman filtering, wavelet transform denoising), data enhancement (such as data resampling, feature transformation, GAN generating fake samples) operations. Next is the feature extraction stage, respectively extracting features from power metering data, transformer characteristic data, and electrical topology data, such as extracting the amplitude and harmonic content of electrical parameters, the change rate of ratio deviation and phase angle error, and the node connection relationship and branch impedance. After that, correlation analysis and deep learning model processing are carried out, including preliminary correlation analysis (Pearson correlation coefficient, grey correlation analysis), optimization based on intelligent algorithms (particle swarm optimization, genetic algorithm), multi-source data fusion (early fusion, late fusion), and model training and optimization. Finally, a judgment is made based on the monitoring results. If the threshold is exceeded, an alarm is triggered; otherwise, monitoring continues.

[0111] Figure 3 It is an architecture diagram of a hybrid model based on a deep convolutional neural network (CNN) and a recurrent neural network (RNN). The input layer of the model receives preprocessed power metering data, electrical topology data, and environmental data. For power metering data, feature extraction is performed through the CNN layer, including convolutional layer, pooling layer, and fully connected layer, to extract the spatio-temporal features of electrical parameters. For electrical topology data, topological features are extracted through the graph convolutional network (GCN) layer or a similar structure. Environmental data is transformed into a feature vector through the fully connected layer. Then, the features of different source data are fused through the feature fusion layer, which can adopt the early fusion or late fusion method. The fused features enter the RNN layer for time series analysis and context learning to handle the dynamic changes of power metering data. Finally, the monitoring results are output through the output layer, such as the error estimation value of the metering device or the judgment result of whether there is an abnormality.

[0112] Embodiment application test

[0113] In the power system of a large industrial park, the online power metering monitoring method based on adaptive filtering and deep convolutional neural network of the present invention has been successfully applied.

[0114] There are multiple large factories in this park, with complex and diverse electrical equipment, and extremely high requirements for the accuracy of power metering. First, high-precision sensors are installed at each power metering device and key electrical nodes to ensure comprehensive collection of voltage, current, power, power factor power metering data, as well as environmental temperature and humidity data. At the same time, detailed parameter tests are carried out on the transformers in the park, and an accurate electrical topology structure model is constructed using advanced power grid analysis technology.

[0115] After data acquisition, Kalman filtering is used to denoise the power metering data. For example, during the peak electricity consumption period in a certain factory, the power grid fluctuates greatly. Kalman filtering effectively removes the noise generated by the fluctuations, significantly improving the data stability. At the same time, data augmentation techniques of data resampling and feature transformation are adopted. For example, the data during some stable operation periods are resampled to change their time intervals, and different random coefficients are multiplied by the electrical parameter features to simulate different electricity consumption conditions, expanding the training data set.

[0116] In terms of correlation analysis and monitoring, through Pearson correlation coefficient analysis, it is found that there is a significant correlation between the voltage amplitude of a certain main power supply line and the ratio deviation of the transformer on this line. The particle swarm optimization algorithm is used to optimize the monitoring model, reducing the power metering monitoring error in this area by 30%. For example, before optimization, the power metering error in this area occasionally reached 5%, and after optimization, it is stably controlled within 3.5%.

[0117] During the training process of the hybrid model based on the deep convolutional neural network and the recurrent neural network, the complex relationships among the power metering data, electrical topology data, and environmental data are fully learned. For example, the model learns the rule that when the temperature is high and the branch load in a certain area of the electrical topology increases, the error of a specific power metering device will increase. After a period of training and optimization, the model performs excellently in actual operation.

[0118] During an actual operation, the system monitors that the error of the power metering device in a certain factory suddenly increases and exceeds the preset threshold. The system immediately issues an alarm signal, and through detailed data analysis and model diagnosis, it is determined that the phase angle error of a nearby transformer has changed due to long-term high-load operation, resulting in inaccurate metering. According to the information provided by the system, the maintenance personnel quickly replaced the faulty transformer, avoiding the incorrect calculation of production costs and waste of power resources caused by metering errors, and ensuring the normal production of the factory and the stable operation of the power system in the park.

[0119] Through the application in this industrial production park, it is fully proved that the online power metering monitoring method of the present invention can effectively improve the accuracy and reliability of power metering, timely discover and solve potential problems of metering devices, and provide a strong guarantee for the stable operation of the power system and the rational use of electricity by enterprises.

[0120] The above specific embodiments are only several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. An online monitoring method for electric energy metering based on adaptive filtering and deep convolutional neural network, characterized in that: The following steps are involved: Based on the monitoring of the association between electrical parameters and equipment characteristics, the electrical parameters, transformer characteristics and electrical topology information collected by the electric energy metering device are used to perform preliminary association analysis using intelligent algorithms, and then processed using global optimization algorithms; Construct a multi-source data deep learning model, using deep convolutional neural network or recurrent neural network, and extract deep features from multi-source data such as power metering data, electrical topology data, and environmental data through training models for adaptive monitoring and analysis; The collected electric energy metering data is adaptively filtered and enhanced.

2. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 1 is characterized in that: The monitoring based on the association between electrical parameters and equipment characteristics specifically includes the following steps: Feature extraction: extracting the amplitude, phase, harmonic content, and power factor fluctuation electrical parameter features from the electric energy metering data; extracting the transformation ratio deviation and phase angle error change rate features from the transformer characteristics; and extracting the node connection relationship and branch impedance features from the electrical topology structure; Preliminary correlation and analysis: Use correlation analysis algorithms to conduct preliminary correlation analysis on the extracted power metering data, transformer characteristics, and electrical topology features; Fine monitoring and optimization, using particle swarm optimization algorithm to optimize the preliminary association analysis results, and using genetic algorithm to globally optimize all monitoring parameters.

3. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 2 is characterized in that: When extracting amplitude features from the electric energy metering data, the mean and variance features are obtained by statistically analyzing the voltage and current amplitudes; when extracting harmonic content features, the fast Fourier transform method is used to analyze the harmonic components in the voltage and current.

4. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 2 is characterized in that: When extracting the ratio deviation feature from the transformer characteristics, the ratio deviation is calculated according to the nominal ratio of the transformer and the actually measured ratio; when extracting the phase angle error change rate feature, the change rate of the transformer phase angle error over time is monitored.

5. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 2 is characterized in that: When extracting node connection relationship features from the electrical topology structure, a graph theory algorithm is used to construct an electrical topology graph to extract the node connection degree and adjacent node information features; when extracting branch impedance features, the impedance value of each branch in the electrical topology is calculated.

6. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 2 is characterized in that: The correlation analysis algorithm is Pearson correlation coefficient or grey correlation analysis, and the Pearson correlation coefficient calculation formula is: Among them, x i and i The sample value sequence of the i-th current amplitude and the transformer ratio deviation, and are the sample means of n sample current amplitudes and transformer ratio deviations respectively, and n is the number of samples.

7. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 1 is characterized in that: The construction of the multi-source data deep learning model specifically includes: Network architecture design: for power metering data, CNN is used to extract the spatiotemporal features of electrical parameters; for electrical topology data, graph convolutional networks or similar methods are used to extract topological structure features; for environmental data, a fully connected layer is used to convert it into a feature vector and fuse it with other data features. The fusion methods include early fusion and late fusion, and attention mechanisms or gated recurrent units are used for contextual learning and fusion of multi-source data information; Training and optimization: Use labeled data sets containing electricity metering data, electrical topology data, and environmental data for training. Use semi-supervised learning or transfer learning methods to optimize the model through the loss function. The training process adopts an end-to-end learning method and uses data enhancement technology.

8. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 7 is characterized in that: The convolutional layer output calculation formula of the CNN is: Among them, x i,j is the input image, k m,n is the convolution kernel, m=1,…,f h ; n=1,…,f w , f h and f w are the height and width of the convolution kernel, b is the bias term, and the convolution layer is followed by the ReLU function: i′,j′ =max(0,y i′,j′ ).

9. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 8 is characterized in that: The formula for calculating the attention weight of the attention mechanism is: Among them, e i =v T tanh(W q q+W h h i ), v is a learnable parameter vector, which participates in the calculation of the intermediate quantity in the attention weight calculation, W q is a learnable weight matrix used to transform the query vector q, W h is a learnable weight matrix used to adjust the hidden state h corresponding to the i-th element in the input sequence i Indicates transformation; Among them, e j =v T tanh(W q q+W h h j ), W h is a learnable weight matrix used to represent the hidden state h of the input sequence j Indicates transformation; h j is the hidden state corresponding to the jth element in the input sequence; in calculating the attention weight α i When j It is used to measure the correlation between the jth element in the input sequence and the query vector q. j After exponentialization and normalization, the context vector c is obtained by weighted summation: α i Formula for calculating attention weights for the attention mechanism.

10. The method for online monitoring of electric energy metering based on adaptive filtering and deep convolutional neural network according to claim 1, characterized in that: The adaptive filtering and data enhancement processing specifically includes: Electric energy metering data denoising: using Kalman filtering or wavelet transform denoising methods to denoise electric energy metering data; Data enhancement: data resampling or feature transformation methods are used to enhance the electric energy metering data.

Images