Research on power quality disturbance classification algorithm based on edge computing

Abstract

Power quality analysis and governance need the identification of power quality issues. With the use of smart meters and various smart collection devices, more and more power quality data are collected, and the massive data collection brings pressure on communication, storage and computation to the conventional algorithm for identifying and classifying power quality disturbances based on cloud computing. In the paper, a classification algorithm for power quality disturbance identification based on edge computing and fusion model is proposed. The algorithm’s key concept is to compress and sense the power quality signals at the edge side, and then transmit the compressed power quality data to the cloud, which uses an improved Dense-Net and LSTM fusion model to identify and classify the compressed power quality data. Through experiments, it is proved that the method can compress the power quality signal to 70% of the original signal size while satisfying the recognition and data on power quality disturbance categorization accuracy, reducing the communication cost of data transmission, lowering the computational pressure and caching pressure on the cloud, and having certain robustness.

Keywords

Edge computing multi-scale parallel dense network power quality

1. Introduction

With the proposed goal of ‘carbon peaking, carbon neutrality’, the rapid expansion of distributed energy resources and micro-small grids, the power grid’s distortion signal will have an impact on the power system’s safety and stability [1], causing power quality disturbance (PQD) disturbance, which may cause serious safety accidents and economic losses. As a result, proper power quality disturbance classification is critical for improving the quality of power supply, monitoring the status of power equipment, and troubleshooting the grid [2, 3].

Traditional power quality perturbation identification shifts edge-side compute-intensive tasks from end-user devices to cloud infrastructure. With the widespread adoption of power electronics, the focus of power quality perturbation signals shifts to communications with high throughput and low latency, and the cloud begins to expose the limits that it has. The physical separation between the terminal device and the cloud is too long, which will cause problems such as high communication costs and insufficient security [4, 5].

To reduce the communication cost of data transmitted from the edge device to the cloud and reduce the storage pressure of the edge device, the computing power of the edge device can be used to adopt compression sensing [6] technology on the edge side, which breaks through the framework of the sampling theorem of Nyquist. The literature [7] uses a combination of empirical modal decomposition and compressive sensing to compress and reconstruct transient power quality signals, which can reduce the need for the memory of the acquisition device and improve the transmission rate. Although the transmission of compressed data to the cloud can reduce the cost of communication, the reconstruction of power quality in the cloud will increase the computational pressure in the cloud, so it is necessary to deploy a model with stronger feature extraction and classification recognition capabilities for power quality signals in the cloud, and disturbances can be identified and classified without reconstructing the signal.

The traditional power quality disturbance recognition classification models include decision trees [8, 9], support vector machines (SVM) [10, 11], artificial neural networks (ANNs) [12] and other methods. The literature [13] proposed an efficient classifier based on curve transformation and support vector machine (SVM) to automatically identify, locate and classify different forms of power quality problems.

However, the traditional machine learning methods have insufficient ability to extract features and process large amounts of data, and the classification accuracy rate also needs to be improved. Convolutional neural networks [14, 15] (CNN) can obtain the characteristics of the input data very well, and can remove noise from the supplied data, and have a large number of applications in the identification of power quality disturbances. Literature [16] proposed a convolutional neural network with a hybrid architecture method for processing power signals using a convolutional neural network with one dimension, and then converted these signals into images, used a two-dimensional convolutional neural network for processing, and combined the generated feature vectors. Finally, the fully linked layer classifies the combinatorial vectors. The convolutional neural network proposed in the literature [17] is a combinatorial method based on variational pattern decomposition, which requires few samples but has high classification accuracy. However, the increase in the quantity of layers of the CNN network will lead to rapid degradation of the accuracy of the model after saturation, and the residual network (ResNet) structure [18] can solve this issue and increase CNN performance. In the literature [19], a ResNet and support vector machine-based model was proposed, which has a higher level of performance than traditional neural network algorithms. ResNet algorithm when extracting 18-layer and 34-layer CNN features and support vector machine classification, and the accuracy rate is improved by 2.01%. The literature [20] proposes a method for identifying and classifying power quality disturbances in microgrids, which consists of a five-layer-one-maintenance positive initial residual network and a three-layer fully connected layer (deep convolutional network), which can quickly extract the features of different types of disturbance signals, identify and classify them. However, the number of ResNet network structure parameters is too large, the computing power requirements are high, and the connection between the network layers is not close enough. Long short-term memory networks [21, 22] (LSTM) are good at extracting timing features from sequence signals.

Therefore, an algorithm based on edge computing and model fusion is proposed, which aims to reduce the communication cost of edge devices sending data to the cloud, reduce the caching and computing pressure in the cloud, and the main innovation points are:

1)
This study proposes a multi-scale parallel dense network based on edge computing, which uses the computing power of the edge-side device to compress the data and reduce the cache pressure and computing pressure in the cloud.
2)
Using the computing power of edge equipment and using compression sensing technology, compressing the power quality disturbance signal, reducing the amount of data acquisition, and then transmitting the compressed data to the cloud for identification and classification, reducing the transmission and communication cost of data, and alleviating the cache pressure in the cloud.
3)
In the cloud using multi-scale parallel dense network and LSTM fusion model, the model uses parallel dense block structure on multiple scales, the extraction of potential spatial features and the fusion and flow of features, the LSTM module using sequence recombination method to extract time features of the sample, and then the features extracted by the two modules are fused, so that the cloud does not reconstruct the signal under the premise of perturbation identification classification, effectively alleviating the computational pressure of the cloud. The network model raises the network’s depth without increasing the amount of network parameters, significantly decreases the number of parameters in the network, as well as lowering the computing pressure in the cloud.

2. Multiscale parallel dense networks based on edge computing

2.1 Compression perception

Compressed Sensing (CS), proposed by Donnho, is a spatially transformed compressive sampling based on a random smooth observation matrix when the original power quality disturbance signal is sparse or becomes sparse through a sparse transform base. The number of points sampled is significantly less than Nyquist sampling, and the original signal is recovered by the reconstruction algorithm. The linear observation model is shown in Eq. (1).

$\displaystyle y=\phi x=\phi\psi S=\Theta S$ (1)

where $y$ denotes the $M\ast 1$ -dimensional observation; $x$ is the $N\ast 1$ -dimensional original signal; $\Phi$ denotes the $M\ast N$ -dimensional observation matrix; $\Psi$ denotes the sparse basis matrix; $\Theta$ denotes the sensing matrix; and $S$ denotes the sparse coefficient.

The sparsity factor $S^{\prime}$ can be derived from the observation $y$ , and thus the original signal $x=\Psi S^{\prime}$ .

To solve the above problem, it can be solved using $L_{0}$ parametric minimization, as in Eq. (2).

$\displaystyle S^{\prime}=\arg\min\left\|S\right\|_{0}\quad(\text{s.t. }Y=\phi% \psi S=\Theta S)$ (2)

where $||{S}||_{0}$ is the zero parameterization of $S$ .

However, the solution to the l0-parametric minimization has been shown to be NP-hard and non-convex. l1-parametric minimization is often used as a convex approximation to l0-parametric minimization, and l1-parametric minimization is given by Eq. (3).

$\displaystyle S^{\prime}=\arg\min\left\|S\right\|_{1}\quad(\text{s.t. }Y=\phi% \psi S=\Theta S)$ (3)

In the meantime, in order to recover $S^{\prime}$ needs to be satisfied.

1) The dimensionality of the observation $y$ should satisfy ${M}={O}[{K\log({N})}]$ , ${O}$ being the computational complexity operator.

2) The observation matrix ${\Phi}$ should satisfy the Constrained Isometry Property (RIP) condition that for any $K$ sparse signals satisfying Eq. (4).

$\displaystyle(1-\delta_{k})\left\|{\phi X}\right\|_{2}^{2}\leqslant\left\|{% \phi X}\right\|_{2}^{2}\leqslant(1+\delta_{k})\left\|{\phi X}\right\|_{2}^{2}$ (4)

Where, $\delta_{k}\in({0,1})$ , which represents constrained equidistance constant.

Therefore, to cut data transmission costs, the data is compressed on the edge side by using the computing power of the edge device to signal compress the sensory acquisition. The first sparse basis used on the edge devices is the Discrete Cosine Transform (DCT), which is highly applicable, simple in structure and easy to implement in hardware. The Bernoulli matrix is then used as the observation matrix, which is more straightforward to implement and store than the random Gaussian observation matrix. The transformation matrix representation of the DCT is shown in Eq. (5), the matrix’s row is $i$ while the matrix’s column is $j$ .

$\displaystyle\textit{DCT}(N\times N)=\left[\textit{ACOS}\left(i\times(1+2j)% \frac{\pi}{2N}\right)\right]$ (5)

Where $i\in\{{0,\ldots,N-1}\}$ and $j\in\{{0,\ldots,N-1}\}$ represent the row number and column number of the matrix, and $A$ is defined by Eq. (6):

$\displaystyle A=\left\{{{\begin{array}[]{ll}\sqrt{\raise 3.01pt\hbox{$1$}\!% \mathord{\left/{\vphantom{1N}}\right.\kern-1.2pt}\!\lower 3.01pt\hbox{$N$}}&% \text{for }i=0\\ \sqrt{\raise 3.01pt\hbox{$2$}\!\mathord{\left/{\vphantom{2N}}\right.\kern-1.2% pt}\!\lower 3.01pt\hbox{$N$}}&\text{for }i\neq 0\\ \end{array}}}\right.$ (6)

2.2 Multi-scale parallel dense networks

The recognition classification network model used in the cloud in the paper is demonstrated in Fig. 1. The network model is comprised of a shallow feature extraction module, a dense block module and a connection layer connecting the dense blocks, and a propagation layer. Within the shallow feature extraction module, the initial convolutional processing operation is mainly performed, as shown in Eq. (6).

$\displaystyle F_{l}=\textit{MaxPool}(\textit{Relu}(Bn(W_{0}\otimes I_{0})))$ (7)

Where $W_{0}$ represents the to-be-trained weight parameter, $\otimes$ represents the convolution kernel as a $3\times 1$ convolution operation, Max-Pool represents the maximum pooling function, while ReLu represents the activation function. For deep feature extraction, the output $F_{l}$ is sent into the deep feature extraction module.

Figure 1.

Multi-scale parallel DenseNet network structure.

The parallel dense block module, as well as the connection and propagation layers that connect the parallel dense blocks, are primarily responsible for deep feature extraction, the sizes of convolution kernels in dense blocks are 3, 5 and 7, respectively. The parallel dense block used in the paper consists of three dense blocks in parallel, Fig. 2 depicts the structure of dense blocks, each dense block consists of five convolutional blocks, the structure of the convolutional blocks is shown in Fig. 3, including two sets of activation, convolutional operations and Batch-Normalization. The convolution kernels of the dense block have scales of 3 $\times$ 1, 5 $\times$ 1, and 7 $\times$ 1, respectively, each convolution block splices the training parameters of all the previous convolution blocks to help with feature fusion, while the number of network parameters can be reduced, the parameters are spliced as shown in Eq. (7), the non-linear change at layer $i$ is noted as ${K}_{i}$ (*), and each layer of the convolution block is connected to all subsequent convolution blocks, so that if there are $i$ convolutional blocks, there would be ${i}\times({{i}-1})/2$ connections.

$\displaystyle X_{i}=K_{i}([X_{0},X_{1},\ldots,X_{i-1}])$ (8)

where are multiple functions combining convolution operations, activation functions and Batch-Normalization; ‘[]’ stands for concatenation, and the dense block is the concatenation of $X_{0}$ to layers together by channel.

Figure 2.

Dense block structure.

Figure 3.

Convolution block structure.

The multi-scale parallel dense network used in the paper in the cloud, the effect of multi-scale dense blocks in parallel can facilitate the flow and fusion of features, which can mine more potential features of the data and reduce the number of network parameters while improving the representation capability of the network.

2.3 Long and short-term memory neural network (LSTM) based on sequence reorganization

The long-short-term memory network (LSTM) is a gating-based recurrent neural network proposed by Hochreiter and Schmidhuber, which can significantly demand the gradient disappearance or explosion problem in RNN. The LSTM augments the RNN with three gates: an input gate, an oblivion gate, and an output gate, which are computed using Eqs (8)–(10). Figure 4 depicts the LSTM’s internal structure.

$\displaystyle i_{t}=\sigma(W_{i}[h_{t-1},x_{t}]+b_{i})$ (9) $\displaystyle f_{t}=\sigma(W_{f}[h_{t-1},x_{t}]+b_{f})$ (10) $\displaystyle o_{t}=\sigma(W_{o}[h_{t-1},x_{t}]+b_{o})$ (11)

The weight matrices of the input, forgetting, and output gates are $W_{i},W_{f}$ , and $W_{o}$ , respectively; the bias terms of the input, forgetting, and output gates are $b_{i},b_{f}$ , and $b_{o}$ , respectively.

Figure 4.

LSTM internal structure.

In this paper, we adopt the method of sequence recombination to cut the signal into multiple sequences according to the period, and recombine the sequences into vectors $X_{1},X_{2}\ldots X_{3},X_{t}$ , reducing the input length of the LSTM, which not only extracts the temporal features of the samples, but also effectively mitigates the gradient disappearance and reduces the computation time.

2.4 Multi-scale parallel dense networks based on edge computing

It is a multi-scale parallel dense network model based on edge computing that is employed in the article, as illustrated in Fig. 5. The model consists of two parts, the data acquisition equipment on the edge side will be a compressed power quality signal for sensing, reduce the amount of transmitted data, reduce the communication cost of data transmission, the discrete cosine transform is employed as a sparse basis on edge equipment, and the observed matrix is a random Bernoulli measurement matrix. The data is transmitted to the cloud, and the cloud fusion model classifies the compressed data. The cloud model has a strong capability of potential feature extraction, and the multi-scale parallel dense network can extract the spatial characteristics of the data well, and the LSTM can extract the temporal features of the samples, and then fuse the spatial features and temporal features to classify them. The module has a strong feature extraction capability, so it can accurately classify the data without restoring it to the original data, reducing the computational complexity of the cloud. The network has fewer training parameters than other network models, thus further reducing the computational complexity of the cloud.

Table 1
Mathematical models of power quality disturbance

Disturbance	Mathematical models	Parameters
Normal	$y(t)=\{1\pm\alpha[u(t-t_{1})-u(t-t_{1})]\}\sin(2\pi ft)$	${\alpha}\leqslant 0.1$ ${T}\leqslant{t}_{2}-{t}_{1}\leqslant 10{T}$ ${f}=$ 50 Hz
Swell	$y(t)=\{1+\alpha[u(t-t_{1})-u(t-t_{1})]\}\sin(2\pi ft)$	$1.1\leqslant{\alpha}\leqslant 1.8$ ${T}\leqslant{t}_{2}-{t}_{1}\leqslant 10{T}$
Sag	$y(t)=\{1-\alpha[u(t-t_{1})-u(t-t_{1})]\}\sin(2\pi ft)$	$0.1\leqslant{\alpha}\leqslant 0.9$ ${T}\leqslant{t}_{2}-{t}_{1}\leqslant 10{T}$
Interruption	$y(t)=\{1-\alpha[u(t-t_{1})-u(t-t_{1})]\}\sin(2\pi ft)$	$0\leqslant{\alpha}\leqslant 0.1$ ${T}\leqslant{t}_{2}-{t}_{1}\leqslant 10{T}$
Fluctuation	$y(t)=[1+\alpha\sin(\beta\omega t)]\sin(2\pi ft)$	$0.1\leqslant{\alpha}\leqslant 0.2$ 5 Hz $\leqslant{\beta}\leqslant$ 10 Hz
Harmonic	$y(t)=\alpha_{1}\sin(2\pi ft)+\alpha_{3}\sin(6\pi ft)+\alpha_{5}\sin(10\pi ft)$	$0.05\leqslant{\alpha}_{1},{\alpha}_{3},{\alpha}_{5}\leqslant 0.15$ $\mathop{\sum}\nolimits{\alpha}_{i}^{2}=1$
Gap	$y(t)=\sin(2\pi ft)+K\times\textit{sign}[\sin(2\pi ft)]\times[\mu(t-t_{1})-\mu(% t-t_{2})]$	$0\leqslant{t}_{1},{t}_{2}\leqslant 0.5$ $0.01{T}\leqslant{t}_{2}-{t}_{1}\leqslant 0.05{T}$ $0.1\leqslant{K}\leqslant 0.4$
Oscillation	$y(t)=\sin(2\pi ft)+[\mu(t_{2})-\mu(t_{1})]\times\alpha^{-c\times(t-t_{1})\div% \tau}\times\sin(2\pi ft)$	$0.1\leqslant{\alpha}\leqslant 0.8$ $0.5{T}\leqslant{t}_{2}-{t}_{1}\leqslant 3{T}$ 8 ms $\leqslant{\tau}\leqslant$ 40 ms 300 Hz $\leqslant{f}_{n}\leqslant$ 900 Hz
Pulse	$y(t)=\{1-\alpha[u(t-t_{1})-u(t-t_{1})]\}\sin(2\pi ft)$	$0.05{T}\leqslant{t}_{2}-{t}_{1}\leqslant 0.1{T}$ $0.1\leqslant{\alpha}\leqslant 0.4$

Figure 5.

Multi-scale parallel DenseNet model structure based on edge computing.

3. Introduction to the data set

Power quality disturbances can be caused by a single disturbance caused by a single cause, or may be caused by a superposition of multiple single causes at the same time, forming a compound disturbance. According to IEEE Std 1159–2019, 35 different categories of raw data are formed in the simulation environment, including 9 categories of single disturbances, 21 categories of double disturbances, and 5 categories of triple disturbances, and the parameters of the data such as duration and disturbance amplitude are set according to Table 1. Each type of A total of 4000 sets of signals were generated for each type of perturbation, using a 6.4 kHz sampling frequency. A single perturbation sampling waveform of 10 cycles, the size of one perturbation data was approximately 23,479 bytes.

The raw data is then perceptually compressed. The training process is cross-validation, and the compressed data set is separated into three parts: training, validation, and testing, in the ratio of 8:1:1. During the training process, the optimal model is saved according to the accuracy of the validation set.

4. Experimental results and analysis

This paper uses Matlab to generate a total of 35 power quality simulation signals; a deep learning model is built in Pytorch to select the appropriate compression ratio, hyper-parameter selection and classification, and to compare with other deep learning models to verify the superiority of the approach presented. In the training process of the model, the learning rate decay method and the early stop mechanism are introduced to improve the training effect: when the accuracy of the model does not increase after 3 rounds of testing on the validation set, the learning rate is adjusted at a rate of 0.1 to improve the training efficiency; when the accuracy of the model does not increase after 5 rounds of testing on the validation set, the training is finished and the optimal model is saved to prevent The model is saved to prevent over-fitting.

The intelligent convergence terminal used in this paper is an edge device with a Cortex-A7 architecture single-core 4-core processor at 1.2DHz, with 2GB DDR3 and 8GB FLASH memory integrated in the periphery.

4.1 Selection of signal compression ratio

In this paper, the accuracy of the model after training is used as an evaluation index to determine the compression ratio of the model. In this paper, the data is compressed for noiseless and noisy environments of 20 dB, respectively, and the model is trained using the compressed data. On the compressed test set, the test results are displayed in Fig. 6. When the data is compressed to 70% of the original data, the compression rate increases and the classification accuracy does not improve significantly. Therefore, considering the cost of data communication as well as the model’s categorization accuracy, the compression rate of the data is set at 70%, and the data mentioned below are all compressed data with 70% compression rate.

Figure 6.

The relationship between compression rate and classification accuracy.

Figure 7.

Training results of various optimizers in a noise-free environment.

Figure 8.

Training results of various batch sizes in different noise environments.

4.2 Hyper-parameter selection

The number of training sessions and classification accuracy are utilized as assessment measures in this study to establish the model’s hyper-parameters. In the process of optimizer selection, 64 batch scales were used for training to select the best optimizer, and the initialisation parameters of each optimizer were the default parameters. The results of comparing the seven common optimizers are shown in Fig. 7.

As shown in Fig. 7, RMSprop has the highest classification accuracy of 99.67%, while its training times are only 15, which is the best classification accuracy and training times among all optimizers, so the model chooses RMSprop as the optimizer.

In order to achieve better training results for the model, the paper chooses to train the model at different batch scales, using data with different noise, and the results are shown in Fig. 8. On the whole, after 64 batch scales, the classification accuracy of the model decreases as the batch scale increases, and the trend is more obvious when the noise of the data becomes larger. Therefore, 64 is chosen as the batch scale size for the model in this paper.

4.3 Noise robustness testing

In this paper, data with 50 dB, 30 dB, and 20 dB Gaussian white noise are used to train the model, and the accuracy of the test set under different ambient noise was obtained as 99.24%, 97.95% and 93.84%, respectively. The accuracy of classification for data with 20 dB of noise is shown in the Table 2. Taking harmonics (C25) as an example, the triple disturbance consisting of the dimensional transient rise, pulse and harmonics, the model’s classification accuracy for this triple disturbance is 99.07%. As the table shows, the detection of composite disturbances containing transient rise, harmonic, notch and oscillation in high noise environment is poor, but overall, the average classification accuracy of the model for various types of single disturbance, double disturbance, triple disturbance and quadruple disturbance in high noise environment reaches 95.67%, 93.99% and 92.15%, respectively, and the model has better robustness in high noise environment (20 dB noise). The model has good robustness and still has high accuracy in a high noise environment (20 dB noise).

Table 2
The classification accuracy of the best model on the test set (SNR $=$ 20 dB)

Label	Disturbance	Accuracy (%)
C1	Normal	95.33
C2	Swell	97.32
C3	Sag	92.56
C4	Interruption	93.55
C5	Pulse	98.43
C6	Oscillation	98.41
C7	Gap	94.66
C8	Harmonic	94.01
C9	Fluctuation	96.72
C10	Swell, Pulse	98.81
C11	Swell, Harmonics	98.82
C12	Swell, Fluctuation	94.20
C13	Sag, Pulse	94.01
C14	Sag, Harmonics	92.61
C15	Sag, Fluctuation	95.22
C16	Interruption, Pulse	95.39
C17	Interruption, Harmonics	92.61
C18	Interruption, Fluctuation	94.58
C19	Pulse, Harmonics	99.35
C20	Pulse, Gap	94.83
C21	Pulse, Fluctuation	98.87
C22	Oscillation, Gap	84.32
C23	Oscillation, Harmonics	86.03
C24	Oscillation, Fluctuation	90.12
C25	Swell, Pulse, Harmonics	99.07
C26	Swell, Harmonics, Oscillation	83.83
C27	Swell, Fluctuation, Gap	94.21
C28	Sag, Pulse, Harmonics	97.32
C29	Sag, Harmonics, Oscillation	74.02
C30	Sag, Fluctuation, Gap	94.81
C31	Interruption, Pulse, Harmonics	98.32
C32	Interruption, Harmonics, Fluctuation	95.29
C33	Interruption, Fluctuation, Gap	84.97
C34	Pulse, Harmonics, Fluctuation	99.03
C35	Pulse, Gap, Fluctuation	92.81

Table 3

Comparison of the number of parameters of the model

Model	Classification accuracy	Number of training parameters	Epoch
LSTM	87.34	2,968,902	12
ResNet50	89.55	16,054,405	29
DenseNet	91.05	2,046,117	28
L-Densenet	93.84	2,245,412	26

4.4 Comparison test

To prove that the method suggested in this research is superior, different models were trained on the same data, with the training process for each model shown in Fig. 9 and then tested using a high noise (SNR $=$ 20 dB) data set. The LSTM, ResNet50, Dense-Net, and LSTM-Dense-Net (L-Dense-Net) were compared. Table 3 shows the outcomes of the comparison. Although the model suggested in this paper requires a longer training period compared to the LSTM, the accuracy is significantly improved, compared to the LSTM alone, the model presented classification accuracy for high noise data compared with the LSTM alone, the suggested model has a 6.5% improvement in classification accuracy on highly noisy data, and requires fewer parameters. The model extracts temporal features from the data and improves the classification accuracy by 2.81% compared to Dense-Net, and requires a shorter training period, but does not require many more training parameters. Compared to ResNet50, the proposed model requires fewer training parameters and achieves a 4.29% improvement in accuracy, further demonstrating the superiority of this model. Combined with Fig. 9, the training accuracy of the proposed model reaches 0.92 at Epoch $=$ 3, which is significantly faster than other models, with faster convergence and the highest classification accuracy, demonstrating that the suggested method is superior to existing deep learning approaches.

Figure 9.

The training process of each model in a noisy environment (SNR $=$ 20 dB).

5. Conclusion

The fusion model on the basis of edge computing proposed in this paper compresses the extracted power quality data on the edge side, and then transmits the compressed data to the cloud for identification and classification. In order to ensure the accuracy of identification and categorization, it can not only reduce the communication cost of data transmission, but also reduce the cache pressure and computing pressure in the cloud. The follow-up work is to increase the types of disturbed data in the data, test whether the model can still maintain a high compression rate and classification accuracy in more complex data sets, and further optimize the network model according to the results.

Footnotes

Acknowledgments

The authors acknowledge the State Grid Shanxi electric power company science and technology project (52053020000X).

References

Gao

Zhu

, et al. Power quality disturbance classification under noisy conditions using adaptive wavelet threshold and DBN-ELM hybrid model. Electric Power Systems Research. 2022; 204: 107682.

Zhang

Zhou

. Classification of power quality disturbances using visual attention mechanism and feed-forward neural network. Measurement. 2022; 188: 110390.

Yue

Feng

Wenjun

, et al. Scalable learning paradigms for data-driven wireless communication. In IEEE Communications Magazine. October 2020; 58(10): 81-87. doi: 10.1109/MCOM.001.2000143.

Saeik

Avgeris

Spatharakis

, et al. Task offloading in Edge and Cloud Computing: A survey on mathematical, artificial intelligence and control theory solutions. Computer Networks. 2021; 195(3): 108177.

Feng

Zhidi

Qinglei

, et al. FedLoc: Federated learning framework for data-driven cooperative localization and location data processing. In IEEE Open Journal of Signal Processing. 2020; 1: 187-215. doi: 10.1109/OJSP.2020.3036276.

Donoho

. Compressed sensing. IEEE Transactions on Information Theory. 2006; 52(4): 1289-1306.

Cao

Yao

, et al. Compressed Sampling and Reconstruction Method of Transient Power Quality Signal. 2020 IEEE 4th Conference on Energy Internet and Energy System Integration (EI2). IEEE, pp. 1043-1048.

Zhong

Zhang

Cai

Yang

Chen

. Power quality disturbance recognition based on multi resolution S-transform and decision tree. IEEE Access. 2019; 7: 88380-88392.

Mahela

Shaik

Khan

, et al. Recognition of complex power quality disturbances using S-transform based ruled decision tree. IEEE Access. 2020; 8: 173530-173547.

10.

Thirumala

Prasad

Jain

, et al. Tunable-Q wavelet transform and dual multiclass SVM for online automatic detection of power quality disturbances. IEEE Transactions on Smart Grid. 2016; 9(4): 3018-3028.

11.

Tang

Qiu

Zhou

. Classification of complex power quality disturbances using optimized S-transform and kernel SVM. IEEE Transactions on Industrial Electronics. 2019; 67(11): 9715-9723.

12.

Moyal

Kothari

Vaghera

, et al. Classifying Power Quality Disturbance using Time and Multiresolution Features through Artificial Neural Network. 2021 International Conference on Intelligent Technologies (CONIT). IEEE, 2021. pp. 1-5.

13.

Liu

Song

Zhao

, et al. Application of improved SVM algorithm and variational mode decomposition in power quality disturbances recognition. 2019; 364-7.

14.

Cai

Cao

, et al. Classifying power quality disturbances based on phase space reconstruction and a convolutional neural network. Applied Sciences. 2019; 9(18): 3681.

15.

Garcia

Grasso

Luchetta

, et al. A comparison of power quality disturbance detection and classification methods using CNN, LSTM and CNN-LSTM. Applied Sciences. 2020; 10(19): 6755.

16.

Sindi

Nour

Rawa

, et al. A Novel Hybrid Deep Learning Approach Including Combination of 1D Power Signals and 2D Signal Images for Power Quality Disturbance Classification. Expert Systems with Applications. 2021.

17.

Deng

, et al. Detection and Classification of Power Quality Disturbances Using Variational Mode Decomposition and Convolutional Neural Networks. 2021 IEEE 11th Annual Computing and Communication Workshop and Conference (CCWC). IEEE, 2021.

18.

Zhang

Ren

Sun

, et al. Deep residual learning for image recognition. IEEE, 2016.

19.

Mahajan

Chaudhary

. Categorical image classification based on representational deep network (RESNET). 2019 3rd International conference on Electronics, Communication and Aerospace Technology (ICECA), Coimbatore, India, 2019.

20.

Gong

Ruan

. A new convolutional network structure for power quality disturbance identification and classification in micro-grids. IEEE Access. 2020; PP(99): 1-1.

21.

Rodriguez Miguel Angel, Sotomonte John Felipe, Cifuentes Jenny, Bueno López Maximiliano. A classification method for power-quality disturbances using hilbert-huang transform and LSTM recurrent neural networks. Journal of Electrical Engineering & Technology. 2020; 16(1).

22.

Rodriguez

Felipe Sotomonte

Cifuentes

Bueno-López

. Power Quality Disturbance Classification via Deep Convolutional Auto-Encoders and Stacked LSTM Recurrent Neural Networks. 2020 International Conference on Smart Energy Systems and Technologies (SEST), 2020. pp. 1-6.

Research on power quality disturbance classification algorithm based on edge computing

Abstract

Keywords

1. Introduction

2.1 Compression perception

Table 1 Mathematical models of power quality disturbance

4. Experimental results and analysis

4.1 Selection of signal compression ratio

4.3 Noise robustness testing

Table 2 The classification accuracy of the best model on the test set (SNR = 20 dB)

Footnotes

Acknowledgments

References

Table 1
Mathematical models of power quality disturbance

Table 2
The classification accuracy of the best model on the test set (SNR $=$ 20 dB)