Skip to main content
PLOS ONE logoLink to PLOS ONE
. 2023 Feb 9;18(2):e0281482. doi: 10.1371/journal.pone.0281482

Power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNet

Lin Lin 1,*, Jie Zhang 1, Xu Gao 2, Jiancheng Shi 1, Cheng Chen 1, Nantian Huang 3
Editor: Mohamed Hammad4
PMCID: PMC9910749  PMID: 36757938

Abstract

In power fingerprint identification, feature information is insufficient when using a single feature to identify equipment, and small load data of specific customers, difficult to meet the refined equipment classification needs. A power fingerprint identification based on the improved voltage-current(V-I) trajectory with color encoding and transferred CBAM-ResNet34 is proposed. First, the current, instantaneous power, and trajectory momentum information are added to the original V-I trajectory image using color coding to obtain a color V-I trajectory image. Then, the ResNet34 model was pre-trained using the ImageNet dataset and a new fully-connected layer meeting the device classification goal was used to replace the fully-connected layer of ResNet34. The Convolutional Block Attention Module (CBAM) was added to each residual structure module of ResNet34. Finally, Class-Balanced (CB) loss is introduced to reweight the Softmax cross-entropy (SM-CE) loss function to solve the problem of data imbalance in V-I trajectory identification. All parameters are retrained to extract features from the color V-I trajectory images for device classification. The experimental results on the imbalanced PLAID dataset verify that the method in this paper has better classification capability in small sample imbalanced datasets. The experimental results show that the method effectively improves the identification accuracy by 4.4% and reduces the training time of the model by 14 minutes compared with the existing methods, which meets the accuracy requirements of fine-grained power fingerprint identification.

1 Introduction

Residential household customers’ electricity consumption continues to rise, and improving residential electricity consumption is of great significance to improving the efficiency of electrical energy utilization [1]. Providing residents with fine-grained energy usage information through Non-Intrusive load decomposition (NILD) has become an important way to effectively reduce energy waste [2]. NILD analyzes the operating status of each appliance from the total load data of residential customers which can guide customers to improve their appliance usage habits and reduce household electricity consumption [35]. At the same time, it helps the grid to build a more accurate model of household electricity consumption, enabling monitoring, control, management, and friendly interaction with electricity-using devices [6].

Power fingerprint identification is a hot issue in the field of Non-intrusive load monitoring (NILM), which relies on power fingerprint features and classifiers to identify different types of devices. Common power fingerprint characteristics typically include voltage, current, harmonics, power, V-I trajectory, etc [79]. The V-I trajectory, as the most common power fingerprint feature, represents the voltage and current waveforms of appliances in the image and has been successfully applied to load identification with good results. In [10], the voltage and current sampling data were plotted as two-dimensional V-I trajectories and then mapped to binary gray images using a normalization method. Reference [11] used V-I trajectory features simplified by elliptic Fourier descriptors, and the classification algorithm was a random forest. Reference [12] used the gramian matrix(GM) color encoding method to construct load markers with color-differentiated load signatures. In [13], a backpropagation (BP) neural network and a convolutional neural network (CNN) are used to extract and fuse the features of the load power and V-I trajectories, respectively. Fused composite features were introduced into the classifier to complete load identification. Reference [14] color-coded the V-I traces to synthesize other load characteristics, but the transient power characteristics are missing and the traces do not reflect the transient characteristics of the device. Research on classification models based on V-I trajectory features is becoming increasingly mature, but the following problems still exist: The original V-I trajectory can only convey trajectory shape information, but cannot reflect other information such as the power of the appliance. Moreover, due to the similarity between different kinds of appliances and working principles, the existing models cannot fully extract the V-I trajectory feature information. Therefore, there is still room for further improvement in the accuracy of its load identification.

Nowadays, scientists and researchers used machine learning (ML) and deep learning (DL) models in several applications including agriculture [15, 16], environment [1720], and power fingerprint identification. Machine learning is often applied to feature extraction and classification of power fingerprints, such as k-nearest neighbors, support vector machines, decision trees, and random forests. These methods are less computationally intensive, but the identification correct rate is lower. Recently, deep learning has achieved good results in the field of power fingerprint identification, such as CNN, RNN, etc. Meanwhile, researchers propose to construct V-I trajectory images with the help of color coding methods to convert power fingerprint identification into an image classification task in which deep learning excels. However, compared to machine learning, deep learning-based classifiers rely on large-scale training data and longer training time, which limits the application of deep learning.

For V-I trajectory classification, the small amount of load data for specific user results in a poorly trained model with low accuracy or low generalization capability [21]. NILD does not have sufficient computational resources to retrain a complex load appliance identification model, which is costly and time-consuming. Training the model requires a large amount of labeled data, which is difficult to achieve in reality. In recent years, transfer learning has been widely used to overcome the limitations of traditional machine learning. Transfer learning does not require the same distribution assumptions for training and test data as traditional machine learning. This avoids the labor and material costs of relabeling the acquired data in traditional machine learning [22]. The main idea of transfer learning is to use the knowledge from the existing source domain and then transfer it to the target domain to complete its classification. Reference [23] introduced transfer learning to NILD and achieved good results for appliance transfer for the same dataset and cross-domain migration for different datasets. Reference [24] used the AlexNet network for the load identification of V-I trajectory images and improved the accuracy of load identification using transfer learning. Reference [25] follow the pretraining to reduce the computation in modeling training, and enhance the transferability of the model. V-I trajectories as a load feature that can be converted to image representation are promising for use in load identification in NILD domains.

In addition, there is a class imbalance in the NILM dataset, which can have an impact on the classification performance of the classifier. When the classifier is more biased toward majority class samples, then the identification accuracy of minority samples decreases and the model lacks generalization [26]. Usually, simple methods such as oversampling and undersampling are often used to solve the problem of training unbalanced datasets [27]. However, the oversampling methods for minority classes change the distribution characteristics of the minority class samples [28]. The undersampling method for the majority class loses valuable information about the minority class samples [29]. The above-mentioned methods are difficult to deal with the data imbalance of the V-I trajectory image samples. Reference [30] used a data balancing method based on PixelCNN++ and sample information entropy and used it for V-I trajectories. However, PixelCNN++ has disadvantages such as random generation of wrong images and long training time. Re-weighting [31] is a method that assigns different weights to different classes during the training process. It not only improves the learning ability of the model for a few classes but also reduces the learning ability of the model for most classes.

To solve these problems, this paper proposes a power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNet34. Experimental results on the PLAID [32] dataset show that the method in this paper improves the accuracy of power fingerprint identification to over 97% compared to other methods. Overall, our main contributions are as follows:

  1. Use color encoding to add more feature information to the V-I trajectory and improve the identification of similar electrical appliances. It solves the problem of insufficient feature information in single feature recognition devices.

  2. The transferred CBAM-ResNet34 model was constructed to fully extract V-I features and was applied to small samples of V-I trajectory samples. The CBAM and model transferred methods effectively improve the extraction capability of V-I trajectories and sufficiently reduce the training time of the model.

  3. For the first time, image-level CB loss is introduced to reweight the loss function to obtain better V-I trajectory identification in unbalanced datasets.

The remainder of this study is organized into the following sections: The second part outlines the overall power fingerprint identification process and introduces the method of acquiring color V-I trajectory. Section 3 explains the construction and improvement process of the proposed model. The validity of the method is verified by the PLAID dataset in Sections 4 and 5. The last section summarizes the work of this paper.

2 Materials and methods

2.1 The power fingerprint identification process

The power fingerprint identification process includes data acquisition, data pre-processing, feature extraction, and load identification. The specific process of power fingerprint identification in this study is shown in Fig 1. First, collect high-frequency voltage and current data of the appliance and extract the original V-I trajectory of the appliance from the voltage and current data. Then, the construction of color V-I trajectory images using color coding. Finally, the color V-I trajectory images are fed into the transferred CBAM-ResNet34 for training. The SoftMax layer classifies the new appliance categories and completes the power fingerprint identification.

Fig 1. The process structure of power fingerprint identification.

Fig 1

2.2 Raw data analysis

The primary function of data acquisition is to sample and store the voltage and current waveforms of various household appliances, including (fridges, air conditioners, fans, microwave ovens, etc.) using a collection system [32]. The collection system is usually composed of a current clamp, voltage probe, and high-frequency oscilloscope. The sampling data for the fridge model during the steady-state operation is shown in Fig 2. Therefore, the data are mainly intercepted when the equipment is in steady-state and transient-state operation for feature extraction and energy use analysis.

Fig 2. One instance of a fridge in PLAID.

Fig 2

(a)The instantaneous voltage of the Fridge over 1 s. (b)The instantaneous current of the Fridge over 1 s. (c)The instantaneous power of the Fridge over 1 s.

2.3 Construction of color V-I trajectory images based on color coding

The original V-I trajectory image is a single-channel, two-dimensional pixel matrix. Much valuable information is lost in the original V-I trajectory images. In this section, the original V-I trajectory image is combined with the color coding technique to construct a unique power fingerprint label.

First, The V-I traces of different cycles vary slightly due to load fluctuations or noise. When only one waveform cycle is used, this phenomenon inevitably leads to misclassification [24]. Therefore, in this paper, multiple cycles of data are extracted to plot V-I traces to accommodate the dynamic changes in the load.

Then, considering that pixel discontinuity may occur in the mapping process of V-I trajectory images, it is not conducive to subsequent load identification. For this purpose, the traditional mapping method is improved using a bilinear interpolation technique in Fig 3. The specific steps of the bilinear interpolation technique are as follows:

Dk=vk+1vkΔv2+ik+1ikΔi2>1 (1)
vk+t=vk+vk+1vkTk+1t (2)
ik+t=ik+ik+1ikTk+1t (3)

where vk and ik represent the kth data point of v and i respectively. Tk = floor(Dk) is the number of interpolation points to be added between the kth and (k+1)th sampling points, floor(⋅) represents rounding down.vk+t,ik+t is the tth interpolation point of the fill t = 1,2,⋯,Tm.

Fig 3. The binary V- I trajectory mapping for a fluorescent lamp(N = 32).

Fig 3

(a) Linear interpolation. (b) Bilinear interpolation.

Finally, To cover the load identification information as much as possible with its characteristic parameters, the current, instantaneous power, and trajectory momentum information are added to the original V-I trajectory image using color coding to obtain a color V-I trajectory image in the RGB color space. Each channel corresponds to a two-dimensional matrix, and each matrix element can vary continuously from zero to one. V-I trajectories were plotted for 11 categories of electrical equipment randomly selected from the PLAID dataset as shown in Fig 4. The specific steps of the color encoding method are as follows.

Fig 4. Color V-I trajectory images of 11 types of appliance loads in the PLAID dataset (N = 32).

Fig 4

  • Red channel: The shape of the V-I trajectory depends heavily on the load current that reflects the physical characteristics of the device. The average current of the device during stable operation is populated in the R channel, aiming to extract a more stable V-I trajectory and ensure the classification effect.
    Rm=imiminimmaximin×N,vmvminvmmaxvmin×N (4)
    Where imax, imin are the maximum and minimum values of the current sampling values, respectively; vmax, vmin are the maximum and minimum values of the voltage sampling value, respectively; ⌊ ⌋ represents rounding down.
  • Green channel: The V-I trajectory characteristics in the transient state are somewhat different from those after the stable operation. Therefore, in this paper, the average instantaneous power of the device in the transient state is filled in the G channel. The transient V-I trajectory characteristics are also an important characteristic for different loads.
    Gm=1K1KWk (5)
    Where K represents the number of cycles of the load in the transient state. Wk represents the corresponding pixel value in the V-I trajectory feature at the kth instantaneous state point value.
  • Blue channel: the rate of change of voltage and current during the stabilization cycle varies from device to device. V-I trajectory has a loop direction, reflecting the phase relationship between current and voltage. To capture the motion information the of V-I trajectory, the blue channel is plotted by the voltage and current of the appliance for continuous 20 cycles in steady-state operation.
    B(m)=arg(vm+1vmvmmax,im+1imimmax)2π*20,m=1,2,3,M (6)
    The above color coding steps fused the Red, Green, and Blue channels to obtain the corresponding color V-I trajectory images in RGB spaces. Color encoding increases the uniqueness of the V-I trajectory and provides better identification.

3 Power fingerprint identification based on transferred CBAM-ResNet34

3.1 CBAM-ResNet34 model

The ResNet model effectively mitigates degradation and gradient disappearance during model training by stacking the residual structures [33]. Networks with fewer layers lack certain feature representation capabilities. On the other hand, the increase in the number of network layers is accompanied by an increase in the number of parameters and computations, which makes the network training slower. However, the residual structure in the ResNet34 neural network has a limited ability to extract V-I trajectory features and cannot ignore irrelevant information in the model training. By applying spatial attention and channel attention, the CBAM module can amplify the weights of effective channels in the feature layer. This enhances the feature representation of V-I trajectories and improves their saliency, laying the foundation for accurate classification of subsequent load recognition.

The CBAM module represents the attention mechanism module of the convolutional module, which is an attention mechanism module that combines spatial and channel [34]. Compared to senet, which only focuses on channels, CBAMs achieve better results. They save parameters and computational power and can be integrated into existing networks as plug-and-play modules.

The details of the channel attention module (CAM) and spatial attention module (SAM) are shown in Fig 5. The CAM focuses more on the more critical parts of the image and ignores extraneous information. First, the input features are processed in parallel by averaging pooling and maximum pooling. Subsequently, the multilayer perceptron (MLP) forwards both types of data with a hidden layer. Finally, the output features are merged by element summation.

Fig 5. The structure of the Convolutional Block Attention Module (CBAM) model.

Fig 5

The CAM can be expressed as follows.

Mc(F)=σMLPAvgPool(F)+MLPMaxPool(F)=σW1W0(Favgc)+W1W0(Fmmaxc) (7)

where W0 and W1 are learnable weights and σ is an S-shaped function.

SAM is a complement to CAM and its main purpose is to discover the most meaningful information after CAM processing. First, the input features are processed serially through the mean and maximum pools. This information is then forwarded by the convolutional layer. The final mathematical representation is as follows.

Ms(F)=σfAvgPool(F);MaxPool(F) (8)

where f denotes the convolution operation.

In this paper, the position of CBAM is added to each residual block to reduce the influence of redundant features with the help of an attention mechanism and improve the accuracy and timeliness of subsequent recognition. As in Fig 6(a) and 6(b), the activation function of ReLU was used before each weighting layer. After extracting the CNN, the CBAM layer extracts the most critical information through the channel and spatial dimensions. The attention mechanism is more likely to learn effective features in shallow networks, and the improvement effect of the attention mechanism module on the performance of deeper networks is less obvious, and it is difficult for the attention mechanism to learn effective features.

Fig 6.

Fig 6

(a)The ResNet network residual structure. (b)The structure of the CBAM-ResNet model.

3.2 Model transfer method

The ResNet model was initially designed to identify the ImageNet datasets. However, the learned high-level abstract features can still be used to assist in identifying the V-I trajectories of different types of appliances. We used a model-transferring method and added the CBAM attention module for the ResNet networks, as shown in Fig 7. The specific training steps were as follows:

Fig 7. The experimental scheme of model transferring.

Fig 7

  • The ResNet34 neural network was pre-trained using the ImageNet dataset, and all layers except the last fully connected layer were extracted from the pre-trained ResNet34.

  • The Simple Attention module of CBAM was introduced, which was added to each residual structure module of ResNet34. The last full connection layer was replaced with a new one, and the layers were transferred to the new electrical equipment identification task.

  • Input color V-I trajectory image. The color V-I trajectory image size was adjusted to the same size as the input neuron of ResNet34.

  • Output the categories of the electric loads. The new full connection layer is adjusted to the same number of electrical equipment identification categories.

  • All the parameters are retained until the termination condition is satisfied.

3.3 Class-balanced loss

In power fingerprint identification, data set imbalance is another issue of concern. The usage frequency of household appliances varies greatly, which is reflected in the data set as a significant variation in the sample sizes of different categories. To solve this problem, it is usually necessary to expand some samples so that the number in different categories is the same. The class-balanced loss is an image-level class-balanced loss function that can be implemented by inverse weighting the loss function by the number of effective classes.

Therefore, to improve the identification accuracy of color V-I trajectory images, we rebalance the loss by weighting the loss function using the number of valid samples per class to improve the loss. A weighting factor is introduced in the class-balanced loss [35]. We apply it to softmax cross-entropy loss as follows:

Assume that the predicted output of the model for all classes is z = [z1, z2, ⋯ zc]T, where C is the total number of classes. The softmax function treats each class as mutually exclusive and calculates the probability distribution of all classes as

pi=exp(zi)j=1Cexp(zj),i1,2,,C (9)

Given a sample with the class label y, the SM-CE loss for this sample is written as:

CE(z,y)=logexp(zy)j=1Cexp(zj) (10)

Suppose the class y has ny training samples, the class-balanced SM-CE loss is:

CB(z,y)=1β1βnylogexp(zy)j=1Cexp(zj) (11)

The class-balanced cross-entropy loss uses a balancing factor ω to increase the weight of minority samples in the target loss and decrease the weight of majority samples in the target loss, which makes the classifier focus more on the features of minority categories in training, thus improving the category imbalance problem in the dataset and increasing the accuracy of recognition.

The 2 appliances in the PLAID dataset, Microwave (majority class) and Washing (minority class), have different sample sizes as shown in Fig 8. The model trained based on these samples is biased towards the majority sample (black solid line). The class balancing loss (Cbloss) is introduced to reweight the loss by reversing the effective number of samples. The generalization performance of the model can be improved when trained with the proposed class balance loss (blue dashed line).

Fig 8. The schematic diagram of the reweighted optimal classifier model.

Fig 8

4 Experiments

4.1 Dataset selection

In the actual example, Python 3.8 and Pytorch 1.3.0 deep learning frameworks are used in the actual arithmetic example. The hardware platform was GeForceRTX3090, and the software platform was Linux OS Ubuntu 18.04. Hardware acceleration using GPUs when training deep learning models. The proposed load identification method was tested using the PLAID dataset. The dataset recorded 1793 sets of voltage and current data of 312 appliance loads in 11 categories from 55 households at a sampling frequency of 30 kHz [32]. The number of class samples and examples of the 11 classes of appliance loads are listed in Table 1. 20% of the data were randomly selected as the test set, and 80% were selected as the training set.

Table 1. The PLIAD dataset of appliance types and instance statistics.

Appliance Type Appliances Instances
Air Conditioner 21 208
Compact Fluorescent Lamp 40 220
Fan 30 210
Fridge 30 90
Hairdryer 37 248
Heater 13 85
Incandescent Light Bulb 30 148
Laptop 41 207
Microwave 37 229
Vacuum 15 73
Washing Machine 18 75
Total 312 1793

4.2 Parameter optimization

The number of iterations and pixels of the color V-I trajectory image was optimized to obtain the best parameters for load identification. When analyzing the influence of one parameter on the identification result, the other parameter was fixed. After determining the input and output of the model transfer, the new network model was trained until the termination condition was reached. The cross-entropy function was chosen as the minimum loss function, and the optimizer was chosen as Adam. The initial learning rate was configured as 1 × 10−4 to reduce the learning rate of the transferred layer. Meanwhile, the learning rate of the new fully connected layer was increased to 1 × 10−3.

To quantitatively analyze the effect of the number of iterations on the method, the variation curves of the identification accuracy and loss value with the number of iterations during a specific test were plotted, as shown in Fig 9. The number of iterations for the entire test was set as 300. It can be seen that as the number of iterations increases, the identification accuracy increases rapidly, and the loss value decreases. When the number of iterations reaches 200, both the loss value and the identification accuracy stabilize, and the classification performance of the method no longer changes significantly. Based on the above analysis, the number of iterations was set to 200, and a more satisfactory identification result could be obtained in a relatively short period.

Fig 9. The comparison of several iterations.

Fig 9

Fig 10 shows the training process of the CBAM-ResNet34 model and the transferred CBAM-ResNet34 model that was used to validate the feasibility and effectiveness of the model transferring method. It can be seen that the convergence rate of the transferred CBAM-ResNet34 model is faster than that of the CBAM-ResNet34 model. The pre-trained model maintains a high accuracy at the beginning of training. When the number of iterations reaches around 10, the identification accuracy exceeds 90% and increases slowly in subsequent training. The method of model transferring improves the training speed and accuracy of power fingerprint identification while retaining prior knowledge, as the abstract feature extraction capability of the pre-trained model can be shared across image identification.

Fig 10. The training and validation accuracy of CBAM-ResNet34 and transferred CBAM-ResNet34 models are illustrated for comparison.

Fig 10

A box plot of the load identification accuracy at different resolutions is shown in Fig 11 to select the appropriate pixel size for the color V-I trajectory image. It can be seen that the overall identification accuracy of load recognition is best when the resolution N of the color V-I trajectory image is 128. However, with increasing image resolution, N, the identification accuracy gradually decreased. When the image resolution is small, the color V-I trajectory image is not clear enough, and there is insufficient information about the features. When the resolution of the image is increased to a specific range, more feature information is contained, and CBAM-ResNet34 can extract the feature information. In addition, an image resolution that is too high can also increase the effect of random noise and perturbations, resulting in sharp color V-I trajectory images that are not conducive to subsequent feature extraction. This causes the accuracy of the load identification to decrease rather than increase. Given this, the resolution N of the color V-I image was set to 128.

Fig 11. The selection of parameter N.

Fig 11

5 Result and discussion

5.1 Evaluation metrics

To analyze the classification results of this method more comprehensively rather than simply using classification accuracy for judgment, a confusion matrix [36, 37] is adopted as the evaluation standard for the classification results. The confusion matrix of each comparison classification result through visual analysis of the data is shown in Fig 12. The numbers in the graph represent the current sample size, and the percentages represent the current sample size as a percentage of the total sample size. The evaluation index Pre represents the precision (shown as a green percentage in the last column in Fig 12). The evaluation index Rre represents the recall (shown as a green percentage in the last row in Fig 12). Fscore represents a reconciled average assessment indicator for precision and recall. These are calculated as follows [38, 39]:

Pre=TpTp+Fp (12)
Rre=TpTp+FN (13)
Fscore=2PreRrePre+Rre (14)

where Tp means that the actual values are valid and classified as positive, Fp means that the real value is false and is classified as positive, and FN means that the actual values are valid and unfavorable.

Fig 12. The confusion matrix for appliance loads of the PLAID dataset.

Fig 12

(a) Confusion matrix based on color V-I trajectory image before model transfer. (b) Confusion matrix based on color V-I trajectory image after model transfer.

5.2 The performance of V-I trajectory with color encoding and CBAM

Fig 12(a) and 12(b) show that when color V-I trajectory images were used for load identification simultaneously, the accuracy rate before the model transfer was only 94.4%. In comparison, the accuracy rate after the model transfer was as high as 97.5%. The strategy of model transfer is to realize the transfer from the multi-classification of large datasets in the source domain (such as the ImageNet dataset) to specific learning tasks in the target domain by adjusting and optimizing the structure and parameters of the pre-trained model. Applying the underlying features learned by ResNet34 on large datasets to the V-I trajectory image classification problem reduces the complexity of the model optimization. This enables the network to perform well even in small sample datasets, improving the accuracy of load identification and reducing the training time.

As shown in Fig 13, the F-scores of the air conditioner, washing machine, and heater are lower than Fmacro (the average of all F-scores of appliance loads). The fridge, air conditioner, and washing machine are multi-state appliances with limited working modes [40]. For example, air conditioners have a variety of working conditions, such as heating and ventilation modes, which are easily confused with hairdryers and heating machines. The fridge has a cooling fan and compressor motor, easily be confused with washing machines and fans. Because these appliances switch between multiple operating modes, their electrical characteristics are complex, making load identification more difficult.

Fig 13. The F-score (%) for appliance loads of the PLAID dataset.

Fig 13

To verify the effectiveness of the CBAM module, the classification results of transferred ResNet34 and transferred ResNet34-CBAM are compared on the PLAID dataset. It can be found that the transferred ResNet34 CBAM has better results than the transferred ResNet34 in all five evaluation metrics as shown in Fig 14. The transferred ResNet model with the CBAM module can better classify the images of V-I trajectories formed by 11 different classes of electrical devices compared to the model without CBAM.

Fig 14. The evaluated metrics between transferred ResNet34 models with and without CBAM.

Fig 14

5.3 Performance comparison of different models

Different models differ in their accuracy of power fingerprint identification. To further illustrate the advantages of the proposed method, AlexNet, VGG16, and GoogLeNet were selected for comparison. The experimental objects all used the color V-I trajectory image of 128 × 128 pixels. As shown in Table 2, the transferred CBAM-ResNet34 improved the power fingerprint identification accuracy significantly, and the training time was shorter than that of the above model. The comparison of the identification accuracy of the above-mentioned three models verifies the validity of CBAM-ResNet34 and the model transfer in power fingerprint identification.

Table 2. Comparison of identification accuracy and training time between different training models on the PLAID dataset.

Training Model PLAID Dataset
Accuracy (%) Time(min)
AlexNet 92.86 27
VGG16 93.50 23
GoogLeNet 94.29 20
ResNet34 95.48 17
CBAM-ResNet34 96.38 18
Transferred CBAM-ResNet34 97.49 13

5.4 Performance comparison of different data balancing algorithms

From Table 1, in the PLAID dataset, Class 5 has 248 samples, while Class 10 has only 73 samples. To solve this problem, it is usually necessary to expand some samples so that the number in different categories is the same. To solve the data imbalance problem in V-I trajectory identification, Class-Balanced(CB) loss is introduced, which is achieved by inverse weighting the loss function by the number of effective classes. Using transferred CBAM-ResNet34 as the base model, the methods in this paper were compared with the original data, random oversampling, SMOTE [41], and SVMSMOTE [42] data balancing algorithms on the PLAID dataset, respectively. Use the F1-score as an evaluation criterion. The experiment was repeated five times to take the mean value and the experimental results are shown in Table 3.

Table 3. The F1 for different data balance algorithms using transferred CBAM-ResNet34.

Original data SMOTE Oversampling SVMSMOTE CBAM-ResNet34
AC 0.896 0.901 0.871 0.939 0.940
Bulb 0.976 0.956 1.000 1.000 1.000
CFL 0.933 0.926 0.921 0.946 0.993
Fan 0.946 0.955 0.963 0.913 0.989
Fridge 0.895 0.912 0.984 0.986 0.914
Hairdryer 0.926 0.936 0.961 0.967 0.980
Heater 1.000 0.988 0.908 0.985 1.000
Laptop 0.932 0.947 0.935 0.965 0.988
Microwave 0.951 1.000 0.983 0.985 0.978
Vacuum 0.939 0.942 0.962 0.964 1.000
Washing 0.928 0.930 0.926 0.929 0.965
Ave F1 0.938 0.944 0.946 0.961 0.977

From Table 3, we can see that the other three methods can effectively improve the F1-score compared with the original data for training. The method in this paper can significantly improve the F1-score of a few sample categories, such as washing machines and vacuum cleaners, compared with the other three methods, the class-balanced loss is used to weight the Softmax maximum cross-entropy loss (SM-CE) function to increase the weight of minority class samples in the target loss, which makes the classifier focus more on the minority class features in training.

5.5 Performance comparison of different methods

To compare the proposed method with existing methods, the critical differences in the load signature, training model, identification accuracy, and other aspects are listed in Table 4. Experiments were conducted on PLAID datasets. Experiment 1 used V-I trajectory features simplified by elliptic Fourier descriptors(EFD), and the classification algorithm was a random forest. Experiment 2 uses color V-I trajectory as loading features, and the classification algorithm is AlexNet. In Experiment 3, the color V-I trajectory features and power features of the loads were fused to form composite features by feature fusion, and the classification algorithm adopted was BPNN. Experiment 4 color encoded the V-I trajectories using the gramian matrix with the improved AlexNet classification algorithm.

Table 4. The comparison of the proposed method and other power fingerprint identification methods.

Load Signature classification algorithm Accuracy(%) Fscore(%) Time(min)
EFD Simplified V-I trajectory RF 77.30 77.60 2
Color V-I trajectory AlexNet 92.86 93.02 27
Power + color V-I trajectory BPNN 94.06 94.40 23
GM-V-I trajectory Improved AlexNet 95.08 95.36 20
Improved color V-I trajectory Transferred CBAM-RestNet34 97.46 97.69 13

As can be seen from Table 4, Experiment 1 only relied on harmonic characteristics for classification and failed to use other features for auxiliary identification. Its identification accuracy was the worst despite the short training time. The color V-I trajectory image adopted in Experiment 2 has a low resolution, and many details are lost. Its identification effect on heaters and other multi-state loads is poor. The load characteristics used in Experiment 3 include the power characteristics of the load, which solves the drawback of losing the power characteristics of the V-I trajectory, but the recognition accuracy and training time is to be further improved. The physical information used in Experiment 4 is the loading characteristics at stable cycles and does not include the loading characteristics at non-stable periods. In the case of color coding using the gramian matrix, each element of the matrix needs to be operated once, which is a relatively large computational effort. Compared to the above experimental methods, the method in this study was improved by feature extraction and model training. Color-encoded V-I trajectory images and transferred CBAM-ResNet34 were used as the load feature and training model, respectively, which significantly enhanced the identification ability of the algorithm and reduced the model training time. Therefore, the load identification effect of the method in this study was better than those of the other four methods.

6 Conclusions

This paper proposes power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNet34. The performance of the algorithm was verified with the PLAID dataset. Compared with the traditional method, the new method has the following advantages:

  • The color encoding is used to construct color V-I trajectory images with different channels in high-dimensional space with higher differentiation. The current, instantaneous power, and trajectory momentum information comprehensively and three-dimensionally reveal the essential properties of various types of electrical equipment and improve the identification accuracy of power fingerprints.

  • The CBAM module is introduced and combined with the model transfer method to construct the ResNet34-CBAM model. The CBAM module can amplify the weights of effective channels in the feature layer and fully extract the relevant features of V-I trajectories. The model transfer reduces the training time required for model training and better improves the accuracy of power fingerprint identification in the case of small load data of specific customers.

  • The CB loss is introduced to reweight the SM-CE loss function to solve the problem of data imbalance in V-I trajectory identification. Compared with other data balancing methods, the improved class-balanced CBAM-ResNet34 in this paper has better classification performance in small-sample unbalanced NILM datasets.

The experimental results show that the method can better improve the accuracy of power fingerprint identification on small sample imbalanced datasets and reduce the time of classification compared with other advanced identification methods. This validates the effectiveness of the method proposed in this paper.

However, the method proposed in this paper still has some problems to be solved. The identification effect of multi-state load in this paper can be further improved, and more advanced power fingerprint identification models are needed to solve the problem in the future. In addition, the actual engineering implementation faces difficulties in constructing sample libraries and high data sampling frequency requirements. Therefore, a lot of related research work is still needed.

Supporting information

S1 Data

(XLSX)

Data Availability

All relevant data are within the paper and its Supporting information files.

Funding Statement

This work was supported by Natural Science Foundation of Jilin Province (YDZJ202101ZYTS189). The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

References

  • 1.Mohamed MA, Eltamaly AM, Alolah AI. PSO-Based Smart Grid Application for Sizing and Optimization of Hybrid Renewable Energy Systems. Du W-B, editor. PLoS ONE. 2016. Aug 11; 11(8): e0159702. doi: 10.1371/journal.pone.0159702 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Moreno Jaramillo AF, Laverty DM, Morrow DJ, Martinez del Rincon J, Foley AM. Load modelling and non-intrusive load monitoring to integrate distributed energy resources in low and medium voltage networks. Renewable Energy. 2021. Dec; 179: 445–66. doi: 10.1016/j.renene.2021.07.056 [DOI] [Google Scholar]
  • 3.Balletti M, Piccialli V, Sudoso AM. Mixed-Integer Nonlinear Programming for State-Based Non-Intrusive Load Monitoring. IEEE Trans Smart Grid. 2022. Jul; 13(4): 3301–14. doi: 10.1109/TSG.2022.3152147 [DOI] [Google Scholar]
  • 4.Zhou M, Shao S, Wang X, Zhu Z, Hu F. Deep Learning-Based Non-Intrusive Commercial Load Monitoring. Sensors. 2022. Jul 13; 22(14): 5250. doi: 10.3390/s22145250 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Zhou X, Feng J, Li Y. Non-intrusive load decomposition based on CNN–LSTM hybrid deep learning model. Energy Reports. 2021. Nov; 7: 5762–71. doi: 10.1016/j.egyr.2021.09.001 [DOI] [Google Scholar]
  • 6.Kang J-S, Yu M, Lu L, Wang B, Bao Z. Adaptive Non-Intrusive Load Monitoring Based on Feature Fusion. IEEE Sensors J. 2022. Apr 1; 22(7): 6985–94. doi: 10.1109/JSEN.2022.3155883 [DOI] [Google Scholar]
  • 7.Cannas B, Carcangiu S, Carta D, Fanni A, Muscas C. Selection of Features Based on Electric Power Quantities for Non-Intrusive Load Monitoring. Applied Sciences. 2021. Jan 7; 11(2): 533. doi: 10.3390/app11020533 [DOI] [Google Scholar]
  • 8.Zheng Z, Chen H, Luo X. A Supervised Event-Based Non-Intrusive Load Monitoring for Non-Linear Appliances. Sustainability. 2018. Mar 28; 10(4): 1001. doi: 10.3390/su10041001 [DOI] [Google Scholar]
  • 9.Mulinari BM, Nolasco L da S, Oroski E, Lazzaretti AE, Linhares RR, Renaux DPB. Feature Extraction of V–I Trajectory Using 2-D Fourier Series for Electrical Load Classification. IEEE Sensors J. 2022. Sep 15; 22(18): 17988–96. doi: 10.1109/JSEN.2022.3194999 [DOI] [Google Scholar]
  • 10.Du L, He D, Harley RG, Habetler TG. Electric Load Classification by Binary Voltage–Current Trajectory Mapping. IEEE Trans Smart Grid. 2016. Jan; 7(1): 358–65. doi: 10.1109/TSG.2015.2442225 [DOI] [Google Scholar]
  • 11.De Baets L, Develder C, Dhaene T, Deschrijver D. Automated classification of appliances using elliptical fourier descriptors. In: 2017 IEEE International Conference on Smart Grid Communications (SmartGridComm) [Internet]. Dresden, Germany: IEEE; 2017 [cited 2022 Nov 4]. p. 153–8.
  • 12.Chen J, Wang X. Non-intrusive Load Monitoring Using Gramian Angular Field Color Encoding in Edge Computing. Chinese J of Electronics. 2022. Jul; 31(4): 595–603. doi: 10.1049/cje.2020.00.268 [DOI] [Google Scholar]
  • 13.Li Y, Wang H, Yang J, Wang K, Qi G. A non-intrusive load monitoring algorithm based on multiple features and decision fusion. Energy Reports. 2021. Nov; 7: 1555–62. doi: 10.1016/j.egyr.2021.09.087 [DOI] [Google Scholar]
  • 14.Huan J., Wang C., Hong H., 2021. Non-intrusive load monitoring method based on color encoding and deep learning. Science Technology and Engineering 21, 8901–8908. [Google Scholar]
  • 15.Anul Haq M. Planetscope Nanosatellites Image Classification Using Machine Learning. Computer Systems Science and Engineering. 2022; 42(3): 1031–46. doi: 10.32604/csse.2022.023221 [DOI] [Google Scholar]
  • 16.Anul Haq M. CNN Based Automated Weed Detection System Using UAV Imagery. Computer Systems Science and Engineering. 2022; 42(2): 837–49. doi: 10.32604/csse.2022.023016 [DOI] [Google Scholar]
  • 17.Anul Haq M. SMOTEDNN: A Novel Model for Air Pollution Forecasting and AQI Classification. Computers, Materials & Continua. 2022; 71(1): 1403–25. doi: 10.32604/cmc.2022.021968 [DOI] [Google Scholar]
  • 18.Anul Haq M. CDLSTM: A Novel Model for Climate Change Forecasting. Computers, Materials & Continua. 2022; 71(2): 2363–81. doi: 10.32604/cmc.2022.023059 [DOI] [Google Scholar]
  • 19.Anul Haq M, Khadar Jilani A, Prabu P. Deep Learning Based Modeling of Groundwater Storage Change. Computers, Materials & Continua. 2022; 70(3): 4599–617. doi: 10.32604/cmc.2022.020495 [DOI] [Google Scholar]
  • 20.Haq MA, Rahaman G, Baral P, Ghosh A. Deep Learning Based Supervised Image Classification Using UAV Images for Forest Areas Classification. J Indian Soc Remote Sens. 2021. Mar; 49(3): 601–6. doi: 10.1007/s12524-020-01231-3 [DOI] [Google Scholar]
  • 21.Yu M, Wang B, Lu L, Bao Z, Qi D. Non-Intrusive Adaptive Load Identification Based on Siamese Network. IEEE Access. 2022; 10: 11564–73. doi: 10.1109/ACCESS.2022.3145982 [DOI] [Google Scholar]
  • 22.Pan SJ, Yang Q. A Survey on Transfer Learning. IEEE Trans Knowl Data Eng. 2010. Oct; 22(10): 1345–59. doi: 10.1109/TKDE.2009.191 [DOI] [Google Scholar]
  • 23.D’Incecco M, Squartini S, Zhong M. Transfer Learning for Non-Intrusive Load Monitoring. IEEE Trans Smart Grid. 2020. Mar; 11(2): 1419–29. doi: 10.1109/TSG.2019.2938068 [DOI] [Google Scholar]
  • 24.Liu Y, Wang X, You W. Non-Intrusive Load Monitoring by Voltage–Current Trajectory Enabled Transfer Learning. IEEE Trans Smart Grid. 2019. Sep; 10(5): 5609–19. doi: 10.1109/TSG.2018.2888581 [DOI] [Google Scholar]
  • 25.Wang L, Mao S, Nelms RM. Transformer for Nonintrusive Load Monitoring: Complexity Reduction and Transferability. IEEE Internet Things J. 2022. Oct 1; 9(19): 18987–97. doi: 10.1109/JIOT.2022.3163347 [DOI] [Google Scholar]
  • 26.Zhang J, Chen X, Ng WWY, Lai CS, Lai LL. New Appliance Detection for Nonintrusive Load Monitoring. IEEE Trans Ind Inf. 2019. Aug; 15(8): 4819–29. doi: 10.1109/TII.2019.2916213 [DOI] [Google Scholar]
  • 27.Park S, Park H. Combined oversampling and undersampling method based on slow-start algorithm for imbalanced network traffic. Computing. 2021. Mar; 103(3): 401–24. doi: 10.1007/s00607-020-00854-1 [DOI] [Google Scholar]
  • 28.Vo MT, Nguyen T, Vo HA, Le T. Noise-adaptive synthetic oversampling technique. Appl Intell. 2021. Nov; 51(11): 7827–36. doi: 10.1007/s10489-021-02341-2 [DOI] [Google Scholar]
  • 29.Koziarski M. Radial-Based Undersampling for imbalanced data classification. Pattern Recognition. 2020. Jun; 102: 107262. doi: 10.1016/j.patcog.2020.107262 [DOI] [Google Scholar]
  • 30.Wang S, Chen H, Guo L, Xu D. Non-intrusive load identification based on the improved voltage-current trajectory with discrete color encoding background and deep-forest classifier. Energy and Buildings. 2021. Aug; 244: 111043. doi: 10.1016/j.enbuild.2021.111043 [DOI] [Google Scholar]
  • 31.Wang L, Wang C, Sun Z, Cheng S, Guo L. Class Balanced Loss for Image Classification. IEEE Access. 2020; 8: 81142–53. doi: 10.1109/ACCESS.2020.2991237 [DOI] [Google Scholar]
  • 32.Gao J, Giri S, Kara EC, Bergés M. PLAID: a public dataset of high-resoultion electrical appliance measurements for load identification research: demo abstract. In: Proceedings of the 1st ACM Conference on Embedded Systems for Energy-Efficient Buildings [Internet]. Memphis Tennessee: ACM; 2014 [cited 2022 Nov 4]. p. 198–9.
  • 33.Liu C, Yin Y, Sun Y, Ersoy OK. Multi-scale ResNet and BiGRU automatic sleep staging based on attention mechanism. V E S, editor. PLoS ONE. 2022. Jun 16; 17(6): e0269500. doi: 10.1371/journal.pone.0269500 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Feng J, Li K, Zhang H, Zhang X, Yao Y. Multichannel Spatio-Temporal Feature Fusion Method for NILM. IEEE Trans Ind Inf. 2022. Dec; 18(12): 8735–44. doi: 10.1109/TII.2022.3148297 [DOI] [Google Scholar]
  • 35.Cui Y, Jia M, Lin T-Y, Song Y, Belongie S. Class-Balanced Loss Based on Effective Number of Samples [Internet]. arXiv; 2019 [cited 2022 Nov 4].
  • 36.De Baets L, Ruyssinck J, Develder C, Dhaene T, Deschrijver D. Appliance classification using VI trajectories and convolutional neural networks. Energy and Buildings. 2018. Jan; 158: 32–6. doi: 10.1016/j.enbuild.2017.09.087 [DOI] [Google Scholar]
  • 37.Sadeghianpourhamami N, Ruyssinck J, Deschrijver D, Dhaene T, Develder C. Comprehensive feature selection for appliance classification in NILM. Energy and Buildings. 2017. Sep; 151: 98–106. doi: 10.1016/j.enbuild.2017.06.042 [DOI] [Google Scholar]
  • 38.Lin Y-H, Tsai M-S. Non-Intrusive Load Monitoring by Novel Neuro-Fuzzy Classification Considering Uncertainties. IEEE Trans Smart Grid. 2014. Sep; 5(5): 2376–84. doi: 10.1109/TSG.2014.2314738 [DOI] [Google Scholar]
  • 39.Jazizadeh F, Becerik-Gerber B, Berges M, Soibelman L. An unsupervised hierarchical clustering based heuristic algorithm for facilitated training of electricity consumption disaggregation systems. Advanced Engineering Informatics. 2014. Oct; 28(4): 311–26. doi: 10.1016/j.aei.2014.09.004 [DOI] [Google Scholar]
  • 40.Hart GW. Nonintrusive appliance load monitoring. Proc IEEE. 1992. Dec; 80(12): 1870–91. doi: 10.1109/5.192069 [DOI] [Google Scholar]
  • 41.Kim K. Noise Avoidance SMOTE in Ensemble Learning for Imbalanced Data. IEEE Access. 2021; 9: 143250–65. doi: 10.1109/ACCESS.2021.3120738 [DOI] [Google Scholar]
  • 42.Rajesh L, Satyanarayana P. Evaluation of Machine Learning Algorithms for Detection of Malicious Traffic in SCADA Network. J Electr Eng Technol. 2022. Mar; 17(2): 913–28. doi: 10.1007/s42835-021-00931-1 [DOI] [Google Scholar]

Decision Letter 0

Mohamed Hammad

22 Dec 2022

PONE-D-22-31416Power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNetPLOS ONE

Dear Dr. Lin,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Feb 04 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

  • A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Mohamed Hammad, Ph.D.

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at 

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and 

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Please note that PLOS ONE has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript. In these cases, all author-generated code must be made available without restrictions upon publication of the work. Please review our guidelines at https://journals.plos.org/plosone/s/materials-and-software-sharing#loc-sharing-code and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse.

3. Our internal editors have looked over your manuscript and determined that it is within the scope of our Smart Energy Systems Call for Papers. The Collection will encompass the latest research in smart grid technologies, including information technologies, device integration, distribution methods, and data mining, all towards improving the efficiency of energy supply networks. Additional information can be found on our announcement page: https://collections.plos.org/call-for-papers/smart-energy-systems/. If you would like your manuscript to be considered for this collection, please let us know in your cover letter and we will ensure that your paper is treated as if you were responding to this call. If you would prefer to remove your manuscript from collection consideration, please specify this in the cover letter.

4. Thank you for stating the following financial disclosure: 

"This work was supported by Natural Science Foundation of Jilin Province (YDZJ202101ZY89)."

Please state what role the funders took in the study.  If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." If this statement is not correct you must amend it as needed. 

Please include this amended Role of Funder statement in your cover letter; we will change the online submission form on your behalf.

5. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see http://journals.plos.org/plosone/s/data-availability.

Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories. Any potentially identifying patient information must be fully anonymized.

Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access.

We will update your Data Availability statement to reflect the information you provide in your cover letter.

6. Please upload a new copy of Figures 1, 3, 4, 9, 10, 11 and 13 as the detail is not clear. Please follow the link for more information:

https://blogs.plos.org/plos/2019/06/looking-good-tips-for-creating-your-plos-figures-graphics/

https://blogs.plos.org/plos/2019/06/looking-good-tips-for-creating-your-plos-figures-graphics/

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: N/A

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Summary:

In this work, a fingerprint identification based on voltage-current (V-I) trajectory with color encoding and transferred CBAM-ResNet34 is proposed. To obtain a color V-I trajectory image, the current, instantaneous power, and trajectory momentum information are added first to the original V-I trajectory image using color coding.

Then, the ResNet34 model was pretrained using the ImageNet dataset and a fully-connected layer. The Convolutional Block Attention Module (CBAM) was added to each residual structure module of ResNet34. Finally, to solve the problem of data imbalance in V-I trajectory identification, Class-Balanced (CB) loss is introduced to reweight the Softmax cross-entropy (SM-CE) loss function.

The manuscript is interesting; however, the following comment should be addressed :

Abstract :

- - - - - - - - - - -

1 – Results in terms of improvement ratio between the proposed and existing work need to be included .

Introduction Section :

- - - - - - - - - - - - - - - - - - - - - -

2 – This section is fine. No comments .

Materials and methods Section :

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

3 – This section is fine. No comments .

Power fingerprint identification based on transferred CBAM-ResNet34 Section :

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

4 – This section is fine. No comments .

Experiments Section :

- - - - - - - - - - - - - - - - - - - - - - - - - -

5 – This section is fine. No comments .

Result and discussion Section :

- - - - - - - - - - - - - - - - - - - - - - - - - - -

6 – This section is fine. No comments .

Conclusion Section :

- - - - - - - - - - - - - - - - - - - - - -

7 – The limitation of this work should be clearly included in the conclusion section .

General Comments:

- - - - - - - - - - - - - - - - -

8 - There are some grammatical errors should be corrected .

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Reviewer #2: Dear Authors

The paper titled “Power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNet” proposed a power fingerprint identification

based on the improved voltage-current(V-I) trajectory with color encoding and transferred CBAM-ResNet34. First, the current, instantaneous power, and trajectory momentum information are added to the original V-I trajectory image using color coding to obtain a color V-I trajectory image. Then, the ResNet34 model was pretrained using the ImageNet dataset and a new fully-connected layer meeting the device classification goal was used to replace the fully-connected layer of ResNet34. The experimental results on the imbalanced PLAID dataset verify that the method in this paper has better classification capability in small sample imbalanced datasets.

The paper is interesting however it needs improvements.

1. There are various types of ResNet models are available including ResNet-18, ResNet-34, ResNet-50, ResNet-101, ResNet-110, ResNet-152, ResNet-164, ResNet-1202, why authors choose ResNet-34.

2. Line 371-372: the SMOTE is not explained, I suggest the author to explain it with proper work e.g. [1].

3. Line 36: Power fingerprint identification is a hot issue in the field of NILM, include the full form when any abbreviation appears first time in the entire manuscript.

4. This paper used ResNet however the role of ML/DL is not discussed to improve the need to utilize the ML/DL.

5. Introduction needs a new paragraph which discuss the AI/DL applications with different domains at once. This paragraph should be added as “Nowadays, scientists and researchers used the machine learning (ML) and Deep learning (DL) models in several applications including agriculture [1, 2], environment [3-6], and power fingerprint identification. 1. Planetscope Nanosatellites Image Classification Using Machine Learning; 2. CNN Based Automated Weed Detection System Using UAV Imagery; 3. SMOTEDNN: A Novel Model for Air Pollution Forecasting and AQI Classification; 4. CDLSTM: A Novel Model for Climate Change Forecasting; 5. Deep Learning Based Modeling of Groundwater Storage Change; 6. Deep Learning Based Supervised Image Classification Using UAV Images for Forest Areas Classification.

6. Figure 1-13 needs improvement, all these figures looks blur.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2023 Feb 9;18(2):e0281482. doi: 10.1371/journal.pone.0281482.r002

Author response to Decision Letter 0


22 Jan 2023

Response to Reviewer 1 Comments

Point 1: Abstract: Results in terms of improvement ratio between the proposed and existing work need to be included.

Response 1: Thank you for underlining this deficiency. In the revised manuscript (page 1), we have added more details about the contribution of the paper in the abstract as the following:

‘’The experimental results show that the method effectively improves the identification accuracy by 4.4% and reduces the training time of the model by 14 min compared with the existing methods, which meets the accuracy requirements of fine-grained power fingerprint identification.’’

Point 2: Conclusion Section: The limitation of this work should be clearly included in the conclusion section.

Response 2: Thank you for underlining this deficiency. In the revised manuscript (page 23), we have added more details about the limitation of this work in the conclusion section as the following:

‘’However, the method proposed in this paper still has some problems to be solved. The identification effect of multi-state load in this paper can be further improved, and more advanced device identification models are needed to solve the problem in the future. In addition, the actual engineering implementation faces difficulties in constructing sample libraries and high data sampling frequency requirements. Therefore, a lot of related research work is still needed.’’

Point 3: General Comments: There are some grammatical errors should be corrected.

Response 3: Thank you for underlining this deficiency. In the revised manuscript (page 4, page 7, and page 21), we have corrected grammatical errors and made detailed changes throughout the text. In addition, changes suggested by another reviewer and the grammar of the paragraphs added to the introduction were checked. The correction of major grammatical errors as the following:

‘’For V-I trajectory classification, the small amount of load data for specific user results in a poorly trained model with low accuracy or low generalization capability [21].’’

‘’(a)The instantaneous voltage of the Fridge over 1 s. (b) The instantaneous current of the Fridge over 1 s. (c) The instantaneous power of the Fridge over 1 s.’’

‘’Color-encoded V-I trajectory images and transferred CBAM-ResNet34 were used as the load feature and training model, respectively, which significantly enhanced the identification ability of the algorithm and reduced the model training time.’’

Response to Reviewer 2 Comments

Point 1: There are various types of ResNet models are available including ResNet-18, ResNet-34, ResNet-50, ResNet-101, ResNet-110, ResNet-152, ResNet-164, ResNet-1202, why authors choose ResNet-34?

Response 1: We are grateful for the suggestion. According to the reviewer’s comment, we have provided more details to describe the possible reasons. We have made additions in the corresponding sections of the manuscript(page 9, and page 11). The reasons for selecting ResNet-34 to construct the classification model in this paper as the following:

The different numbers of ResNet networks represent the number of layers of the network. Compared to ResNet-18, the network with the least number of layers lacks certain feature representation abilities. Compared to ResNet50, ResNet-101, ResNet-110, ResNet-152, ResNet-164, and ResNet-1202, the increase in the number of layers of the network is accompanied by an increase in the number of parameters and computations, which makes the network training slower.

Also, with the addition of the CBAM attention module, ResNet34 can achieve not only higher accuracy but also higher accuracy at the beginning of training. This indicates that the network can learn the target features more efficiently and speed up the training of the network. This is since the introduction of the attention mechanism inside the residual block can learn the features of V-I image samples more efficiently.

Compared with other layers of ResNet networks, the accuracy improvement of the attention mechanism module on the network is inferior to that of ResNet34. This is because the attention mechanism is more likely to learn effective features in shallow networks, and the improvement effect of the attention mechanism module on the performance of deeper networks is less obvious, and it is difficult for the attention mechanism to learn effective features.

Therefore, considering the hardware cost limitation and the difficulty of training V-I images, and to reflect the feature extraction capability and convergence speed of ResNet34-CBAM, ResNet34-CBAM is selected as the classification model for V-I images in this paper.

Point 2: Line 371-372: the SMOTE is not explained, I suggest the author explain it with proper work e.g. [1].

Response 2: The authors thank the reviewer for pointing out this problem. The required explanation has been made. In the revised manuscript (Line 388, page 19), the revised contents are as follows:

‘’Using transferred CBAM-ResNet34 as the base model, the methods in this paper were compared with the original data, random oversampling, SMOTE [41], and SVMSMOTE [42] data balancing algorithms on the PLAID dataset, respectively.’’

[41] Kim K. Noise Avoidance SMOTE in Ensemble Learning for Imbalanced Data. IEEE Access. 2021; 9: 143250–65. https://doi.org/10.1109/ACCESS.2021.3120738

[42] Rajesh L, Satyanarayana P. Evaluation of Machine Learning Algorithms for Detection of Malicious Traffic in SCADA Network. J Electr Eng Technol. 2022 Mar; 17(2): 913–28. https://doi.org/10.1007/s42835-021-00931-1

Point 3: Line 36: Power fingerprint identification is a hot issue in the field of NILM, include the full form when any abbreviation appears first time in the entire manuscript.

Response 3: Thank you for underlining this deficiency. In the revised manuscript (Line 39, page 2), the revised contents are as follows:

‘’Power fingerprint identification is a hot issue in the field of Non-intrusive load monitoring (NILM), which relies on power fingerprint features and classifiers to identify different types of devices. Common power fingerprint characteristics typically include voltage, current, harmonics, power, V-I trajectory, etc [7-9].’’

Point 4: This paper used ResNet however the role of ML/DL is not discussed to improve the need to utilize the ML/DL.

Response 4: We deeply appreciate the reviewer’s suggestion. According to the reviewer’s comment, we have added more details to describe and discuss the role of ML/DL. In the revised manuscript ( page 3), the revised contents are as follows:

‘’Nowadays, scientists and researchers used machine learning (ML) and deep learning (DL) models in several applications including agriculture [15, 16], environment [17-20], and power fingerprint identification. Machine learning is often applied to feature extraction and classification of power fingerprints, such as k-nearest neighbors, support vector machines, decision trees, and random forests. These methods are less computationally intensive, but the identification correct rate is lower. Recently, deep learning has achieved good results in the field of power fingerprint identification, such as CNN, RNN, etc. Meanwhile, researchers propose to construct V-I trajectory images with the help of color coding methods to convert power fingerprint identification into an image classification task in which deep learning excels. However, compared to machine learning, deep learning-based classifiers rely on large-scale training data and longer training time, which limits the application of deep learning.’’

Point 5: Introduction needs a new paragraph which discuss the AI/DL applications with different domains at once. This paragraph should be added as “Nowadays, scientists and researchers used the machine learning (ML) and Deep learning (DL) models in several applications including agriculture [1, 2], environment [3-6], and power fingerprint identification. 1. Planetscope Nanosatellites Image Classification Using Machine Learning; 2. CNN Based Automated Weed Detection System Using UAV Imagery; 3. SMOTEDNN: A Novel Model for Air Pollution Forecasting and AQI Classification; 4. CDLSTM: A Novel Model for Climate Change Forecasting; 5. Deep Learning Based Modeling of Groundwater Storage Change; 6. Deep Learning Based Supervised Image Classification Using UAV Images for Forest Areas Classification.

Response 5: We are grateful for the suggestion. According to the reviewer’s comment, we have provided more details to discuss the AI/DL applications with different domains at once in the introduction. In the revised manuscript ( page 3), the revised contents are as follows:

‘’Nowadays, scientists and researchers used machine learning (ML) and deep learning (DL) models in several applications including agriculture [15, 16], environment [17-20], and power fingerprint identification. Machine learning is often applied to feature extraction and classification of power fingerprints, such as k-nearest neighbors, support vector machines, decision trees, and random forests. These methods are less computationally intensive, but the identification correct rate is lower. Recently, deep learning has achieved good results in the field of power fingerprint identification, such as CNN, RNN, etc. Meanwhile, researchers propose to construct V-I trajectory images with the help of color coding methods to convert power fingerprint identification into an image classification task in which deep learning excels. However, compared to machine learning, deep learning-based classifiers rely on large-scale training data and longer training time, which limits the application of deep learning.’’

[15] Anul Haq M. Planetscope Nanosatellites Image Classification Using Machine Learning. Computer Systems Science and Engineering. 2022; 42(3): 1031–46. https://doi.org/10.32604/csse.2022.023221

[16] Anul Haq M. CNN Based Automated Weed Detection System Using UAV Imagery. Computer Systems Science and Engineering. 2022; 42(2): 837–49. https://doi.org/10.32604/csse.2022.023016

[17] Attaallah A, Ahmad Khan R. SMOTEDNN: A Novel Model for Air Pollution Forecasting and AQI Classification. Computers, Materials & Continua. 2022; 71(1): 1403–25. https://doi.org/10.32604/cmc.2022.021968

[18] Anul Haq M. CDLSTM: A Novel Model for Climate Change Forecasting. Computers, Materials & Continua. 2022; 71(2): 2363–81. https://doi.org/10.32604/cmc.2022.023059

[19] Anul Haq M, Khadar Jilani A, Prabu P. Deep Learning Based Modeling of Groundwater Storage Change. Computers, Materials & Continua. 2022; 70(3): 4599–617. https://doi.org/10.32604/cmc.2022.020495

[20] Haq MA, Rahaman G, Baral P, Ghosh A. Deep Learning Based Supervised Image Classification Using UAV Images for Forest Areas Classification. J Indian Soc Remote Sens. 2021 Mar; 49(3): 601–6. https://doi.org/10.1007/s12524-020-01231-3

Point 6: Figure 1-13 needs improvement, all these figures looks blur.

Response 6: Thank you for underlining this deficiency. In the revised manuscript, we have adjusted the clarity of Figures 1-13.

Attachment

Submitted filename: Response to Reviewers.docx

Decision Letter 1

Mohamed Hammad

25 Jan 2023

Power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNet

PONE-D-22-31416R1

Dear Dr. Lin,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Mohamed Hammad, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Summary:

In this work, a fingerprint identification based on voltage-current (V-I) trajectory with color encoding and transferred CBAM-ResNet34 is proposed. To obtain a color V-I trajectory image, the current, instantaneous power, and trajectory momentum information are added first to the original V-I trajectory image using color coding.

Then, the ResNet34 model was pretrained using the ImageNet dataset and a fully-connected layer. The Convolutional Block Attention Module (CBAM) was added to each residual structure module of ResNet34. Finally, to solve the problem of data imbalance in V-I trajectory identification, Class-Balanced (CB) loss is introduced to reweight the Softmax cross-entropy (SM-CE) loss function.

The authors have addressed the raised comments .

Abstract :

- - - - - - - - - - -

1 – The abstract is fine. No further comments .

Introduction Section :

- - - - - - - - - - - - - - - - - - - - - -

2 – This section is fine. No further comments .

Materials and methods Section :

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

3 – This section is fine. No further comments .

Power fingerprint identification based on transferred CBAM-ResNet34 Section :

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

4 – This section is fine. No further comments .

Experiments Section :

- - - - - - - - - - - - - - - - - - - - - - - - - -

5 – This section is fine. No further comments .

Result and discussion Section :

- - - - - - - - - - - - - - - - - - - - - - - - - - -

6 – This section is fine. No further comments .

Conclusion Section :

- - - - - - - - - - - - - - - - - - - - - -

7 – This section is fine. No further comments .

General Comments:

- - - - - - - - - - - - - - - - -

8 – No further comments .

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Reviewer #2: Dear Authors

I have now completed the review of the revised manuscript, titled " Power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNet”. I have observed that the authors put in good efforts to address most of the comments satisfactorily. The improvements in the figures is also seems excellent. Ref 17 details are incorrect, please correct.

Best wishes

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: Yes: Dr Mohd Anul Haq

**********

Acceptance letter

Mohamed Hammad

30 Jan 2023

PONE-D-22-31416R1

Power fingerprint identification based on the improved V-I trajectory with color encoding and transferred CBAM-ResNet

Dear Dr. Lin:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Mohamed Hammad

Academic Editor

PLOS ONE


Articles from PLOS ONE are provided here courtesy of PLOS

RESOURCES