Deep learning and signal processing based algorithm for autorecognition of harmonic loads

Abstract

A convolution neural network (CNN) based deep learning method has been proposed for automatic classification and localization of nonlinear loads present in an interconnected power system. The identification of nonlinear loads has been previously dealt with the use of Nonlinear Auto Regression neural network with eXogenous inputs (NARX), Backpropagation Neural Network (BPNN), Probabilistic Neural Network (PNN), Artificial Neural Networks (ANN) and Fuzzy Logic (FL). However, these techniques had not explored the area of classification of industrial and domestic nonlinear loads in an interconnected power system. Also, a Deep learning-based solution for identification of the type of nonlinear load has not been reported in the literature to date. Hence, to address these shortcomings, an IEEE-9 Bus system with industrial nonlinear loads has been used to obtain various current waveforms with distortions. The recorded current waveforms are transformed into a time-frequency (TF) domain plane, and the obtained images are then fed to the deep learning algorithm. The colored images of the TF plots of each type of nonlinear load in Red-Green-Blue (RGB) index provide the best visual features for extraction. The TF domain signatures of individual events are scaled to a standard size before feeding to the algorithm. Through these TF signatures, unique features were extracted with the deep learning algorithm, and then passed on to different stages of convolution and max-pooling with fully connected layers. The softmax classifier at the end classifies the input data into the type of nonlinear present in the power system. The algorithm, when run at different buses, also identifies the location of the nonlinear load. The proposed methodology avoids the usage of any additional fusion layer for obtaining unique features, reduces the training time and maintains the highest accuracy of 100%.

Keywords

Nonlinear loads localization identification deep learning time-frequency representation

1 Introduction

Artificial intelligence (AI) has become the most widely used technique in the last decade. Due to its ability to make decisions like humans and classify complex events, AI is being found in every industrial sector, including education, medical and commerce. The application of AI in the area of engineering is much vaster and more efficient. Both types of AI, i.e. behavioural and cognitive approach, are much applicable in the field of engineering. Ranging from expert systems, evolutionary computing system to the latest deep learning algorithms, AI has been widely used in engineering. However, in recent times, the application of Deep Learning (DL) is found to be more suitable, easy to learn and efficient. The reason being the availability of many online training modules, robust learning infrastructure and highly accurate practical results produced by deep learning algorithms.

A lot of research has been reported in the literature in the area of artificial intelligence techniques for power system monitoring and protection. The major problem statements in the power system is the identification and location of disturbances originating due to various sources such as malfunctioning equipment, system faults, nonlinear loads and others. Out of all such sources, identifying the source of disturbance due to nonlinear loads is a challenging task. It is due to the reason that the system is fully interconnected and nonlinear loads are widely spread. Hence, always there is a need to predict the type of nonlinear load along with the location. Such information helps the utility to identify the area of harmonic pollution and quickly act to avoid future events. Extensive research has been done in this area wherein researchers have proposed methodologies with and without feature extraction techniques. The results are encouraging and improving till date, which can be proofread from the literature of the past five years. Various methods ranging from the design of standard indices to neural networks have been utilized for nonlinear load identification.

M.F. Romero-L et al. have proposed Sensitivity index, Linearity index and Asymmetry index to identify the type of nonlinear load in domestic Low Voltage grids. They have studied the typical penetration ratios of specific nonlinear household loads viz. Microwave oven, Conventional Television, Video Player and other such loads used in Colombia. Based on their equivalent Norton equivalent circuits, few indices have been developed, and the type of load was identified. However, the paper discussed only the LV grid in isolation [1].

Similarly, Indhana Sudiharto et al. have proposed a Fast Fourier Transform (FFT) and Levenberg Marquardt Backpropagation (LMBP) network based algorithm for identification of nonlinear load in residential buildings. They have considered an Air Conditioner, Personal Computer, Laptop and a fluorescent lamp for analysis and identification. A Power Quality (PQ) meter has been used to record the signal. The THDv and THDi of the obtained data are then fed to the LMBP network with 5 and 19 neurons to identify the type of load based on the harmonic signature [2]. Nonintrusive load monitoring method was also proposed for the identification of domestic nonlinear loads [3]. However, both the papers have limited identification to residential nonlinear loads.

AY Hatata et al. proposed a NARX neural network based identification of nonlinear loads. They have used Khalda –Main Razzak (MRZK) power station 1.2MW capacity isolated microgrid power systems to obtain the waveforms of different nonlinear loads. The microgrid system contains Electrical Submersible pump driven by Adjustable Speed Drive (ASD)/ Variable Frequency Drive (VFD). After the waveform data is collected, the trained NARX network is fed with the sinusoidal waveform to separate the harmonics and identify the pattern of the nonlinear load [4]. However, the discussion was limited to microgrid only.

Rakhmanov N.R. et al. have proposed probabilistic characteristics of the amplitudes of harmonic components of the power harmonics-based method to identify the type of nonlinear load. They have used a 110kV Azerbaijan power system for recording various type waveforms of current and voltage. Accordingly, they have adopted the Nonlinear model (NLM) for identifying the harmonics present in the system. However, various types of nonlinear load present in an interconnected power system were not considered [5].

True Root Mean Square (RMS) Current measurements based on the Gradient Boosted Regression trees were also used to predict the type of nonlinear loads. For this purpose, the authors Flávia P. Monteiro et al. have applied the proposed methodology in a campus of Brazilian University. They have shown that based on the Mean Absolute Percent Error (MAPE) and Permutation importance, the presence of nonlinear load on a specific phase can be identified. However, they have not discussed the type of nonlinear load classified [6].

Current’ physical components (CPC) methodology was proposed by Yuval Beck et al. for identification of various harmonic loads. For validation, they have recorded the current waveforms of multiple loads and have demonstrated their respective CPCs. These features are then fed to the ANN and nearest neighbour algorithms for identification. An accuracy of 87% was shown. However, the testing was done with the loads in isolation but not with an interconnected power system. Also, the location of the harmonic loads was not explored [7]. Yaroslav Shklyarskiy et al. have proposed the application of compensating devices for identifying the origin of the nonlinear loads, i.e. to predict the load presence is from the User or the utility. However, they have not discussed the type of nonlinear load present in any of the side [8].

Srikanth et al. have proposed a signal processing based Fuzzy logic methodology for identification of industrial nonlinear loads in isolation [9, 10]. Also, they have introduced a similar method to an interconnected power system. The authors have used Stockwell transform and fuzzy logic to identify and locate the type of nonlinear load in an IEEE-5 Bus system [11]. However, it is still felt that the complete time-frequency spectrum is not effectively used for the identification. Also, the review article by Asiye K. Ozcanli et al. shows that the application of deep learning algorithm for nonlinear loads identification is not reported [12].

Summarily, a methodology which can identify both industrial and domestic nonlinear loads in an interconnected power system with the highest accuracy is still needed. This requirement forms the principal motivation of this article. In view of the above reported facts, a signal processing and Deep learning-based method have been proposed in this paper. The potential contribution of the proposed methodology is that a deep learning-based algorithm has been explored for classification of nonlinear loads which will lead to efficient substation monitoring systems. Secondly, the proposed algorithm can be implemented for identification of industrial nonlinear loads of any voltage level, as the algorithm takes the input data in per-units.

1.1 Novelty of the work

As mentioned above a Deep learning method based on the TF analysis of the harmonic currents of a power system has been proposed for classification of the nonlinear loads. Though there are many articles which had reported results with reasonably fair accuracy, there are few unexplored areas which have been addressed in this article. They are listed as below: -

Industrial nonlinear loads in an interconnected power system have been considered. Till date, much research was done in residential harmonic loads or industrial nonlinear loads in microgrids but not in a fully interconnected power system.

Deep learning methodology for the classification of the type of harmonic loads has been used. Though researchers had reported the identification of power quality issues, including harmonics, further analysis of the source of harmonics has been reported here.

An accuracy of 100% has been reported despite considering harmonic loads of similar nature. Even though the dominant frequencies, i.e. 3rd, 5th, 7th and 9th of Arc Furnace, Traction Model and Low Magnitude harmonics are very close, the obtained results, in this case, is promising.

With these points, it is mentioned here that this article will add more value to the literature and bridges the unexplored gaps in the area of nonlinear load identification. This paper is organized into seven sections, including the introduction. Section 2 explains the type of nonlinear loads considered for identification. The details about the signal processing method are provided in Section 3. The Deep learning algorithm architecture and design are explained in Section 4. The overall process is described in Section 5. Results and discussions are provided in Section 6 with conclusions in Section 7.

2 Selection of harmonic loads

Nonlinear industrial harmonics loads have been modelled and fed into the IEEE-9 Bus power system for identification, as shown in Fig. 1. The details of the power system are mentioned in Appendix-A. An Arc Furnace, Electric Traction load, 12 Pulse VFD, Nonlinear loads causing interharmonics, and low magnitudes have been considered for identification and location. Arc furnace has been modelled using the data as defined by KU. Vinayaka et al. [13]. Electric traction load has been modelled based on the available library of Electro Magnetic Transient Program (EMTP). These loads are found very commonly in any 11kV and 33kV industrial systems. The localization of harmonics sources like Arc Furnace and Traction loads is significant as they cause voltage flickers and unbalanced currents with harmonics into the systems.

Fig. 1

IEEE 9 Bus system showing the location of harmonic loads as per the composition shown in Table 1.

In addition to the above, detection of the presence of Alternating Current (AC) drive loads is also considered, as industries like cement and steel use huge VFDs/AC drives with negligence towards PQ filtering. These cause harmonic injection along with sags in the power supply. ABB Guide to Harmonics in AC drives [14] has been used to model the 12 Pulse VFD and low magnitude harmonic load. Other than these, nonlin ear loads like cycloconverters which cause interharmonics [15] are also considered as per the predefined models.

These low magnitude harmonic loads are those which inject harmonics just above the defined limits at the Point of Common Coupling (PCC) due to the presence of loads like IT Complex, LED lighting etc. Mostly such loads are found in office complexes and residential areas due to the operation of single-phase loads. They inject low magnitude harmonics in the power system as there is a step-up in voltage level. Thus, all types of harmonics present in the industrial power system are considered. The data of all the nonlinear loads are provided in Table 1. The system, as shown in Fig. 1 is an IEEE standard nine bus system, where all the quantities are represented in per unit (PU). EMTP has been used for simulating the model. Each type of nonlinear loads has been fed at three different buses, i.e. Bus 5, 6 and 8, to identify the type and location of the nonlinear load through the recordings at Bus 4, 7, and 9.

Table 1

Nonlinear load details

S No.	Type of Load		Harmonic percentage (%)
1	12-pulse with double wound transformer [14]	5^th 3.6	7^th 2.6	11^th 7.5	13^th 5.2
2	Arc Furnace [13]	3^rd 4.47	5^th 1.10	7^th 0.66	9^th 0.40
3	Traction model of EMTP	3^rd 22.86	5^th 22.02	7^th 4.62	9^th 11.39
4	Low Magnitude Harmonic load model [14]	3^rd 0.12	5^th 0.24	7^th 1.19	9^th 0.9
5	Interharmonic Harmonics load from Section 09 of [15].	3.5^th 0.10	5^th 1.11	7^th 2.26	9^th 2.26

3 Stockwell transform

The present section discusses the signal processing method adopted for identifying and locating the type of nonlinear loads in the IEEE 9 Bus power system. The S transform (ST) has been used for obtaining the time-frequency signatures for the nonlinear loads which are as defined in Equations (1) & (3) shown below $S (a, b) = \int_{- \infty}^{+ \infty} h (τ) (\frac{| b |}{\sqrt{2 π}} . exp (- \frac{{(τ - a)}^{2} b^{2}}{2}) . e^{(- i 2 π b τ)} . d τ$ (1) with a, b being the time and frequency in ST domain and τ is the time of the input signal. The width of the Gaussian window is $σ = \frac{1}{| b |}$ (2)

The Discrete S–Transform (DST) of Equation (1) is a representation of the local spectra, which can be obtained by the shift operation on the Fourier spectrum and is expressed as Equation (3), $S [kT, \frac{n}{NT}] = \sum_{k = 0}^{N - 1} H [\frac{m + n}{NT}] . e^{\frac{- 2 π^{2} m^{2}}{n^{2}}} e^{\frac{i 2 π mk}{N}} for n \neq 0$ (3) where k, n, m=0 .... N-1, T=Sampling time and N=Number of samples [8, 13].

It is known that S-Transform does localization of both phase and amplitude spectrum. The Gaussian window defined in ST varies with both time and frequency of the input signal. Due to such definition ST maintains good time and frequency resolution. The power system signals being non-stationary, S-transform can effectively be applied. Next section discusses the extraction of discriminatory features of ST matrix coefficients obtained from Equation (3) for identification of nonlinear loads.

4 Deep learning algorithm design

Currently, many pretrained Convolution Neural Network (CNN) based Deep Learning Algorithms (DLAs) are available for application in various fields of research. However, in recent times AlexNet, GoogleNet, VGGNet and ResNet are found more frequently in most of the applications in the field of engineering [16]. Out of the above, Google Inception network and ResNet-152 have occupied the top two slots in terms of accuracy [17]. However, the computations of ResNet-152 is 11Billion Floating Point Operations (FLOPS) which too high compared to the other networks. In contrast, a GoogleNet takes only 2B FLOPS with an accuracy of more than 90%, as shown in Fig. 2. However, the architecture is slightly complicated due to a lot of parallel layers with different filter sizes. It would be difficult for a researcher with beginners’ knowledge in Deep Learning algorithms to edit the layers as per their convenience. Hence, the AlexNet has been used as the baseline network in this work.

Fig 2

Performance of various pretrained networks.

The AlexNet is loaded, and the last three layers are modified to improve the suitability of the network to the considered problem statement. To perform transfer learning, MATLAB software has been used [17]. The network architecture of the AlexNet is shown in Fig. 3. It can be observed that the network comprises of three stages of maxPoolings, two stages of dropouts and three fully connected networks which are as per the standard AlexNet architecture. Figure 3 is slightly different from the standard AlexNet since three more layers are added to the network for classifying the type of nonlinear loads. A fully connected layer, softmax layer and a classification output layer has been added.

Fig. 3

Various stages of the AlexNet based Deep Learning network designed for the identification of the nonlinear loads.

The input data to this network is an image of the size of 227x227x3 pixels. As described in section 2, five types of nonlinear loads which are frequently found in the power system industry are considered for identification. The images obtained after applying the ST technique is processed with a fixed size of 227x227x3 to match the input size of the network. A set of 100 images of each type of nonlinear loads is prepared as the database for feeding the network out of which 40 images are used for validation. The training parameters set for this network are as shown in Table 2.

Table 2

Training Parameters

S No.	Parameter	Case 1 Value	Case 2 Value
1	Weight Learn Rate Factor	20	20
2	Bias Learn Rate Factor	20	20
3	Minimum Batch Size	120	120
4	Max Epochs	6	6
5	Initial Learn Rate	1e-3	1e-4
6	Validation Frequency	5	3

5 Overall process

The overall process chart is shown in Fig. 4. The IEEE-9 Bus system explained in section 2 is simulated, and three cycles of the obtained current waveforms at each bus are passed through the ST technique. This provides a time-frequency plot of the input current based on the defined sampling rate. The ST image is then resized to a size of 227x227 pixels to feed the deep learning algorithm. The algorithm takes these images as the input and based on the past training; the type of nonlinear loads is predicted. Based on the prediction and predefined suggestions, the utility can decide the type of action to be taken.

Fig. 4

Step by step representation of Overall process of identifying the type of nonlinear load using Deep Learning.

6 Results and discussions

Based on the step-by-step process explained in Fig. 4, the results have been obtained. The recording of the current waveform and obtaining a 227x227 pixeled image for three (03) cycles of the input current is shown in Fig. 5. The left side of the figure shows the current waveform recorded at Bus 4 for a different type of loads at Bus No. 5 of Fig. 1. The corresponding ST for three cycles, i.e. from t=60 to 120 samples (0.06 to.12s) with a sampling frequency of 1000, is shown on the right side of Fig. 5. It is visually evident that the signatures of ST are different for each load. The loads with very low amplitudes of harmonic content are also distinguished. However, it is difficult to predict by the visual process that what is the exact nature of loads as the signatures for 5(b), 5(c) and 5(d) are almost similar.

Fig. 5

Recorded current with their corresponding ST plots for three cycles of (a) Arc Furnace (Arc) (b) Interharmonic loads (SwHarm) (c) low magnitude harmonic loads (Harm) (d) Electric traction load (Trac) (e) 12 Pulse drive (TwPDw).

The MATLAB model of the Deep learning network is shown in Fig. 6. The stages of convolution, max-pooling and fully connected layers are clearly shown with the added layers of softmax and classifier output. The filter size is set at 11x11, and the stride is set at 4, Zero padding is used with the number of filters as 96 in the first layer of convolution. The filter size is set to 5x5 in the next convolution layer and to 3x3 in all the consecutive convolution layers. The dropout layer has been introduced at the end of the network with 40% value at both the dropout layers. Other parameters are kept unaltered as per the predefined AlexNet.

Fig. 6

AlexNet based Deep Network designed in MATLAB for the identification of nonlinear loads in an IEEE-9 Bus power system.

After these initializations, the Deep learning algorithm is run with the training parameters, as shown in Table 2 previously. The matrix of the design analysis of the Deep learning network is shown in Table 3. In total, the network processes 5.69x10⁷ data points which are slightly lesser than the original number of parameters due to redesign of the network. The training progress of the Deep learning network is shown in Fig. 7.

Table 3

The activations, learnables and Total Learnables at each stage of the network

Name	Dimension	Learnables	Total Learnables
Data	227×227×3	–	0
Convolution 1	55×55×96	Weights 11×11×3×96;
Bias 1×1×96	34944
ReLU 1	55×55×96	–	0
Normalisation 1	55×55×96	–	0
Max Pooling 1	27×27×96	–	0
Convolution 2	27×27×256	Weights 5×5×48×128×2; Bias 1×1×128×2	307456
ReLU 2	27×27×256	–	0
Normalisation 2	27×27×256	–	0
Max Pooling 2	13×13×256	–	0
Convolution 3	13×13×384	Weights 3×3×256×384; Bias 1×1×384	885120
ReLU 3	13×13×384	–	0
Convolution 4	13×13×384	Weights 3×3×192×192×2; Bias 1×1×192×2	663936
ReLU 4	13×13×384	–	0
Convolution 5	13×13×256	Weights 3×3×192×128×2; Bias 1×1×128×2	442624
ReLU 5	13×13×256	–	0
Max Pooling 5	6×6×256	–	0
Fully
Connected 6	1×1×4096	Weights 4096×9216; Bias 4096×1	37752832
ReLU 6	1×1×4096	–	0
Drop 6	1×1×4096	–	0
Fully
Connected 7	1×1×4096	Weights 4096×4096; Bias 4096×1	16781312
ReLU 7	1×1×4096	–	0
Drop 7	1×1×4096	–
Fully
Connected	1×1×5	Weights 5×4096; Bias 5×1	20485
SoftMax	1×1×5	–	0
Classification Output	–	–	0
		Total Learnables =	56888709

Fig. 7

Training progress plots of the Deep Learning network with (a) learning rate = 1×10^–3 & validation frequency = 5 (b) learning rate = 1×10^–4 & validation frequency = 3.

The algorithm is run for two different training variables i.e. learning rate=1x10^–3 & validation frequency=5 and learning rate=1x10^–4 & validation frequency=3. The progress of both cases is shown in Fig. 7. The reason for changing these two parameters is to minimize loss and improve the accuracy in identification. It is known that changing these two parameters impacts the computation time. However, to study the performance of the Deep learning network for the types of nonlinear loads, these parameters are changed.

Figure 7(a) shows that the algorithm reached an accuracy of 100% after Epoch 5 with a loss value of approx. 5 for a learning rate of 1x10^–3. The network has taken approximately 42sec for reaching the final stage of the training process. The loss has minimized almost to zero only after the 10^th iteration. Whereas, in Fig. 7(b) accuracy of 100% has reached early, i.e. at Epoch 4. The loss value is approx. 3, which has reached zero at the 9^th iteration, which is also early compared to Fig. 7(b). The training time for achieving these results is 46sec which is 4sec higher than the previous. This difference clearly shows that the change in parameters affects the computational time. The summary of performance is shown in Table 4.

Table 4

Performance Comparison for different values of Initial Learn Rate (ILR) and Validation Frequency (VF)

S No.	Parameter	Case 1 of Table 2	Case 2 of Table 2
		ILR = 1e-3 VF = 5	ILR = 1e-4 VF = 3
1	Accuracy (in %)	100	100
2	Training time (in sec)	42	46

It is visible from this table that the accuracy for the case of Initial learning rate (ILR) and Validation Frequency (VF) of 1e-3 and 5 is 100% with 42s of training time. Whereas, for Case –2 with ILR=1e-4 and VF=3, the accuracy is 100% with a training time of 46s. The increase in training time is due to the decrease in learning rate leading but early reach of the accuracy of 100% stage. Hence, Case-2 is more effective at the cost of 4seconds of additional training time.Since, the deep learning algorithm passes the input image through various layers of convolutions and provides the output after the final fully connected layers, the input image changes at every layer. This can be visualized by using the deepDreamImage function provided in MATLAB®. To visualize this function, the deep dream images at every layer is obtained after running the algorithm for Case-2.

Figure 8 in the next page shows the Deepdream images at every stage of the convolution and fully connected layers. The reason for choosing these stages only is that the size of the matrix changes at these layers due to convolution operations and the values change due to weight and bias factors. This leads to a change in the image representation. The first two layers resemble their input images as the algorithm has just started learning the pattern. By the passing of stages, the image totally changes.

Fig. 8

Deepdream Images of the network at various layers.

The results of the deep learning algorithm in the identification of the type of nonlinear loads along with the location are shown in Fig. 9. Since the confusion values define the robustness and efficiency of any learning and classification algorithm, a confusion matrix plot has been provided in Fig. 9 with both the training parameters. It is observed that both the options have identified the type of with 100% accuracy and without any confusion in the validated samples of 40 numbers for Bus 5. Similarly, the proposed method has identified the nonlinear loads at all the Buses considered with 100% accuracy.

Fig. 9

Confusion plots for the Deep learning network with (a) learning rate = 1×10^–3 & validation frequency = 5 (b) learning rate = 1×10^–4 & validation frequency = 3.

However, in order to establish the efficacy of the proposed methodology, the same set of images have been used for classification using Support Vector Machine (SVM). Before feeding to the SVM, two features have been extracted for each image, then the same has been used for training and validation. It was observed that per feature per image, the ST based signal processing technique had consumed 0.32s, which in total for 500 images had consumed 335 seconds, including training. The results are presented in Table 5, and the accuracy of the SVM based technique is 99.2%. In contrast, the proposed method of signal processing based deep learning has produced 100% accuracy. Also, the time of training in SVM is 335 seconds which is eight times the total time of 46 seconds taken by the proposed SP based DLA. This shows the superiority of the proposed method.

Table 5

Comparison with the results obtained from Support Vector Machine based identification

S. No	Techniques	Time for generating time-frequency signatures using S-transform	Features Extracted	Total no. of images	Time for feature extraction per feature per image (in sec)	Avg. Time for Training and Identification (in sec)	Total time (in sec)	Accuracy (in %)
1	Support Vector Machine based method	0.8	Two	500	0.32	15	(50020.32)+15 = 335 seconds	99.2
2	Proposed Deep Learning Method		Nil	500	Not Applicable		46	100

Also, a comparison of the results presented in the references [1] to [11] is given in Table 6. The table also provides the details of the research gap, which has been addressed in the present article. Briefly, the proposed methodology has addressed the identification of both industrial and domestic loads in an interconnected power system using the SP based DLA with 100% accuracy. Since two of the loads have been modelled using the original data sheets of the drive manufacturer and two loads are predefined standardized models, the results are very near to real-time. The time for identification is at par compared to all other algorithms which make the proposed SP based DLA is applicable for online identification for any voltage level system after a one-time training.

Table 6

Comparison with the results reported in the literature in the last 5 years using various methods

Ref. No.	Methodology Used	Interconnected power system used (Yes/No)	Noise considered (Yes/No)	Category of Nonlinear loads considered	Neural network-based methodology proposed (Yes/No)	Fastness of the algorithm (in sec)	Accuracy reported (%)	Research Gap
[1]	Sensitivity Indices	No	Yes	Residential	No	Not Mentioned	Not Reported	Speed and Accuracy not reported
[2]	FFT+LMBP Neurons 10 Neurons	No	Yes	Residential and Office	Yes		96.88 99.89	Interconnected system not considered
[3]	Nonintrusive load monitoring	No	Yes	Residential and Office	No	Not Mentioned	97.46
[4]	NARX Neural Network	No	Yes	Industrial	Yes	<1	97.88
[5]	ELM Model F Model	Yes	Yes	Industrial	Yes	Not Mentioned	97.111 6.032	Only Arc Furnace is considered
[6]	GBRT Based RMS Measurements	No	Yes	Office	No	Not Mentioned	>95.0	Interconnected system not considered
[7]	Current Physical Components	No	Yes	Residential and Office	Yes	Not Mentioned	87.64
[8]	Compensating Devices based	No	No	Industrial	No	Not Mentioned	Not Reported	Only Traction load is considered
[9]	Modified Inverse S-transform + Fuzzy	No	No	Industrial	No	0.42	96.8	Interconnected system not considered
[10]	Stockwell Transform + Fuzzy	Yes	Yes	Industrial	No	Not Mentioned	95.0	Small system has been considered
[11]	Modified Inverse S-transform + Fuzzy	No	Yes	Office	No	0.42	100	Interconnected system not considered
This article	Stockwell Transform + Deep Learning	Yes	Yes	Industrial + Office/Residential	Yes	0.8	100	All major gaps are addressed

7 Conclusions

A new methodology based on signal processing and deep learning algorithm has been used to identify the type of nonlinear loads present in an interconnected power system. Realistic modelling has been done to obtain the current waveforms of various nonlinear loads. Frequently found nonlinear loads viz. Arc Furnaces, Electric Traction, VFD/ASD and two different loads which contribute interharmonics like cycloconverters and low magnitude harmonics due to ageing are considered for identification. Three cycles of the current waveform recorded at the source are then processed through the ST methodology to obtain the time-frequency plot. This TF plot is then scaled to 227x227 pixels size to suit the input data size of the Deep learning algorithm.

All the five types of nonlinear loads are processed with ST, and 100 images of the TF plots of each load are then fed to the algorithm. After training with 60 images of each nonlinear load, it was observed that the algorithm with both learning rates and validation frequency had produced 100% accurate results. Each case has taken 42s and 46s for training for case 1 and case 2. Since, there is only a difference of 4s between both the cases, the network with a learning rate of 1e-4 and validation frequency of 3 is considered for identification due to lesser loss and early 100% accuracy stage.

The deepdream images of this network have been visualized and found that the network has expertly trained with the provided images. The confusion plot for both the cases shows Zero overlap and 100% accuracy in all the cases. The proposed technique, in comparison with SVM and other methods reported in the literature, had outperformed in terms of fastness and accuracy. Such a system can be used to predict the type and location of nonlinear loads by the utilities to make necessary actions like penalizing, revising the contract demands and improving the system performance by installing the filters.

Footnotes

Appendix -A

IEEE-9 Bus System Data

Parameter	G1	G2	G3
Operation Mode	Swing	Voltage Control	Voltage Control
Rated MVA	80	220	110
kV	16.5	18	13.8
Power factor	0.9	0.85	0.85
Type	Hydro	Thermal	Thermal
Speed	1500	1500	1500
Tdo’	5.6	5.6	5.6
Tqo’	3.7	3.7	3.7

Line Data
Line	Resistance	Reactance	Susceptance
		(in PU)
1-4	0	0.0576	0
4-5	0.017	0.092	0.158
5-6	0.039	0.17	0.358
3-6	0	0.0586	0
6-7	0.0119	0.1008	0.209
7-8	0.0085	0.072	0.149
8-2	0	0.0625	0
8-9	0.032	0.161	0.306
9-4	0.01	0.085	0.176

Load Data
Bus	Type of Bus	Voltage		Load		Generation
		Voltage	δ	P	Q	P	Q
		(PU)	(θ)	(MW)	(MVAr)	(MW)	(MVAr)
1	Slack	1.03	0	0	0
2	Load	1	0	10	5
3	Load	1	0	25	15
4	Load	1	0	60	40
5	Load	1.06	0	10	5	80
6	PV	1	0	100	80
7	Load	1	0	80	60
8	PV	1.01	0	40	20	120
9	Load	1	0	20	10

References

Romero

M.F.

, Gallego

, Meyer

and Müller

, Characterization of nonlinear household loads for frequency domain modeling, Ingeniería E Investigación 35(1Sup) (2015), 65–72. DOI: 10.15446/ing.investig.v35n1sup.53895.

Sudiharto

, Anggriawan

D.O.

, Tjahjono

, Harmonic Load Identification Based on Fast Fourier Transform and Levenberg Marquardt Backpropagation, (2017). http://www.jatit.org/volumes/Vol95No5/11Vol95No5.pdf

Djordjevič

and Simic

, Nonintrusive identification of residential appliances using harmonic analysis, Turkish Journal of Electrical Engineering and Computer Sciences 26 (2018), 780–791. DOI: 10.3906/elk-1705-262

Hatata

A.Y.

and Eladawy

, Prediction of the true harmonic current contribution of nonlinear loads using NARX neural network, Alexandria Engineering Journal 57 (2017), 1509–1518. http://dx.DOI.org/10.1016/j.aej.2017.03.050

Rakhmanov

N.R.

, Kurbatsky

V.G.

, Guliyev

, Tomin

N.V.

, Rakhmanov

, Analysis and modeling of harmonic distortions in consumer’s distribution networks, Containing a Powerful Nonlinear Load, (2018). https://doi.org/10.1051/e3sconf/20185803015

Monteiro

F.P.

, Monteiro

S.A.

, Tostes

M.E.

and Bezerra

U.H.

, Using True RMS Current Measurements to Estimate Harmonic Impacts of Multiple Nonlinear Loads in Electric Distribution Grids, Energies 12 (2019), 4132. DOI: 10.3390/en12214132

Beck

and Machlev

, Harmonic Loads Classification by Means of Currents’ Physical Components, Energies 12 (2019), 4137. DOI: 10.3390/en12214137

Shklyarskiy

Y.E.

, Skamyin

A.N.

, Vladimirov

and Gazizov

F.Y.

, Distortion Load Identification Based on the Application of Compensating Devices, Energies 13 (2020), 1430. DOI: 10.3390/en13061430

Pullabhatla

, Koley

, Fuzzified Inverse S-Transform for Identification of Industrial Nonlinear Loads, In: Satapathy S., Govardhan A., Raju K., Mandal J. (eds) Emerging ICT for Bridging the Future - Proceedings of the 49th Annual Convention of the Computer Society of India (CSI) Volume 1. Advances in Intelligent Systems and Computing, vol 337, (2015), Springer, Cham. https://doi.org/10.1007/978-3-319-13728-5_39.

10.

Srikanth

, Koley

, Identification of Industrial Nonlinear Loads Using S-Transform Aided Fuzzy Classifier, In: Mishra S., Sood Y., Tomar A. (eds) Applications of Computing, Automation andWireless Systems in Electrical Engineering. Lecture Notes in Electrical Engineering, vol 553. Springer, Singapore, First Online 01 June 2019. DOI: https://doi.org/10.1007/978-981-13-6772-4_27.

11.

Srikanth

, Koley

, Intelligent System for Monitoring and Recognizing the Type of Domestic Harmonic Loads, In: Malik H., Iqbal A., Yadav A. (eds) Soft Computing in Condition Monitoring and Diagnostics of Electrical and Mechanical Systems, Advances in Intelligent Systems and Computing, vol 1096, (2020), Springer, Singapore First Online 18 January 2020. DOI: https://doi.org/10.1007/978-981-15-1532-3_11.

12.

Ozcanli

A.K.

, Yaprakdal

, Baysal

, Deep learning methods and applications for electrical power systems: A comprehensive review, International Journal of Energy Research, (2020). DOI: 10.1002/er.5331

13.

Bharath

B.S.

, Vinayaka

K.U.

, Investigation of power quality disturbances in an electric arc furnace, 2017 International Conference on Energy, Communication, Data Analytics and Soft Computing (ICECDS), (2017), 2268–2273. 10.1109/ICECDS.2017.8389856

14.

ABB DRIVES, Technical Guide No. 6. Guide to harmonics with AC drives.

15.

http://ps.eei.eng.osaka-u.ac.jp/jaug/haginomori/index.html

16.

Sapijaszko

, Mikhael

W.B.

, An Overview of Recent Convolutional Neural Network Algorithms for Image Recognition, 2018 IEEE 61st International Midwest Symposium on Circuits and Systems (MWSCAS), (2018). 743–746. DOI: 10.1109/MWSCAS.2018.8623911

17.

https://towardsdatascience.com/the-w3h-of-alexnet-vggnet-resnet-and-inception-7baaaecccc96#: :text=AlexNet%20and%20ResNet%2D152%2C%20both,training%20time%20and%20energy%20required.