A deep CNN approach for islanding detection of integrated DG with time series data and scalogram

The ever increasing demand for electricity leads to the advancement of distributed generation (DG). Almost all DG sources are renewable nature. One of the major complications with the high penetration of DG sources is islanding. The islanding may damage the clients and their equipment. As per the IEEE 1547 DG interconnection standards, the islanding will be identified in two seconds and the DG must be turned off. In this paper, an advanced islanding detection process stands on a deep learning technique with continuous wavelet transforms and convolution neural networks implemented. This approach transforms the time series information into scalogram images, and later, the images are used to train and test the islanding and non-islanding events. The outcomes are correlated with the artificial neural networks and fuzzy logic methods. The comparison shows that the proposed deep learning approach efficiently detects the islanding and non-islanding events.


Introduction
The high integration of DG systems makes the power system network further complex. One of the major complications as a result of such DG assimilation is islanding. It is a situation where DG feeds the regional loads after disconnecting from the utility grid (Reddy et al. 2022). It can be intentional or unintentional. The intentional islanding arises with the maintenance of utility. The unintentional islanding may cause due to utility grid failure or uncertainties in the power network (Cui et al. 2018). It not only damages the customer's appliances and personal but also makes the grid cumbersome (Raju et al. 2021). Considerable islanding detection approaches are recommended by the researchers. They are briefly described here.
The passive methods encounter the situation by regularly auditing the passive parameters at the point of common coupling (PCC) and comparing it with the predefined threshold value (Reigosa et al. 2017). The passive parameters are voltage, current, frequency, impedance, phase angle, etc. If the parameter exceeds the specified value, the method affirms the islanding (Rami et al. 2021). However, they have been suffering from massive non-detection zone (NDZ) and complexity in fixing threshold values (Salles et al. 2015;Rami and Harinadha Reddy 2019a). To overcome these demerits, active methods are suggested. In active methods, a low-frequency harmonic signal is continuously injected and the parameters at PCC are monitored (Rami and Harinadha Reddy 2019b). In grid-connected affair, the injected signal will not affect the monitored parameters, but in the islanding case, it leads to the discrepancy in the observed guidelines. The particular discrepancies have been used to find the islanding (Murugesan and Murali 2020;Sivadas and Vasudevan 2020). These recommendations have no NDZ, but they are degrading the quality of power (Rami and Harinadha Reddy 2018). To eliminate the drawbacks of active methods, hybrid methods are proposed. They are the aggregate of active and passive approaches (Kermany et al. 2017). When the passive method suspects the islanding case, the active approach confirms it (Ch and Harinadha Reddy 2018). These methods have no NDZ and affect power quality less compared to active methods (Chen et al. 2019). The remote islanding approach finds the islanding by collecting data from utility and DG (Xu et al. 2007). Various signal processing approaches have been proposed by the researchers which reduce the NDZ and enhance the performance of the passive methods by extracting the hidden features from the passive parameters (Reddy et al. 2020;Reddy and Harinadha Reddy 2019;Do et al. 2016). Artificial intelligence learning models classify the islanding and non-islanding events without threshold settings efficiently (Khamis et al. 2018). They do not have NDZ, but large data are required for training the models (Kermany et al. 2017). It is compulsory to produce an accurate islanding detection technique due to advancements in smart grid technology and the complexity of the power system network in the future. This paper presents a new IDM based on deep learning. This method uses CWT and CNN. First, the time series data obtained at PCC are transformed toward the scalogram illustrations with CWT which contain the data of various islanding and non-islanding events. Later the scalogram images will be used to train the proposed CNN model. This is the second attempt of applying image processing techniques for the classification of islanding cases. The remaining part of the paper is organized as per the following aspects. Segment 2 describes the practice of transforming time series input toward scalogram illustrations. Segment 3 describes the test system and data set preparation. In Sect. 4, the design and training of CNN are presented. The results and discussions are illustrated in Sect. 5. Section 6 presents the conclusion.

Time series data to Scalogram image conversion
This section presents the operation of transforming the time series signal toward the scalogram appearances. The signal data of (1) are used to prepare the basic scalogram image (Manikonda and Gaonkar 2019). It is one second duration composed of two different frequencies 10 Hz and 200 Hz near amplitudes 15 and 25, respectively. The amplitudes and frequencies are randomly selected for illustrating the explanation. This approach uses the wavelet transform of a signal.
The wavelet transform of any signal f ðtÞ can be specified as: In wavelet transforms, the time frequency energy density of a signal is a scalogram. In simple words, a scalogram is an observable impersonation of the wavelet transform, to what end x, y and z axes produce the time, frequency and magnitude in color gradient, respectively (Sejdic et al. 2008). The scalogram of time series results represented in Eq. (1) is depicted in Fig. 1. It is obtained by applying the CWT with Morse wavelets. Figure 1 shows two frequencies 10 Hz and 200 Hz and two amplitudes 15 and 25, respectively. In this manner, any time series data can be converted into scalogram images. It is generally known that any supervised learning requires data set for training of the network and testing. In this paper, the data set is prepared with scalogram images of different time series events. The next section describes the test system and data set preparation for the training of CNN in detail.

Test system and data set preparation
Large training information is needed for testing any supervised learning methods. For problems related to image classifications, standard data sets are available. No such standard data sets are available for islanding detection methods. Hence, a standard test system is appropriate for developing a sufficient data set. A 100 KW grid integrated PV source shown in Fig. 2 is considered to create such a data set. This model has been adopted in such a way as to satisfy the proposed work. The simulations are born in the MATLAB/Simulink platform. At t = 0.4 s, by opening the CB (circuit breaker), the islanding event is created. The phase angle between the positive sequence component of voltage and current at PCC is acquired for 6 cycles at 1000 samples per second. A PC with an i5 processor, 8 GB RAM and Windows 10 operating system is used to get these simulations. For producing the image data set, different islanding and non-islanding events are validated and their results are recorded as time series plots.
CWT is applied to each time series data, for the generation of scalogram images. The scalogram of the phase angle between the positive sequence component of voltage and current at PCC for grid integrated and disconnected operations are shown in Fig. 3. It is observed against the scalogram illustrations that there is a good variation among the islanding and non-islanding images. The image classification technique is applied to these images for the detection of events. Most of the passive approaches are failed to detect the islanding cases when there is a zero or small power variation among the DG and the load in the islanding situation. This situation is taken into account, and different islanding capsules at nearly worst power mismatches are studied and included in the data set. The data set also includes several islanding cases and non-islanding cases such as switching of loads, capacitor banks, shortcircuit faults and motor switching events. A total of 300 islanding and non-islanding are generated for data set creation, which include 150 islanding and 150 non-islanding events that are listed in Table 2.

Methodology and CNN design
This segment presents the methodology, architecture and training particulars of CNN. Figure 4 represents the steps in the proposed islanding detection process. The phase angle between the positive sequence component of voltage and current is acquired at PCC in time series form. This knowledge is transformed into scalogram pictures.
The scalogram pictures are given as input to the already experienced CNN for the classification of events. For any supervised learning method, feature extraction is crucial for workouts and examinations. The accuracy of the approach depends on these features. In deep learning, CNN naturally extracts these features from the input pictures. It has multiple layers, most of the layers are used for feature extraction and only the concluding minority layers are used for analysis. The general structure of CNN is depicted in

Convolution layer
In deep learning, convolution is a mathematical operation on two functions. Among the two functions, one function is an image in the form of pixels at the point on the picture and the other function is the kernel. Both are characterized as a cluster of numbers. The multiplication of these two arrays accords to the outcome. The filter is now moved to another position on the image which is decided through the stride duration. The convolution is continued as far as the total picture has been covered. The output of these computations is an activation map. Unlike the artificial neural networks where all input neurons are connected to all the output neurons, CNN has sparse connections, which means only the input neurons have only a few connections with the next layer neurons. The convolution activity is represented by the Ã operator. Output f ðxÞ is characterized when the input IðxÞ is convoluted with the kernel KðxÞ as (3): If x takes only integer attitudes, the discretized convolution can be defined as (4), which assumes the one-dimensional convolution The two-dimensional convolution with input Iða; bÞ and filter Kðm; nÞ is illustrated as (5): By commutative law, filter is flipped and Eq. (5) is corresponding to (6): Neural networks appliance the cross-correlation operation, it is the same as the convolution operation without flipping the filter and the Eq. (6) changes to (7). Figure 6 shows the convolution operation in detail

Rectified liner unit (ReLu) layer
The activation function at the yield of the convolution lamination is linear naturally. The activations commonly happen through the ReLu unit, for getting the nonlinear transformation. There are different types of activation functions; a few among the familiar functions are tanh, sigmoid and rectified linear unit (ReLu). In this CNN architecture, ReLu activation function is used at the output of previous layers. It can be represented in Fig. 7. ... Here, x is the input to the neuron. It gives the output as zero if the input is negative and it gives the same output if the input is a positive value. This layer simplifies the calculations and accelerates the design, and it advises escaping the fading gradient problems.

Pooling layer
The pooling sheet lowers the resolution of the extractions. This layer produces the extractions strong counter to distortion and noise. Here are four types of pooling: They are max pooling, average pooling, L2 normalization and sum pooling. In these classifications, the input is separated into non-overlapping two-dimensional zones. For max pooling, the maximum value of zone values is considered as output. For average pooling, the average of zone values is considered as output, and for sum pooling, the sum of all values in the zone is considered. The proposed approach uses the max pooling layer (Fig. 8).

Softmax layer
Softmax layer provides the probabilities of all classes for ndimensional input real numbers vector. These probabilities are used for classification. Mathematically, it can be represented as in Eq. (8): All the determined contingencies are in the dimension of zero and one. The importance of this function is that it can add the entire probabilities up to one.

Fully connected layer
These are the output layers of the CNN. This layer produces the output classification. Every neuron in a fully connected layer has a connection with all neurons in the last layers. All the features received from the previous layers are weighted together to produce the specific

Design of CNN for islanding detection approach
In this paper, the CNN is constructed for the classification of different islanding and non-islanding events. Several aspects are taken into account while constructing the CNN. The subsequent steps are initially supported. During the training process, all the hyperparameters are uninterrupted initially. This will help in identifying the number of layers required for good efficiency. Once the statistics of slabs are identified, the variation of hyper parameters is identified for optimal values and they are fixed while designing CNN. It is initially started with a single layer. Every layer of CNN implements three operations such as convolution, ReLu activation and max pool operation. Once the CNN is designed and executed successfully for a single layer, the other layer is added and the same operations are repeated until it gets high accuracy. The response for the number of layers on accuracy found that eight layers architecture has good accuracy compared to five and seven layers. Hence,  9 Islanding case for 100% of load Fig. 10 Islanding case for 80% of load eight layers architecture is fixed for the CNN design for the classification of islanding and other events. Once it is fixed, the next step is the investigation of the size of the filters. It is found that 3 9 3 kernel has good output compared to 5 9 5 and 11 9 11 kernels. The variation of learning rate and momentum with stochastic gradient descent with the momentum method is verified. The learning rate of 0.001 accords good outputs in terms of accuracy and loss. The CNN design parameters and data set information and customized CNN parameters are listed in Tables 1, 2, 3, respectively.

Results and discussion
The constructed structure is experienced with 75% of data and tested with 25% of data. 25% of data are completely unseen by the designed network. The data set contains the islanding events and various non-islanding events. The non-islanding events include load switching, capacitor switching, feeder switching, and fault switching for ON/ OFF cases. In all these cases, the time series data are transformed into scalogram pictures. The voltage variations for islanding and non-islanding events are reflected as color gradients in the scalogram images. The few testing scalograms are depicted in Figs. 9, 10, 11, 12, 13. A total of 65 (25 islanding? 40 non-islanding) cases are tested. Out of all the testing cases, only 3 cases are wrongly predicted. The accuracy and loss plots for training and validation are depicted in Fig. 14.

Conclusion
This paper presents a novel islanding detection method with CWT and CNN. The time series data of phase angle between positive sequence component of voltage and current obtained from Simulink are transformed into scalogram images. The data set is prepared with 258 events of islanding and non-islanding cases. 75% of the data set has been used for training the CNN, and the remaining 25% (65 cases) is used for testing. Out of tested 25 islanding and 40 non-islanding cases, only three non-islanding cases are wrongly predicted. This method has an accuracy 95.4%. It has been found that the deep learning-based CNN can  detect the islanding classifications effectively compared to machine learning approaches.
Funding No funding received from external funding agencies.

Declaration
Conflict of interest The authors declare that they have no conflict of interest.