Portal de Periódicos da CAPES

Evaluating the effectiveness of a machine learning approach based on response time and reliability for islanding detection of distributed generation

2017; Institution of Engineering and Technology; Volume: 11; Issue: 11 Linguagem: Inglês

10.1049/iet-rpg.2016.0987

1752-1424

Mollah Rezaul Alam, Kashem M. Muttaqi, Abdesselam Bouzerdoum,

Power Systems Fault Detection

IET Renewable Power GenerationVolume 11, Issue 11 p. 1392-1400 Research ArticleFree Access Evaluating the effectiveness of a machine learning approach based on response time and reliability for islanding detection of distributed generation Mollah Rezaul Alam, Corresponding Author Mollah Rezaul Alam mra015@aiub.edu Department of Electrical & Electronic Engineering, American International University-Bangladesh (AIUB), Banani, Dhaka, 1212 BangladeshSearch for more papers by this authorKashem M. Muttaqi, Kashem M. Muttaqi School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, NSW, 2500 AustraliaSearch for more papers by this authorAbdesselam Bouzerdoum, Abdesselam Bouzerdoum School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, NSW, 2500 Australia College of Science and Engineering, Hamad Bin Khalifa University (HBKU), Education City, Doha, QatarSearch for more papers by this author Mollah Rezaul Alam, Corresponding Author Mollah Rezaul Alam mra015@aiub.edu Department of Electrical & Electronic Engineering, American International University-Bangladesh (AIUB), Banani, Dhaka, 1212 BangladeshSearch for more papers by this authorKashem M. Muttaqi, Kashem M. Muttaqi School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, NSW, 2500 AustraliaSearch for more papers by this authorAbdesselam Bouzerdoum, Abdesselam Bouzerdoum School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, NSW, 2500 Australia College of Science and Engineering, Hamad Bin Khalifa University (HBKU), Education City, Doha, QatarSearch for more papers by this author First published: 11 July 2017 https://doi.org/10.1049/iet-rpg.2016.0987Citations: 25AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract Conventional relays, such as vector surge relay, frequency relay and rate-of-change-of-frequency relay, are usually employed for islanding detection; however, these conventional relays fail to detect islanding incidents in the presence of small power imbalance inside the islanded system. This study presents an islanding detection approach for synchronous type distributed generation using multiple features extracted from network variables and a support vector machine (SVM) classifier. Features are extracted from a sliding temporal window, whose width is selected so as to achieve the highest detection rate at a fixed false alarm rate. The SVM classifier is trained with linear, polynomial and Gaussian radial basis function kernels, and the parameters of the kernels are tuned to improve the classification performance. The application of the proposed method is illustrated for islanding cases associated with different power imbalance conditions, including small power imbalance conditions associated with the non-detection zone of conventional relays. Furthermore, variation of detection time as a function of power imbalance scenarios, which involve all probable combinations of deficit of active/reactive and excess of active/reactive power imbalance, is assessed in the testing phase. The performance of the proposed approach is evaluated and compared with those of conventional relays in terms of reliability and response time of islanding detection. 1 Introduction The penetration of distributed generation (DG) is forcing the electricity system planners and operators to develop standards, referred to as interconnection guidelines (IG) of distributed resources with electric power system (EPS) [1, 2]. An essential requirement for IG is the protection of DG islanding, also known as the protection of 'loss of mains'. Islanding is a situation when a portion of EPS is energised solely by one or more local DG systems while that portion of the EPS is electrically isolated from the rest of the EPS [1]. This electrical isolation may occur due to switching of feeder, operation of switchgear and/or action of fault clearing and so on. According to IG [1, 2], islanded DG must be disconnected quickly in order to avoid possible hazardous conditions, such as power quality degradation and damage of utility and customer equipment. The IEEE 1547–2003 standard recommends that the DG disconnection time should be <2 s. However, recent trend of fast automatic reclosing may have adverse effect (e.g. causing serious damage) on the islanded synchronous DGs as well as on neighbouring utility equipment if 2 s of DG disconnection time range is practiced [3]. Therefore, this disconnection time range may need to be reduced to permit the disconnection of islanded DG prior to completing the first reclosing operation [1, 2, 4]. Normally, islanding detection is performed by a special protection device, namely islanding detection relay whose operating principles are based on active, passive, hybrid and remote communication methods [5]. Active methods are reliable, but they incur a power quality issue as periodic disturbance is introduced in this method. Remote communication based methods, such as power line signalling and transfer trip, are the most reliable scheme, but they require an enormous cost on infrastructure development. Passive methods have low cost but display poor performance when power imbalance in the islanded network is very small [5-8]. The conventional relays, for instance, frequency relay (FR), vector surge relay (VSR) and rate-of-change-of-frequency (ROCOF) relays, are operated on the principle of passive islanding detection methods. ROCOF and VSR relays have a strong correlation with active power imbalance of the islanded system. The performance of VSR and ROCOF relays deteriorates when the power imbalance drops below a specified threshold, which gives rise to the limitation of non-detection zone [9]. Hybrid detection scheme is a combination of active/passive methods; it aims to maximise the advantages and minimises the disadvantages of both methods. Considering cost and reliability, machine learning approach has been shown recently as a potential tool for islanding detection [3, 5, 10-13]. To this end, this paper proposes and assesses the effectiveness of a passive islanding detection method which involves multiple features and a support vector machine (SVM) classifier. In [5, 12], an SVM-based approach was proposed to demonstrate the reliability of the method in terms of accuracy of islanding detection with a minor risk of nuisance tripping [false alarm (FA)]. The classification performance was assessed based on features extracted from a window spanning ten cycles in time [5, 12]. In [12], the SVM-based approach was compared with a decision tree based approach presented in [13]. The same test network and test cases were considered for the comparative study. Test results indicate that the SVM-based technique can detect all islanding events in the test cases, unlike the method presented in [13], which fails to detect three islanding cases in the presence of 5% power imbalance as reported by the authors. However, for real-time applications the SVM-based relay (SVMR) needs to be investigated for its detection accuracy and speed, i.e. response time or detection time. Therefore, this study incorporates a sliding time window, and investigates the performance of the SVMR as a function of the window length. The window size imposes a tradeoff between the speed and detection accuracy: increasing the width of the window increases the detection accuracy at the expense of response time. In order to achieve an optimum tradeoff, the smallest window size is selected that achieves the highest detection rate (DR) at a given FA rate. The SVMR is assessed considering the islanding cases with all possible combinations of deficit and excess of active/reactive power imbalance. Moreover, a majority voting rule is proposed to assess the effectiveness of SVMR under real-time environment considering response time and accuracy. The rest of the paper is structured as follows. Section 2 presents the proposed methodology describing the feature extraction, theory of SVM, training and testing of SVM for the detection of islanding. Section 3 investigates the evaluation of the method using the features obtained from different window sizes. In this section, the islanding cases involving all possible combinations of power imbalance scenarios are also tested. In Section 4, the proposed approach is examined in various realistic scenarios that may be encountered under real-time operation. Comparative analysis with conventional relays, in terms of reliability and detection time, is also carried out in this section. Section 5 concludes the paper. 2 Machine learning approach for islanding detection The proposed methodology is narrated in three sub-sections. Section 2.1 describes the process of feature extraction from plausible islanding and non-islanding events. The theory of SVM, relevant to the classification of two classes, is briefly discussed in Section 2.2. Finally, Section 2.3 narrates the training and testing procedure of SVM. 2.1 Feature extraction A 3-bus radial distribution network containing distribution line, upstream grid and a synchronous generator (SG) type DG (see Fig. 1) is considered for demonstrating the process of feature extraction. Fig. 1Open in figure viewerPowerPoint 3-bus radial distribution network with SG type DG The behaviour of network variables can be examined from the vector diagram shown in Fig. 2. The diagram (i.e. Fig. 2) shows the voltage behaviour at DG connection point; it is drawn based on the values obtained by simulating the islanding scenarios generated by opening the circuit breaker (CB) in the system shown in Fig. 1. During islanding period, the system is composed by the SG and load only. At this instant, the SG starts to feed a smaller (or larger) load because the current injected into (or provided by) the utility side is suddenly interrupted. Thus, the generator begins to accelerate (or decelerate) its rotor speed to reduce the gap of power mismatch. Therefore, the terminal voltage and angle with respect to a reference are affected; which are illustrated in one pre-islanding (solid line) and two post-islanding scenarios (dashed and dash-dotted lines) of Fig. 2. The dashed and dash-dotted lines represent the behaviour of voltage at the connection point of DG during sixth and seventh cycles after the onset of islanding. In islanded mode, change of voltage behaviour in each cycle is observed, which is influenced by the dynamics of SG. Note that voltage and current phasors of Fig. 2 are extracted at each cycle by applying discrete Fourier transform on the instantaneous voltage and current signals. Fig. 2Open in figure viewerPowerPoint Phasor diagram representing the voltage behaviour at DG connection point of the system shown in Fig. 1 during pre- and post-islanding condition From Fig. 2, it is noticeable that islanding provokes the variations of voltage magnitude and phase angle. Moreover, a change in frequency is also observed from the change of period of voltage cycle. Therefore, in the proposed method, five variables are employed for feature extraction: frequency (f), rate of change of frequency , rotor angle (δ), voltage (V) in pu and rate of change of voltage . Five features are extracted from these five variables, by taking standard deviation (SD) inside a sliding data-window having a width of ΔT. For instance, feature from a signal s(t), which can be any of the five network variables, are extracted by taking SD inside the ΔT width of a sliding data-window. Following this procedure, five features are extracted from five network variables, which are obtained during islanding and non-islanding situations such as capacitor switching, load switching and so on (see Fig. 3 for illustration). Fig. 3Open in figure viewerPowerPoint Illustration of feature extraction from five network variables (a) V, (b), (c) f, (d) , (e) δ using a sliding data-window of ΔT width Mathematically, the five features can be presented as follows: (1) (2) (3) (4) (5) where represent the features extracted from the network variables: voltage (V), frequency (f), rotor angle (δ), rate-of-change-of-voltage and ROCOF , respectively. Therefore, the feature vector is given by (6) where [.]T denotes the transpose operator. 2.2 Support vector machine classifier In the present analysis, two groups of data, extracted from islanding and non-islanding events, are classified using SVMs. Therefore, a brief overview of the theory of SVM, proposed by Vapnik and co-workers [14, 15], is presented in this sub-section. For a two-class classification problem, a real valued input or feature vector can be labelled as , which indicates the class of. The SVM separates the two classes by establishing a decision boundary hyperplane defined by its normal vector w and a scalar bias b(7) The function can be used as a decision function to obtain an output , indicating the class of the input feature vector . The optimal separating hyperplane is the one with minimum Euclidean norm of w, satisfying . Training of the SVM amounts to solving the following quadratic programming (QP) problem: (8) where is a slack variable and the scalar C is a regularisation parameter, which determines the tradeoff between the maximisation of the margin and the minimisation of classification errors. To solve the QP problem of (8), Lagrange multipliers and are introduced, which yields w in the form (9) Clearly, w is determined by the training data corresponding to Lagrange multipliers which are non-zero and for which the constraints in (8) are exactly met; these training samples are known as support vectors (SVs). The final decision boundary g(x) can be expressed as (10) where x is the input test vector. As indicated earlier, the decision function can be taken as , with a tuning parameter C. The kernel trick is also incorporated with SVMs to classify non-linearly separable classes. Therefore, replacing the inner product in (10) with the kernel function yields the decision function (11) Some popular choices of kernel functions used with SVMs include: Radial basis function (RBF) kernel: . Polynomial kernel of degree p: The kernel parameters p and σ, along with C, are used as input parameters to the SVM training process; they are tuned to achieve the desired tradeoff between training and generalisation performance. 2.3 Training and testing of SVM for islanding detection Training of SVM is conducted off-line by extracting the input features from all possible scenarios of islanding and non-islanding events which may occur in DG networks. For each non-islanding and islanding events, the feature vector x, as presented in Section 2.1, is obtained from a window of width ΔT. In the training phase, the location of the onset of islanding is known a priori (ground truth); therefore, any window that includes the islanding onset is considered as islanding case. Thus, two groups of labelled data: islanding and non-islanding, are stored in a feature matrix. Then, soft-margin SV classification algorithm employing quadratic programming (QP) optimisation is applied. Cross-validation is conducted on the training set to obtain the optimum values of the regularisation parameter C, kernel parameters σ (for Gaussian RBF kernel) and order of polynomial p (for polynomial kernel). Testing of SVM is carried out by using the features extracted from a different set of islanding and non-islanding events, generated separately from the training set. Performance of the SVM classifier is evaluated using the DR and FA, where DR indicates the ratio of the number of successfully detected islanding events to the total number of islanding events, and FA indicates the ratio of the number of misclassified non-islanding events to the total number of non-islanding events. 3 Test results of the proposed approach The proposed machine learning approach is assessed for the islanding detection of a test network energised with synchronous type DG units at distribution feeder. The detailed description of the test system, generation of plausible islanding and non-islanding test cases or events, and their classification results are presented in the following three sub-sections. 3.1 Test system The SVM-based approach is tested by the events generated through the simulation of a test network of Fig. 4. MATLAB/SIMPOWER software is used to build the test system model. The sampling frequency during the simulation study is kept at 2 kHz, and the relays of CBs are placed at the connection points of transformers adjacent to SG1, SG2 and SG3. These relays are used to collect the voltage signal during islanding and non-islanding conditions. Note that the simulation study involves the sampling rate of 2 kHz, therefore, the relays at CB's end would receive the voltage at 2 kHz sampling rate (i.e. 40 samples/cycle for 50 Hz system); this sampling rate is realistic in power system, since for phasor measurement unit, which is a reliable device for measuring the voltage in electricity grids, can process 10–256 samples in each cycle for 50 Hz system. Fig. 4Open in figure viewerPowerPoint Single line diagram of a 10-bus radial distribution network under study As illustrated in Fig. 4, the test system is a radial distribution network having base power of 18 MVA. The distribution network is connected with 132 kV, 50 Hz, sub-transmission system with fault level of 1000 MVA, shown by a Thévenin equivalent (Sub). A 33 kV distribution system is connected with the sub-transmission system or grid side through a 132/33 kV transformer. There are three 6 MW, 1.2 MVAr SGs connected to the distribution system through 33/0.69 kV transformers at buses 6, 8 and 10, respectively (see Fig. 4). For this simulation study, three-phase models of all network components are used. The π section lines are modelled as distribution lines. Loads are modelled as dynamic loads which are of constant impedance, constant current and constant power type. The SG is represented by a sixth-order three-phase model in the dq rotor reference frame [16] and it is equipped with an automatic voltage regulator (AVR) represented by the IEEE Type 1 model. More information about the test system can be obtained from [5]. 3.2 Test cases/events The network events that result in the isolation of the DG energised distribution network from the supply of the upstream network (or grid side) are considered as islanding conditions. Normal events that may persist in real power systems due to normal operation or disturbances, such as, capacitor switching, loss of lines, load addition, load disconnection, faults and so on, for which DG energised network is not isolated, are considered as non-islanding conditions. In this study, simulation of non-islanding cases include: (i) normal operation or normal condition, (ii) temporary faults which include balanced three-phase faults, unbalanced single- and double-phase faults, (iii) switching of capacitor banks, (iv) switching due to addition and/or disconnection of loads, (v) disconnection of DGs apart from the target DG, (vi) loss of any branch in the radial distribution feeder, which is away from the distribution line connected to the target DG. The islanding cases are simulated by opening the CB or feeder breaker (B1) of Fig. 4 under the following network conditions [17]: (i) Containing wide range of active power imbalance (deficit and excess, varying from 0 to 100%) in the islanded portion, (ii) Containing wide range of active power imbalance (deficit and excess, varying from 0 to 100%) and reactive power imbalance (deficit and excess, varying from 0 to 50%) in the islanded portion, (iii) Containing three types of loads: constant impedance, constant current and constant power. To generate the islanding events with a wide range of active and reactive power imbalance, the load-generation profile in the islanded segment is varied by applying the procedure presented in [17]. 3.3 Classification of events using SVM A total of 2817 events (see Table 1) are generated for assessing the classification performance of SVM. It is worth noting that the generated islanding events have covered all four probable combinations or groupings of power imbalance scenarios; which include, (a) deficit of ΔP and deficit of ΔQ, (b) deficit of ΔP and excess of ΔQ, (c) excess of ΔP and excess of ΔQ, (d) excess of ΔP and deficit of ΔQ [9]. Table 1 shows the numbers of training and test events, which reveal that the number of generated islanding events is greater than the number of non-islanding events. Note that the probability of occurring islanding incidents is rare in the distribution network in comparison to the non-islanding incidents. However, in this work, the probability of islanding occurrence is not considered; rather the different scenarios of islanding and non-islanding cases are taken into account. Furthermore, in comparison to testing data less number of training data is used. As stated earlier, the training data is selected as a different subset from the test data. For the generation of islanding training data, all possible combinations of power imbalance scenarios, where ΔP ranges from 0 to 100% (step size 4.8%) and ΔQ ranges from 0 to 50% (step size 10%), are considered. Thus, a total of 249 islanding events are generated for training. The training data is used to obtain the optimum parameters of the SVM classifier (see Section 2.3 for further illustration). In order to investigate the effectiveness of the proposed method, the trained SVM is tested on a large number of separate islanding and non-islanding events (consisting of 1848 islanding and 471 non-islanding): the test events are generated by varying ΔP from 0 to 100% (step size 0.5%) and ΔQ from 0 to 50% (step size 10%). Table 1. Generated islanding and non-islanding events under different combinations of power imbalance level Scenarios Islanding events Non-islanding events Training Test Training Test (a) deficit ΔP and deficit ΔQ 63 462 249 471 (b) deficit ΔP and excess ΔQ 62 462 (c) excess ΔP and excess ΔQ 62 462 (d) excess ΔP and deficit ΔQ 62 462 total 249 1848 249 471 The performance of SVM classifier is assessed from the classification results of numerous events of Table 1, by using the five features extracted through a window of ΔT width (see Section 2.1). The width of the window is optimally selected by conducting the SVM-based classification, using the features extracted from the situations having window width (ΔT) of one-cycle, five-cycle, eight-cycle and ten-cycle. For each situation, the threshold value of SVM classifier is varied gradually to obtain the DR and FA. Thus, the receiver operating characteristics (ROC) curve, as shown in Fig. 5, is obtained. Fig. 5Open in figure viewerPowerPoint ROC curve of the proposed approach for one-cycle, five-cycle, eight-cycle and ten-cycle data-window length The ROC curve of Fig. 5 indicates that the classifier's performance using eight-cycle and ten-cycle data-window are almost similar; and their performances are comparatively better than the five-cycle and one-cycle data-window for FA ≤ 5%. However, considering the response time and performance, eight-cycle data-window is selected as optimal data-window to investigate the performance of the proposed method. Thus, using the extracted features through eight-cycle data-window and applying the SVM classifier with linear, polynomial and Gaussian RBF kernels, the test islanding and non-islanding events are classified. The bound on the regularisation parameter 'C' is selected as 10 after performing the grid search using the SVM classifier. The test results presented in Table 2 show that linear and polynomial kernel yield almost similar performance for scenarios (a) and (b), considering DR and FA. However, taking all four power imbalance scenarios into account, polynomial kernel shows the best classification performance among the three kernels, as indicated in Table 2. Table 2. Performance of SVM classifier including all four combinations of deficit of active/reactive and excess of active/reactive power imbalance Kernels Kernel pa­rameter Scenario (a) Scenario (b) Scenario (c) Scenario (d) DR, % FA, % DR, % FA, % DR, % FA, % DR, % FA, % linear — 100 3.08 100 3.08 94.7 3.08 94.1 3.08 Gaussian RBF σ = 1.5 94.6 3.34 93 3.34 98.2 3.34 98.8 3.34 polynomial p = 3 100 3.29 100 3.29 99.2 3.29 99.7 3.29 4 Performance evaluation and comparative analysis Islanding detection tool has to be implemented in real-time application. Therefore, speed and detection time of the SVM-based algorithm embedded relay (SVMR) needs to be investigated. In this context, speed implies the processing speed of SVMR and it is expected to be fast, given the fact that the real-time extracted features are passed through the trained SVM containing a small number of SVs. Detection time of relay is defined as the time delay, which starts soon after the onset of islanding and finishes as soon as islanding is detected. In the following sub-sections, first, the performance of the proposed approach is evaluated considering the reliability, which comprises the DR, FA and the response time or detection time of islanding. Then, SVM-based machine learning approach is compared with conventional relays under all possible power imbalance scenarios that could be present during islanding. Finally, a general discussion is presented in Section 4.3. 4.1 Performance assessment of the proposed machine learning approach It should be noted that the test results of Table 2 were achieved with the features extracted from a fixed window width of eight-cycle, since eight-cycle data-window was selected as an optimum tradeoff window considering response time, DR and FA as illustrated in Fig. 5. In this investigation, the event (islanding/non-islanding) inception time was known a priori; therefore, the starting point of the window was considered from the event inception time and the end point of the window was eight cycles after the event onset. However, in practice, the event (islanding/non-islanding) has to be detected in real time without prior knowledge of inception time. Therefore, to examine the performance of SVM more critically, an eight-cycle long sliding window is considered with a step-size of one-cycle and classification results are recorded (see Fig. 6 for illustration). Note that in simulations, islanding and non-islanding events are mutually exclusive in an eight-cycle window; in other words, islanding and non-islanding events cannot occur simultaneously within an eight-cycle window. If a sliding window contains at least one cycle post-islanding, it is considered as an islanding event. Similarly, if the window contains at least one cycle after a non-islanding event, then it is considered as a non-islanding event. Taking this definition into consideration, classification results containing one-cycle to eight-cycle of post-islanding and non-islanding events are illustrated in Figs. 7a–d for four possible power imbalance scenarios. Moreover, in order to assess the classification performance in detail, F-measure is calculated and presented in Fig. 7. Note that F-measure is an indicator for the classifier's performance; it is given by F-measure = 2TP/(2TP + FP + FN), where TP is true positive, FP is false positive, FN is false negative and TN is true negative [18]. In this work, TP indicates the successful classification of islanding events, FP implies the misclassification of non-islanding events, TN indicates the successful classification of non-islanding events and FN specifies the misclassification of islanding events. Fig. 6Open in figure viewerPowerPoint Data stream for illustrating the concept of moving 8-cycle data-window and its classification at each step (one-cycle to eight-cycle) Fig. 7Open in figure viewerPowerPoint Classification results (F-measure, DR, FA) using eight-cycle moving data-window including data stream of one-cycle to eight-cycle post-non-islanding and islanding events under scenarios (a) Deficit ΔP and deficit of ΔQ, (b) Deficit of ΔP and excess of ΔQ, (c) Excess of ΔP and excess of ΔQ, (d) Excess of ΔP and deficit of ΔQ From the results of Figs. 7a–d, it is observed that when eight-cyle data-window includes five-cycle data of post-islanding and non-islanding events, then F-measure and its associated DR and FA show comparatively better performance under all four power imbalance scenarios. Hence, in order to summarise the classification results under all possible scenarios, F-measure, DR, FA, FN and TN are shown in Fig. 8. Fig. 8Open in figure viewerPowerPoint Classification results using eight-cycle data-window including data stream of five-cycle post-non-islanding and islanding events under scenarios (a) Deficit of ΔP and ΔQ, (b) Deficit of ΔP and excess of ΔQ, (c) Excess of ΔP and ΔQ, (d) Excess of ΔP and deficit of ΔQ From Fig. 8 it is evident that 100% DR with 0% FA is not achieved for all possible power imbalance scenarios. However, for practical application, if 100% DR and 0% FA is required, then a majority voting rule along with the SVM is proposed. The proposed rule guarantees 100% DR and 0% FA with a minor r