Indonesian J our nal of Electrical Engineering and Computer Science V ol. 36, No. 1, October 2024, pp. 115 126 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v36.i1.pp115-126 r 115 Enhanced fault identification in grid-connected micr ogrid with SVM-based contr ol algorithm Di vya Shoba Nair , Thankappan Nair Rajee v, Sindhura Miraj Department of Electrical Engineering, Colle ge of Engineering T ri v andrum, APJ Abdul Kalam T echnological Uni v ersity , K erala, India Article Inf o Article history: Recei v ed Feb 2, 2024 Re vised Mar 28, 2024 Accepted May 12, 2024 K eyw ords: D WT F ault detection Machine learning Microgrids Rene w able ener gy SVM ABSTRA CT The penetration of rene w able ener gy sources, electric v ehicles (EVs) and load dynamics, and netw ork comple xities often lead to nuisance tripping in grid- connected microgrids. T raditional protection methods f ail to discriminate f ault and other dynamic v olatilities in the system. The paper presents a no v el tw o- le v el adapti v e relay algorithm to a v oid nuisance tripping in a grid-connected microgrid under v arying grid dynamics. The no v elty of the adapti v e relay algorithm is that nuisance tripping is eliminated by precisely determining normal system-le v el dynamics at the first le v el using a phase de viation reference block. The first le v el determines the necessity for acti v ating the second le v el, which consists of a detection scheme combining a multiclass support v ector machine (SVM) and discrete w a v elet transform (D WT). The h ybrid D WT - SVM methodology ensures ef fecti v e f ault diagnosis, adapting to v ariations in ener gy sources, load fluctuations, and f ault scenarios. Real-time hardw are- in-the-loop (HIL) simulation v alidates the system’ s ef fecti v eness in dynamic microgrid en vironments. Extensi v e e xperiments on scenarios, including f aults, fluctuations in rene w able ener gy generation, and intermittent simulations of EV char ging and capacitor switching, were conducted to test the ef ficac y of the adapti v e relay algorithm. Finally , e xperiments using OP AL-R T HIL real- time simulator and the Raspberry Pi microcontroller v alidated the adapti v e relay algorithm in a grid-connected microgrid under v arying grid dynamics. This is an open access article under the CC BY -SA license . Corresponding A uthor: Di vya Shoba Nair Department of Electrical Engineering, Colle ge of Engineering T ri v andrum APJ Abdul Kalam T echnological Uni v ersity Thiruv ananthapuram, K erala, India Email: di vyaperoor@gmail.com 1. INTR ODUCTION The era of profound change for po wer systems has be gun with the e xtensi v e inte gration of rene w- able ener gy sources, electric v ehicle (EV) proliferation, and dynamic load changes. This inte gration led to the microgrid concept, an interc o nnec ted system of sources and loads with controllable attrib utes. Micro- grids e xhibit the fle xibility to operate in both grid-connected and isolated modes. Ho we v er , when operating in grid-connected modes, microgrids are vulnerable to f aults occurring on the grid side [1]. In such instances, immediate disconnection becomes imperati v e due to the limited capacity of lo w-rating distrib uted ener gy re- sources (DER) to withstand high f ault currents, particularly true for in v erter -based DERs [2], which ha v e a lo wer current-carrying capacity compared to con v entional synchronous-based units [3]. Furthermore, the f ault current le v el in an microgrid is highly reliant on the netw ork layout and changes dramatically across operation J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
116 r ISSN: 2502-4752 modes (grid-connected/islanded) [4], [5]. Hence, an adequate protection strate gy is essential for microgrids, considering the distincti v e characteristics of their sources, dynamic operating conditions, and t he need to ac- commodate rene w able ener gy sources [6], [7]. EVs are g aining popularity as fossil fuel prices k eep rising and en vironmental concerns about greenhouse g as emissions from the transporta tion sector increase. Ho we v er , there are more security and protection issues due to the e xtensi v e EV inte gration into the grid [8]. The traditional protection strate gies used in distrib ution systems become inef fecti v e due to b i direc- tional f ault currents in microgrids. Furthermore, the participation of DERs in f ault currents interferes with protection de vice trip times [9], which deteriorates the coordinat ion [10]. Immediately disconnect ing the DERs in case of a problem is a simple w ay to deal with these issues. This strate gy guards ag ainst problems lik e sympathetic tripping and blinding protection while maintaining the ef ficac y of le g ac y protection programmes [11]. Ho we v er , when there is a significant penetration of DERs, disconnection of DERs may cause a decrease in grid v oltage duri ng f ault conditions, resulting in instability [12]. De v eloping protection strate gies that con- sider t he impact of DERs and v ariations in grid t o pol ogy becomes i mperati v e to address the dif ficulties abo v e. In order to do this, a v arie ty of machine learning and computational intelligence techniques, including fuzzy systems, multi-agent systems, artificial neural netw orks (ANNs) [13], and metaheuristics, ha v e been proposed as microgrid protection mechanisms [14]. The inte gration of rene w able ener gy , coupled with v ariations in l oad patterns, poses significant chal- lenges for microgrid protection [15]. Centralized protection schemes used in con v entional po wer systems are deemed insuf ficient for dynamic microgrids [16]. The widespread adoption of rene w able ener gy alters protec- tion characteristics, requiring a rob ust f ault prediction scheme that considers dynamic v ariations in the system [17]. A real-time f ault detection and isolation protection s ystem is crucial to pre v ent cascading f ailures and sympathetic tripping. Numerous ongoing research ef forts are dedicated to de v eloping comprehensi v e protec- tion systems that address the unique challenges grid-connected microgrids f ace [18]. The critical component in these protection systems is the relay , b ut traditional relays with single-threshold and current-dependent func- tions are unsuitable for dynamic microgrids. Thus, there is a pressing need to de v elop ne w dynamic protection schemes [19]. V arious schemes ha v e been proposed in the literature to address microgrid challenges, including data mining-based dif ferential protection, time-frequenc y-based dif ferential schemes, f ault clearing methods for con v erter -dominant microgrids [20], and machine learning approaches [21]. Recent studies ha v e e xplored the application of con v olutional neural net w orks (CNN) [22], random forest, k-nearest neighbour algorithms, and ANN for f ault identification. Among these classifiers, support v ector machine (SVM)-based classifiers ha v e demonstrated superior performance [23]. V ijayachandran and Sheno y [24], a relay coordination scheme utilizing SVM for distrib ution systems with rene w able ener gy sources is introduced. Li et al. [25] discusses a learning approach that incorporates f ault detection and diagnosis, considering v ariations in irradiance. Aisw arya et al. [26] presents a unique adapti v e scheme based on SVM for precise f ault identification in microgrids. SVM yield se v eral benefits: reduced outlier impact, f aster prediction, higher accurac y , shorter e x ecution time, and a v oided o v er -fitting. The discrete w a v elet transform (D WT) is utilized for feature e xtraction to impro v e prediction speed and accurac y while reducing the amount of data the machine learning model must handle [27]. Numerous studies ha v e made a substantial contrib ution to diagnosing f aults in grid-connected micro- grids; ho we v er , considerable research g aps still open up possibilities for further in v estig ation. When it comes to selecting and representing data for analysis, there remains a crucial g ap. The dependability and quality of the data will influence ho w well the protection mechanism functions. More res earch on dif ferent f ault scenarios, load patterns, and fluctuations in rene w able ener gy sources should be conducted methodically to impro v e the fle xibility of the f ault diagnosis system. Unco v ering the comple x dynamics of DERs and dynamic demand changes is also necessary . In order to close these research g aps, it is recommended that D WT be used for feature e xtraction from the current and v oltage signals in f ault identification methodologies. Extracting signifi- cant features from the input signal-based approaches simplifies the SVM classification process and reduces the data training requirements. The approach precisely dif ferentiates the f ault cases from system-le v el dynamics. The approach will reduce the data handled by t he machine learning model and impro v e prediction speed and accurac y . This paper proposes a tw o-le v el adapti v e relay algorithm to a v oid nuisance tripping in a grid-connected microgrid under v arying grid dynamics. The relay uses multiclass SVM in conjunction with D WT feature e xtraction for f ault-type detection. Real-time e xperiments using the prototype relay significantly impro v ed Indonesian J Elec Eng & Comp Sci, V ol. 36, No. 1, October 2024: 115–126 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 117 microgrid resilience and o v erall performance. A grid- conn e cted microgrid f ault prediction and identification method based on sequential multiclass machine learning is de v eloped. The relay w as trained using a h ybrid D WT and SVM-based approach, which pro vided quick and precise training information for f ault diagnosis. The major contrib utions of this paper are outlined belo w: The no v el tw o-le v el adapti v e f ault identi fication scheme is tailored for grid-connected microgrids, address- ing challenges lik e bidirectional f ault currents, dynamic load v ariations, and DER participation. The pro- posed approach combines the D WT and multiclass SVM to enhance f ault diagnosis ef ficienc y . An adapti v e relay system based on D WT and multiclass SVM ensures quick, accurate, and reliable protec- tion, enhancing microgrid resilience under di v erse operating conditions. A k e y de v elopment is the h ybrid D WT and SVM methodology , which impro v es f ault diagnosis speed and accurac y e v en with limited training data. Real-time HIL simulations v alidate the proposed scheme, utilizing a Raspberry Pi microcontroller inter - f aced with the OP AL-R T si mulator . This e xperimental v alidation demonstrates the practical implementation and real-time prediction capabilities, ensuring reliability and precision in dynamic microgrid scenarios. This paper is structured as follo ws: section 2 discusses the de v elopment of a f ault identification scheme with multiclass SVM and D WT . Section 3 describes the analysis and outcomes of the simulation. The e xperi- mental v alidation of the approach utilising real-time HIL simulation is co v ered in section 4. 2. DEVELOPMENT OF A F A UL T DETECTION ALGORITHM UTILIZING MUL TI CLASS SVM AND D WT TECHNIQ UES The microgrid promotes rene w able ener gy with DERs b ut f aces challenges from source v ariability and load changes lik e EV char ging, causing v oltage instability and po wer quality issues. Existing protecti v e relays struggle, risking islanding and safety hazards. The proposed adapti v e relaying method incorporates a tw o-tier adapti v e relay algorithm to pre v ent nuisance tripping. The first le v el is a phase de viation reference block, which uses a PLL-based control strate gy to continuously monitor the phase angle between the main grid and the microgrid. If the main grid and microgrid lose synchronization, the phase de viation reference block sends a initiating signal to the second le v el, which is a h ybrid detection scheme that combines SVM and D WT techniques. This h ybrid detection scheme is responsible for detecting an y anomalies in the system. Therefore, the phase de viation reference block determines whether the h ybrid detection scheme needs to be acti v ated, allo wing the system to a v oid unnecessary disconnections and unintentional islanding while maintaining oper - ational speed. Figure 1 depicts a tw o le v el h ybrid SVM and D WT method to detect system issues ef ficiently . W ith minimal data, it identifies f aults, source dynamics, and load fluctuations reliably . Multiclass SVM distin- guishes f aults using v oltage and current measurements processed through D WT .D WT -based feature e xtraction captures fine signal details, enabling f ast and accurate SVM decisions with minimal data. Signal under goes tw o-le v el decomposition via lo w-pass and high-pass filters. SVM classifier is trained on four essential features per scenario. 2.1. Discr ete wa v elet transf orm The process of D WT in v olv es di viding a signal into w a v elets that are localized in both time and fre- quenc y domains. This di vision is achie v ed through a combination of filtering and do wnsampling techniques. T o establish a hierarchical representation, the signal is systematically decomposed into high-frequenc y detail coef ficients and lo w-frequenc y approximation coef ficients. Features that capture v arious aspects of the signal are then e xtracted from these coef ficients at dif ferent le v els. The Daubechies w a v elet, specifically the db4 w a v elet, is selected a s the mother w a v elet for its adv antageous properties suc h as good time-frequenc y local- ization, similarity to signals observ ed during f ault conditions, and its ability to ef fecti v ely capture both lo w and high-frequenc y components of the signal. In this c o nt e x t , a tw o-le v el decomposition in D WT is chosen for feature e xt raction. This decision is made because the initial tw o le v els typically contain the most critical information, leading to reduced noise and computational comple xit y . By e xamining both the approximation and detail coef ficients obtained through D WT , rel e v ant features are e xtracted to f acilitate f ault detection within the signal. This methodology enables the identification of k e y characteristics that can indicate the presence of f aults or anomalies, thereby enhancing the ef fecti v eness of f ault detection and diagnosis processes. The features selected for the D WT analysis are the mean of detailed v oltage coef ficients, the mean of approximate Enhanced fault identification in grid-connected micr o grid with ... (Divya Shoba Nair) Evaluation Warning : The document was created with Spire.PDF for Python.
118 r ISSN: 2502-4752 v oltage coef ficients, the mean of detailed current coef ficients, and the L1 norm. V appr ox = P n i =1 V appr ox i n (1) V det = P n i =1 V det i n (2) I det = n X i =1 I det i (3) k V k det 1 = n X i =1 V det i (4) F or e v ery microgrid e v ent in this w ork, four features are g athered from thirteen b uses, for a total of fifty-tw o D WT features. The total number of data points collected is calculated as 52 (selected D WT features) multiplied by the sum of 556 data points from 245 f ault cases and 311 normal cases, resulting in a total of 28,912 data points. The g athered data under go preprocessing and are classified with suitable labels to train the supervised learning model. F or this training, 80% of the data is utilized, while the remaining 20% is allocated for e v aluating the model’ s ef fecti v eness. Figure 1. F ault identification scheme with D WT and SVM 2.2. Multiclass SVM classifier SVM present a practical option for f ault diagnosis in comple x data en vironments where t raditional machine learning and deep learning methods may f all short. Their sound theoretical base mak es it possible to manage high-dimensional data ef fecti v ely , which is a common feature of f ault identification tasks in v olving man y measurements and indicators. SVM’ s resilience to o v erfitting is one of its main adv antages in f ault de- tection, helping to a v oid costly errors. This resilience results from its statistical learning principles, which are particularly helpful when the number of dimensions e xceeds the num ber of samples. The k ernel trick enables SVM to na vig ate and classify a wide range of comple x and v ariable data structures by con v erting them into spaces where the y become separabl e. SVM pro vides transparenc y in its decision-making process, dif ferenti- ating it f rom the frequently opaque deep learning models. The process is a crucial aspect of a f ault diagnosis, where it is crucial to comprehend the reasoning behind predictions. SVM may also of fer computational ef fi- cienc y for smaller to medium-sized datasets, of fering a f aster solution without compromising performance in comparison to its deep learning peers. Furthermore, SVM’ s skill in managing imbalanced datasets, which are Indonesian J Elec Eng & Comp Sci, V ol. 36, No. 1, October 2024: 115–126 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 119 pre v alent in f ault identification, ensures its sensiti vity to minority classes, such as f aults, due to its thoughtful k ernel and param eter selection. As a result, SVM are a standout option for problem diagnosis, supported by their capacity to handle com p l icated and high-dimensional data as well as their open and theoretically sound decision-making process. The multiclass SVM classifier is practical for tasks with multiple classes. Originally for binary clas - sification, SVM has been adapted for multiclass tasks. SVMs ef ficiently cate gorize occurrences into multiple classes, utilizing support v ectors near the decision boundary . F or multiclass SVMs, there are tw o approaches: one-vs-one (OvO) and one-vs-rest (OvR). F or each pair of classes, OvO trains a binary classifier , producing K : ( K 1) = 2 classifiers for K classes. OvR produces K classifiers by training a binary classifier for each class ag ainst a ll others. Comput ationally , OvR usually performs better , especially when dealing with se v eral classes.The radial basis function (RBF) k ernel adopted demonstrates e xceptional performance when dealing with o v erlapping data. Specifically , the most influential f actors in classifying ne w observ ations are the closest data points, while those situated at a greater distance ha v e minimal impact on the classification process. W ith a set of training samples ( x 1 ; y 1 ) ; ( x 2 ; y 2 ) : : : :: ( x m ; y m ) , where y i represents the associated class label and x i represents the feature set of the i th sample, the decision-making function f ( x ) for a binary SVM with labels +1 and -1 is as follo ws: f ( x ) = sig n   n X i =1 w i :x i + b ! (5) in this case, the maximum number of features is n , the bias term is b , the input features are x i , and the weight parameters are w i . The decision function for class k in a multiclass SVM is represented by f k ( X ) , where K is the total number of classes.An SVM that is binary and trained t o dif ferentiate class k from the others is represented by each f k ( X ) . The predicted class y for a data point x is the one with the maximum decision function v alue: ^ y = ar g max k   n X i =1 w k i :x i + b k ! (6) Figure 2 sho ws the multiclass SVM training and prediction flo wchart using a O VR strate gy . F ault states are classified, whi le normal conditions and dynamic fluctuati ons are grouped together . This aids in distinguishing when relay operation is required. It also enables f ault detection and identification. Figure 2. Multiclass SVM training and prediction Enhanced fault identification in grid-connected micr o grid with ... (Divya Shoba Nair) Evaluation Warning : The document was created with Spire.PDF for Python.
120 r ISSN: 2502-4752 3. IMPLEMENT A TION OF F A UL T DIA GNOSIS METHODOLOGY USING MUL TICLASS SVM AND D WT 3.1. Modelling of system under study A radial microgrid with 13 b uses, operating at 13.8 kV and 50 Hz, is simulated using the MA T - LAB/Simulink platform. Three distrib uted generation (DG) sources are part of the microgrid configuration: tw o diesel generators and photo v oltaic (PV) units at b uses 1, 2, and 10. Furthermore, b uses 1, 2, 7, and 8 ha v e a dynamic EV char ging load. T able 1 contains the microgrid’ s comprehensi v e specifications. T able 1. Microgrid parameters Component P arameter Three phase grid supply V oltage = 13.8 kV (transformer 69/13.8 kV ) Photo v oltaic system 70 kW each at b us 1,2 and 10 Generator (diesel) 500 MV A ,440 V , at b us 8,13 EV station 22 kW each at Bus 1,2,7,8 Load 20 kW load at 1, 2, 7, 8, 9, 11, 12, 13 10 kW load at b us 7 and 10 3.2. Design of pr ediction scheme An adapti v e protection scheme is de v eloped for f ault identification in a grid-connected microgrid, particularly when inte grating dynamic sources lik e PV and loads lik e EV . As the system beha vior becomes dy- namic, a reliable f ault detection technique is crucial to minimize ener gy loss and monitoring costs in a gro wing microgrid. This section introduces a f ault detection system based on D WT and multiclass SVM . The flo wchart depicting of fline training and online prediction processes is sho wn in Figure 3. Figure 3. Multiclass SVM and D WT scheme A Simulink model of a 13-b us system is constructed for this objecti v e. Sensing de vices g ather the v oltage and current signals, which are then supplied to the phase de viation reference block-basically a PLL. The D WT and SVM module will recei v e an i nitiating signal when the reference v alue abo v e a predetermined limit. In this scenario, a pretrained SVM classifier recei v es the data. V oltage and current readings from e v ery Indonesian J Elec Eng & Comp Sci, V ol. 36, No. 1, October 2024: 115–126 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 121 b us under a v ariety of conditions, such as f aults, load v ariations, and re gular grid operations, are included in the training dataset. These measurements are used to e xtract features using the D WT , whi ch is the n input into SVM models. In the course of testing, b us dat a is sent in real-time communication to the transformation block, which uses it to e xtract features for the SVM models that are e x ecuted in microcontrollers. These models forecast microgrid problems and sho w them to operators through a graphical user interf ace (GUI). This or g anized process ensures systematic defect detection in dynamic micr og r id settings, aiding quick and ef fecti v e decision-making. 3.2.1. Data collection and classifier modelling The IEEE 13-b us system is modified and modeled in MA TLAB to simulate the microgrid under study . The D WT is used to e xtract features from the three-phase v oltages and currents from each of the thirteen b uses. W ith Google Colab’ s assistance, SVM models are b uilt and trained using Python’ s sci-kit learn module. The sv c function is utilized for training. Each microgrid e v ent g athers four features from each of the 13 b uses, resulting in a total of 52 D WT features. This comprehensi v e dataset comprises 28,912 data points co v ering v arious f ault and normal scenarios, enabl ing a thorough e xamination of microgrid beha vior . T able 2 displays the obtained D WT characteristics for the b us at the point of common coupling, along with the decision made by the SVM classifier , feature v alues, and corresponding label for each scenario. T able 2. Extracted D WT -based features Sl no Ev ent Feature 1 Feature 2 Feature 3 Feature 4 Label Decision 1 LG f ault at 0.2 sec 19,854 6.4732 -0.0086 647.32 LG f ault -1 2 LLG f ault at 0.2 sec 11,928 19.28 -0.0183 1,928 LLG f ault -1 3 LLLG f ault at 0.2 sec 150.18 56.904 -0.1286 5,690.4 LLLG f ault -1 4 Normal 20,701 13.587 0.0014 1,358.7 Normal 1 5 EV1 and EV2 char ging at 0.2 sec 20,556 13.575 -0.0115 1,357.5 Normal 1 6 T ransformer ener gisation at 0.2 sec 20,695 13.593 0.0017 1,359.3 Normal 1 7 Capacitor switching at 0.2 sec 20,855 -29.325 0.2759 -2932.5 Normal 1 8 Irradiation v ariation in PV unit 1 ( 1,000 W/m 2 to 100 W/m 2 ) 20,700 13.588 0.0014 1358.8 Normal 1 9 Load v ariation of 10% at 0.2 sec 20,688 13.573 -0.0008 1,357.3 Normal 1 10 Load v ariation of 20% at 0.2 sec 20,629 13.327 0.0066 1,332.7 Normal 1 4. RESUL TS AND DISCUSSION 4.1. Simulation r esults The microgrid model is tested under v arious scenarios, including f aults, fluctuations in rene w able ener gy generation, and intermittent simulations of EV char ging and capacitor switching. These simulations aim to enhance the machine learning model’ s understanding of microgrid dynami cs. F ault scenari os such as three-phase to ground (LLLG), double line to ground (ABG, BCG, A CG), and single line to ground f aults in three-phase lines (A G, BG, CG) are incorporated into both islanded and grid-connected modes. Belo w are e xamples of simulated test scenarios used for data collection. F ault in line: single line to ground (LG), double line, and triple line f aults are simulated on Line 1 in both grid-connected and islanded modes of the microgrid. F aults start at 0.4 seconds and end by 0.8 seconds, causing a tenfold increase in f ault c urrent. V oltage and current w a v eforms during the LLLG f ault are depicted in Figure 4, reflecting changes in output, which will serv e as features for SVM analysis. PV irradiation v ariations and PV outages: PV units are disconnected sequentially at 0.2 seconds, with v oltage and current measurements recorded. Irradiati on decreases abruptly from 1,000W/m 2 to 100W/m 2 for 0.2 seconds during each disconnection. Significant po wer and current v ariations are observ ed at b us 1. Figure 5 sho ws v oltage and current w a v eforms during PV irradiation fluctuation, reflecting changes in output. These v ariations serv e as features for SVM analysis. Dynamic v ariations in system: dynamic simulations induce v ariations such as load fluctuations, f ast EV char ging, and transformer/capacitor switching. Data is collected, wi th loads arbitrarily altered by +/- 10% to simulate fluctuations. F ast-char ging EVs are acti v ated at b uses 1, 2, 7, and 8 for 0.2 seconds, noting re- sulting changes. Adding EV load causes v oltage drop and increased current. T ransformer ener gization and capacitor switching simulations illustrate system dynamics further . Switching e v ents occur at 0.2 seconds, Enhanced fault identification in grid-connected micr o grid with ... (Divya Shoba Nair) Evaluation Warning : The document was created with Spire.PDF for Python.
122 r ISSN: 2502-4752 observing v oltage and current fluctuations. Relay operation is unnecessary as it produces transient changes considered a natural part of operation. Figure 4. V oltage, current and po wer output at b us 1 during LLLG f ault Figure 5. V oltage and current output at b us 1 during irradiation v ariation at PV1 Simulations conducted for v arious e v ents present prediction outcomes in T able 3. The multiclass SVM D WT model’ s performance summ ary and e v aluation metrics is illustrated in Figure 6. W ithin the confusion matrices illustrated in Figure 6(a), the diagonal entries represent the count of accurately predicted observ ations, kno wn as true positi v es (TP), while the of f-diagonal entries indicate incorrect predictions, referred to as f alse positi v es (FP). The comparison report sho wing the accurac y , precision and performance matrix components are sho wn in Figure 6(b). The computation time tak en by classifier for both testing and training is illustrated in Figure 6(c). The proposed multiclass SVM with D WT approach outperforms pre v i ous SVM models based on RMS v oltage and current v alues [26], of fering higher accurac y as sho wn in T able 4, which denotes e xcellent predic- tion alignment. Moreo v er , a comparison betwe en the SVM classifier and the Decision T ree (DT) classifier across v arious e v ents as sho wn in T able 5 demonstrates SVM’ s superior performance in predicting microgrid parameter v ariations. T able 3. Predictions by the de v eloped SVM based scheme Ev ent simulated Specification Label Predicted e v ent No of cases simulated Prediction accurac y LG f ault F ault at line 1 and 5 LG f ault LG 85 100% LLG f ault LLG f ault LLG 75 99.50% LLLG f aults LLLG F ault LLLG 85 99.10% Normal operating condition Normal Normal 101 99.50% PV irradiation V ariation and outages At b us 1, 2 and 10 Normal Normal 35 99.30% Load v ariations 10% and 20% at b us 1, 2, 7, 9 Normal Normal 70 100% EV load additions At b us 1, 2, 7, 8 Normal Normal 60 99.40% T ransformer ener gisation At b us 3 Normal Normal 25 100% Capacitor switching At b us 3 Normal Normal 20 100% Indonesian J Elec Eng & Comp Sci, V ol. 36, No. 1, October 2024: 115–126 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 123 (a) (b) (c) Figure 6. Ev aluation metrics and performance o v ervie w of the multi-class SVM D WT model: (a) confusion matrix displaying classification accurac y and errors; (b) comparison report with precision, recall, and F1-score metrics for each class; and (c) compilation time sho wing the model’ s training ef ficienc y T able 4. Comparison with prior w ork Ev ent No of cases tak en SVM with Vrms and Irms SVM and D WT LG f ault 85 98.70% 100% LLG f ault 75 98.20% 99.50% LLLG f ault 85 98.20% 99.30% Normal 311 98.50% 99.10% T able 5. Comparisons SVM vs DT ML model Accurac y SVM 99.1 DT 97 5. EXPERIMENT V ALID A TION OF D WT B ASED SCHEME USING REAL-TIME HARD W ARE IN LOOP SIMULA TION The microcontroller and OP AL-R T w ork together to simulate in real time. The simulated 13 b us microgrid model comprises tw o main subsystems: SM Master and SC Console. All computational components are inte grated into the SM Master , which is inserted into the OP AL-R T simulator OP 4510. Con v ersely , the SC Console operates within the host system, f acilitating user interaction and inte grating Simulink blocks for data collection and display . Enhanced fault identification in grid-connected micr o grid with ... (Divya Shoba Nair) Evaluation Warning : The document was created with Spire.PDF for Python.
124 r ISSN: 2502-4752 F or real-time operation, the micro grid model in the SM Master subsystem is con v ert ed to C and put into the OP 4510 OP AL-R T real-time simulator . The Raspberry Pi model 4 b, kno wn for its lar ge RAM and f ast onboard processor , is the optimal choice for machine learning applications, pro viding ef ficient tools for running machine learning programs. T rained machine learning models, sa v ed as Python Pickle files (.pkl) from Google Colab, are included in the Python pre diction code for seamless inte gration. This method enables the easy deplo yment of trained models in the Python en vironment for accurate and ef ficient predictions, ultimately installed onto the Raspberry Pi for real-time use. In real-time simulation outputs are for f ault cases are depicted in Figure 7, the v oltage and current w a v eforms under single line to ground f ault is sho wn in Figure 7(a) and double line to ground f ault case is sho wn in Figure 7(b). The f aults are created while system is running in real time mode with OP AL R T unit and corresponding v ariati ons are observ ed in the w a v eforms. The occurence of LLG f ault in line 1 is predicted by the in around 11.79 ms and LG f ault in line 1 is predicted in around 7.26 ms. The corresponding e xperimental set up, GUI display , 13 b us v oltage and current measurements are displayed in the host PC. (a) (b) Figure 7. Real time v oltage and current output at b us 1 from PCC for (a) LG f ault and (b) LLG f ault The real-time simulator transmits thirteen b us measurements as a 1D array to the Raspberry Pi using the UDP/IP protocol. Tkinter Python library constructs a user -friendly GUI on the Raspberry Pi, displaying SVM model predictions c learly . The microgrid model is loaded into the OP AL-R T OP4510 simulator for HIL simulation. The Raspberry Pi’ s GUI and control are displayed on the PC for the R T Lab . The Raspberry Pi controller e x ecutes the trained SVM models as part of the system implementation. OP AL-R T and the Raspberry Pi microcontroller communicate via ethernet and UDP/IP . Figure 8 sho ws the HIL simulation configuration for real-time prediction. Figure 8(a) illustrates the HIL e xperimen- tal setup, whereas Figure 8(b) sho ws the output at the GUI interf ace. OP AL-R T’ s analog output range is -16 V to +16 V , so the DSO displays scaled v oltage and current outputs. Real-time output serv es as test data, with classifier output displayed on the Raspberry Pi GUI. (a) (b) Figure 8. Hardw are in loop simulation of system (a) e xperimental setup and (b) prediction output at user interf ace of Raspberry Pi Indonesian J Elec Eng & Comp Sci, V ol. 36, No. 1, October 2024: 115–126 Evaluation Warning : The document was created with Spire.PDF for Python.