Adaptive Fault Detection Based on Neural Networks and Multiple Sampling Points for Distribution Networks and Microgrids

J. Marín-Quintero; C. Orozco-Henao; A. S. Bretas; J. C. Velez; A. Herrada; A. Barranco-Carlos; W. S. Percybrooks

网刊加载中。。。

使用Chrome浏览器效果最佳，继续浏览，你可能不会看到最佳的展示效果，

确定继续浏览么?

复制成功，请在其他浏览器进行阅读

Adaptive Fault Detection Based on Neural Networks and Multiple Sampling Points for Distribution Networks and Microgrids PDF

- ORCID：
J. Marín-Quintero
✉
- ORCID：
C. Orozco-Henao
✉
- ORCID：
A. S. Bretas
✉
- ORCID：
J. C. Velez
✉
- ORCID：
A. Herrada
✉
- ORCID：
A. Barranco-Carlos
✉
- ORCID：
W. S. Percybrooks
✉

the Electric and Electronic Engineering Program, Universidad Tecnológica de Bolívar, Cartagena, Colombia； the Electrical and Electronic Engineering Department, Universidad del Norte, Barranquilla, Colombia； the Distributed Systems Group, Pacific Northwest National Laboratory, Gainesville, USA

Updated：2022-11-20

DOI：10.35833/MPCE.2021.000444

OUTLINE

Abstract

Smart networks such as active distribution network (ADN) and microgrid (MG) play an important role in power system operation. The design and implementation of appropriate protection systems for MG and ADN must be addressed, which imposes new technical challenges. This paper presents the implementation and validation aspects of an adaptive fault detection strategy based on neural networks (NNs) and multiple sampling points for ADN and MG. The solution is implemented on an edge device. NNs are used to derive a data-driven model that uses only local measurements to detect fault states of the network without the need for communication infrastructure. Multiple sampling points are used to derive a data-driven model, which allows the generalization considering the implementation in physical systems. The adaptive fault detector model is implemented on a Jetson Nano system, which is a single-board computer (SBC) with a small graphic processing unit (GPU) intended to run machine learning loads at the edge. The proposed method is tested in a physical, real-life, low-voltage network located at Universidad del Norte, Colombia. This testing network is based on the IEEE 13-node test feeder scaled down to 220 V. The validation in a simulation environment shows the accuracy and dependability above 99.6%, while the real-time tests show the accuracy and dependability of 95.5% and 100%, respectively. Without hard-to-derive parameters, the easy-to-implement embedded model highlights the potential for real-life applications.

Keywords

Distribution network; microgrid; adaptive protection; intelligent electronic device; machine learning

I. Introduction

TECHNOLOGICAL advances in non-conventional energy sources have enabled high penetration of distributed energy resources (DERs) in modern power systems. In recent years, concepts such as active distribution network (ADN) and microgrid (MG) have been introduced to address the technological challenges imposed by such developments [

1].

ADNs are networks with a high penetration of distributed generation (DG) and capabilities of automation and control [

2]. Besides, MGs combine DERs, energy storage systems (ESSs), and flexible loads. Additionally, they can operate in grid-connected or islanded mode [3].

Both MGs and ADNs facilitate operational goals, as to reduce power losses and improve voltage profile [

4]. Nevertheless, MG and ADN operations present new technical challenges associated with system stability and protection [5], [6]. Considering system protection, some of the technical challenges are associated with bidirectional power flow, short-circuit level variation, false tripping, blinding protection, and network topology changes [7]. Recently, solutions have been presented in the specialized literature addressing these challenges [8]. Adaptive protection-based approaches, which typically require robust communication links, have been presented as a viable solution. Most of the presented system protection solutions for MGs and ADNs consider a centralized communication architecture. With such architecture, a central controller makes operational decisions considering dynamic system conditions [9]. The solutions to decentralized adaptive system protection work in an autonomous mode, using only the available local observation [10]-[12].

The solutions to adaptive system protection can be classified according to their functionality. Two approaches are highlighted: fault detector and fault coordination. The first one considers an event-based system, where the operation is initialized when a state of fault is detected by an intelligent electronic device (IED). The second one, otherwise, detects a state of fault while supporting neighboring devices through backup protection [

13].

Recently, the solutions to adaptive system protection based on artificial intelligence (AI) models have been presented. These solutions show outstanding performance in simulated environments [

11], [12]. It should be noted that these methods use available local observations, e.g., current and voltage signals, to train and validate AI models [14], [15]. For example, [14] and [15] use two 60 Hz cycles of the voltage and current signals for decision making. The solutions are not validated in a physical system, thus no aspects regarding the implementation or physical validation are reported.

Among the solutions of adaptive system protection based on AI, [

11], [16], and [17] are highlighted. In these research works, learning-based techniques are applied to reduce the dependency on robust communication systems. Reference [11] presents an intelligent fault detection scheme that provides fast fault type, phase, and location information for MG protection. The solution uses the discrete wavelet transform and deep neural networks (NNs) to develop the fault detector. However, the method does not consider some operating scenarios like cut-off generation. Also, the proposed method is not validated in a physical system. In [16], a Chu-Beasley meta-heuristic-based solution is implemented to optimize feature selection for a machine learning (ML) model. Still, the solution is only validated in a simulated environment. A data-mining-based intelligent differential protection scheme for MGs is presented in [17]. However, it considers the communication between the devices at feeder-ends, and the scheme is not validated as well in a physical system. In [18], a voltage-restrained classifier-based relaying scheme is proposed to detect the faults and their respective zone. The scheme uses a decision tree that is ensembled to develop the protection scheme, but it considers a few operating scenarios. Reference [19] proposes an intelligent protection scheme for MGs using a combination of wavelet transforms and decision trees. The solution addresses the protection problem through two steps: fault detection and fault classification. Reference [20] uses the differential energy calculated for both three-phase current and extreme gradient boost to detect and classify a fault in an MG. Although both solutions are validated in a real-time simulator, they consider the no-fault states that do not allow a reliable validation of the ML model. Most of the proposed adaptive protection methods based on AI do not validate their models in physical systems, which could cause some bias that affects the interpretation of the results [21]-[23].

This paper presents the implementation and validation aspects of a fault detection solution of adaptive system protection based on AI models embedded on a physical-edge device. The solution uses only local observations, voltage, and current signals on the physical-edge device to detect network faults. The main contributions of this work towards the state-of-the-art are as follows.

1) Formulation of an autonomous adaptive fault detection solution based on AI models and multiple sampling points.

2) Real-life implementation and validation aspects of adaptive fault detection formulation on an edge device.

The remaining of this paper is organized as follows. Section II presents the main technical challenges of system protection of ADN and MG. Section III describes the adaptive fault detector embedding based network model. Section IV presents the experimental setup for the case study. Section V presents the test results and the discussion. Finally, Section VI presents the conclusion.

II. Main Technical Challenges of System Protection of ADN and MG

A. Formulation Challenges

The integration of high penetration of DER with control and automation functionalities in ADNs and MGs makes the system protection more complex. DER connection changes the typical operational characteristics of conventional distribution networks, and classical protection strategies are no longer efficient and reliable. As a result, new protection strategies have been proposed recently [

24].

High penetration of DER changes one of the main characteristics of conventional distribution networks: the presence of a unidirectional power flow. With DER integration, the power flow is bidirectional. The classical protection device used in distribution networks does not consider bidirectional flow, which can affect the selectivity of the system [

25]. Other aspects associated with DERs that can affect the performance of protection strategies are the blinding protection and the false tripping. The blinding protection occurs when the fault current observed by a relay changes due to the connection of a DER between the fault point and the relay itself [26]. In this scenario, the relay will not see the DER current contribution to the short circuit, causing coordination problems with neighboring relays. Further, false tripping occurs when a relay, connected to a feeder supplied by a DER, activates before the primary protection relay of an adjacent line in fault [27].

Additionally, high-power injection variance in the primary renewable source of the DER may impose DER connection and disconnection at some intervals during the day. All these changes in the operation of the ADNs and MGs cause significant variations in short-circuit levels, which are traditionally used to determine the settings of the protection devices. Therefore, selecting a setting that considers these variations is a complex task, and some designs can be infeasible with classical protection strategies [

28], [29].

B. Implementation Challenges

The main implementation challenges to the protection strategies are related to the method used. When the fault detection is based on ML models, there are relevant challenges to consider such as decision speed and accuracy [

30].

Decision speed relates to the hardware requirements to run the ML models at acceptable response time levels for the protection task. This includes the processing time and estimation of the attributes according to the inputs required by the ML models. Other relevant aspects are the sampling rate, processing capacity, and speed to process the information.

True accuracy indicates the performance of ML models in real-time application. For models obtained from synthetic databases such as the fault detection techniques proposed in [

31], the adequate modeling and simulation of the system under study is one of the main aspects that determine the performance of the ML model. If the simulated system represents the behavior of the real physical system, the performance of the model will be closer to the value obtained in a simulated environment. Other relevant aspects include the number of scenarios considered in the training and how well those scenarios cover the number of operating states in a real physical system [32].

III. Adaptive Fault Detector Embedding Based Network Model

The adaptive fault detector is based on an AI model and can operate at both terminals of a three-phase distribution line. The method uses an IED that measures the voltage and current signal to differentiate between a fault state and a no-fault state in the ADN or MG.

The adaptive fault detector contains six data processing steps grouped into two categories. The first category includes offline data processing steps, while the second category includes the online steps. The offline data processing steps have the goal of training and validating the adaptive fault detector model. The objective of the online steps is to estimate the faults in real time. The data processing steps are explained in detail in the following subsections.

A. Step 1: Database Simulation

The main bottleneck to obtain a suitable database from a real ADN or MG is gathering fault-condition data. This can involve either: continuous monitoring towards recording infrequently occurring faults or introducing faults to speed up data collection, both of which can be time-consuming or unfeasible. A synthetic (simulated) database, if well modeled, can be a suitable solution to train the adaptive fault detector model. The last one is the method considered in this paper. Hence, an electromagnetic transient (EMT) software is used to generate synthetic faults under no-fault states. The pseudo-code presented in Algorithm 1 shows the process used to obtain the synthetic database.

Algorithm 1 : database simulation
Input information of topologies of ADN, list of operating states, and thelocation of IED
Output fault and no-fault databases
Begin
1:	Set the time of simulation and time step
2:	Set the time of fault
3:	For to each condition i do
4:	Modify the ADN to the i^th operating state
5:	Run an optimal power flow
6:	Obtain the power delivered by each DG
7:	Set the DG power
8:	If the i^th operating state is in fault state
9:	Set the information of the fault
10:	Else if the i^th operating state is not in fault state
11:	Choose no-fault operating state
12:	Set the information of the new operating state
13:	End if
14:	Run an EMT simulation
15:	Save v(t) and i(t) signals for the i^th combination in a CSV
16:	End for
End

B. Step 2: Adaptive Fault Detection Method and Training of NN Model

The goal of the proposed method is to protect the three-phase section connected to the IED. Also, it protects the three-phase section lines adjacent to itself. Figure 1 presents an ADN with five three-phase section lines, six nodes (N1-N6), three loads (Load1-Load3), one DG, and eight IEDs (IED1-IED8).

Fig. 1 Protection zone for a particular IED in a distribution network.

If we focus on IED1, located between N1 and N2, it is possible to observe that this device will protect the lines between the following nodes: N1-N2, N2-N3, and N2-N6. IED1 protects the area delimited by the dotted red line. The faults located outside the protection zone will be considered as no-fault states for the model running on IED1. The stages associated with Step 2 are explained in detail as follows.

1)　Stage 1: Feature Selection

The main criteria to select the features used in the classifier are calculation simplicity and prediction accuracy. Voltage and current signals measured at the location of the IED are considered. Table I lists the features used by the classifier.

Table I Operating Scenarios for Adaptive Fault Detection Training and Validation

Item	Feature	Description of feature
1-3	$\|V_{a b c}\|$	Magnitude of voltage per phase
4-6	$θ V_{a b c}$	Angle of voltage per phase
7-9	$\|I_{a b c}\|$	Magnitude of current per phase
10-12	$θ I_{a b c}$	Angle of current per phase
13-15	$S_{a b c}$	Apparent power per phase
16-18	$Z_{a b c}$	Impedance per phase

Here, an NN model is used to train the adaptive fault detector. The NN is fitted as a binary classifier that chooses between a fault state or a no-fault state. This procedure will be executed until the predefined maximum number of iterations is reached.

Thus, the number of features presented in Table I could change according to the combination of features that give the best prediction accuracy.

2)　Stage 2: Signal Processing

Once the features have been selected, they must be processed. In the proposed method, this is done by the fast Fourier transform (FFT) applied to one cycle of the signal. For the synthetic database obtained from simulation, we can determine a priori the fault inception instant, where the operating states change from no-fault state to fault state. However, in a physical system, under real operating states, this instant is unpredictable. For this reason, the proposed method includes a new metric that introduces the uncertainties in current and voltage signals into the synthetic database. This metric is named multiple sampling. Figure 2 presents how multiple sampling works for a current signal with a fault inserted at time F_p.

Fig. 2 Analysis of multiple sampling for a fault operating state.

Figure 2 contains three 60 Hz cycles of analysis. The blue window shows the reference cycle where the fault point is in the middle of the cycle. Generally, this is the scenario used for the training of the classifier in a simulated environment. The green window shows the second analyzed cycle, where the window has a delay $Δ S_{1}$ of three samples concerning the reference window. Similarly, the red window has an advance $Δ S_{2}$ of three samples concerning the reference window. Using these three windows during the training, multiple sampling $Δ S$ can simulate an unknown fault inception instant. This helps to obtain a database that better represents the real behavior of a physical system.

3)　Stage 3: Database Processing

The fault detection model must use the database obtained in stage 2. However, the data needs to be preprocessed, which consists of two operations: standardization and randomization. These operations are executed according to the description in [

14].

Mean vectors and standard deviation matrices are also computed. In this paper, these values are used to perform the standardization of input data during both training and inference.

4)　Stage 4: Database Splitting

In this stage, 80% of the available database is used to train the fault detector model while the remaining 20% of the available database is put aside for validation [

33].

5)　Stage 5: Training of NN Model

The main goal of this stage is to obtain the adaptive fault detector model for the chosen IED. Here, the proposed method uses a multilayer NN model to detect fault states in the ADN or MG, as presented in [

15]. The NN model is formulated according to (1).

\begin{array}{l} y_{k} (x_{k}) = σ (\sum_{j = 1}^{m} w_{j} h (\sum_{i = 1}^{d} w_{j i} x_{i k}) + W_{0}) \\ x_{k} \in R^{d}, y_{k} \in {1,2, \dots, l} \end{array}

(1)

where $w_{j}$ and $w_{j i}$ are the weight parameters; $m$ is the number of neurons in the hidden layer; and $σ$ and $h$ are the activation functions. The most commonly used activation functions are logistic sigmoid, hyperbolic tangent, and radial basis function (RBF) [

34].

C. Step 3: Validation of NN Model on Simulation

In this step, the fault detector model is validated. As mentioned previously, 20% of the available database is reserved for validation. Two performance indicators are used to assess the performance of the model: accuracy ACC and dependability DP as defined in (2) and (3).

A C C = \frac{T N F + T F}{T N S} \times 100 %

(2)

D P = \frac{T F}{T F + F N F} \times 100 %

(3)

where TNF is the number of operating states correctly predicted as no-fault; TF is the number of operating states correctly predicted as a fault; FNF is the number of operating states wrongly predicted as no-fault; and TNS is the total number of scenarios simulated.

D. Step 4: Tuning of NN Model

In this step, the NN model is tuned according to the combination of features and hyperparameters. If the current iteration is smaller than the maximum number of iterations, the methodology must propose a strategy to modify the combination of features and hyperparameters. The adjustment strategy is made according to [

15].

E. Step 5: Data Acquisition

The presented AI model requires current and voltage signals at the IED location. The observation is preprocessed by a dedicated acquisition module adjusted for a sampling rate between 32-128 samples per cycle. The data are framed into a window equal to one cycle.

The samples are processed by an FFT and then standardized using mean and standard deviation values from Step 2.

F. Step 6: Embedding of NN Model for Online Fault Detection

In this step, the adaptive fault detection model is embedded into a single-board computer (SBC), which is appropriate to perform real-time inference with AI models. A Jetson Nano board is used in this step. The final model is set after accomplishing the maximum number of iterations in Step 2 to Step 4. Figure 3 illustrates the adaptive fault detector model on an edge device, where $Z_{m}$ is the hidden unit.

Fig. 3 Adaptive fault detector model on edge device.

IV. Case Study

The adaptive fault detector model is tested in the ADN laboratory of the College of Engineering at Universidad del Norte (Uninorte), Colombia. This network is based on the IEEE 13-node test feeder scaled down to 220 V. The IEEE 13-node test feeder presents an unbalanced operation characteristic and topology with different lateral types [

35]-[37]. Figure 4 illustrates the ADN laboratory at Uninorte. Note that the laboratory has been conceived as a modular so that different topological configurations can be assembled. The ADN laboratory has a synchronous generator connected to node N671. Additionally, it can change its topology by a circuit breaker located between the nodes N671 and N692 and a normally open line located between the nodes N680 and N675. The adaptive fault detection method is assessed for two scenarios, i.e., offline and online, which are described in the following subsections.

Fig. 4 ADN laboratory at Uninorte.

A. Offline Test Setup

The offline test includes Steps 1 to 4 from Section III. Table II presents the operating states for the training database. The numbers of combination under no-fault and fault states are 716800 and 336000, respectively.

Table II Operating States for Training Database

State	Factor	Level
No-fault	Change of topology	Original and alternative
	Load variation (%)	10-20, 20-30, 30-40, 40-50, 60-80, 80-100, 100-120, 120-130
	Cut-off generation	On and off
	Capacitor switching	On or off
	Voltage unbalance	+2% or -2% per phase (5 random scenarios)
	Sample stamp	-12, -6, -3, 0, 3, 6, 12
Fault	Fault resistance (Ω)	0, 15, 35, 45, 60
	Lines in fault	5
	Fault type	ag, bg, cg, ab, bc, ca, abg, bcg,cag, and abc
	Change of topology	Original and alternative
	Load variation (%)	40-50, 60-80, 80-100, 100-120
	Cut-off generation	On and off
	Voltage unbalance	+2% or -2% per phase (4 random scenarios)
	Capacitor switching	On and off
	Sample stamp	-12, -6, -3, 0, 3, 6, 12

The EMT software used to obtain the database is the Power Factory Digsilent [

38]. The initial settings used to obtain the database are simulation time of 100 ms and a sampling frequency of 1920 Hz [39].

Three case studies are considered in the offline test with the main goal of highlighting the contributions of the method to improve the accuracy and protection responses in a physical system. The first case considers a fault detection model without multiple sampling. The second case includes multiple sampling. However, the model does not consider feature selection and the tuning of NN model. The third case presents a fault detection model with multiple sampling, feature selection, and tuning of NN model.

For the first and second cases, the NN is set up as follows: input features as listed in Table I, the number of hidden layers as 1, and the numbers of neurons in the hidden layer as 100. Then, for the third case, the method uses the algorithm proposed in [

15] to obtain a better combination of features and hyperparameters. The maximum numbers of hidden layers and neurons per layer are set to be 3 and 150, respectively [33].

B. Online Test Setup

The online test includes Steps 5 and 6 from Section III. The adaptive fault detector model obtained from the offline test is validated in real time using the ADN laboratory at Uninorte. The experimental conditions for the implementation of the NN model are as follows.

1) The acquisition module is an SEL 751 relay, which performs signal acquisition and pre-processing. It is also in charge of sending the trip signal to the circuit breaker when a fault state is detected.

2) The SBC is a Jetson Nano running the NN model. Modbus TCP/IP protocol is used for the communication between the relay and the Jetson. Appendix A Fig. A1 illustrates the physical assembly for the test, which is at the head of the test system between nodes N650 and N632 in the ADN laboratory at Uninorte.

Due to the high effort required to reproduce every fault and no-fault states in the ADN laboratory at Uninorte, it is necessary to select some relevant operating states. Tables III and IV list the no-fault and fault states selected. Table III shows the states where there are no faults in the network. These operating states are applied in the same way for a connection or disconnection of the DG. Table IV includes the conditions where there is a fault in the network but outside the protection zone of IED. Therefore, no-fault state should be detected for the model.

Table III No-fault State to Reproduce for Physical Validation

Topology	Load variation when capacitor switching is on (%)			Load variation when capacitor switching is off (%)
Topology	1	2	3	1	2	3
Topology 1	67	100	120	From 100 to 67	From 100 to 120	From 100 to 100
Topology 2	67	100	120	From 100 to 67	From 100 to 120	From 100 to 100

Table IV Fault State to Reproduce with Faults Outside of Protection Zone on N675

Fault type	Fault resistances under topology 1 (Ω)	Fault resistances under topology 2 ( $Ω$ )
Single-phase fault to ground (phase c)	40	40
Double-phase fault to ground (phase a-b)	40	40

Similarly, Table V presents the fault state to reproduce with faults into protection zone. These conditions are chosen according to what is viable to replicate in the laboratory without compromising its safety. The total numbers of fault and no-fault states reproduced in the ADN laboratory at Uninorte are 18 and 28, respectively.

Table V Fault State to Reproduce with Faults into Protection Zone

Fault type	Fault resistance under topology 1 (Ω)			Fault resistance under topology 2 (Ω)
Fault type	N634	N634	N671	N634	N634	N671
Single-phase-to-theground fault	40	40	40	40	40	40
Double-phase-to-the ground fault	40	40	40	40	40	40
Three-phase faults	40	40	40	40	40	40

V. Test Results and Discussion

A. Offline Test Validation

The offline test validation for the adaptive fault detector model is made at three different lines of the ADN laboratory at Uninorte: lines containing N650-N632, lines containing N632-N633, and lines containing N632-N671.

Table VI shows the confusion matrix for the first case described in Section IV-A, where the first cell of the confusion matrix represents the number of no-fault states correctly predicted, and the second cell represents the number of no-fault states wrongly predicted. The accuracies for the IEDs 1, 2, and 3 are 49.42%, 92.58%, and 57.07%, respectively. The fault detection models classify a significant number of no-fault states with fault states and vice versa. Table VII presents the confusion matrix for adaptive fault detector model using default hyperparameters with multiple sampling, as described in Section IV-A.

Table VI Confusion Matrix for Adaptive Fault Detector Model Using Default Hyperparameters Without Multiple Sampling

IED	Actual state	Predictive value with no-fault	Predictive value with fault	ACC (%)	DP (%)
1	No fault	308485	20879	49.42	55.61
1	Fault	321514	26161	49.42	55.61
2	No fault	621868	2889	92.58	63.15
2	Fault	47331	4951	92.58	63.15
3	No fault	366374	26982	57.07	42.64
3	Fault	263625	20058	57.07	42.64

Table VII Confusion Matrix for Adaptive Fault Detector Model Using Default Hyperparameters with Multiple Sampling

IED	Actual State	Predictive value with no-fault	Predictive value with fault	ACC (%)	DP (%)
1	No fault	629999	2653	99.60	94.36
1	Fault	0	44387	99.60	94.36
2	No fault	669199	0	100.00	100.00
2	Fault	0	7840	100.00	100.00
3	No fault	629850	1419	99.76	96.98
3	Fault	149	45621	99.76	96.98

It can be observed that multiple samplings result in a dramatic improvement in the performance of the models, even with default hyperparameter values. The average accuracy and dependability of the three IEDs are now 99.78% and 97.11%, respectively. This shows a 50% increase in the average accuracy and an 80% increase in the average dependability concerning the first case.

Table VIII shows the combination of features and hyperparameters obtained once the tuning strategy is applied, which is described in case 3 from Section IV-A. Additionally, Table IX shows the confusion matrix for adaptive fault detector model with an adjustment process and multiple sampling. From those two tables, it can be argued that hyperparameter tuning and feature selection achieve only marginally improved performance compared with multiple samplings alone for the problem under study.

Table IX Confusion Matrix for Adaptive Fault Detector Model with an Adjustment Process and Multiple Sampling

IED	Actual state	Predictive value without fault	Predictive value with fault	DP (%)
1	No fault	629999	413	99.12
1	Fault	0	46627	99.12
2	No fault	669199	0	100.00
2	Fault	0	7840	100.00
3	No fault	629999	387	99.17
3	Fault	0	46653	99.17

Table Ⅷ Combination of Features and Hyperparameters Obtained per Each Adaptive Fault Detector Model

IED	ACC (%)	Hyperparameter	Feature selection combination
1	99.93	(109, 135, 75)	$S_{a b c}, Z_{a b c}, V_{a b c}, θ V_{a b c}, I_{a b c}, θ I_{a b c}$
2	100.00	(100, 0, 0)	$S_{a b c}, Z_{a b c}, V_{a b c}, θ V_{a b c}, I_{a b c}, θ I_{a b c}$
3	99.94	(130, 93, 84)	$S_{a}, Z_{a c}, V_{a b c}, θ V_{a c}, I_{a c}, θ I_{b}$

The final average accuracy and dependability of the three IEDs increase to 99.95% and 99.43%, respectively. The main improvement is on the average dependability, with an increase of 2% with respect to the second case.

B. Online Test Validation

The online test analyzes the behavior of the IED located between N650 and N632. The model is embedded into the SBC board and connected to the ADN laboratory at Uninorte to reproduce the operating states proposed in Tables III, IV, and V. Table X presents the confusion matrix obtained from the model validation.

Table X Confusion Matrix Obtained from Model Validation

IED	Actual state	Predictive values without fault	Predictive value with fault	ACC (%)	DP (%)
1	No fault	26	0	95.65	90
1	Fault	2	18	95.65	90

Considering Table X, it is possible to note that the adaptive fault detector model achieves performance metrics on a physical network comparable to those reached on simulated tests. Only two cases of no-fault states are detected as fault states.

Figures 5 and 6 show the behavior of different features and the trip signal considering several no-fault states. Figure 5 presents an no-fault state with no DG connected, no capacitor switching, original topology, and a change in load from 100% to 67%. Figure 6 reproduces a no-fault state with DG connected, capacitor switching, alternative topology, and 67% of the load.

Fig. 5 No-fault state without DG connected. (a) S_a. (b) Z_a. (c) I_a. (d) Trip signal.

Fig. 6 No-fault state with DG connected. (a) S_a. (b) V_a. (c) $θ I_{a}$ . (d) Trip signal.

Figure 5 shows a change in the injection of apparent power in the power grid between the interval from 11 s to 17 s. However, the trip signal remains zero, which means that the model can identify the signal variations in the power grid as part of a no-fault state, represented by a load change. Similarly, Fig. 6 presents a variation of the measure of the features between 8 s and 10 s, which can be associated with capacitor switching. Despite the observed variation, we can note that the trip signal remains zero, meaning a no-fault detection is given by the model.

Figure 7 shows the fault state with a three-phase fault in N632 and a resistance of $40 Ω$ . Similarly, Fig. 8 shows the fault state with a single-phase fault (phase a) in N671 and a resistance of $40 Ω$ . In Fig. 7, an increase at the current level in the interval between 8 s and 10 s is presented which matches with a trip signal equal to 1, meaning that the model detects a fault in the ADN. Similarly, in Fig. 8, a variation of the measured features in the interval between 8 s and 10 s is presented which can be associated with the presence of a single-phase-to-the-ground fault (phase a). Also, we can observe that the model detects the fault when the trip signal changes from 0 to 1, which means that the model takes just one cycle or 16.6 ms to identify a fault state. Finally, one should note that the average time from fault detection to open the CB is nearly 800 ms. This reaction time includes the average interrupting time of the circuit breaker itself, which is approximately 780 ms according to the datasheet of the manufacturer. Significant improvements on the total reaction time will only be achieved by replacing the current circuit breaker with a faster solution.

Fig. 7 Fault state with a three-phase fault in N632 and a resistance of 40 $Ω$ . (a) S_a. (b) V_a. (c) I_a. (d) Trip signal.

Fig. 8 Fault state with a single-phase (phase a) fault in N671 and a resistance of $40 Ω$ . (a) I_a. (b) I_b. (c) V_a. (d) Trip signal.

VI. Conclusion

This paper describes an adaptive fault detector based on NN model running on an edge device for ADN and MG system protection. The proposed method uses a local protection strategy where each device is responsible for detecting the faults over its three-phase section lines and their immediately adjacent lines. The method uses voltage and current observations as inputs to the AI model and trains an artificial NN model embedded into an SBC for real-time fault detection in a physical system. The proposed method shows the results considering the effect of several operating conditions such as cut-off generation, change of topology, voltage unbalance, and different fault conditions. Besides, the training process uses a multiple sampling factor to improve the robustness in fault detection. The NN model decides to trip the protection device in just one signal cycle. The method highlights the value of considering multiple samplings in the training of the fault detection model for physical implementation, where the results show an increase in the average accuracy and dependability of 50% and 80%, respectively, with respect to the same models trained without multiple samplings. Simulation-based testing shows the results with an accuracy of 99.5% and dependability of 99.5% for the fault detector model with feature selection and tuning of NN model. Final online tests are made in physical ADN. The results show the effectiveness and robustness of the adaptive detection model with an accuracy of 95% and dependability of 100%. Test results, together with an easy-to-implement embedded model, indicate the potential of the method for real-life applications.

J. C. Velez received the B.S. and M.S. degrees in radio electronic systems from the O. S. Popov Odessa National Academy of Telecommunications, Odessa, Ukraine, in 1994, and the Ph.D. degree from the Moscow Power Engineering Institute, Moscow, Russia, in 2004. He is currently a Titular Professor with the Electrical and Electronics Engineering Department, Universidad del Norte, Barranquilla, Colombia. His research interests include radio electronic systems, remote sensing, digital signal processing, and stochastic processing.

Appendix

Appendix A

Fig. A1 Adaptive fault detector assembly into ADN laboratory at Uninorte.

References

M.-G. Choi, S.-J. Ahn, J.-H. Choi et al., “Adaptive protection method of distribution networks using the sensitivity analysis for changed network topologies based on base network topology,” IEEE Access, vol. 8, pp. 148169-148180, Aug. 2020. [Baidu Scholar]

S. Chowdhury, S. P. Chowdhury, and P. Crossley, Microgrids and Active Distribution Networks. London: Institution of England Technology, 2009. [Baidu Scholar]

N. Hatziargyriou, Microgrids: Architectures and Control. New York:Wiley, 2014. [Baidu Scholar]

L. Mariam, M. Basu, and M. F. Conlon, “Microgrid: architecture, policy and future trends,” Renewable and Sustainable Energy Review, vol. 64, pp. 477-489, Oct. 2016. [Baidu Scholar]

A. Monti, F. Milano, E. Bompard et al., Converter-based Dynamics and Control of Modern Power Systems. Cambridge: Academic Press, 2020. [Baidu Scholar]

S. A. Gopalan, V. Sreeram, and H. H. C. Iu, “A review of coordination strategies and protection schemes for microgrids,” Renewable and Sustainable Energy Review, vol. 32, pp. 222-228, Apr. 2014. [Baidu Scholar]

T. S. Ustun, C. Ozansoy, and A. Zayegh, “Modeling of a centralized microgrid protection system and distributed energy resources according to IEC 61850-7-420,” IEEE Transactions on Power Systems, vol. 27, no. 3, pp. 1560-1567, Feb. 2012. [Baidu Scholar]

A. A. Memon and K. Kauhaniemi, “An adaptive protection for radial AC microgrid using IEC 61850 communication standard: algorithm proposal using offline simulations,” Energies, vol. 13, no. 20, p. 5316, Aug. 2020. [Baidu Scholar]

T. S. S. Senarathna and K. T. M. U. Hemapala, “Review of adaptive protection methods for microgrids,” AIMS Energy, vol. 7, no. 5, pp. 557-578, Sept. 2019. [Baidu Scholar]

P. Mahat, Z. Chen, B. Bak-Jensen et al., “A simple adaptive overcurrent protection of distribution systems with distributed generation,” IEEE Transactions on Smart Grid, vol. 2, no. 3, pp. 428-437, Jun. 2011. [Baidu Scholar]

J. Yu, Y. Hou, A. Y. S. Lam et al., “Intelligent fault detection scheme for microgrids with wavelet-based deep neural networks,” IEEE Transactions on Smart Grid, vol. 10, no. 2, pp. 1694-1703, Nov. 2019. [Baidu Scholar]

S. Rahman, S. K. Sarker, S. M. Muyeen et al., “A deep learning based intelligent approach in detection and classification of transmission line faults,” International Journal of Electric Power Energy System, vol. 133, p. 107102, Jan. 2021. [Baidu Scholar]

A. Draz, M. M. Elkholy, and A. A. El, “Soft computing methods for attaining the protective device coordination including renewable energies: review and prospective,” Archives of Computational Methods in Engineering, vol. 28, no. 7, pp. 4383-4404, Dec. 2021. [Baidu Scholar]

J. Marin-Quintero, C. Orozco-Henao, W. S. Percybrooks et al., “Toward an adaptive protection scheme in active distribution networks: intelligent approach fault detector,” Applied Soft Computing, vol. 98, p. 106839, Jan. 2021. [Baidu Scholar]

J. Marín-Quintero, C. Orozco-Henao, J. C. Velez et al., “Micro grids decentralized hybrid data-driven cuckoo search based adaptive protection model,” International Journal of Electric Power & Energy Systems, vol. 130, p. 106960, Jan. 2021. [Baidu Scholar]

C. Cepeda, C. Orozco-Henao, W. Percybrooks et al., “Intelligent fault detection system for microgrids,” Energies, vol. 13, no. 5, pp. 1-21, Jan. 2020. [Baidu Scholar]

S. Kar, S. R. Samantaray, and M. D. Zadeh, “Data-mining model based intelligent differential microgrid protection scheme,” IEEE Systems Journal, vol. 11, no. 2, pp. 1161-1169, Jan. 2017. [Baidu Scholar]

K. Saleh and A. Ayad, “Fault zone identification and phase selection for microgrids using decision trees ensemble,” International Journal of Electric Power & Energy Systems, vol. 132, p. 107178, Oct. 2021. [Baidu Scholar]

D. P. Mishra, S. R. Samantaray, and G. Joos, “A combined wavelet and data-mining based intelligent protection scheme for microgrid,” IEEE Transactions on Smart Grid, vol. 7, no. 5, pp. 2295-2304, Oct. 2016. [Baidu Scholar]

B. Patnaik, M. Mishra, R. C. Bansal et al., “MODWT-XGBoost based smart energy solution for fault detection and classification in a smart microgrid,” Applied Energy, vol. 285, p. 116457, Jan. 2021. [Baidu Scholar]

Y. Naitmalek, M. Najib, M. Bakhouya et al., “Embedded real-time battery state-of-charge forecasting in micro-grid systems,” Ecological Complexity, vol. 45, p. 100903, Nov. 2021. [Baidu Scholar]

E. Principi, D. Rossetti, S. Squartini et al., “Unsupervised electric motor fault detection by using deep autoencoders,” IEEE/CAA Journal of Automatica Sinica, vol. 6, no. 2, pp. 441-451, Feb. 2019. [Baidu Scholar]

H. Li, G. Hu, J. Li et al., “Intelligent fault diagnosis for large-scale rotating machines using binarized deep neural networks and random forests,” IEEE Transactions on Automation Science and Engineering, vol. 19, no. 2, pp. 1-11, Apr. 2021. [Baidu Scholar]

G. Burt and F. C. C. B. A. Dys, “Quantitative analysis of network protection blinding for systems incorporating distributed generation,” IET Generation, Transmission & Distribution, vol. 6, pp. 1218-1224, Jan. 2012. [Baidu Scholar]

S. Amir, H. Askarian, S. Hossein et al., “An overview of microgrid protection methods and the factors involved,” Renewable and Sustainable Energy Reviews, vol. 64, pp. 174-186, Oct. 2016. [Baidu Scholar]

K. Zheng and M. Xia, “Impacts of microgrid on protection of distribution networks and protection strategy of microgrid,” in Proceedings of 2011 International Conference on Advanced Power System Automation and Protection, Beijin, China, Apr. 2011, pp. 356-359. [Baidu Scholar]

F. Mumtaz and I. S. Bayram, “Planning, operation, and protection of microgrids: an overview,” Energy Procedia, vol. 107, pp. 94-100, Sept. 2017. [Baidu Scholar]

H. F. Zhai, M. Yang, B. Chen et al., “Dynamic reconfiguration of three-phase unbalanced distribution networks,” International Journal of Electric Power & Energy Systems, vol. 99, pp. 1-10, Dec. 2018. [Baidu Scholar]

S. K. Bhattacharya and S. K. Goswami, “Distribution network reconfiguration considering protection coordination constraints,” Electric Power Components and Systems, vol. 36, no. 11, pp. 1150-1165, Mar. 2008. [Baidu Scholar]

F. Aminifar, S. Teimourzadeh, A. Shahsavari et al., “Machine learning for protection of distribution networks and power electronics-interfaced systems,” The Electricity Journal, vol. 34, no. 1, p. 106886, Jan. 2021. [Baidu Scholar]

H. Khalid and A. Shobole, “Existing developments in adaptive smart grid protection: a review,” Electric Power Systems Research, vol. 191, p. 106901, Jun. 2021. [Baidu Scholar]

M. Alonso, H. Amaris, and D. Alcala, “Smart sensors for smart grid reliability,” Sensors, vol. 20, no. 8, pp. 1-23, Apr. 2020. [Baidu Scholar]

M. A. Shahin, H. R. Maier, and M. B. Jaksa, “Data division for developing neural networks applied to geotechnical engineering,” Journal of Computing in Civil Engineering, vol. 18, pp. 105-114, Apr. 2004. [Baidu Scholar]

C. M. Bishop, Pattern Recognition and Machine Learning. New York: Springer, 2006. [Baidu Scholar]

C. Orozco-Henao and J. Marín-Quintero, “Active distribution networks laboratory: a case of experiments in power quality,” in Proceedings of 4th International IEEE Workshop on Power Electronics and Power Quality Applications, Manizalez, Colombia, Jun. 2019, pp. 1-6. [Baidu Scholar]

W. H. Kersting, Distribution System Modeling and Analysis. Boca Raton: CRC press, 2002. [Baidu Scholar]

Pérez-Londoño, S. Garcés, A. Bueno-López et al. (2020, Jan.). Components modelling in AC microgrids. [Online]. Available: https://doi.org/https://doi.org/10.22517/97895 [Baidu Scholar]

User Manual Powerfactory v15.0, DIgSILENT GmbH, Germany, 2013. [Baidu Scholar]

L. Castro, M. F. B. López, M. Á. R. Ocampo et al. (2021, Jun.). Control jerárquico en micro-redes AC. [Online]. Available: https://hdl.handle.net/11059/13701 [Baidu Scholar]

Address:No.19 Chengxin Avenue, Jiangning District, Nanjing 211106, China

E-mail: mpce@alljournals.cn

Tel:86-25-81093060

Fax:86-25-81093040

Home

Introduction

Editorial Board

For Author

Call For Papers

APC

Sponsor & Publisher