1. Introduction
An MVDC system can supply power for shipping services, propulsion loads in shipboard electrical systems, onboard DC marine vessels, and remote area mining sites. This is attributed to several advantages, including reduced losses, lower complexity with fewer conversion stages, relaxed synchronization requirements, and the ease of integrating multiple energy storage systems [
1,
2,
3]. However, when such multi-terminal MVDC systems experience a fault contingency, they are subjected to a rapid rise in current due to the discharge of the converter capacitor [
4]. If a fault is not promptly isolated and cleared, it can potentially damage the power electronic devices within the voltage source converter (VSC) due to the substantial fault current. Moreover, a non-faulty line and a broader MVDC system may also suffer adverse effects. Therefore, faults and their locations in DC systems must be quickly and accurately determined to eliminate them [
5,
6].
Protecting DC grids is particularly challenging due to the typically high magnitude of the DC fault currents, which often peak within milliseconds of a fault’s onset [
7]. Fault detection, diagnosis, and clarification are also challenging. When identifying a fault, the line on which it occurred can be isolated; this also ensures its rapid reintegration back into the DC system [
8,
9]. Thus, fault diagnosis within DC systems, including MVDC systems, is crucial for enhancing system safety and reliability, facilitating efficient operation. Fault diagnosis is a prominent area of research in DC systems, with numerous studies conducted to improve their effectiveness [
10].
The traditional fault detection algorithm in DC systems, such as overcurrent, undervoltage, and rate-of-change among the current or voltage methods, are widely used [
11,
12,
13,
14]. However, these approaches often cannot detect faults with low or high resistance, and their selectivity can be problematic in meshed MVDC configurations. To achieve greater sensitivity and selectivity, some studies employed fault detection algorithms based on differential and directional currents, which rely on communication [
15,
16]. Despite the advantages of these schemes, they allow for fault detection, but do not enable the accurate classification of the fault type and location.
The studies on fault detection and diagnosis in HVDC systems are extensive [
17,
18,
19]. One study [
17] reviews the fault detection and location estimation techniques in multi-terminal HVDC systems, focusing particularly on protection strategies for systems utilizing voltage source converter (VSC) technology. However, this study only provides a comprehensive overview of the fault detection and location estimation techniques in multi-terminal HVDC systems. In a different study [
18], a new method combining Bayesian-optimized long short-term memory (LSTM) networks with the discrete wavelet transform (DWT) is proposed to detect faults in modular multilevel converter (MMC)-based HVDC systems. Another study [
19] proposes a novel protection scheme that combines the DWT with artificial neural networks (ANNs) for fault detection and classification in meshed multi-terminal HVDC grids. However, the approaches in these studies [
18,
19] are specifically focused on HVDC systems; their applicability to MVDC systems is yet to be explored.
Several other studies proposed fault detection and classification methods specifically for MVDC systems. For example, in [
20], an active impedance estimation (AIE) technique was proposed for fault classification and localization. However, obtaining an accurate estimation within a short timeframe remains a significant challenge. Another study [
21] proposed a method combining the wavelet transform (WT) and the self-organized adaptive resonance theory (ART) neural network to utilize both low- and high-frequency fault signals for advanced line diagnosis. However, selecting the appropriate wavelet function and decomposition level is crucial to balance accurate fault detection and computational efficiency. In ref. [
22], a self-healing procedure was proposed, which considers communication delays and measurement unit failures to identify unusual data behavior, whether caused internally or externally, to prevent incorrect computations. However, the method applied in [
22] does not determine the fault’s location and type.
Recently, AI-based algorithms were explored for fault diagnosis in DC systems [
23,
24,
25,
26,
27,
28,
29]. In ref. [
23], an artificial neural network (ANN)-based method for fault detection in VSC-HVDC which involved using only the measured voltage waveforms at the rectifier substation was presented. Another paper [
24] proposed a diagnosis scheme based on an ANN and high-frequency components to detect DC fault currents. In ref. [
25], a method combining WT multi-resolution analysis (MRA) with ANNs was proposed for fault detection and classification in MVDC systems. In ref. [
26], discrete wavelet analysis and a fuzzy neural pattern recognition approach were employed to detect the fault lines and locations by considering the interactions of short-circuit-induced traveling waves among the lines within the same loop. A different paper [
27] proposed a transient signal identification method based on wavelet entropy and a support vector machine (SVM) using the high-frequency signals generated by fault transients in HVDC. In ref. [
28], a novel fault diagnosis method based on short-time wavelet entropy was proposed, integrating the long short-term memory (LSTM) algorithm with an SVM. In ref. [
29], a DC line fault diagnosis method based on LSTM was proposed to improve the response and accuracy of the transient-based methods. However, the proposed diagnosis method in [
23,
24,
25,
26,
27,
28,
29] was applied only to DC fault classification, excluding the inverter faults in DC systems.
Further studies on AI-based fault diagnosis for multi-terminal DC systems were conducted [
30,
31,
32]. One paper [
30] proposed a Bi-GRU-based end-to-end fault diagnosis method that eliminated the need for manual feature extraction and classifier selection while offering a strong feature extraction capability without complex data preprocessing. Another paper [
31] presented a data-driven protection framework designed to improve the fault detection accuracy and reliability in multi-terminal DC networks, while ensuring strong interpretability by clearly presenting the extracted features and the softmax regression classifier mathematics. In ref. [
32], a novel and straightforward fault detection algorithm was presented, combining the ANN and discrete wavelet transform (DWT) methods. However, the proposed diagnosis method in [
30,
31,
32] was unable to detect faults across all the regions of the DC system.
An AI-based fault diagnosis scheme is proposed in this paper to enhance the accuracy of fault diagnosis in MVDC systems. The proposed scheme analyzes data on both the DC line and internal inverter faults, with digital signal processing performed using the discrete wavelet transform (DWT) algorithm. The processed data are then used to develop a classification model of fault location and type using a bidirectional long short-term memory (Bi-LSTM) network, targeting both the DC line and internal inverter faults. Additionally, a resistor-based fault current limiter (R-FCL) and a direct current circuit breaker (DCCB) are modeled and applied to the system to suppress the fault current peak value. A case study is simulated to verify the ability of the proposed diagnosis scheme to accurately diagnose faults in MVDC systems.
The main contributions of this paper are summarized as follows:
To the best of our knowledge, this study is the first to develop a fault diagnosis scheme using the Bi-LSTM algorithm in an MVDC system. The Bi-LSTM is used to develop a highly accurate classification model for the location and type of faults, including both DC line and internal inverter faults.
To quickly alleviate faults in the MVDC system, an R-FCL and a DCCB are implemented at both the ends of each DC line, and the two devices are appropriately combined to ensure system stability in the event of a fault.
The proposed fault diagnosis method is verified for reliability and accuracy by comparison with the existing methods, such as the CNN and LSTM in a four-terminal MVDC system.
3. Structure of the R-FCL and DCCB
In this paper, R-FCLs and DCCBs are implemented at both ends of each DC line to swiftly mitigate faults in the MVDC system. The R-FCL is a protection system comprising resistors and switches arranged in parallel to mitigate the fault currents. The R-FCL can promptly adjust the current limiter impedance without the need for additional fault detection and control systems, thus serving as an effective technical solution to address short-circuit current issues in MVDC systems [
35,
36]. Properly setting the resistance value (
Rf) in the R-FCL is crucial for reducing the peak fault values.
Figure 2 illustrates the structure of the R-FCL and provides a comparison of the
Idc1_peak values corresponding to various
Rf values to determine the optimal
Rf value. In the R-FCL system, the switch remains closed under normal conditions; however, in the event of a fault, the switch opens, and the fault current is influenced by
Rf, thereby limiting the fault current. In the system depicted in
Figure 1, when a PTP fault with a fault resistance of 1 Ω occurs at a distance of 25 km along the DC12 line, the
Idc1_peak value is determined after applying
Rf values ranging from 3 Ω to 30 Ω. These resultant values represent the outcomes when the R-FCL and the DCCB are jointly applied. Consequently, it is verified that the lowest
Idc1_peak value of 0.95 kA is achieved when
Rf is set to 13 Ω.
Figure 3 illustrates the structure of the DCCB, as modeled in this paper. DCCBs are available in various structural configurations, including arc-based and bidirectional series Z-source circuit breakers (ACBs and ZCBs) utilizing Z-source inductors and thyristors and hybrid CBs [
37,
38]. In this paper, a simple power electronic circuit breaker (PECB) employing two MOSFETs connected in series are employed to ensure rapid operation and fault current limitation. Note that the main purpose of this study is not to design an efficient and fast DCCB, but to develop a fault diagnosis scheme with a high accuracy. In
Figure 3, each MOSFET’s drain is connected to a DC line, and the sources of the two MOSFETs are interconnected. In the event of a fault, an operation signal is transmitted from ‘1’ to ‘0’ to the gate of each MOSFET, causing both the MOSFETs to open and mitigate the fault.
4. Proposed Bi-LSTM Based on a Fault Diagnosis Scheme
In this section, we present the proposed fault diagnosis scheme designed to enhance the reliability of MVDC systems in the event of a fault. Accuracy and the diagnosis time are critical factors in evaluating the effectiveness of a fault diagnosis scheme. Insufficient accuracy may result in incorrect system responses during a fault event. Additionally, in MVDC systems, particularly in the case of DC line faults, such as pole-to-pole (PTP) and pole-to-ground (PTG) faults, slow diagnosis can have severe consequences because the absence of a zero-crossing point complicates fault clearance.
As illustrated in
Figure 4, the proposed fault diagnosis scheme, developed to improve both accuracy and diagnosis, consists of the following two steps:
Data signal processing transforms the fault data (Idc, Vdc, iac, vac, and vpcc) measured during MVDC system faults, which are modeled in MATLAB/Simulink R2024a, using the discrete wavelet transform (DWT).
Fault classification involves utilizing the Bi-LSTM network to classify the location and type of faults, including DC line (PTP, P-PTG, and N-PTG) and internal inverter faults (inverter short-circuit), based on the processed digital data.
Finally, based on the fault diagnosis results predicted in steps 1 and 2, the faults in the MVDC system are detected and classified.
4.1. Digital Signal Processing Based on the DWT with Normalization
The DWT is widely utilized in wavelet analysis due to its advantageous characteristics, including its enhanced speed and accuracy, efficient selectivity, and higher resolution for transient signals. The DWT enables a comprehensive analysis of transient signals, thereby facilitating precise fault detection in MVDC systems.
In the DWT, the parameters ‘n’ and ‘m’ are integers that control the translation and dilation of the mother wavelet, respectively, corresponding to the sampling number and the decomposition level. In addition, ‘Ψ’ means the mother wavelet. WT analysis encompasses various families of wavelets, such as the Daubechies (
dbN), Symlets (
symN), and Coiflets families (
coifN), among others, which can be employed as the mother wavelet. In this context, ‘N’ denotes the order of the wavelet function.
dbN wavelets are often preferred over the other wavelet families due to their superior capabilities in signal recognition and noise removal [
39]. Accordingly,
dbN wavelets are considered as particularly suitable for use as the mother wavelet in this study. Within the framework of the DWT, decomposing a fault signal allows for the detailed evaluation of the fault conditions. The approximation (
A) and detail coefficients (
d) in the DWT are typically expressed as in Equations (2)–(4).
Here,
CN−1−k represents the scaling coefficient,
N is the total number of scaling coefficients, and
bk is the reconfigured coefficient. For feature extraction, multi-resolution signal decomposition is utilized to decompose the fault signal into both approximate and detailed components.
Figure 5 presents the
Idc results, converted from real data under normal and fault conditions using the DWT with a
db3 mother wavelet. The fault type is PTP, which was initiated at 1.0 s at 25 km along DC line 12 and lasting for 0.1 s. As shown in
Figure 5b, the wavelet-filtered signal (WFS) under fault conditions is observed to increase up to 10.0 and increase 10 times compared to that under normal conditions. This confirms that the DWT effectively processes abnormally fluctuating fault data in time series analysis.
4.2. Fault Clarification Based on the Bi-LSTM Algorithm
Figure 6 illustrates the cell structure of Bi-LSTM. Bi-LSTM enhances standard LSTM by processing the input data bidirectionally, incorporating both the forward and backward directions. It comprises forward LSTM that processes the input sequence from start to finish and backward LSTM that processes the sequence in reverse [
40]. The hidden states from both directions are typically concatenated or combined at each time step to yield a comprehensive representation of the input. The primary advantage of employing Bi-LSTM lies in its ability to capture the context and long-term dependencies in both the directions, thereby facilitating a deeper understanding of the input sequence. This algorithm is particularly well suited for accurate fault classification using the extensive and long-term processed data generated in the earlier stage. In
Figure 6, Bi-LSTM consists of multiple LSTMs, which is composed of three key gate structures: the forget, input, and output gates. The first gate, known as the forget gate (
ft), is responsible for discarding irrelevant information. The second gate, referred to as the input gate (
it), extracts relevant information from the preprocessed data and stores it in the cell state. The third gate, the output gate (
ot), classifies the information, specifically the fault types in the context of the proposed study. The function for each gate, including the update (
ct) and new cells (
ct*), is as follows:
In the LSTM architecture, four internal layers effectively propagate information, in addition to the three gates that regulate ct information. In the first layer, ft selects the previous cell state (ct−1) and determines the amount of information from ct−1 to retain in ct. ft reads the previous block’s output (ht−1) and the feature vector input at time instance t (xt), yielding a value between 0 and 1 for each element in the previous cell state. A value of 0 indicates firm discarding, while a value of 1 signifies firm reservation. In the second layer, the input gate (it) determines how much information from xt is stored in the cell state. Concurrently, tanh generates a new candidate vector (ct*), incorporating all the potential values to be added to the cell state. Subsequently, it and ct* are multiplied and added to the current state through an addition operation, progressing to the third layer. In the third layer, the cell state is updated to integrate new information and discard the previous state. Finally, the output layer selects the pertinent information and presents it as the output result. This layer employs a Sigmoid function for output selection. Subsequently, the cell state is processed through the tanh function and multiplied by ot. In this scenario, the variable ot transforms ct into the vector ht, which exposes the cell-state memory to the remainder of the network. At each time step of the sequence data, Wf, Wi, Wc, and Wo represent the learnable weights, while bf, bi, bc, and bo denote the biases, respectively.