Fault Diagnosis Method for MMC-HVDC Based on Bi-GRU Neural Network

Abstract

The Modular Multilevel Converter-High Voltage Direct Current (MMC-HVDC) system is recognized worldwide as a highly efficient strategy for transporting renewable energy across regions. As most of the MMC-HVDC system electronics are weak against overcurrent, protections of the MMC-HVDC system are the major focus of research. Because of the insufficiencies of the conventioned fault diagnosis method of MMC-HVDC system, such as hand-designed fault thresholds and complex data pre-processing, this paper proposes a new method for fault detection and location based on Bidirectional Gated Recurrent Unit (Bi-GRU). The proposed method has obvious advantages of feature extraction on the bi-directional structure, and it simplifies the pre-processing of fault data. The simplified pre-processing avoids the loss of valid information in the data and helps to extract detailed fault characteristics, thus improving the accuracy of the method. Extensive simulation experiments show that the proposed method meets the speed requirement of MMC-HVDC protections (2 ms) and the accuracy rate reaches 99.9994%. In addition, the method is not affected by noise and has a high potential for practical applications.

1. Introduction

With the gradual exhaustion of traditional fossil energy, the pressure on the ecological environment becomes serious, and the development of new energy is imperative. In the past few years, China has become the country with the largest wind power capacity and the fastest-growing photovoltaic power generation in the world [1,2]. The rapid development of new energy power generation in China has brought many technical challenges. The contradiction between the explosive growth of new energy and the slowdown in the growth of electricity demand leads to the consumption problem in energy development [3]. Moreover, the distribution of renewable energy and load centers in China are extremely unbalanced. At present, the construction of multi-terminal Modular Multilevel Converter-High Voltage Direct Current (MMC-HVDC) system is considered as one of the solutions to solve these problems [4].

MMC-HVDC protection as a line of defense to ensure the safe and reliable operation of the power system has become the focus of research [5,6]. At present, the challenge of MMC-HVDC system protection is that most electronic devices are weak to overcurrent, and the protection device must detect DC faults within a few milliseconds to ensure security [7,8]. In MMC-HVDC system, fault diagnosis considering both accuracy and rapid response is a very challenging task [9].

At present, parts of the fault diagnosis research are based on the fault mechanism model of the system. He Zhen et al. [10] propose a fault analysis method, which is based on Common- and Differential-Mode (CDM) network. In this method, the authors decompose the HVDC network into CDM network, which is balanced and decoupled. Then, a transfer impedance is defined and calculated based on the impedance matrices of the CDM networks. Based on the analysis method, analytical expressions of the PTG fault characteristics that vary with space and time are obtained. Liu Ruoping et al. [11] establish an MMC-HVDC equivalent model suitable for AC protection setting calculation and propose a quantitative index for evaluating the influence of MMC-HVDC on AC protection setting calculation. Li Bin et al. [12] analyze the transient characteristics of DC faults in an MMC-based DC system combined with the numerical method. Yang Haiqian et al. [13] establish an equivalent circuit model and performs the MTDC line fault location by analyzing the line inductance. The above research follows the process that first analyze the system fault mechanism and establish the equivalent circuit model, then obtain the expression of fault current or voltage, finally realizing the fault diagnosis. However, in the process of establishing the fault mechanism model, the simplifying circuit will result in errors. In addition, the method of determining the fault threshold through mathematical expressions is affected by line parameter changes, load changes, harmonic wave, short-circuit transition resistance, and other factors. These methods are difficult to determine accurate fault thresholds, resulting in low accuracy and lack of generalization ability.

Parts of the fault diagnosis research is based on the signal analysis approach. Wang Mian et al. [14] propose a frequency domain-based methodology to analyze the influence of HVDC system configurations and parameters on the traveling wave during a DC fault. Wang Shuai et al. [15] use wavelet time entropy to extract fault features of MMC-HVDC system to realize fault classification. Li Jianwei et al. [16] propose a novel protection scheme for MTDC systems based on high-frequency components detected from the fault current signal. The above approaches require manual determination of fault thresholds after extracting signal fault characteristics and are not accurate enough.

With the rapid development of AI algorithms in recent years, AI algorithms have been gradually applied to the electrical field. Compared with the traditional DC fault detection method, the application of AI algorithm in the fault diagnosis of MMC-HVDC system has significant advantages. The AI algorithm does not need to manually determine the threshold. It only needs a large amount of data and can avoid the accumulation of human design errors and enhance accuracy. The fault diagnosis based on AI algorithm is fast and accurate and meets the requirements of power grid protection [17]. Luo Guomin et al. [18] propose a transient identification method based on a deep belief network. This method trains the network with the normalized line mode component of transient currents and identifies faults and lightning disturbances with an accuracy of 96.4%. Wang Jun et al. [19] construct a neural network using the convolutional neural network (CNN) to detect and classify DC faults. CNN has obvious advantages in image feature extraction. When extracting fault features with CNN, 1-D time series data need to be composed into 2-D grayscale images. Preprocess of the grayscale image may loss some valid fault information, and the accuracy of this method is 92.5%. Wang Hao et al. [20] construct a Parallel Convolutional Neural Network (P-CNN) with a dual branching structure for fault classification and fault localization. The P-CNN is similar to CNN in terms of the data preprocessing process, which still requires data grayscale graphing. In addition to CNN, some scholars use Recurrent Neural Networks (RNN) to diagnose faults. Wang Qinghua et al. [21] establish a neural network based on Long Short-Term Memory (LSTM) to detect and classify MMC faults with high accuracy. This article illustrates that RNN neural networks have advantages over CNN in processing sequential data.

In the above paper, the neural networks adopted have some limitations in feature extraction. For better extraction of fault features, the authors need to pre-process the fault feature quantities, such as fast Fourier transform, wavelet transform, grayscale image processing, etc. However, in the process of data preprocessing, improper preprocessing may lose several valid information, resulting in a poor final training effect of the neural network. Therefore, it is necessary to simplify the preprocessing of fault data, which needs higher requirements on the performance of the neural network model.

This paper analyzes the fault characteristics under different fault conditions in MMC-HVDC system and proposes a Bi-GRU based deep learning fault diagnosis method. The main contributions are listed below:

(1)

An end-to-end fault diagnosis method is proposed for distinguishing fault types and fault location. The part of artificial design such as feature extraction and classifier selection is not needed.

(2)

The Bi-GRU based method achieves excellent performance without complex pre-processing of the raw data(positive and negative voltage and current), and it has strong feature extraction capability.

(3)

The accuracy of the fault diagnosis method reaches 99.9994% with 2 ms data length. Both speed and reliability meet the requirements for MMC-HVDC fault diagnosis.

2. Simulation Model and Fault Analysis

2.1. Simulation Model

The topology diagram of the MMC-HVDC system is shown in Figure 1. A four-terminal MMC-HVDC electromagnetic transient simulation model was established based on PSCAD/EMTDC. The model consists of four converter stations, named as M, N, P, and Q, respectively, and four DC transmission lines. Each transmission line connects two converter stations through the current-limiting reactor and the DC circuit breaker. R stands for DC line relay protection. The measuring device of voltage and current are located at the line side of the current-limiting reactors.

The main parameters of the MMC-HVDC system are shown in

Table 1. Main Parameter Of MMC-HVDC.

Parameters	Converter M	Converter N	Converter P	Converter Q
Capacity/MVA	1200	1200	3000	3000
AC voltage/kV	500	500	220	500
DC voltage/kV	±500	±500	±500	±500
Submodule Capacitance/mF	10	10	10	15
Number of submodules	625	625	625	625
Bridge arm reactor/mH	150	150	150	150

, and the line parameters are shown in

Table 2. Line Parameters.

Line	Line Length/km	Current Limiting Reactor/mH
LineMN	227	200
LineMP	66	300
LinePQ	219	200
LineNQ	126	200

. To accurately describe the transient process of the fault traveling wave, the DC transmission line adopts the frequency-varying parameter model. The AC system utilizes the constant power supply model. The converter station adopts a double loop vector control mode. The P converter station adopts the constant DC voltage control method, and the other converter stations adopt the constant active power control method.

2.2. Fault Classification and Analysis

For the four-terminal MMC-HVDC system, line protection needs to quickly and accurately distinguish between internal faults and external faults. Take the protection of the line MN as an example. Line MN is between protection relay devices ${{R}}_{MN}$ and ${{R}}_{NM}$, as shown in Figure 1. There are two operation modes for DC transmission systems: single-pole operation and double-pole operation [7]. If a single-pole fault occurs on the line MN, such as positive-pole-to-ground fault (PG) and negative-pole-to-ground fault (NG), the system can still operate as a single pole. The relay protection of line MN only needs to remove the fault pole line. If a positive-pole-to-negative-pole fault (PN) occurs on the line MN, the relay protection of line MN needs to clear the positive line and negative line [12]. The external fault occurs outside the line MN, the relay protection of line MN does not act. Because it is difficult to distinguish between line faults and DC bus faults, we subdivide external faults into faults at the busbar and other faults [9].

Integrating the above relay protection requirements, we divide the MMC-HVDC system into three parts, the red box, the blue box, and the green box in Figure 2. The red box indicates the internal faults, the blue box indicates the DC bus faults, the green box indicates the other external faults.

In the existing fault diagnosis methods, whatever the data preprocessing methods are, the fault characteristics are all derived from the positive and negative voltage (Up, Un), and positive and negative current (Ip, In). These signals complement each other with more obvious fault characteristics, thus achieving better diagnostic accuracy. The system parameters are shown in

Table 3. System Parameters.

Fault Number	Fault Location	Fault Type
f1	The internal fault
f2	M-side DC Bus	PG, NG, PN
f3	N-side DC Bus
f4, f5, f6	The external fault	None

. The time window is set as 2 ms after the fault occurs.

The fault waveforms of the internal fault and the external fault under PG fault are shown in Figure 3. The difference between Ip and In are not apparent for internal faults and external faults. The steepness of descent of Up and Un are more obvious for internal faults than for external faults. Therefore, Up and Un can be used as a characteristic quantity to distinguish between internal fault and external fault.

The fault waveforms under different faults of the 50% Line MN are shown in Figure 4. If a single pole-to-ground fault occurs, the current corresponding to the fault pole rises rapidly and the current of another pole almost remains unchanged. If a pole-to-pole fault occurs, both Ip and In rise swiftly. Therefore, Ip and In can be used as a characteristic quantity to distinguish fault types.

The PG fault waveforms of the Line MN at different locations are shown in Figure 5. For the different fault locations, the fluctuations of Ip, In, Up, and Un are different. It is difficult to design the fault threshold artificially.

At 50% of the line MN. The PG fault waveforms of the DC transmission line under different transition resistance are shown in Figure 6. The fault current distorts differently when the transition resistance is different. The larger the transition resistance is, the smaller the waveform distortion is. Therefore, it is difficult to design the fault threshold artificially when the transition resistance is too large.

It can be seen from Figure 3, Figure 4, Figure 5 and Figure 6 that the raw voltage and current signals can distinguish the fault characteristics. The system parameters affect the decay rate and magnitude of the waveform. Therefore, the traditional method of determining the threshold value of the fault waveform signal has limitations in fault diagnosis of MMC-HVDC systems. The artificial intelligence-based method can avoid the manual design of thresholds and prevent the error of manual design thresholds. In addition, consider that the raw voltage and current signals are able to avoid the possible loss of features caused by signal pre-processing, aim to propose a fault diagnosis method based on neural network using the original voltage and current signals as input quantities.

3. The Proposed Method

3.1. Reasons for Choosing Bi-GRU

GRU is a new type of RNN. The gradient of the conventional RNN is apt to explode when the time step is too large and is apt to decay when the time step is too small. Therefore, RNN is difficult to capture the large dependence of time steps in time series. GRU can capture the relationship well between long steps in a time series because it has gates to control the flow of information.

GRU is a simplified version of LSTM, but in practice, the performance of GRU in sequence processing is similar to LSTM. GRU instead of LSTM will greatly reduce computation and ensure the performance of information extraction at the same time. The structures of GRU and LSTM are shown in Figure 7. GRU has two gates: the reset gate and the update gate. LSTM has three gates: the input gate, the output gate, and the forget gate. Comparing with LSTM, the main simplification of GRU is to combine the forget gate and the input gate into an update gate and merge the unit state with the hidden state. Due to this simplification, the number of parameters for GRU training is reduced and the training convergence is accelerated.

In the GRU, the input of the reset gate and update gate is the input $X_t$ of the current time step and the hidden state $H_{t-1}$ of the previous time step; the output is calculated by the fully connected layer, which has a sigmoid activation function. Suppose the number of hidden units is h, the small batch input $X_t\in R^{n\times d}$ is (the number of samples is n, the number of inputs is d) at the time step t, and the hidden state is $H_{t-1}\in R^{n\times h}$ at the previous time step t−1. The calculation of reset gate $R_t\in R^{n\times d}$ and update gate $Z_t\in R^{n\times d}$ are follows:

$$\begin{array}{c}{R}_t=\sigma (X_t{H}_{xr}+H_{t-1}{W}_{hr}+b_r)\hfill \\ Z_t=\sigma (X_t{H}_{xz}+H_{t-1}{W}_{hz}+b_z)\hfill \end{array}$$

(1)

where $H_{xr}$, $H_{xz}\in R^{d\times h}$ and $W_{hr}$, $W_{hz}\in R^{h\times h}$ are the weight parameters, $b_r,b_z\in R^{1\times h}$ are the deviation parameters. $\sigma$ is the sigmoid function.

The candidate hidden state ${\tilde{H}}_t\in R^{n\times h}$ and the hidden state $H_t\in R^{n\times h}$ at time step t are calculated as follows:

$$\begin{array}{c}{\tilde{H}}_t=tanh(X_t{W}_{xh}+(R_t\otimes H_{t-1}{W}_{hh}+b_h))\hfill \\ H_t=Z_t\otimes H_{t-1}+(1-Z_t)\otimes {\tilde{H}}_t\hfill \end{array}$$

(2)

where $W_{xh}\in R^{d\times h}$, $W_{hh}\in R^{d\times h}$ are the weight parameters, and $b_h\in R^{1\times h}$ are the deviation parameters.

In GRU, the reset gate helps to capture the short-term dependencies in the time series, and the update gate helps to capture the long-term dependencies in the time series. In theory, GRU can store and retrieve information in long steps. However, GRU has limitations in terms of structure. GRU only uses historical information in the process of feature extraction but does not make use of future information. To overcome this limitation, we use Bi-GRU to build a neural network. The basic structure of Bi-GRU is shown in Figure 8. In this structure, the final output of Bi-GRU at time t depends not only on the previous frame T − 1 at time t but also on the next frame T + 1 at time t. In other words, the Bi-GRU stacked by two GRUs can benefit from the information in the two-time directions. Specifically, one GRU flows forward to calculate the forward hidden state, $({\overrightarrow{h}}_1,{\overrightarrow{h}}_2,{\overrightarrow{h}}_3\dots {\overrightarrow{h}}_n)$ at the same time, another GRU flows backward and calculates the reverse hidden state (${\overleftarrow{h}}_1$,${\overleftarrow{h}}_2$,${\overleftarrow{h}}_3$…${\overleftarrow{h}}_{\mathrm{n}}$).Therefore, at each time step, Bi-GRU can simultaneously capture historical information and future information [22].

3.2. Proposed End-to-End Model

Based on the above information, we choose Bi-GRU as the basic neural network to design a diagnosing MMC-HVDC faults model. As shown in Figure 9, this model has the following blocks: input layer, hidden layer, output layer, and training network. To demonstrate the model in detail, an example is shown as follows:

(1)

The input layer is the bottom part of the model. The current and voltage raw signals are normalized into two-dimensional arrays using batch processing techniques and are labeled. The two-dimensional array T*4 is reorganized by [x₁, x₂, x₃…x_n], where n is the length of the sequence. [x₁, x₂, x₃…x_n] is the input sequence. The shape of input is identified as [batch_size, steps, dimension], where the batch_size is the period number of the input sequence in one batch, the Steps refers to sampling points in one cycle and the Dimension is the number of fault features. The batch_size is 20 considering the computing power of the computer hardware, and Steps is 100 considering the time window.

(2)

The hidden layer is the central part of the model and the main part of the feature extraction. At each time step, multiple Bi-GRU layers are stacked and used for feature extraction of the input sequence [x₁, x₂, x₃…x_n]. As a bidirectional network, multi-layer Bi-GRU will affect the prediction weight with both future and historical information. The output sequence of the hidden layer is expressed as [y₁, y₂, y₃…y_α]. Similarly, the output is reorganized into [batch sequence size, step size, dimension].

(3)

The output layer is the top component of the model. It determines the type of each element in the input sequence according to the output by the hidden layer. However, the output of the hidden layer is a three-dimensional tensor and the input of the fully connected layer is required to be a two-dimensional tensor. The output needs to be shaped before it is sent to the fully connected layer. Therefore, [y₁, y₂, y₃…y_α] is reshaped from [batch_size, step_size, number of GRU neurons] to [batch_size×step_size, number of GRU neurons]. Then, the fully connected layer outputs [o₁, o₂, o₃…o_β] are sent directly to the SoftMax layer.

The output corresponds to a label with a discrete value. As the error between the discrete value and the uncertain range of the output value is difficult to measure, we solve this problem by the SoftMax operation. SoftMax regression is the equivalent of linear regression. The input features and weights are linearly superimposed. The output values are transformed into a probability distribution with positive values by the following formula:

$$\begin{array}{c}{\widehat{y}}_1,{\widehat{y}}_2,{\widehat{y}}_3\dots {\widehat{y}}_n=softmax(o_1,o_2,o_3\dots o_{\beta })\hfill \\ {\widehat{y}}_1=\frac{exp\left(o_1\right)}{{\sum }_{i=1}^{{}_{\beta }}exp\left(o_i\right)},{\widehat{y}}_2=\frac{exp\left(o_2\right)}{{\sum }_{i=1}^{{}_{\beta }}exp\left(o_i\right)},{\widehat{y}}_n=\frac{exp\left(o_{\beta }\right)}{{\sum }_{i=1}^{{}_{\beta }}exp\left(o_i\right)}\hfill \end{array}$$

(3)

where ${\widehat{y}}_1,{\widehat{y}}_2,{\widehat{y}}_3\dots {\widehat{y}}_n$ and $o_1,o_2,o_3\dots o_{\beta }$ represents the fully connected layer output and Softmax layer output.

The shape of the SoftMax layer output is [batch_size, steps, fault types and locations]. The total output of SoftMax is 1, and the output value is regarded as the probability of the predicted fault types and locations.

The strategy to train the above model is through a backpropagation algorithm in time. Specifically, first, the output of SoftMax is obtained based on the above network and the error term is calculated by backpropagation. Then, the gradient of the model parameters is calculated based on the error term. Finally, the loss function selects cross-entropy and Adam algorithm as the optimization method. The cross-entropy loss function is shown in the following formula:

$$l(\Theta )=\frac{1}{n}\sum _{i=1}^{n}H(y^{\left(i\right)},\widehat{y}{}^{\left(i\right)})$$

(4)

where $\Theta$ and n represents the model parameters and the number of samples in the training datasets. Once the parameters of the feature extractions and decision part are well trained, the model can be used for MMC-HVDC fault diagnosis in the actual field.

3.3. Fault Datasets Preparation

In the four-terminal MMC-HVDC system, we simulate different faults on DC lines, busbars, and converter stations. DC faults include PG, NG, PN. DC transmission line fault locations are set at 1%, 10%, 30%, 50%, 70%, 90%, and 99% of each line. Fault transition resistance is randomly set as 0.01∼300 $\Omega$. AC faults include single-phase short circuit (1PSC), two-phase short circuit (2PSC), three-phase short circuit (3PSC). Fault transition resistance is randomly set as 0.01 $\Omega$. The sampling frequency is 50 kHz. A measuring device at N converter stations collects positive and negative voltages and currents to produce fault datasets.

In data processing, we normalize the data to be distributed between 0 and 1. Normalization can accelerate the speed of convergence. The proposed method is an end-to-end deep learning structure. The labeling process is shown in

Table 4. Tagging label.

Fault Location	Fault Type	Label
f1	PG NG PN	(1,5)(1,6)(1,7)
f2	PG NG PN	(2,5)(2,6)(2,7)
f3	PG NG PN	(3,5)(3,6)(3,7)
f4,f5,f6	None	(4,8)
None	None	(0,0)

. For the coordinate (x, y), x represents the fault type and y represents the fault location. Due to the protection strategy, internal faults are divided in detail in Table 4. For the coordinate x, 1 represents internal faults, 2 and 3 represent the faults at the DC buses. For the coordinates y, 5 represents PG fault, 6 represents NG fault, and 7 represents PN fault. The coordinates of all external faults are (4, 8). For the coordinate (0, 0) repsents the noraml state.

We then fully shuffle the fault dataset to prevent training over-fitting. The dataset includes a training set, validation set, and test set. A reasonable data set division is very important for the effect of the trained model, and we empirically divide the training set, validation set, and test set into 3:1:1. If the training data is not enough, the trained neural network is probably under-fitting. We fully consider different fault types and locations and produce more than 4000 sets of training datasets. The overall methodology is implemented by using the Tensorflow 2.0. The adopted software environment is Anaconda Python 3.7.1. All the experiments are carried out on a workstation equipped with an Intel i7-8700 K processor, a 32-GB memory, and a GTX 1080 Ti graphics processing unit.

3.4. Parameters and Hyperparameter Design

We choose the accuracy and the testing time to judge the performance of the neural network. The accuracy rate is expressed as follows:

$$Accuracy=\frac{numberofcorrectclasscifation}{totalnumberoftests}$$

(5)

The hyperparameters of the neural network are the number of Bi-GRU layers, batches, the number of neurons, and the learning rate. We determine the optimal hyperparameters based on the comparative performance.

When the number of Bi-GRU neurons is set to 256 and the small batch is set to 20, the accuracy and testing time of different Bi-GRU layers are shown in Figure 10a. When the number of Bi-GRU layers is set to 3, the classification accuracy reaches the peak value of 99.99% and the corresponding testing time is 1.24 ms. As the number of Bi-GRU layers increases, the accuracy remains the same and the testing time increases to 1.856 ms. The accuracy difference between 3-Bi-GRU layers and 4-Bi-GRU layers is not significant, and the testing time increases to 1.74 ms. Therefore, we set the optimal number of Bi-GRU layers to 3.

When the number of Bi-GRU layers is set to 3 and the number of Bi-GRU neurons is set to 256, the corresponding accuracies and testing times for the different number of batches are shown in Figure 10b. It can be seen that the precision and testing time increase significantly as the Batches increase from 1 to 20. When the number of batches increases from 20 to 25, the precision hardly changes, but the testing time increases sharply. Hence, we set the number of batches to 20.

When the number of Bi-GRU layers is set to 3 and the number of batches is set to 20, the precision and testing time corresponding to the different numbers of Bi-GRU neurons are shown in Figure 10c. It can be seen that the precision curve increases significantly as the number of Bi-GRU neurons increases from 16 to 256. And the precision remains the same as the number of Bi-GRU increases from 256 to 512. The testing time curve always increases with the increase of the number of Bi-GRU neurons. Considering both the accuracy and testing time, the number of GRU neurons is set to 256.

Plentiful test results show that increasing the number of Bi-GRU layers and GRU neurons in the model improves the performance of the method, but the complexity and computational effort of the neural network increase correspondingly. The learning rate is mainly used to control the step size of gradient descent. If the learning rate is too low, the algorithm can only converge after several iterations, and the training time is long, which reduces the optimization speed. If the learning rate is too high and the gradient descent step is too large, the neural network will eventually diverge rather than converge. The neural network adopts the Adam optimization method, we set the learning rate as 0.0001.

A lot of experiments have shown that 3 Bi-GRU layers, 20 batches, 256 GRU neurons, and a 0.001 learning rate can achieve a great balance between accuracy and testing time. The training process of the fault diagnosis model is shown in Figure 11. The X-axis represents the number of iterations in the network training process, the left Y-axis represents the training accuracy, and the right Y axis represents the training loss. The full epoch is 200. After the 27th epoch of training, the accuracy of the model reaches the maximum. To minimize the training time, the number of iterations of the fault detection model is set as 27 to obtain the best training effect. It can be seen that the accuracy of the fault diagnosis model reaches 99.9994%.

4. Simulation Results

In this section, the performance of the proposed fault diagnosis method is tested by the test set and compared with a variety of mainstream AI algorithms.

4.1. Proposed Model Performance

4.1.1. Testing Sets Performance

To test the reliability of the fault diagnosis method, each sample set is randomly divided into a training set and a test set with appropriate proportions.

Table 5. Test results of fault classification and location.

Fault Location	Fault Type	Transition Resistance/Ω	Samples Number	Correct Number
Line MN	PG		60	60
Line MN	NG		60	60
Line MN	PN		60	60
Line NQ	PG		60	60
Line NQ	NG		60	60
Line NQ	PN		60	60
M-side DC bus	PG	0.01∼300	60	60
M-side DC bus	NG		60	60
M-side DC bus	PN		60	60
N-side DC bus	PG		60	60
N-side DC bus	NG		60	60
N-side DC bus	PN		60	60
M converter station	1PSC		60	60
M converter station	2PSC		60	60
M converter station	3PSC	0.01	60	60
N converter station	1PSC		60	60
N converter station	2PSC		60	60
N converter station	3PSC		60	60

lists the detailed sample distributions. For each fault location, 180 samples are produced. The method is independent of fault types and transition resistance. It can distinguish between internal faults and external faults, as well as DC bus faults. The fault diagnosis method has an excellent performance in distinguishing internal faults from external faults.

4.1.2. Anti-Noise Interference Capability

In practical engineering, environmental noise and measuring device errors can interfere with fault diagnosis. To evaluate the performance of the model in a noisy environment, we compared the performance of the model under normal conditions and a 40db Gaussian noise environment. The results of the model’s interference immunity are shown in

Table 6. Test results of noise interference.

Fault Location	Fault Type	Transition Resistance/Ω	Samples Number	Correct Number
Line MN	PG		30	30
Line MN	NG		30	30
Line MN	PN		30	30
Line NQ	PG		30	30
Line NQ	NG		30	30
Line NQ	PN		30	30
M-side DC bus	PG	0.01∼300	30	30
M-side DC bus	NG		30	30
M-side DC bus	PN		30	30
N-side DC bus	PG		30	30
N-side DC bus	NG		30	30
N-side DC bus	PN		30	30
M converter station	1PSC		30	30
M converter station	2PSC		30	30
M converter station	3PSC	0.01	30	30
N converter station	1PSC		30	30
N converter station	2PSC		30	30
N converter station	3PSC		30	30

. From the simulation results, it can be seen that the proposed fault diagnosis model has great robustness.

4.1.3. Capability to Adapt to Different Operating Conditions

In practical engineering, it is necessary to shut down the converter station and nearby transmission lines during converter station maintenance. Therefore, we test the effect of different operating conditions of the system on the accuracy of the model. To verify whether the model can achieve fault diagnosis under different operating conditions of the system, new fault data for the case of disconnected P-conversion stations and surrounding lines are added to the fault data set. The Bi-GRU based neural network is trained with the new fault data set. The test results are shown in

Table 7. Test results of fault classification and location during P converter station maintenance.

Fault Location	Fault Type	Transition Resistance/Ω	Samples Number	Correct Number
Line MN	PG		30	30
Line MN	NG		30	30
Line MN	PN		30	30
Line NQ	PG		30	30
Line NQ	NG		30	30
Line NQ	PN		30	30
M-side DC bus	PG	0.01∼300	30	30
M-side DC bus	NG		30	30
M-side DC bus	PN		30	30
N-side DC bus	PG		30	30
N-side DC bus	NG		30	30
N-side DC bus	PN		30	30
M converter station	1PSC		30	30
M converter station	2PSC		30	30
M converter station	3PSC	0.01	30	30
N converter station	1PSC		30	30
N converter station	2PSC		30	30
N converter station	3PSC		30	30

. After a number of test sets, the neural network outputs were tested to be accurate. The fault diagnosis method has great performance under different system operating conditions.

4.1.4. Capability at Low Sampling Rate

To verify whether the model can distinguish different faults accurately at a lower sampling rate, we tested this model at a sampling rate of 10 kHz. We change the number of neurons in the input layer of the Bi-GRU-based neural network to 64 (the sampling rate is reduced to 1/5 of the previous one, and the sampling data is also reduced), and then import the new data and train them. After 60 test sets, the neural network outputs were tested to be accurate. The fault diagnosis method has great performance at the low sampling frequency.

4.2. Comparison

We compare the GRU based fault diagnosis method with the proposed Bi-GRU based fault diagnosis method in parameters and hyperparametersin

Table 8. The parameters and hyperparameters of Bi-GRU and GRU.

Category	Name	Parameter Setting
Basic architecture	Ratio of training set to test set and validation set	1:3:1
	Bi-GRU or GRU layers	3
	Bi-GRU or GRU neurons	256
	Batches	20
	Initial learning rate	0.0001
	Maximum number of epochs	200
	Activation function	ReLU
Learning criterion	Loss function	Cross entropy
Softmax Algorithm	Optimization function	Adam

.

The comparison performance results are shown in

Table 9. The results comparison of GRU with Bi-GRU.

Parameter	Bi-GRU	GRU
Location Accuracy/%	100	100
Classification Accuracy/%	99.99	95.45
Training Time/min	68	29
Testing time/ms	1.31	0.69

. The results illustrate that both GRU and Bi-GRU have 100% fault location accuracy. This great performance is because the localization dataset is much larger than the classification dataset and the model is easier to extract location features. Bi-GRU has a bi-directional structure compared to GRU, its fault classification accuracy reaches 99.99%. The GRU has limitations and the fault classification accuracy is 95.45%. Although Bi-GRU requires more training time and testing time, its testing time of 1.31 ms fully meets the requirements of the protection action speed of a flexible DC transmission system. In contrast, the Bi-GRU based fault diagnosis method, which has higher accuracy, has better performance.

We also compare the proposed Bi-GRU based fault diagnosis method with other fault diagnosis method based on K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Parallel Convolutional Neural Network (P-CNN), and Bidirectional Long Short-Term Memory (Bi-LSTM) [20,21]. As shown in

Table 10. Fault diagnosis results of artificial intelligence algorithms.

The Neural Network	Testing Accuracy/%	Testing Time/ms
KNN	90.82	0.704
SVM	95.23	0.800
P-CNN	99.24	1.856
Bi-LSTM	98.7	0.97
Bi-GRU	99.99	1.313

, these AI algorithms meet the fault diagnosis speed requirements (2 ms). Although the testing time of KNN reaches 0.704 ms, its accuracy is only 90.82%. The accuracy of P-CNN reaches 99.24%, but its testing time up to 1.856 ms. These AI algorithms do not have a good balance of accuracy and testing time. Our method achieves 99.9994% accuracy with proper testing time.

5. Conclusions

This paper proposes a Bi-GRU-based fault diagnosis method for MMC-HVDC transmission systems. This method provides fast and reliable fault diagnosis without complex data processing. We analyze the fault characteristics of positive and negative currents and voltages on transmission lines and utilize current and voltage raw signals to train an efficient neural network. Simulation results illustrate the effectiveness of the proposed method. The following conclusions can be achieved.

(1)

The proposed end-to-end method performs excellently in fault diagnosis without hand-designed components such as feature extraction and classifier selection. The speed meets the requirement of MMC-HVDC fault diagnosis (2 ms). The overall accuracy is higher than 99.9994%. With more training samples, the accuracy can be further improved.

(2)

The method achieves excellent performance without complex pre-processing of the raw data, and it has a strong feature extraction capability.

(3)

The proposed fault diagnosis model has strong robustness and has a high potential for practical applications.

In MMC-HVDC system with overhead transmission lines, lightning disturbances may cause the misoperation of the line protection, which must be distinguished from line fault cases. Deep learning based lightning disturbance identification method is to be studied.