Fault Diagnosis Method of Intelligent Substation Protection System Based on Gradient Boosting Decision Tree

Abstract

In order to improve the efficiency of the devices’ fault diagnosis of the protection systems of intelligent substation, a fault diagnosis method based on a gradient boosting decision tree (GBDT) was proposed. Using the integrated alarm information, the device self-checking information, the link information of generic object-oriented substation event (GOOSE) and sampled value (SV) and the sampling value information generated during the fault of the protection system, the fault feature information set is constructed. According to different fault characteristics, the protection system faults are classified into simple faults and complex faults to improve the diagnosis efficiency. Using GBDT training rules, a fault diagnosis model of protection system based on GBDT is established and fault diagnosis steps are given. This study takes a 110 kV intelligent substation in southern China as an example, to verify the effectiveness and accuracy of the proposed fault diagnosis method, and compared it with the existing methods in terms of the accuracy. The diagnostic accuracy in the case of false alarms and the case of multiple faults are verified. The results show that the method can meet the practical engineering application.

1. Introduction

The protection system is an important part of the smart substation, which mainly includes the merging unit, intelligent terminal, and protection device. It is of great significance for the safe operation of the whole power grid to detect the faults or abnormal conditions in the power system, so as to send an alarm signal, or directly isolate and remove the faulty part. Therefore, it is necessary to ensure the reliable operation of the protection system itself. Even if it fails, it should also quickly eliminate the fault and ensure its support for the normal operation of the power grid to the maximum extent. With the rapid development of smart grid technology, intelligent substations or digitally transformed smart stations have gradually become popular. Substations have the new features of intelligent equipment, data networking, and overall station informatization [1]. Compared with the traditional substation, the fault types and fault characteristics of its protection system have changed a lot, and it is necessary to continuously develop new fault diagnosis methods on the basis of traditional fault diagnosis methods [2]. Improvements in data acquisition, storage, and analysis in intelligent substations provide us with new ideas for developing new fault diagnosis methods. The function of the secondary system in the substation is more complete and the amount of monitoring information is larger, for example, the operation event record of the secondary system, the main station information, the key information of the protection device, the integrated alarm information, the network operation information, the device self-checking information, etc., which provide solid data support for the condition assessment and fault diagnosis of the substation. Reference [3] analyzed the operation behavior of the secondary system of intelligent substation, and sorted out 18 core evaluation sub-items, including sampling accuracy, internal environment, self-checking state, protection start, input consistency, real-time input and output, operation time of the whole group, port flow, fault-free time, correct action, correct control, station-control layer communication, process-layer communication, port function, port performance, on-time receiving, and on-time output. A Big Data mining technology for intelligent substations was proposed by Reference [4]. By deeply mining the integrated alarm information, the device self-checking information, the link information of GOOSE and SV, the sampling value information provided data support for fault diagnosis.

With regard to the research on the fault diagnosis of the protection system of the intelligent substation, Reference [5] proposed a method for evaluating the performance of secondary equipment in smart substations based on availability, dependability, and capability (ADC). The accuracy of its evaluation needs to be improved. Reference [6] proposed an online monitoring and fault diagnosis method of the secondary circuit of relay protection based on multi-parameter information. Through the monitoring and analysis of the SV, GOOSE, and manufacturing message specification (MMS) messages, the online state monitoring method, abnormal sampling value, switch monitoring, and abnormal alarm strategy for relay protection devices was proposed. At the same time, the typical alarm information of the protection device and secondary circuit when faults occur was collected, analyzed, and uploaded to construct a set of online monitoring and fault diagnosis systems for a secondary circuit of relay protection. References [7,8,9,10] proposed a method of locating the secondary equipment fault based on the substation configuration description (SCD) of the intelligent substation. However, the actual workflow of intelligent substations is highly dependent on the configuration tools of integrators and manufacturers, and the difference in configuration tools leads to a poor standardization of the files. Therefore, there are still many shortcomings in the method of locating faults simply using SCD files. With the research in machine learning and deep learning algorithms, their application in fault diagnosis is increasingly employed [11]. Reference [12] proposed a research method for the fault location of the secondary device in intelligent substations based on deep learning. According to the device self-checking information, a fault location model for the secondary device based on a recurrent neural network (RNN) was established and the fault location steps were given; however, the data source used had certain limitations, and due to the limitations of the time and accuracy of the algorithm, this method needs to be improved. Reference [13] proposed an intelligent state assessment of the protection systems based on random forest algorithm, but the prediction accuracy and robustness need to be improved, and the requirements for the parameters are high.

The above method can basically meet the needs of fault diagnosis, but there are still various problems in its practical application. In summary, the current research on the protection system of intelligent substations still needs to face these problems: There are many types of faults in the protection system, but the correlation between the fault characteristics is weak; the complex components of the protection system equipment and the connection relationship between the different devices cause a large amount of data to be generated when a fault occurs [14]. Conventional methods cannot efficiently and quickly analyze the massive multi-dimensional data; since the fault-feature information may be distorted and lost during the acquisition process, the results obtained by the conventional method fluctuate with the confidence of the fault feature information; In addition, the accuracy of the algorithm also needs to be improved.

In order to solve the problems of data source, data processing, fault diagnosis logic, diagnosis method, and fault accuracy rate faced in fault diagnosis, a new fault diagnosis method of an intelligent substation protection system based on a gradient boosting decision tree is proposed. The GBDT algorithm is a supervised ensemble learning method. Through the continuous iteration of the weak prediction model composed of decision trees, the strong prediction model is trained with the goal of minimizing the prediction errors of the previous round. It has extremely high accuracy and a fast convergence speed. Taking the protection system merging unit, intelligent terminal, and protection device as the main body of fault diagnosis, this method used the integrated alarm information, device self-checking information, link information of GOOSE and SV, and the sampling value information as the judgment basis to form the fault feature information set. According to the historical fault feature data and maintenance records, the faults of the protection system are divided into simple faults and complex faults. At the same time, the gradient boosting decision tree (GBDT) intelligent algorithm is used as a diagnostic tool, and the fault diagnosis process of the protection system is proposed to realize the diagnosis of complex faults of the protection system. The effectiveness of the method proposed in this paper is verified by example analysis.

2. Fault Type and Fault Feature Information of the Protection System

2.1. Classification of Protection System Fault Types

When the fault diagnosis of the protection system is carried out, the fault types are properly classified, which can ensure the accuracy of judgment, reduce the amount of calculation in the process of fault diagnosis, reduce the amount of computer resources, and improve the response speed and convergence speed [12].

By analyzing the alarm information, self-checking information, sampling value information, and fault maintenance-record data of the device of the protection system, we can divide the faults into two categories. One is the simple fault, that is, there is an obvious mapping relationship between the fault type and the fault feature information. After the fault occurs, the fault type can be simply deduced according to the fault feature information. For example, if the fault feature information is “Power failure alarm of merging unit”, it can be directly deduced that the fault is “Power module fault of merging unit”. Another type is the complex fault, which means that the mapping relationship between the fault type and fault feature information is weak, and cannot be directly deduced by simple reasoning of fault feature information. An intelligent algorithm is needed for the reasoning and diagnosis.

According to the equipment manual, fault data, and fault characteristics, the high-frequency faults are classified as shown in

Table 1. Simple fault types and corresponding fault feature information.

Number	Fault Types	Fault Feature Information
1	Power module fault of merging unit	Power failure alarm of merging unit
2	The RX1 receive Bus fault of the merge unit	4–8 Bus MU frame Loss/4–8 Bus check error
3	GPS timing signal of merging unit not accessed	B01 synchronous anomaly alarm of merging unit
4	Device board card configuration and specific engineering design drawings do not match	Board card configuration error
5	Error in merging unit memory check	Merging unit memory checking error
6	Incorrect configuration of send text for merge unit	Incorrect configuration of send text for merge unit
7	Errors in HTM data exchange between merging unit board boards	B01 HTM error of merging unit
8	Device stall error of merging unit	BO1 report BO3 Stall error of merging unit
9	B11 input power loss of merging unit	B11 optocoupler power loss alarm of merging unit
10	BO3 sampling plate anomaly of merging unit	BO3 sampling plate anomaly of merging unit
11	The sampling of BO3 dual AD in merging unit is inconsistent	The sampling of BO3 dual AD in merging unit is inconsistent
12	GOOSE-A network storm of merging unit	GOOSE-A network storm of merging unit
13	The sampling voltage of merging unit BO2 drops below the set value	BO2 sampling power drop of merging unit
14	Error in intelligent terminal memory check	Intelligent terminal memory checking error
15	BO1_A network link N of intelligent terminal disconnected	BO1_A network link N of intelligent terminal disconnected
16	BO1_GOOSE-A network storm of intelligent terminal	BO1_GOOSE-A network storm alarm of intelligent terminal
17	BO1_GOOSE configuration error of intelligent terminal	BO1_GOOSE configuration error of intelligent terminal
18	BO2_A network link N of intelligent terminal disconnected	BO2_A network link N of intelligent terminal disconnected
19	GPS timing signal of intelligent terminal not accessed	B01 synchronous anomaly alarm of intelligent terminal
20	Error in HTM data exchange between intelligent terminal device boards	B01 HTM error of intelligent terminal
21	Device stall fault of intelligent terminal	BO1 report BO2 stall error of intelligent terminal
22	The power supply of the intelligent terminal AD chip exceeds the normal range	BO3 sampling power supply anomaly of intelligent terminal
23	B09 input power loss of intelligent terminal	B09 optocoupler power loss alarm of intelligent terminal
24	Abnormal opening of intelligent terminal outlet power supply	Bus QD Signal Anomaly
25	GOOSE text configuration error of intelligent terminal	Link N configuration error
26	GPS input signal of intelligent terminal is lost or abnormal	GPS clock abnormal
27	Abnormal timing of the protection device	Abnormal timing of the protection device
28	Spring device of protection device does not store energy	No energy storage alarm for spring device of protection device
29	CT disconnection of protection device	CT disconnection alarm
30	PT disconnection of protection device	PT disconnection alarm
31	Synchronous voltage circuit disconnection	Abnormal synchronous voltage

and

Table 2. Complex fault types and corresponding main fault feature information.

Number	Fault Types	Fault Feature Information
f1	Main DSP module failure of merging unit	Sampling anomaly of merging unit, Synchronization anomaly of merging unit, SV alarm of merging unit/protection device, Protection locking, etc.
$f_2$	Sampling DSP module failure of merging unit
$f_3$	$\mathrm{I}/\mathrm{O} \mathrm{module} \mathrm{fault} \mathrm{of} \mathrm{merging} \mathrm{unit} (\mathrm{Merging} \mathrm{unit} \leftrightarrow$ Switch)	SV alarm of protection device, SV interruption of protection device/merging unit/measurement and control device, GOOSE interruption of merging unit, SV/GOOSE total alarm, Protection locking, etc.
$f_4$	$\mathrm{I}/\mathrm{O} \mathrm{module} \mathrm{fault} \mathrm{of} \mathrm{merging} \mathrm{unit} (\mathrm{Bus} \mathrm{merging} \mathrm{unit} \to$ Line merging unit)
$f_5$	$\mathrm{I}/\mathrm{O} \mathrm{plug}-\mathrm{in} \mathrm{fault} \mathrm{of} \mathrm{protection} \mathrm{device} (\mathrm{Protection} \mathrm{device} \leftrightarrow$ Intelligent terminal)	SV/GOOSE alarm of protection device, GOOSE alarm of intelligent terminal, GOOSE interruption of protection device, Protection locking, Reclosing device locking, etc.
$f_6$	$\mathrm{I}/\mathrm{O} \mathrm{plug}-\mathrm{in} \mathrm{fault} \mathrm{of} \mathrm{protection} \mathrm{device} (\mathrm{Switch} \to$ Protection device)
$f_7$	$\mathrm{I}/\mathrm{O} \mathrm{plug}-\mathrm{in} \mathrm{fault} \mathrm{of} \mathrm{protection} \mathrm{device} (\mathrm{Merging} \mathrm{unit} \to$ Protection device)
$f_8$	Longitudinal channel fault	Channel differential exit, Longitudinal channel abnormal, In/out communication interruption of protection device, etc.
$f_9$	CPU plug-in fault of protection device	Protection device parameter error, CPU X exception, Memory self-check error, Protection locking, etc.
$f_{10}$	$\mathrm{I}/\mathrm{O} \mathrm{board} \mathrm{fault} \mathrm{of} \mathrm{intelligent} \mathrm{terminal} (\mathrm{Intelligent} \mathrm{terminal} \leftrightarrow$ Switch)	GOOSE interruption of merging unit/measurement and control device/intelligent terminal/protection device, GOOSE alarm of merging unit/intelligent terminal/protection device, etc.
$f_{11}$	$\mathrm{I}/\mathrm{O} \mathrm{board} \mathrm{fault} \mathrm{of} \mathrm{intelligent} \mathrm{terminal} (\mathrm{Protection} \mathrm{device} \leftrightarrow$ Intelligent terminal)
$f_{12}$	GOOSE plug-in fault of protection device	SV/GOOSE alarm of protection device, GOOSE alarm of intelligent terminal, SV/GOOSE interruption of protection device, Protection locking, Reclosing lock, etc.

.

Set the complex fault set as Formula (1):

$$F=\left\{f_1,f_2,\dots ,f_{12}\right\}$$

(1)

In Formula (1), $f_1-f_{12 }$, respectively, represent the 12 faults in Table 2.

For the simple faults in Table 1, the expert system can be used for fault diagnosis according to the fault feature information summarized in Table 1. For space reasons, the performance of the method proposed in this paper is explored based on the faults in the complex fault set F. In addition, with the development of intelligent substations and the improvement in field complexity, the fault set F will be further expanded, and the method proposed in this paper is still applicable to subsequent faults.

2.2. Fault Feature Information of the Protection System

Based on the complex fault types of the protection system summarized in Table 2, this paper selects four features of integrated alarm information, device self-checking information, link information of GOOSE and SV, and sampling value information as the feature information of fault diagnosis, which can comprehensively reflect the change in feature quantity caused by the fault of the protection system [15].

The main function of the integrated alarm information is to reflect whether the protection system fails. If a fault occurs, the equipment will issue alarm information and upload it to the monitoring terminal, which can be used as one of the bases for equipment fault diagnosis while realizing fault warning.

Device self-checking is an important function of an intelligent protection system. When any abnormality occurs in the operation process of the device, the device will record the abnormal information through the event-recording function for the operator to query.

GOOSE (Generic Object-Oriented Substation Event) is equivalent to the DC control and signal cables in traditional substations, which transmit control instructions and signals. It mainly includes a switch/knife switch position, control switch position, abnormal/alarm signal, blocking signal, etc. SV (Sampled Value) is equivalent to the secondary AC cable in the traditional substation, which transmits the sampled instantaneous values of voltage and current, including the instantaneous value of voltage and current on the secondary side of the transformer. The link information of GOOSE and SV are important indicators to indicate whether the information links between the protection system equipment and between the equipment and the monitoring terminal work normally, reflecting the link connection state of the equipment.

The sampling value information is the sampling value of three-phase voltage and current transmitted by two channels, which can reflect whether the voltage and current-sampling function of the protection system are normal.

The protect system fault feature information, as shown in

Table 3. Protect system fault feature information.

Type of Fault Feature Information	Fault Feature Information
Integrated alarm information	Protection locking, Abnormal alarm of merging unit, Abnormal alarm of intelligent terminal, Abnormal alarm of protection device, SV total alarm, GOOSE total alarm, SV alarm of merging unit, SV alarm of protection device, GOOSE alarm of merging unit, GOOSE alarm of protection device, GOOSE alarm of intelligent terminal
Link information of GOOSE and SV	GOOSE interruption of merging unit, GOOSE interruption of protection device, GOOSE interruption of intelligent terminal, SV interruption of merging unit, SV interruption of protection device, SV interruption of measurement and control device
Device self-checking information	Sampling anomaly of merging unit, Synchronization anomaly of merging unit, Board card configuration error, Merging unit memory checking error, Incorrect configuration of send text for merge unit, B01 HTM error of merging unit, BO1 report, BO3 Stall error of merging unit, BO3 sampling plate anomaly of merging unit, Sampling of BO3 dual AD in merging unit is inconsistent, GOOSE-A network storm of merging unit, BO2 sampling power drop of merging unit, Intelligent terminal, Memory-checking error, B09 optocoupler power loss of intelligent terminal, BO1_A network link N of intelligent terminal disconnected, BO1_GOOSE-A network storm alarm of intelligent terminal, BO1_GOOSE configuration error of intelligent terminal, BO2_A network link N of intelligent terminal disconnected, B01 synchronous anomaly alarm of intelligent terminal, B01 HTM error of intelligent terminal, BO1 report BO2 stall error of intelligent terminal, BO3 sampling power supply anomaly of intelligent terminal, Channel differential exit, Longitudinal channel abnormal, Protection device parameter error, Abnormal timing of protection device, In/out communication interruption of protection device, Abnormal synchronous voltage, CPU X exception, Memory self-check error, 4–8 Bus MU frame loss, 4–8 Bus check error
Sampling value information	Double-channel voltage sampling value, Double-channel current sampling value

:

2.3. Fault Feature Information Set of the Protection System

According to the fault feature information in Table 3, the fault feature information set is established to provide data support for the subsequent fault diagnosis of the protection system of an intelligent substation.

The integrated alarm information set $A_i$ of the protection system of an intelligent substation in the i-th fault event is established as shown in Formula (2):

$$A_i=\left\{a_1,a_2,a_3,\dots ,a_{11}\right\}$$

(2)

a₁–a₁₁ in the above formula are the 11 kinds of fault feature information contained in the integrated alarm information in Table 3. When the monitoring host receives the alarm information, the element at the corresponding position is set to 1, otherwise it is set to 0.

The link information of the GOOSE and SV set $I_i$ of the protection system of an intelligent substation in the i-th fault event is established as shown in Formula (3):

$$I_i=\left\{i_1,i_2,i_3,\dots ,i_6\right\}$$

(3)

i₁–i₆ in the above formula are the 6 kinds of fault feature information contained in the link information of GOOSE and SV in Table 3. When the secondary monitoring system receives the alarm information, the element at the corresponding position is set to 1, otherwise it is set to 0.

The device self-checking information set $C_i$ of the protection system of an intelligent substation in the i-th fault event is established as shown in Formulas (4)–(7):

$$C_i=\left\{C_{MU},C_P,C_{IT}\right\}$$

(4)

$$C_{MU}=\{C_{MU1},C_{MU2,}\dots , C_{MUa}\}$$

(5)

$$C_P=\{C_{P1},C_{P2,}\dots , C_{Pb}\}$$

(6)

$$C_I=\{C_{I1},C_{I2,}\dots , C_{Ic}\}$$

(7)

In the above Formula (4), $C_i$ contains the device self-checking information in Table 3, and it is divided into three parts: merging unit self-checking information $C_{MU}$, protection device self-checking information $C_P$, and intelligent terminal self-checking information $C_{IT}$, where Formulas (5)–(7) subscripts a, b, and c represent the number of these three types of device in the protection system of an intelligent substation. When the secondary monitoring system receives the alarm information, the element in the corresponding position is set to 1, otherwise, it is set to 0.

The sampling value information set $S_i$ of the protection system of an intelligent substation in the i-th fault event is established as shown in Formula (3).

$$\left\{\begin{array}{c}{S}_i=\left\{M_1,M_2\right\}\\ M_1=\left\{I_{A1},I_{B1},I_{C1},U_{A1},U_{B1},U_{C1}\right\}\\ M_2=\left\{I_{A2},I_{B2},I_{C2},U_{A2},U_{B2},U_{C2}\right\}\end{array}\right.$$

(8)

In the above Formula (8), $M_1$ and $M_2$ represent the three-phase voltage and current sampling values in Channel 1 and Channel 2, respectively. I and U represent the three-phase current and voltage values of the dual channel.

To make the sample data of the different units comparable, improve the convergence speed of the model, and improve the accuracy of the model, the sampling value information is preprocessed by the Min-Max method, and the original value $S_i$ in the dataset is mapped to the value ${S_i}^{\prime }$ in the interval [0,1]. The conversion formula is shown in Formula (9):

$${S_i}^{\prime }=\frac{S-S_{min}}{S_{max}-S_{min}}$$

(9)

In the Formula (9), $S_{max}$ and $S_{min}$ are the maximum and minimum values of the sampled values, respectively.

3. Fault Diagnosis of the Protection System Based on Gradient Boosting Decision Tree (GBDT)

The gradient boosting decision tree intelligent algorithm belongs to the ensemble algorithm, which has a good processing ability for discrete data and is very prominent in dealing with small sample data. Gradient boosting is the core idea and step of this intelligent algorithm for the classification task. When it carries out ‘multi-classification’ work, it is based on ‘two classifications’ and adopts the idea of ‘one positive class, multiple negative classes’. The training process of the gradient boosting decision tree intelligent algorithm is the main work of this section, which mainly includes selecting the optimal value of the learning rate and the number of iterations, and finally gives the fault diagnosis process of the protection system based on the gradient boosting decision tree.

3.1. Principle and Training Steps of Gradient Boosting Decision Tree

GBDT intelligent algorithm belongs to the ensemble learning algorithm. The ensemble learning algorithm is a hot topic in the field of engineering applications. It is a method to improve the learning ability through the combination of multiple weak learners [16]. Compared with conventional methods, it has a good performance in terms of accuracy and generalization ability. Bagging and Boosting algorithms are two typical ensemble learning algorithms. The schematics are shown in Figure 1 and Figure 2.

The bagging algorithm generates n training sample sets from the total sample library according to the random sampling method with playback. Each sample set trains a weak learner and uses the sample set to train n weak learners. The weak learners run in parallel. According to different combination strategies, n weak learners are combined to generate strong learners. The boosting algorithm is an inherited algorithm, in which the weak learners operate in a serial manner. The data weight in the training set of each iteration is changed by the learning results of the weak learners. The learning results are fitted according to the residuals, and then n weak learners are combined according to different combinations to generate strong learners [17].

The gradient boosting decision tree (GBDT) intelligent algorithm is one of the most widely used boosting algorithms in the engineering field, and it combines the sampling idea of the bagging algorithm, allowing sampling samples and features to increase the independence between weak learners. The GBDT intelligent algorithm does not change the sample weight in the iteration process, but continuously learns the negative gradient of the loss function, generates multiple new weak learners, and combines multiple weak learners into strong learners. Compared with the traditional machine learning algorithm, the GBDT intelligent algorithm can achieve higher accuracy in many of the application scenarios, and has a faster operation speed, stronger generalization ability, and lower requirements for parameter adjustment.

Regardless of whether the GBDT intelligent algorithm performs regression tasks or classification tasks, its core idea is “gradient boosting”, and the negative gradient of the loss function in the iterative process is shown in Formula (10):

$${\tilde{y}}_i=-\frac{\partial L\left(y_i,F\left(x_i\right)\right)}{\partial F\left(x_i\right)},\hspace{1em}i=1,2,\dots ,\mathrm{K}$$

(10)

In the above formula, $\widetilde{y_i}$ is the negative gradient of the current loss function, namely the fitting target of the next iteration, $L\left(y_i,F\left(x_i\right)\right)$ is the current loss function, $y_i$ is the learning target value of the current weak learner, $F\left(x_i\right)$ is the output value of the current weak learner, x is the input variable (refers to the fault feature information in this paper), K is the number of training samples, and i is the current training sample.

GBDT multi-classification is an organic combination of the GBDT binary classifier. In the training process, the idea of ‘one positive class, multiple negative classes’ is adopted. There are 12 kinds of fault types in the fault set. In the training process of a single sample, when the sample is the i-th type of fault ( $i=1,2,\dots ,12$), it is assumed that the fault type of this sample is from 1 to 12, and each time it is assumed that the other 11 types of fault are unified as the negative samples of this sample, and 12 binary classifiers are trained to generate independent classification. Then, according to the real fault type of the sample, only the output result of one binary classifier (assuming that the i-th class is a positive class) is correct, and the rest are errors, then the final classification result is the fault type corresponding to this binary classifier. The schematic diagram of the GBDT multi-classification algorithm is shown in Figure 3.

In this paper, when conducting 12-classification training for complex faults in the protection systems of intelligent substations, the fault feature information of the protection system in Table 3 is used as a variable, and the complex fault types of the protection system in Table 2 are used as the fitting target. The training process is as follows:

Step 1

Select the fault sample i in the training set. The fault feature information set of this sample is = $X_i=\left\{A_i ,I_i,C_i,{S_i}^{\prime }\right\}$, and the fault type is $f_1$—“Main DSP module failure of merging unit” in Table 2. Then, the true classification label (probability) of the fault sample in the 12 binary classifiers is the fitting target y = (1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0). Because the fault sample belongs to the first fault $f_1$, ‘1’ is used to indicate that it belongs to the fault, and ‘0’ is used to indicate that it does not belong to the other 11 faults, forming the input consisting of the fault feature information set and the fitting target: ($X_i$,1)($X_i,0$)($X_i,0$)…($X_i,0$);

Step 2

Input the input into 12 weak classifiers and obtain the output results: $F_1\left(X_i\right), F_2\left(X_i\right), F_3\left(X_i\right),\dots ,F_{12}\left(X_i\right);$

Step 3

Convert the output result into probability, as shown in Formula (11):

$$p_n\left(X\right)=\frac{\exp \left(F_n\left(X_i\right)\right)}{{{\displaystyle \sum }}_{k=1}^{12}\exp \left(F_n\left(X_i\right)\right)}, n=1, 2, 3, \dots , 12$$

(11)

Step 4

Calculate the loss function and solve the negative gradient of the loss function.

The loss function formula is shown in Formula (12):

$$L\left(y,p\left(X\right)\right)=-{{\displaystyle \sum }}_{n=1}^{12}{y}_n \log p_n\left(X\right), n=1, 2, 3, \dots , 12$$

(12)

The negative gradient formula of the loss function is shown in Formula (13):

$${\tilde{y}}_n=-\frac{\partial L\left(y,p\left(X\right)\right)}{\partial F_n\left(X_i\right)}, n=1, 2, 3, \dots ,12$$

(13)

Step 5

Generate a new fitting target, namely ($X_i$, ${\tilde{y}}_1$)($X_i,{\tilde{y}}_2$)($X_i,{\tilde{y}}_3$)…($X_i,{\tilde{y}}_{12}$), and repeat steps 2 to 5.

Iterate M times according to the above steps, and after generating M weak learners, one training is completed. The training set contains 12 kinds of complex faults in the fault set and has several corresponding training samples. After each sample of the training set is trained once, the GBDT 12 classification model with a high accuracy can be obtained. It should be noted that the values of the number of iterations M and the learning rate σ require multiple verifications, and this process is described in detail in Section 3.2.

3.2. Training of Fault Diagnosis Model Based on Gradient Lifting Tree

The GBDT model is trained by using the fault information of the protection system of a typical 110 KV intelligent substation in southern China. Figure 4 shows the information topology between the devices of the protection system of intelligent substation. The protection system includes a line merging unit, a line protection device, a Bus protection device, and an intelligent terminal. The GOOSE/SV message-receiving form between devices is shown in

Table 4. Interval information flow in intelligent substations.

Message Number	Transmit Port	Receive Port	Form of Message Transmission
1	Line merging unit	Line protection device	Point-to-point SV Communication
2	Line merging unit	Measurement and control device	Networking SV Communication
3	Line merging unit	Bus protection device	Point-to-point SV Communication
4	Intelligent terminal	Line protection device	Point-to-point GOOSE Communication
5	Intelligent terminal	Line merging unit	Networking GOOSE Communication
6	Bus protection device	Intelligent terminal	Point-to-point GOOSE Communication
7	Bus protection device	Line protection device	Networking GOOSE Communication
8	Intelligent terminal	Measurement and control device	Networking GOOSE Communication
9	Line protection device	Intelligent terminal	Point-to-point GOOSE Communication
10	Measurement and control device	Line merging unit	Networking GOOSE Communication
11	Measurement and control device	Intelligent terminal	Networking GOOSE Communication
12	Intelligent terminal	Bus protection device	Point-to-point GOOSE Communication

.

Select 4200 actual fault samples of this intelligent substation, and the distribution of fault types in the samples is shown in

Table 5. Distribution of fault samples.

Fault Types	Number of Samples	Number of Training Samples	Number of Test Samples
$f_1$	351	316	35
$f_2$	371	334	37
$f_3$	347	312	35
$f_4$	344	310	34
$f_5$.	345	310	35
$f_6$	353	318	35
$f_7$	352	317	35
$f_8$	348	313	35
$f_9$	346	311	35
$f_{10}$	354	319	35
$f_{11}$	340	306	34
$f_{12}$	349	314	35

.

The GBDT fault diagnosis model is trained by using the fault sample set of the protection system, with 75% of the samples as the training set and 25% as the test set. Use the training set to train the model according to the steps in Section 3.1. Taking the diagnostic accuracy of the test sample set as the optimization index, the model is optimized by adjusting the learning rate σ and the number of iterations M, because these two parameters have the greatest impact on the accuracy of the model. The training results are shown in

Table 6. Diagnostic accuracy of GBDT under different parameters.

M		0.05	0.1	0.3	0.5	0.7	1
	σ
10		0.97619	0.98333	0.98095	0.98095	0.97857	0.97381
30		0.97857	0.99048	0.98333	0.98095	0.97857	0.97142
50		0.97619	0.98809	0.98571	0.98333	0.97857	0.97381
70		0.97619	0.98809	0.98809	0.98333	0.98095	0.97619
100		0.97619	0.98571	0.98571	0.98571	0.98095	0.97857

.

It can be seen from Table 6 that the accuracy of the GBDT model for fault diagnosis of the protection system is quite high, and when the number of iterations is 30 and the learning rate is 0.1, the accuracy is the highest, reaching 99.048%. The specific diagnosis results of the test set samples at this time are shown in

Table 7. Diagnostic results of test set at $\mathsf{\sigma }$ = 0.1, M = 30.

	${\mathit{f}}_1$	${\mathit{f}}_2$	${\mathit{f}}_3$	${\mathit{f}}_4$	${\mathit{f}}_5$	${\mathit{f}}_6$	${\mathit{f}}_7$	${\mathit{f}}_8$	${\mathit{f}}_9$	${\mathit{f}}_{10}$	${\mathit{f}}_{11}$	${\mathit{f}}_{12}$
Number of samples	35	37	35	34	35	35	35	35	35	35	34	35
Number of correct diagnoses	35	37	34	33	35	34	34	35	35	35	34	35

.

Compared with the existing research methods, such as recurrent neural network (RNN) [12] and random forest algorithm (RF) [13], the fault diagnosis accuracy under the same dataset is shown in

Table 8. Comparison of diagnostic accuracy of different models.

Model	Diagnostic Accuracy
GBDT	0.9905
[12]	0.9871
[13]	0.9609

:

It can be seen from Table 8 that the GBDT algorithm has the highest accuracy compared with the other two algorithms when dealing with the same dataset due to its excellent performance on small sample sets, and GBDT has fewer iterations in training, faster-running speed, and training process.

To explore the influence of the number of samples in the training set on the accuracy of the model and compare the accuracy of the three methods, according to the distribution ratio of fault samples in Table 5, the number of samples in the training set is changed for training, and the test results are shown in Figure 5

Figure 5 shows that when the number of samples in the training set reaches 3800, the accuracy of the model reaches 99%, and with the increase in the number of samples in the training set, the accuracy of the model does not improve much. Therefore, in practical application, higher accuracy can be achieved when the number of samples in the training set reaches 3800.

3.3. Fault Diagnosis Process of the Protection System

Based on the above content, the fault diagnosis process of the protection system of an intelligent substation based on the gradient boosting decision tree is constructed, as shown in Figure 6.

The specific steps are:

Step 1

To avoid the false start of the diagnosis process, set the minimum number $k_0$ of alarm messages within 30 s after receiving the first alarm message. When the number of alarm messages received by the secondary monitoring system within the specified time is greater than or equal to $k_0$, the fault diagnosis of the protection system of this intelligent substation is triggered.

The intelligence and integration of the secondary system make it produce a lot of alarm messages when a fault occurs. When the maintenance personnel repair the equipment incorrectly or the equipment is disturbed by environmental factors, alarm information is also generated, but the alarm information is single and small in number. In this case, the fault diagnosis of the protection system should not be started. To avoid the false start of diagnosis, according to the actual fault data analysis and field experience of the intelligent substation, within 30 s after the first alarm message appears, whether the number of alarm messages received by the secondary equipment monitoring system is greater than or equal to $k_0$ is the trigger diagnosis condition, and set $k_0$ = 3;

Step 2

If the number of alarm messages is greater than $k_0$, the fault feature information of the protection system of the intelligent substation is extracted to form a set of fault feature information. The feature information in Table 3 is collected, including the integrated alarm information, link information of GOOSE and SV, device self-checking information, and sampling value information from the secondary monitoring system. After data processing, the fault feature information set X$=\left\{A ,I,C,S^{\prime }\right\}$ in Section 2.3 is generated, which prepares the data for the fault diagnosis of the protection system of this intelligent substation;

Step 3

Input the processed fault feature information set X$=\left\{A ,I,C,S^{\prime }\right\}$ into the fault diagnosis system based on GBDT for diagnosis. The specific process is: Input X into the binary classifiers in the GBDT model, respectively, and calculate the probability $P_{f_i}($i = 1, 2, …) that this fault belongs to each complex fault. The one with the highest probability determines that the fault belongs to this type of complex fault and outputs the diagnostic result set R $= \left\{f_1,f_2,\dots \right\}=\left\{1,0,\dots \right\}$ (Suppose the fault is $f_1$).

4. Case Analysis

To verify the effectiveness of the fault diagnosis method based on the gradient boosting decision tree proposed in this paper, some fault cases are selected from the historical fault data of the intelligent substation described in Section 3.2 for analysis. According to the three special cases of complete fault feature information, false alarm, and multiple faults, the method in this paper is used for fault diagnosis. Finally, the effectiveness and general applicability of the method are proved.

4.1. Cases with Complete Fault Feature Information

Taking $f_4$-“I/O module fault of merging unit (Bus merging unit → Line merging unit)” as an example, when the fault occurs, the fault feature information is shown in

Table 9. Fault feature information of $f_4$.

Fault Type	Fault Feature Information
I/O module fault of merging unit (Bus merging unit → Line merging unit)	Abnormal alarm of merging unit
	SV total alarm
	GOOSE total alarm
	SV interruption of merging unit
	GOOSE interruption of merging unit
	SV interruption of protection device
	SV alarm of protection device
	GOOSE alarm of intelligent terminal
	Protection locking

:

Establish the integrated alarm information set $A$ as shown in (14):

$$A=\left\{a_1,a_2,a_3,\dots ,a_{16}\right\}=\left\{1,1,1,0,1,\dots ,0\right\}$$

(14)

Establish the link information of GOOSE and SV set $I$ as shown in (15):

$$I=\left\{I_1,I_2,I_3,\dots ,I_9\right\}=\left\{1,0,0,1,1,\dots ,0\right\}$$

(15)

Establish device self-checking information set $I$ as shown in (16):

$$\left\{\begin{array}{c}C=\left\{C_{MU},C_P,C_{IT}\right\}\\ C_{M{U}_{ZT}}=\left\{0,0,\dots ,0\right\}\\ C_{P_{ZT}}=\left\{0,0,\dots ,0\right\} \\ C_{I{T}_{ZT}}=\left\{0,0,\dots ,0\right\} \end{array}\right.$$

(16)

The voltage and current sampling values are shown in (17):

$$\left\{\begin{array}{c}S=\left\{M_1,M_2\right\}\\ M_1=\left\{I_{A1},I_{B1},I_{C1},U_{A1},U_{B1},U_{C1}\right\}\\ M_2=\left\{I_{A2},I_{B2},I_{C2},U_{A2},U_{B2},U_{C2}\right\}\end{array}\right.=\left\{\begin{array}{c}S=\left\{M_1,M_2\right\}\\ M_1=\left\{\begin{array}{c}3.023,3.028,3.016,\\ 57.488,57.558,57.549\end{array}\right\}\\ M_2=\left\{\begin{array}{c}2.969,2.958,2.894,\\ 57.402,58.218,56.473\end{array}\right\}\end{array}\right.$$

(17)

After preprocessing by the Min-Max method, the sampling data are transformed into (18):

$$\left\{\begin{array}{c}{S}^{\prime }=\left\{{M_1}^{\prime },{M_2}^{\prime }\right\}\\ {M_1}^{\prime }=\left\{\begin{array}{c}0.9781,0.9768,0.9729,\\ 0.9810,0.9822,0.9821\end{array}\right\}\\ {M_2}^{\prime }=\left\{\begin{array}{c}0.9577,0.9542,0.9335,\\ 0.9796,0.9935,0.9637\end{array}\right\}\end{array}\right.$$

(18)

After the fault occurs, a total of 9 alarm messages are received within 30 s after receiving the alarm message of “abnormal alarm of the merging unit “, which is greater than 3, so the fault diagnosis of the protection system of the intelligent substation is triggered. The fault feature set X$=\left\{A ,I,C,S^{\prime }\right\}$ is constructed by calling fault feature information and processing data, as shown in Formulas (14)–(18). Input into the GBDT-based fault diagnosis system for diagnosis, input X into the 12 binary classifiers in the GBDT algorithm, respectively, and calculate the probability P as shown in

Table 10. The output probability of the binary classifier under f_4 fault.

	${\mathit{f}}_1$	${\mathit{f}}_2$	${\mathit{f}}_3$	${\mathit{f}}_4$	${\mathit{f}}_5$	${\mathit{f}}_6$	${\mathit{f}}_7$	${\mathit{f}}_8$	${\mathit{f}}_9$	${\mathit{f}}_{10}$	${\mathit{f}}_{11}$	${\mathit{f}}_{12}$
P	0.02	0.09	0.41	0.96	0.03	0.04	0.11	0.09	0.14	0.14	0.11	0.04

.

It can be seen that this fault is most likely to be $f_4$, and the output diagnostic result set R $=\left\{f_1,f_2,f_3,f_4,\dots \right\}=\left\{0,0,0,1,\dots \right\}$, then it is considered that this fault belongs to the “I/O module fault of merging unit (Bus merging unit → Line merging unit)”, so the diagnosis is correct.

To verify whether the method in this paper has a generally high accuracy under the condition of complete fault feature information, 1125 groups of samples that have not been used were selected from historical fault samples. The fault diagnosis is carried out by this method. The steps are shown in the above, and the distribution of fault samples and diagnosis results are shown in

Table 11. Fault sample distribution and diagnosis results when fault feature information is complete.

	${\mathit{f}}_1$	${\mathit{f}}_2$	${\mathit{f}}_3$	${\mathit{f}}_4$	${\mathit{f}}_5$	${\mathit{f}}_6$	${\mathit{f}}_7$	${\mathit{f}}_8$	${\mathit{f}}_9$	${\mathit{f}}_{10}$	${\mathit{f}}_{11}$	${\mathit{f}}_{12}$
Number of samples	97	89	96	95	97	88	96	96	89	94	95	93
Number of correct diagnoses for GBDT	96	86	96	94	95	87	96	96	89	94	95	93

.

The final accuracy rate is 99.29%, which has high accuracy under the condition of complete fault feature information.

4.2. Cases of False Alarms in Fault Feature Information

The fault feature information of the protection system of intelligent substation may be misreported or missed, especially the integrated alarm information, the link information of GOOSE and SV, and the device self-checking information. The sampling value information is obtained by the dual-channel method, and the reliability is high, with almost no false positives or omissions. To improve the diagnostic accuracy in the case of falsely reported fault feature information, this paper extends the training set by adding an appropriate proportion of noise information in the feature information of the fault sample-set data to enhance the generalization ability of the model and reduce the influence of interference information on the final diagnosis. This process was completed in the training process of Section 3.2. The verification shows that when the proportion of noise in each feature information reaches 5%, the generalization ability of the model can meet the requirements.

To verify the reliability of the method proposed in this paper when the information is falsely reported, randomly select a piece of information from the fault feature information A or C or I of the case in Section 4.1 to invert (the original “1” is set to “0” or the original “0” set to “1”) to simulate the situation that the fault feature information is falsely reported. In this case, the “SV alarm of protection device” is set to “0” to simulate information missed, and the “Sampling anomaly of merging unit” is set to “1” to simulate information misreported. The diagnosis steps are the same as those in Section 4.1, and the output result set R $=\left\{f_1,f_2,f_3,f_4,\dots ,f_{12}\right\}=\left\{0,0,0,1,\dots ,0\right\}$. The fault is diagnosed as the fault of “I/O module fault of merging unit (Bus merging unit → Line merging unit)”, and the diagnosis is correct.

To verify whether the proposed method has a generally high accuracy in the case of false alarms of fault feature information, 120 sets of fault data are selected from the test samples in Section 4.1. These 120 sets of data can obtain the correct diagnosis results through the fault diagnosis of the protection system of an intelligent substation based on GBDT. In each sample of fault feature information, one, two, or three information samples are randomly selected to invert, respectively, to simulate the situation that the fault feature information has one, two, or three false reports, and provides a comparison of the diagnostic results of GBDT with RNN and RF for one false report. According to the diagnosis process in Section 4.1, the 120 faults are re-diagnosed, and the results are shown in

Table 12. Fault sample distribution and diagnosis results when one false alarm occurs.

	${\mathit{f}}_1$	${\mathit{f}}_2$	${\mathit{f}}_3$	${\mathit{f}}_4$	${\mathit{f}}_5$	${\mathit{f}}_6$	${\mathit{f}}_7$	${\mathit{f}}_8$	${\mathit{f}}_9$	${\mathit{f}}_{10}$	${\mathit{f}}_{11}$	${\mathit{f}}_{12}$
Number of samples	10	10	10	10	10	10	10	10	10	10	10	10
Number of correct diagnoses for GBDT	9	10	10	9	10	10	10	10	10	9	10	10
Number of correct diagnoses for RNN	7	8	9	9	9	10	9	9	8	7	9	10
Number of correct diagnoses for RF	5	6	9	9	10	10	9	7	9	6	9	9

,

Table 13. Fault sample distribution and diagnosis results when two false alarms occur.

	${\mathit{f}}_1$	${\mathit{f}}_2$	${\mathit{f}}_3$	${\mathit{f}}_4$	${\mathit{f}}_5$	${\mathit{f}}_6$	${\mathit{f}}_7$	${\mathit{f}}_8$	${\mathit{f}}_9$	${\mathit{f}}_{10}$	${\mathit{f}}_{11}$	${\mathit{f}}_{12}$
Number of samples	10	10	10	10	10	10	10	10	10	10	10	10
Number of correct diagnoses for GBDT	9	8	10	9	9	9	10	10	9	9	10	10

and

Table 14. Fault sample distribution and diagnosis results when three false alarms occur.

	${\mathit{f}}_1$	${\mathit{f}}_2$	${\mathit{f}}_3$	${\mathit{f}}_4$	${\mathit{f}}_5$	${\mathit{f}}_6$	${\mathit{f}}_7$	${\mathit{f}}_8$	${\mathit{f}}_9$	${\mathit{f}}_{10}$	${\mathit{f}}_{11}$	${\mathit{f}}_{12}$
Number of samples	10	10	10	10	10	10	10	10	10	10	10	10
Number of correct diagnoses for GBDT	8	8	9	8	9	7	10	9	9	7	10	10

.

It can be seen from Table 12, Table 13 and Table 14 that when the fault feature information is falsely reported, the GBDT model has a very strong anti-overfitting ability. When there are one and two false alarms in the fault feature information, the diagnostic accuracy is high, which is 97% and 92%, respectively. When the false alarm information reaches three, the diagnostic accuracy is 84%. Therefore, this method has a high accuracy when the number of false alarms is less than or equal to three. With the increasing number of false alarms, the lack of feature information will seriously affect the diagnosis of GBDT.

4.3. Case of Multiple Faults

Due to the high integration of the device of the protection system, the fault of a component in the merging unit, protection device, and intelligent terminal is likely to cause other component faults in the current device. To verify the accuracy of the GBDT diagnostic model under multiple faults, the multiple fault types shown in

Table 15. Type of multiple fault.

Number	Complex Faults Included
$f_{13}$	Main DSP module failure of merging unit Sampling DSP module failure of merging unit
$f_{14}$	$\mathrm{I}/\mathrm{O} \mathrm{module} \mathrm{fault} \mathrm{of} \mathrm{merging} \mathrm{unit} (\mathrm{Merging} \mathrm{unit} \leftrightarrow$ Switch) $\mathrm{I}/\mathrm{O} \mathrm{module} \mathrm{fault} \mathrm{of} \mathrm{merging} \mathrm{unit} (\mathrm{Bus} \mathrm{merging} \mathrm{unit} \to$ Line merging unit)
$f_{15}$	$\mathrm{I}/\mathrm{O} \mathrm{plug}-\mathrm{in} \mathrm{fault} \mathrm{of} \mathrm{protection} \mathrm{device} (\mathrm{Protection} \mathrm{device} \leftrightarrow$ Intelligent terminal) $\mathrm{I}/\mathrm{O} \mathrm{plug}-\mathrm{in} \mathrm{fault} \mathrm{of} \mathrm{protection} \mathrm{device} (\mathrm{Switch} \to$ Protection device) $\mathrm{I}/\mathrm{O} \mathrm{plug}-\mathrm{in} \mathrm{fault} \mathrm{of} \mathrm{protection} \mathrm{device} (\mathrm{Merging} \mathrm{unit} \to$ Protection device)
$f_{16}$	I/O board fault of $\mathrm{intelligent} \mathrm{terminal} (\mathrm{Intelligent} \mathrm{terminal} \leftrightarrow$ Switch) I/O board fault of $\mathrm{intelligent} \mathrm{terminal} (\mathrm{Protection} \mathrm{device} \leftrightarrow$ Intelligent terminal)

are considered.

Four types of fault samples in Table 15 are selected from the sample set for training based on the GBDT model formed in Section 3.2. The sample distribution and diagnosis results are shown in

Table 16. Multiple fault sample distribution and diagnosis results.

Fault Type	Number of Training Samples	Number of Test Samples	Number of Correct Diagnoses for Test Samples
$f_{13}$	75	24	23
$f_{14}$	79	25	24
$f_{15}$	74	24	20
$f_{16}$	80	25	23

.

It can be seen from Table 16 that the diagnostic accuracy of this method for multiple faults is as high as 91.8367%. Considering the low probability of multiple faults in the actual operation environment, this method has a reliable fault diagnosis ability.

Taking $f_{13}$—“Main DSP module failure of merging unit, Sampling DSP module failure of merging unit” as an example, when the fault occurs, the fault feature information is shown in

Table 17. Fault feature information of this case.

Fault Type	Fault Feature Information
Main DSP module failure of merging unitSampling DSP module failure of merging unit	Sampling anomaly of merging unit, Synchronization anomaly of merging unit, SV alarm of merging unit/protection device, Protection locking, etc.

.

The diagnosis steps are the same as those in Section 4.1, and the output result set R $=\left\{f_1,f_2,f_3,f_4,\dots ,f_{13},f_{14},f_{15},f_{16}\right\}=\left\{0,0,0,0,\dots ,1,0,0,0\right\}$. The fault is diagnosed as the fault of “Main DSP module failure of merging unit, Sampling DSP module failure of merging unit”, and the diagnosis is correct.

5. Conclusions

This paper sorts out the common faults of the protection system and proposes simple fault types and corresponding fault feature information and complex fault types and corresponding main fault feature information. The integrated alarm information, link information of GOOSE and SV, device self-checking information, and sampling value information that can be used as fault feature information of the protection system of an intelligent substation are sorted out to form a set of fault feature information. The model parameter adjustment of GBDT is completed according to the fault data. The fault diagnosis model of the protection system of an intelligent substation based on GBDT is studied and verified.

The method proposed in this paper has a high diagnostic accuracy and stronger generalization ability and is more suitable for processing the fault feature data of the protection system of the intelligent substation. The calculation example shows that the overall accuracy of the method proposed in this paper can reach 99.0476%. Compared with the existing methods based on recurrent neural networks and random forest algorithms, the method proposed in this paper has a higher fault diagnosis accuracy. In the case of one false alarm in the fault feature information data, the accuracy rate of the proposed method can reach 97%. In the case of two false alarms in the fault feature information data, the accuracy rate of the proposed method can reach 92%. In the case of three false alarms in the fault feature information data, the accuracy rate of the proposed method can reach 84%. In multiple fault diagnosis, the accuracy of the proposed method is 91.8367%. Through the above analysis, it can be concluded that the method proposed in this paper gives full play to the high accuracy and anti-overfitting ability of the GBDT algorithm when dealing with device faults in the protection system. Compared with the RNN and RF algorithms, this method is more convenient to adjust algorithm parameters in addition to higher accuracy. Compared with the existing methods, this method also performs very well when faced with bad data (false alarms of fault information, multiple faults). In conclusion, the method proposed in this paper can play a better role in practical applications.