An Ancillary Decision-Making Method for Hydropower Station Failure Handling Based on Case-Based Reasoning and Knowledge Graph

Abstract

This paper proposes an ancillary decision-making method for hydropower station failure handling based on knowledge graph and case-based reasoning. The proposed method assists the power station dispatcher to carry out accurate and timely failure handling after an accident. First, the main steps of case-based reasoning are introduced. The main difficulties and their corresponding solutions when applying case-based reasoning to hydropower station failure handling are discussed. Then, an ancillary decision-making method for hydropower station failure handling is proposed. Key steps such as case construction, case retrieval, and case revision are designed. In the proposed method, each case is represented in the form of multiple knowledge graphs, i.e., a system topology knowledge graph, a dispatching regulation knowledge graph, and an accident case knowledge graph. The flexibility of case knowledge extraction, management, and retrieval is greatly enhanced. Finally, the simulation analysis is carried out on a large-scale cascade hydropower station in China. The simulation results show that the proposed method can provide reasonable and reliable ancillary decision-making for the power station dispatcher in the failure handling process, and greatly improve the intelligence level of emergency management at a hydropower station.

1. Introduction

In order to achieve the “carbon peak by 2030, carbon neutralize by 2060” goal, the urgency of constructing a new power system is greatly increasing. In the new power system, which has characteristics of clean and low-carbon, renewable energy is the main resource form on the supply side [1]. Due to the intermittent and fluctuating nature of renewable energy, the demand for power resources with flexible adjustment ability is increasing after a large amount of renewable energy integration into the new power system. Hydropower, as a clean and sustainable energy resource with excellent regulation performance, is playing an increasingly important role in the flexible and efficient new power system [2,3].

With the construction of large-capacity hydropower stations, such as cascade hydropower stations, failure handling for hydropower stations is becoming more and more important. After an accident occurs, if not handled properly, the outage of large-capacity hydropower generators may lead to a huge power shortage in the power system, which may cause severe consequences such as low frequency, cascading failures, and even the destruction of system stability. To avoid this, accurate and timely failure handling can help to block the cascading failure chain, reduce the affected scope of the accident, and guarantee the security and reliability of the power system. Currently, in the process of accident handling, power station dispatchers mainly make decisions by considering the system situation, dispatching regulations, and their own experience. The automation system can only aid in basic functions such as device status monitoring, alert signal reporting, and data querying. The whole process of accident handling is entirely dependent on manual decision-making [4,5]. As the complexity of the power system and the loss for power outage are both increasing, a more intelligent ancillary decision-making tool is urgently needed to carry out accurate and rapid failure handling, liberate the dispatchers from redundant information, and provide them with decision-making support [6,7,8].

Case-based reasoning (CBR) is a knowledge management method that learns from historical cases and uses them to solve new problems. CBR is a type of analogical reasoning technique [9]. As a method of mining and reuse, CBR is very similar to the human cognitive learning process. When CBR is used in decision-making, compared with other traditional methods such as rule-based reasoning (RBR) and expert systems (ESs), its advantages include the following [10,11,12,13,14]:

• Wide applicability. For many complex problems, it is difficult to summarize the decision-making basis into clear and generalized rules, which makes it difficult to construct RBR and ES systems. But there are often many historical cases addressing how to solve the problems. Knowledge can be extracted from these cases for CBR decision-making;
• Abundant sources of learning. Many domain experts often struggle to summarize their experiences into words, which makes the application of RBR difficult, and the experts also lack the time and patience to help to construct an ES. However, the experts are usually happy to tell “stories” about past decision-making experiences, which can be used as the learning materials for CBR;
• Suitable for machine learning. The cases in the CBR database are stored independently in the same knowledge management mode. When a new case is to be added, it can be directly incorporated into the CBR case database by machine learning. The rules of RBR and ES, on the other hand, are stored together in the database with many logical relations existing between them. When a new rule is to be added, contradictions may occur between new and old knowledge in the update process, which usually requires human interface, and makes machine learning difficult;
• Easy to understand. When CBR is used in decision-making, in addition to providing assistance in decision-making strategies, it can also provide related historical cases for reference. Black-box models can only provide failure handling strategies, and decision-makers are prone to question the basis of their decisions.

From the above analysis, it can be concluded that CBR is suitable for decision-making in fields with abundant knowledge integration, and are experiencing drastic changes, which is precisely the case for the failure handling of hydropower stations discussed in this paper. In previous studies, CBR has been applied to multiple scenarios such as situational awareness, fault diagnosis, and optimal scheduling [15,16,17,18,19]. In the field of ancillary decision-making for failure handling, a decision-making model with a self-learning mechanism based on CBR is proposed in [20], which realizes a rapid retrieval and intelligent matching of emergency plans under an emergent event. In [21], a cognitive framework for formulating power grid restoration strategies is proposed. Combined with CBR and heuristic algorithms, the framework achieves the rapid formulation of large-scale power grid restoration strategies. In [22], a CBR-based model is designed to effectively identify the alert signals and push them accurately after a failure happens. In [23], a six-fold ontology model is used to retain the accident cases of steam turbine generators. A knowledge retrieval algorithm is designed to help the maintenance personnel with fault diagnosis in steam turbine generators. In [24], combining CBR with rule-based reasoning, a method to obtain a traction transformer maintenance plan is proposed. In the method, CBR is used to retain accident cases, and rule-based reasoning is used to obtain key parameters reflecting the status of the traction transformer.

The basic data of CBR are taken in this case. Knowledge expression and a management method for the case are crucial for CBR. In recent years, the knowledge graph (KG) has received extensive attention as an intelligent and flexible knowledge management tool in the field of machine learning, and it has been applied to various engineering problem-solving scopes such as decision-making, system performance enhancing, and knowledge exploration [25,26,27,28,29]. Some previous studies combined KG with CBR and used KG to manage the knowledge in CBR cases [30,31,32]. In [33], based on the topology entity KG and the failure handling plan KG, an ancillary decision-making strategy was designed to generate the failure handling scheme of the distribution network automatically. In [34], a two-layer KG for power grid accident handling based on the deep learning method was constructed. In the two-layer KG, the upper layer stores the accident handling rules learned from historical cases, and the lower layer stores the specific network topology and operation information of the historical cases. From the above literature review, we see that the research on the combination of KG and CBR is still in the preliminary stage, and its application scope has not yet been extended to the failure handling of hydropower stations, which is the scope of this paper. By combining KG and CBR, this paper proposes an ancillary decision-making method for hydropower station failure handling. The objectives of the proposed method are as follows:

• KG and CBR are both commonly used knowledge management techniques. By combining KG and CBR, the proposed method is fit for managing the rapidly growing and frequently updated knowledge required for failure handling;
• Traditional CBR is usually constructed via manual work. By deeply integrating CBR and KG, the proposed method is fit for machine learning, which greatly increases the case construction and retrieval efficiency of CBR, and improves the performance of decision-making;
• By taking advantages of both CBR and KG, the proposed method can provide reasonable and reliable ancillary decision-making strategies. The strategies are provided in the form of KGs and are clear enough to assist the power station dispatcher to carry out reasonable and reliable accident handling.

The contributions of this paper are as follows:

• In this paper, KG is used as a knowledge management technique for CBR. By fully utilizing the characteristics of KG and CBR, an ancillary decision-making method is proposed;
• In the proposed method, each case is represented in the form of multiple KGs, i.e., the system topology KG, the dispatching regulation KG, and the accident case KG. The key steps, i.e., case construction, case retrieval and case revision, are designed;
• The proposed method is verified with the practical data of a large-scale cascade hydropower station in China. Simulation results show that the method can assist the dispatcher to carry out reasonable and reliable failure handling.

The rest of the paper is organized as follows. First, the main steps of CBR are introduced, and the difficulties and their corresponding solutions when applying CBR to the failure handling of hydropower stations are analyzed. Then, the ancillary decision-making method is proposed. The algorithms and processes of key steps, such as case construction, case retrieval and case revision, are designed. Finally, the simulation analysis is carried out to verify the practicability of the proposed method.

2. The Main Steps, Application Difficulties and Solutions of CBR

2.1. The Main Steps of CBR

In the classical 4R cognitive model, the main steps of CBR include case retrieve, case reuse, case revise, and case retain, as shown in Figure 1 [10].

The main steps of CBR in Figure 1 include:

• Case retrieve—retrieve one or more cases that are most similar to the new problem from the case database;
• Case reuse—reuse solutions in similar cases;
• Case revise—according to the specific situation of the new problem, make necessary revisions to the solutions in similar cases to better meet the current situation;
• Case retain—build a new case with the new problem and the revised solution, and save it in the case database.

After combining CBR with machine learning methods, the basic step of the 4R model is to design a data structure to manage the knowledge in the case in order to realize automatic knowledge extraction and efficient knowledge retrieval. This paper uses KG as the knowledge management tool and focuses on the case construction and retrieval method that intricately combines KG and CBR.

2.2. Difficulties in the Application of CBR

When applying CBR to the ancillary decision-making of hydropower station failure handling, the difficulties are as follows:

• Complex and changeable knowledge. In the new power system, the variety of knowledge required for failure handling is complex and the update frequency is increasing, and the reasons for this are as follows. (1) The dynamic characteristics of the power system are becoming more and more complicated, leading to multiple constraints on power supply and hydraulic regulation. The dispatching and control of hydropower stations are becoming more and more complex. (2) The measurement and communication devices are upgraded, and the resolution, accuracy, and transmission bandwidth of the operation data of hydropower stations are increasing. (3) The electricity market is growing all around China, and the operation and management of hydropower stations must consider their market performance. Due to the complexity and changeability of knowledge, CBR needs a more intelligent and flexible knowledge management tool to extract and retrieve knowledge;
• Few available cases. Since accidents generally cause large losses, hydropower stations and the power grid take a variety of measures to ensure safety, such as following N-1 rules, obeying operating regulations, setting up safety boundaries, and conducting strict training for dispatchers. These measures ensure the power system operates within a relatively conservative safety boundary, and greatly reduce the probability of serious accidents. At the same time, despite the large number of devices in hydropower stations, the failure rate for a single device is not high. This may lead to a situation whereby a certain number of historical cases are accumulated, but the cases are still scarce when it comes to specific device failure. The small number of available cases may lead to a small number of cases for CBR retrieval, making it difficult to find similar cases;
• Limited guidance. CBR can provide historical cases that match the current accident to increase the persuasiveness of the failure handling strategy. However, it still lacks necessary explanations, and to a certain extent, the dispatcher may still “only know what to do but not know why”. At the same time, the ancillary decision-making effect of CBR depends entirely on whether historical cases that match the current situation can be found. It is difficult for CBR to give helpful suggestions when similar cases are hard to find.

2.3. Ideas for Solution

In view of the difficulties when applying CBR to the failure handling of hydropower stations, the following ideas are proposed:

• Using KG to manage knowledge. In KG, entities are connected to each other to form a network structure. KG can achieve the effective organization, storage, and querying of multiple heterogeneous knowledge types, and adapt to the needs of frequent knowledge updates;
• Expanding the scope of case matching. This involves, on one hand, expanding the scope of case mining, using the historical accidents, failure handling plans, and emergency drill plans of hydropower stations as the source data for building a case database. On the other hand, the cases are generalized based on the system topology to increase the versatility of case retrieval, i.e., for cases that are not of the same faulty device, but if their topology is similar, these cases can be generalized into one category;
• Integrating the dispatching regulations into cases. When constructing cases, in addition to historical accidents, the dispatching regulations of the hydropower station are also used to build KG. In case retrieval, both the accident case KG and the dispatching regulation KG are retrieved, in order to provide the failure handling strategy as well as related dispatching principles. Through mutual corroboration, the persuasiveness of the assisted decision-making scheme is increased.

3. An Ancillary Decision-Making Method for Hydropower Station Failure Handling

In this section, the ancillary decision-making method for hydropower station failure handling is proposed.

3.1. Framework of the Decision-Making Method

The framework of the proposed ancillary decision-making method for hydropower station failure handling based on CBR and KG is shown in Figure 2. The main steps of the proposed method in Figure 2 include the following:

• Case construction. Each CBR case is composed of three KGs, i.e., system topology KG, dispatching regulation KG, and accident case KG. To construct a case is to construct the three KGs by use of machine learning;
• Case retrieval. After a failure occurs, retrieve one or more historical cases that are most similar to the new one from the case database. In the proposed method, case retrieval includes accident case retrieval and dispatching regulation retrieval, which uses all three KGs to find a similar case;
• Case revision. Modify the KGs in the similar case to make them closer to the current situation of the hydropower station after the accident;
• Strategy output. The failure handling strategy is composed of the revised accident case KG and dispatching regulation KG, and they can assist the power station dispatcher to carry out reasonable and reliable accident handling;
• Case retention. Construct a new case with the new problem and its failure handling strategy, and save it in the CBR case database.

Figure 2. Framework of the ancillary decision-making method for hydropower station failure handling based on CBR and KG.

Figure 2. Framework of the ancillary decision-making method for hydropower station failure handling based on CBR and KG.

3.2. Case Construction

The key to case construction is to design a unified and proper model to represent the case, which is easy for machine learning and reasoning. A top-down construction method is used to construct each KG, i.e., system topology KG, dispatching regulation KG, and accident case KG, separately.

3.2.1. System Topology KG Construction

The learning source of system topology KG is the topology of the hydropower station. The system topology is stored as structured data in the automation system. The KG data layer can be directly constructed by extracting key information from the automation system after the scheme layer is designed. The ontology types of the scheme layer are device types such as generators, transformers, breakers, busbars, lines, and other device types. The relation between ontologies is the connect relation. To construct the system topology KG, we must first extract the entities from the topology data according to their device types and names, then add connect relations between entities if the two devices are connected in topology. Figure 3 shows an example of constructing a system topology KG.

3.2.2. Dispatching Regulation KG Construction

The learning sources of dispatching regulation KG are the dispatching regulations in the dispatching handbooks of the hydropower station. The dispatching regulations are stored in the form of unstructured text. To construct the dispatching regulation KG, it is necessary to use the deep learning method to extract entities from the text based on scheme layer designing. The ontology types of the dispatching regulation KG include four categories, i.e., device, event, action and status. The interconnected relation types between ontologies also include four types, i.e., happen, decide, cause, and adjacent, as shown in

Table 1. Relation categories of the schema layer of dispatching regulation KG.

Relation Type	Explanation	Example
Happen	Indicating that the event occurs on the device or after a pre-ordered event.	Device → Event: a generator → trips abnormally Event → Event: the top cover (of a generator) is leaking → water bearing is about to be flooded
Decide	Indicating that decisions on whether to perform actions are based on device attributes, event attributes, pre-order operations, or system status.	Device → Action: (if the faulty device is) the power grid dispatching device → execute the command of the power grid dispatchers Event → Action: (The generator) emits a continuous abnormal sound → shutdown the generator Operation → Action: (after a system accident occurs) ensure the power supply of hydropower station → carry out failure handling Status → Action: If the fault cannot be eliminated immediately → the faulty device should be isolated
Cause	Indicating that the occurred event or a performed action cause the system to enter a certain status.	Event → Status: (if the generator is) asynchronously oscillating → system frequency increases Action → Status: Do troubleshoot → (but) the fault is not eliminated
Adjacent	Indicating that the ontology is adjacent in the dispatching regulation text, but there is no direct logical relation.	/

.

After defining the scheme layer, the deep learning model is used to extract the entities from the dispatching regulation text of the hydropower station to construct the data layer. In this paper, a BiLSTM+CRF model with word vectors embedded is used as the deep learning model, and its framework is shown in Figure 4. In the training process, the annotated training text is divided into a training set and a verification set, and is used to train the deep learning model. After the model is trained with the training data, the dispatching regulation text is input into the model to obtain the annotated data, which are used to construct the dispatching regulation KG. Figure 5 shows an example of dispatching regulation KG construction.

3.2.3. Accident Case KG Construction

The accident case KG is the core of CBR, which is used to store the knowledge extracted from historical accidents, failure handling plans, and emergency drill plans. The learning sources of an accident case KG are the historical accident records and failure handling plans mentioned above. Since such cases are often stored in the form of semi-structured text, extraction rules are designed under the guidance of the schema layer to extract key information from the cases. The extracted knowledge is used to form the data layer. Considering the characteristics of accident cases, the ontology types of the scheme layer basically cover the ontology types of the system topology KG and dispatching regulation KG, with some minor adjustment. The relation type set adds two relation types based on the dispatching regulation KG, as shown in

Table 2. Ontology categories of the schema layer of accident case KG.

Source	Ontology Type	Explanation
System topology KG	Generators, transformers, breakers, busbars, lines, and other device types	The same.
Dispatching regulation KG	Device, action, status, fault signals (events)	After an accident has happened, the occurred event can only be judged based on the fault signals, so event is replaced by the fault signals ontology.
New	Time	Indicating the specific time when a fault signal is reported or an action is taken.
New	Fault reason	Indicating the cause of the accident.

and

Table 3. Relation categories of the schema layer of accident case KG.

Source	Relation Type	Explanation	Example
Dispatching regulation KG	Happen, Decide, Cause	As shown in Table 1	As shown in Table 1
New	Report	Indicating the fault signal is reported at a certain point in time.	Time → Device: (at) 22:50:31 → line A (report fault signals)
New	Next	Indicating the connection between time instants.	Time → Time: (at) 22:50:31 (a failure occurs) → (at) 22:50:35 (the failure continues, or an action is taken)

.

In the accident case text of the hydropower station, the main content usually includes the system status before the accident, the fault process, the failure handling actions, the fault cause and analysis, and review and suggestions. Among them, the fault process and handling part is usually written with the format of “time + alert signals/actions”, which can be used to directly extract key information from the case under the guidance of the scheme layer and build the data layer. To construct the accident case KG, we must first extract the entities from the text according to their types and names, then add relations between entities according to the types of the starting and terminal entities. Figure 6 shows an example of building an accident case KG.

3.3. Case Retrieval

The key to CBR case retrieval is to design a similarity function to measure the similarity between cases, and to find the historical case that is most similar to the current case from the case database. In the proposed ancillary decision-making method, case retrieval includes accident case retrieval and dispatching regulation retrieval.

3.3.1. Accident Case Retrieval

The accident case retrieval mainly measures the similarity of the fault signal parts of two accident case KGs, as shown in Figure 7.

The accident case KG shown in Figure 7 can be represented as:

$${\mathit{c}}_i=\left\{{\mathit{t}}_i,{\mathit{d}}_i,{\mathit{s}}_i,{\mathit{R}}_i^{td},{\mathit{R}}_i^{ds}\right\}$$

(1)

where c_i denotes case i in the form of a KG; t_i, d_i, s_i, represent the set of time, device and fault signal entities in the KG, respectively; ${\mathit{R}}_i^{td}$ denotes the set of relations between time and device entities, and ${\mathit{R}}_i^{ds}$ denotes the set of relations between device and fault signal entities, whereby if the two entities are connected, the value of the set element is 1, and otherwise it is 0.

The similarity function between two accident cases, i.e., c_iandc_j, is defined as

$$\mathit{Sim}\left({\mathit{c}}_i,{\mathit{c}}_j\right)={\displaystyle \frac{1}{N_t}}\sum _{n=1}^{N_t}\mathit{Sim}\left({\mathit{t}}_i\left(n\right),{\mathit{t}}_j\left(n\right)\right)\cdot \mathit{Sim}\left(\begin{array}{c}\left\{{\mathit{d}}_i,{\mathit{s}}_i,{\mathit{R}}_i^{ds}|{\mathit{R}}_i^{td}\left({\mathit{t}}_i\left(n\right)\right)=1\right\},\\ \left\{{\mathit{d}}_j,{\mathit{s}}_j,{\mathit{R}}_j^{ds}|{\mathit{R}}_j^{td}\left({\mathit{t}}_j\left(n\right)\right)=1\right\}\end{array}\right)$$

(2)

where $\mathit{Sim}\left(x_1,x_2\right)$ is the similarity function, which is used to represent the similarity between two elements/graphs, $\mathit{Sim}\left(x_1,x_2\right)\in \left[\mathrm{0,1}\right]$, and a larger value means the similarity is higher, while 1 means it is exactly the same, and 0 means it is completely different; $N_t$ is the number of time entities in the accident case KG; ${\mathit{t}}_i\left(n\right)$ represents the nth time entity.

In Equation (2), case similarity is defined as the average value of the product of the time similarity and section similarity at each time instant. The time similarity function is defined as

$$\mathit{Sim}\left({\mathit{t}}_i\left(n\right),{\mathit{t}}_j\left(n\right)\right)=\left\{\begin{array}{c}1,{\displaystyle \frac{\left|{\mathit{t}}_i\left(n\right)-{\mathit{t}}_i(n-1)\right|}{\left|{\mathit{t}}_j\left(n\right)-{\mathit{t}}_j(n-1)\right|}}\le {\alpha }_t\\ 0,{\displaystyle \frac{\left|{\mathit{t}}_i\left(n\right)-{\mathit{t}}_i(n-1)\right|}{\left|{\mathit{t}}_j\left(n\right)-{\mathit{t}}_j(n-1)\right|}}>{\alpha }_t\end{array}\right.$$

(3)

where ${\alpha }_t$ is the threshold value.

When the fault signals occur at multiple time instants in a row, Equation (3) limits the time interval. Only when the time interval of the fault signals reported in the two cases is relatively close are the two cases considered as the same type of fault.

The similarity function of the section at each time instant in Equation (2) is defined as

$$\begin{array}{c}\mathit{Sim}\left(\begin{array}{c}\left\{{\mathit{d}}_i,{\mathit{s}}_i,{\mathit{R}}_i^{ds}|{\mathit{R}}_i^{td}\left({\mathit{t}}_i\left(n\right)\right)=1\right\},\\ \left\{{\mathit{d}}_j,{\mathit{s}}_j,{\mathit{R}}_j^{ds}|{\mathit{R}}_j^{td}\left({\mathit{t}}_j\left(n\right)\right)=1\right\}\end{array}\right)\\ \hspace{1em}\hspace{1em}\hspace{1em}=\mathit{max}{\displaystyle \frac{1}{N_d}}{\displaystyle \sum _{m=1}^{N_d}}\mathit{Sim}\left({\mathit{d}}_i\left(m\right),{\mathit{d}}_j\left(m\right)\right)\cdot \mathit{Sim}\left(\begin{array}{c}\left\{{\mathit{s}}_i|{\mathit{R}}_i^{ds}\left({\mathit{d}}_i\left(m\right)\right)=1\right\},\\ \left\{{\mathit{s}}_j|{\mathit{R}}_j^{ds}\left({\mathit{d}}_j\left(m\right)\right)=1\right\}\end{array}\right)\end{array}$$

(4)

where $N_d$ is the number of device entities in the graph at ${\mathit{t}}_i\left(n\right)$; ${\mathit{d}}_i\left(m\right)$ represents the mth device entity.

In Equation (4), the section similarity at each time is defined as the maximum value after averaging the product of the similarity of each device and the similarity of the fault signal. The similarity function of each device is composed of the similarity function of the device type and the device location,

$$\mathit{Sim}\left({\mathit{d}}_i\left(m\right),{\mathit{d}}_j\left(m\right)\right)=P\left(\mathit{Sort}\left({\mathit{d}}_i\left(m\right)\right),\mathit{Sort}\left({\mathit{d}}_j\left(m\right)\right)\right)\cdot P\left(Loc\left({\mathit{d}}_i\left(m\right)\right),Loc\left({\mathit{d}}_j\left(m\right)\right)\right)$$

(5)

where $\mathit{Sort}\left({\mathit{d}}_i\left(m\right)\right)$ is the device type corresponding to the device entity in the graph; $Loc\left({\mathit{d}}_i\left(m\right)\right)$ is the distance between the device and the central device in the system topology KG; $P\left(x_1,x_2\right)$ is a function used to determine whether two elements are equal, i.e.,

$$P\left(x_1,x_2\right)=\left\{\begin{array}{c}1,x_1=x_2\\ 0,x_1\ne x_2\end{array}\right.$$

(6)

To increase the universality of the case, the names of the two devices are not simply compared in Equation (5), but the characteristics of the two devices in the system topology KG, i.e., the device types and positions, are compared. In cases where the topologies of the faulty devices are very similar, although the faulty devices are not the same, a higher similarity function value can also be obtained.

The similarity function of the fault signals in Equation (4) is defined as

$$\mathit{Sim}\left(\begin{array}{c}\left\{{\mathit{s}}_i|{\mathit{R}}_i^{ds}\left({\mathit{d}}_i\left(m\right)\right)=1\right\},\\ \left\{{\mathit{s}}_j|{\mathit{R}}_j^{ds}\left({\mathit{d}}_j\left(m\right)\right)=1\right\}\end{array}\right)={\displaystyle \frac{1}{N_s}}\sum _{l=1}^{N_s}L\left({\mathit{s}}_i\left(l\right),{\mathit{s}}_j\left(l\right)\right)$$

(7)

where $N_s$ is the number of fault signals that occur on device ${\mathit{d}}_i\left(m\right)$ at time ${\mathit{t}}_i\left(n\right)$; ${\mathit{s}}_i\left(l\right)$ represents the lth fault signal entity; $L\left(s_1,s_2\right)$ denotes the Jaccard similarity coefficient for the two elements/sets, i.e.,

$$L\left(s_1,s_2\right)={\displaystyle \frac{\left|s_1\cap s_2\right|}{\left|s_1\cup s_2\right|}}$$

(8)

3.3.2. Dispatching Regulation Retrieval

The dispatching regulation retrieval is carried out to retrieve the dispatching principles related to the current fault from the dispatching regulation KG, and to provide guidance for decision-making. The dispatching regulation retrieval mainly measures the similarity between the failure reason of the new accident and the dispatching regulation KGs. Specifically, the similarity of dispatching regulation is defined as the Jaccard similarity coefficient between the fault reason entity in the accident case KG and the dispatching regulation KG, i.e.,

$$\mathit{Sim}({\mathit{r}}_i,{\mathit{F}}_j)=L({\mathit{r}}_i,{\mathit{F}}_j)$$

(9)

where ${\mathit{r}}_i$ is the set of fault reason entities in the accident case KG i; ${\mathit{F}}_j$ is the set of event entities in the dispatching regulation KG j.

3.4. Case Revision

After obtaining a similar case, the purpose of CBR case revision is to modify the solutions in the similar case to make them closer to the current situation of the hydropower station after the accident, and to increase the practicability of the ancillary decision-making method. In the proposed method, the case revision is performed to revise the accident case KG, i.e., the devices in the similar case are replaced with current faulty devices according to the matching result in the similarity calculation (Equation (4)).

4. Case Study

4.1. Testing Environment and Parameters

Table 4. Testing environment of the case study.

Testing Environment	Configuration/Version
Python Operating Platform	BML Codelab
CPU	2 Cores
GPU	Tesla V100 16 GB
Memory	16 GB
Python	3.9.16
PaddlePaddle	2.4.1
Neo4j	5.17.0

lists the parameters of the testing environment of the case study.

When constructing the dispatching regulation KG, the BiGRU+CRF deep learning model based on pre-trained word vectors is used for entity extraction. The training parameters are shown in

Table 5. Training parameters of the deep-learning model.

Parameters	Value
Maximum sequence length	300
Embedding word vector dimensions	635,968
Number of neurons	300
Learning rate	0.0008
Training rounds	3
Optimizer	Adam

.

4.2. Case Contruction

In the case study, the proposed method is applied to a large cascade hydropower station in China. In this section, the system topology KG, accident case KG and dispatching regulation KG are constructed.

4.2.1. System Topology KG

The 500 kV hydropower station adopts a 4/3 wiring mode, with a total of nine generators and four outgoing lines. By extracting the information from the system topology with the method proposed in Section 3.2.1, the system topology KG of the hydropower station is constructed, as shown in Figure 8. The number of entities of each type in the KG is shown in

Table 6. Number of entities in the system topology KG of the hydropower station.

Entity Type	Entity Numbers	Entity Type	Entity Numbers
Generator	9	Line	4
Busbar	2	Transformer	9
Breaker	27
Total	51

.

4.2.2. Accident Case KG

Based on the accident events handling records and emergency drill plans of the hydropower station, the accident case KG is constructed with the method proposed in Section 3.2.3. Figure 9 shows the accident case KG of a line tripping event. The numbers of entities and relations of each type are shown in

Table 7. Number of entities in the accident case KG of the hydropower station.

Entity Type	Entity Numbers	Entity Type	Entity Numbers
Time	3	Line	3
Fault Signal	14	Action	3
Breaker	3	Status	3
Fault Reason	1
Total	30

and

Table 8. Number of relations in the accident case KG of the hydropower station.

Relation Type	Relation Numbers	Relation Type	Relation Numbers
Report	6	Happen	14
Next	2	Decide	3
Cause	3
Total	28

, respectively.

4.2.3. Dispatching Regulation KG

The method proposed in Section 3.2.2 is used to construct the dispatching regulation KG. There are 91 principles in the dataset of the hydropower station dispatching regulation book. The training, verification and test sets are allocated according to a ratio of 8:1:1. A BiGRU+CRF deep learning model with pre-defined word vector is used for entity extraction. After training, the precision, recall and F1 of the trained model are 0.935, 0.935, and 0.935, respectively. It is proven that the deep learning model can achieve an accurate entity extraction for the dispatching regulation of hydropower stations. The number of entities and relations of each type of the constructed dispatching regulation KG is shown in

Table 9. Number of entities in the dispatching regulation KG of the hydropower station.

Entity Type	Entity Numbers	Entity Type	Entity Numbers
Device	18	Action	160
Event	43	Status	19
Total	240

and

Table 10. Number of relations in the dispatching regulation KG of the hydropower station.

Relation Type	Relation Numbers	Relation Type	Relation Numbers
Happen	40	Cause	19
Decide	187	Adjacent	2
Total	248

, respectively.

4.3. CBR Application Examples

To analyze the application scenario of the proposed ancillary decision-making method, an example is demonstrated in this section. It is assumed that an accident occurs on the outgoing GX-A Line of the hydropower station. The fault signals after the accident are shown in

Table 11. The failure signals after GX-A Line failure of the hydropower station.

Time	Fault Signals
14:21:59	“5253 breaker relay phase A tripping”, “5252 breaker relay phase A tripping”, “5252, 5253 breaker relay three-phase inconsistent tripping”, “GX-A Line relay 1 phase A tripping, GX-A Line relay 2 phase A tripping”.
14:22:00	“5253 breaker relay reclosing”, “5253 breaker relay comprehensive fault”, “5253 breaker relay phase A, B, C tripping”, “5253 breaker relay phase A, B, C communication tripping”, “GX-A Line relay 1 phase A, B, C tripping, GX-A Line relay 2 phase A, B, C tripping”.
14:22:15	“GX-A Line relay 1 operating normally”, “GX-A Line relay 2 operating normally”.

.

According to the fault signals shown in Table 11, the accident case KG is constructed with the method proposed in Section 3.2.3, as shown in Figure 10. By retrieving the accident case KG shown in Figure 10 in the case database, and calculating the values of the similarity function (Equation (2)) between the new case and historical cases, it can be found that the most similar case is a historical case of an XH-C Line tripping accident that happened in 2013, with a similarity function value of 0.954. The comparison of the two cases is shown in

Table 12. The failure signals of GX-A Line failure and its similar case (HX-C Line failure).

Time	Fault Signals	Time (Similar Case)	Fault Signals (Similar Case)
14:21:59	“5253 breaker relay phase A tripping”, “5252 breaker relay phase A tripping”, “5252, 5253 breaker relay three-phase inconsistent tripping”, “GX-A Line relay 1 phase A tripping, GX-A Line relay 2 phase A tripping”.	22:56:38	“5243 breaker relay phase B tripping”, “5242 breaker relay phase B tripping”, “5242, 5243 breaker relay three-phase inconsistent tripping”, “XH-C Line relay 1 phase B tripping, XH-C Line relay 2 phase B tripping”.
14:22:00	“5253 breaker relay reclosing”, “5253 breaker relay comprehensive fault”, “5253 breaker relay phase A, B, C tripping”, “5253 breaker relay phase A, B, C communication tripping”, “GX-A Line relay 1 phase A, B, C tripping, GX-A Line relay 2 phase A, B, C tripping”.	22:56:39	“5242 breaker relay reclosing”, “5242 breaker relay comprehensive fault”, “5242 breaker relay phase A, B, C tripping”, “5242 breaker relay phase A, B, C communication tripping”, “XH-C Line relay 1 phase A, B, C tripping, XH-C Line relay 2 phase A, B, C tripping”.
14:22:15	“GX-A Line relay 1 operating normally”, “GX-A Line relay 2 operating normally”.	22:56:53	“XH-C Line relay 1 operating normally”, “XH-C Line relay 2 operating normally”.

. Figure 11 shows the topology of the two failure lines.

It can be seen in Table 12 that the current accident case and its similar case are both line tripping accidents, where the failure happens on a single phase, and then three phases are tripped after an unsuccessful reclosure. The fault signals of the two cases are basically the same, except for the faulty devices and phase. The GX-A Line locates on an incomplete string of 4/3 wiring in the hydropower station, and the breakers on both sides are 5252 and 5253, respectively. 5253 is the bus-side breaker (reclosing switch), and 5252 is a middle breaker. In the similar case, the XH-C Line is located on a complete string, and the breakers on both sides are 5242 and 5243, respectively. Both breakers are middle switches, and 5242 is the reclosing breaker, as shown in Figure 11. The simulation results show that the proposed case retrieval method can retrieve the most similar case to the current accident from the historical case database, and has a good, generalized recognition ability for cases with similar topologies.

By revising and reusing the accident handling strategy in the similar case KG with the method proposed in Section 3.4, the accident handling strategy for the current accident is obtained, as shown in Figure 12.

In Figure 12, for an outgoing line tripping accident, the power station dispatcher should first report to the power grid dispatcher. Then, he should notify the personnel on duty at the station to check the relay acting information. After careful inspection, the personnel on duty at the station report the relay acting information and confirm that there is no abnormality in the primary and secondary devices in the station, and the conditions for forced power transmission are met. Then, the power station dispatcher should report to the power grid dispatcher and cooperate with them to carry out the forced power transmission of the faulty line, i.e., the GX-A Line. After the breakers on both sides are closed, the GX-A Line comes back online. The accident handling strategy is clear and accurate, and effectively deals with the current accident.

According to the fault reason set out in the similar case, the dispatching regulation KG is retrieved, and the dispatching principles that are most closely related to the current accident are shown in Figure 13.

The two dispatching principles in Figure 13 perfectly explain the accident handling strategy shown in Figure 12. On one hand, after the line trips, it is necessary to report to the power grid dispatcher, and to perform failure handling actions according to the dispatching command. On the other hand, if the power grid dispatcher judges that the conditions for forced power transmission are met, the opposite-side breaker should close first, and then the local side breaker. The dispatching principles can help the power station dispatcher to better understand the reason behind the failure handling strategy and increase the credibility of ancillary decision-making.

It can be seen that the proposed method can perform the automatic retrieval and matching of the system topology KG, dispatching regulation KG, and accident case KG, and generate a feasible strategy for failure handling as well as providing related dispatching principles as guidelines. In practical circumstances, the KG entities can associate with the multi-dimensional data obtained from SCADA, OCS, OMS and other automation systems. Ancillary decision-making can be realized through automatic search and matching within only milliseconds, and the retrieval accuracy can be greatly enhanced. In contrast, the manual retrieval of the expert system is not only slow (generally minutes or longer) but it is also less accurate (which is closely related to the experience of the dispatcher). The above simulation results show that the proposed method can provide reasonable and reliable ancillary decision-making for the station dispatcher during the failure-handling process, and greatly improve the intelligence and automation level of the emergency management of hydropower stations.

5. Conclusions

CBR is a knowledge management method that learns from existing cases and uses them to solve new problems. Based on CBR and KG, this paper proposes an ancillary decision-making method for hydropower station failure handling. The conclusions are as follows:

• In view of the main difficulties met in applying CBR to the failure handling of hydropower stations, KG is used as a knowledge management technique for CBR. By fully utilizing the characteristics of KG, such as its ability to store interrelated heterogeneous data and it being easy to construct by machine learning, the corresponding solutions are proposed;
• The ancillary decision-making method for the failure handling of hydropower stations is proposed. Each case is represented in the form of multiple KGs, i.e., a system topology KG, a dispatching regulation KG, and an accident case KG. The key steps, i.e., case construction, case retrieval and case revision, are designed, which greatly increases the flexibility of case knowledge extraction, management and retrieval;
• The proposed method is verified with the practical data of a large-scale cascade hydropower station in China. After a failure occurs, the proposed method can provide a practical and accurate ancillary decision-making scheme as well as related dispatching principles to assist the power station dispatcher in carrying out reasonable and reliable accident handling, which effectively improves the intelligence level of the emergency management of the hydropower station.

Further research on this topic will focus on the improvement of the deep learning method for KG construction and reasoning, so as to improve its speed and accuracy. Future work will also include proposing a reasoning method that can deduce failure-handling principles from historical accident cases (in the form of KG). The deduced principles can be used to enrich the dispatching regulation KG and increase the practicability and intelligence of the ancillary decision-making method based on CBR and KG.