Enhanced Named Entity Recognition and Event Extraction for Power Grid Outage Scheduling Using a Universal Information Extraction Framework

Abstract

To enhance online dispatch decision support capabilities for power grid outage planning, this study proposes a Universal Information Extraction (UIE)-based method for enhanced named entity recognition and event extraction from outage documents. First, a Structured Extraction Language (SEL) framework is developed that unifies the semantic modeling of outage information to generate standardized representations for dual-task parsing of events and entities. Subsequently, a trigger-centric event extraction model is developed through feature learning of outage plan triggers and syntactic pattern entities. Finally, the event extraction model is employed to identify operational objects and action triggers, while the entity recognition model detects seven critical equipment entities within these operational objects. Validated on real-world outage plans from a provincial-level power dispatch center, the methodology demonstrates reliable detection capabilities for both named entity recognition and event extraction. Relative to conventional techniques, the F₁ score increases by 1.08% for event extraction and 2.48% for named entity recognition.

1. Introduction

Power grid scheduling texts contain a wealth of knowledge crucial for ensuring the safety and stability of the grid, and they serve as a guide for control and operation personnel [1]. In the field of grid control, scheduling texts are often unstructured, which makes them difficult to directly apply in control systems or provide real-time decision support for dispatchers, unlike structured data. A power outage plan is a crucial emergency response document, pre-arranged to handle power shortages, equipment maintenance, or unforeseen situations [2]. The plan includes arrangements across various time dimensions, such as yearly, monthly, weekly, and daily schedules, aimed at improving the stability of the power system, ensuring grid stability, and minimizing the impact on users. These plans form the foundation for applications such as load forecasting, economic dispatching, and fault prevention in the grid [3]. With the rapid development of power systems and the increasing reliance on electricity, factors such as power shortages, equipment failures, and natural disasters may lead to power outages. The need for digitalized and object-oriented outage plans, enhancing their online application and service capabilities, has become urgent.

Entity recognition and event extraction from outage plan texts are essential tasks in natural language processing (NLP) and serve as the foundation for downstream applications such as intelligent retrieval, plan verification, and scheduling operations [4,5]. Traditional entity recognition and event extraction methods, such as rule-based and dictionary-based models, Conditional Random Fields (CRFs), and Support Vector Machines (SVMs) [6,7,8], often rely on a large number of manually defined rule templates. When the characteristics of text data change, these rules may fail, resulting in a sharp decline in model performance. Additionally, the flexibility of rule templates is relatively low, with poor generalization, making it challenging to handle large-scale, complex text data. To reduce manual intervention and achieve automatic recognition, deep learning techniques have introduced Recurrent Neural Networks (RNNs) as effective methods for extracting sequence features [9]. However, traditional RNNs suffer from the vanishing gradient problem, which makes it difficult for the model to maintain dependencies over long text sequences, affecting its ability to capture long-range dependencies. To address this, Long Short-Term Memory (LSTM) networks with gating mechanisms and Bidirectional LSTM (BiLSTM) methods have been proposed for feature extraction [10] to overcome the shortcomings of traditional RNNs in capturing long-sequence features.

Event extraction has further developed based on entity recognition. It not only requires identifying entities in the text but also extracting all relevant event elements associated with triggering words, ensuring the completeness of the events. With continuous advances in deep learning, methods based on LSTM, BiLSTM, and other network architectures for entity recognition and event extraction have become important research directions in text information processing [11,12,13]. These studies have shown that BiLSTM can effectively recognize event trigger words and entities. However, most research in entity recognition and event extraction relies on publicly available datasets, and the performance in real-world applications still needs improvement.

In the power grid field, studies have also been conducted on entity recognition and event extraction for power grid scheduling texts. Research [14,15] indicates that entity recognition and event extraction for power scheduling texts are key technologies for building intelligent dispatching applications. The authors of [16] achieved entity extraction from power equipment defect texts by integrating BERT, Bidirectional Long Short-Term Memory (BiLSTM), and Conditional Random Field (CRF). The authors of [17] employed a BiLSTM-CRF model to achieve intelligent label classification for power industry accident report texts. The authors of [18] proposed a relation extraction framework that utilizes GloVe embeddings and a bidirectional GRU network for vector representation of sentences in power distribution network operation corpora. The approach further incorporates an Attention-enhanced Graph Convolutional Network (GCN) to model syntactic dependencies and nonlinear textual relationships. The authors of [19] introduced a deep learning method for knowledge extraction in grid control texts. In summary, current research on power text parsing mainly focuses on rule-based and deep learning-based entity recognition models, with less emphasis placed on event extraction, and no research specifically targeting power outage plan event extraction. Simply recognizing entity types has limited application in actual power grid scheduling; only the accurate recognition of both entity types and operational events can enhance the online application and service capabilities of scheduling systems.

To solve the problem of complex logic and nested entities in power grid outage plans, this paper proposes a UIE-based framework for entity and event recognition in outage plans. This approach utilizes the Structured Extraction Language (SEL) method to unify the modeling of outage plan information extraction tasks and generate the required structured language for both event extraction and entity recognition. Using ERNIE 3.0 [20] as the pretrained language model for feature encoding and combining it with a dual-pointer decoding mechanism to replace traditional framework models, the UIE framework adapts the model parameters and transforms the input-output format of outage plan data into a sequence labeling task. The method builds a mapping relationship from domain knowledge to semantic space, allowing the extraction of relevant events based on syntactic features and trigger words in outage plans.

The model was validated using historical outage plan texts from a power control center, achieving high entity and event extraction accuracy. It significantly improved the recognition of nested structures in complex outage plans, providing a strong foundation for online comprehensive optimization of outage plans.

2. Power Outage Plan Information Recognition Method

The proposed power outage plan recognition method was developed based on the Universal Information Extraction (UIE) framework, featuring a unified architecture that concurrently executes entity extraction, relation extraction, and event extraction tasks, thereby significantly enhancing model adaptability and generalization across heterogeneous tasks. This multi-task learning system, built upon the Transformer architecture with ERNIE 3.0-enhanced representations, demonstrates robust capability in identifying and extracting domain-specific information from unstructured textual data.

The UIE framework uses Structured Extraction Language (SEL) to uniformly describe the structure of four different information extraction tasks, ensuring that the structure of the output remains consistent when the model performs different tasks. The framework uses a structural schema instructor (SSI), which can guide the model to perform specific tasks for multiple tasks via prompts. The overall architecture of the UIE framework is shown in Figure 1.

By using a structured pattern prompt and the original text, the UIE model generates SEL-structured information. The relevant formulas are as follows:

$$s\oplus x=\left[s_1,s_2,\dots ,s_{\left|s\right|},x_1,x_2,x_{\left|x\right|}\right]$$

(1)

s represents SSI, x represents the input original sentence, and ⊕ represents the concatenation symbol. Expanding its content fully, it becomes

$$s\oplus x=\left[\left[spot\right],spotNam{e}_1,\dots ,\left[spot\right],SpotNam{e}_n,\left[text\right],x_1,\dots ,x_{\left|x\right|}\right]$$

(2)

In the formula, [spot] represents the different entity category markers, SpotName represents the name of different entity categories, and [text] represents the original text content that follows.

$$input=s\oplus x$$

(3)

$$y=UIE(input)$$

(4)

Here, it represents the results extracted and generated by UIE based on Structured Extraction Language.

Specifically, the input is first fed into the encoder layer to obtain the hidden layer representation of each token:

$$H=Encoder(H)$$

(5)

The encoder layer adopts the ERNIE 3.0 knowledge-enhanced pre-trained model, which essentially consists of a two-layer Transformer network. The first layer is a general information extraction layer, primarily used to capture the basic feature information of general data. The second layer is a fine-tuning layer, used to recognize and extract specific power outage plan entities and event information. The two-layer Transformer network structure in this framework ultimately enables secondary training based on the existing pre-trained model, effectively improving training results while accelerating the training process.

The decoder layer uses a dual-pointer method for outputting the entity’s start and end positions. The two pointers receive input information from the encoder layer, and after parameter calculations in the decoder layer, the start and end positions of the entities are obtained.

Training the parameters of the UIE framework is mainly divided into pre-training tasks and fine-tuning tasks.

The pre-training tasks are divided into three types. The first task involves inputting SSI and the original text x, with the output being structured data y. The loss function is as follows:

$$f_{pair\_loss}={\displaystyle {\sum }_{(x,y)\in D_{pair}}-\log p(y | s,x;{\theta }_e,{\theta }_d)}$$

(6)

The second task involves only structured data y, with input being the first half and the output being the remaining portion. Only the decoder part of the UIE is trained, allowing it to learn SEL syntax. The loss function is as follows:

$$f_{record\_loss}={\displaystyle {\sum }_{y\in D_{record}}-\log p(y^{\prime } | y^{″};{\theta }_d)}$$

(7)

The third task mainly focuses on unsupervised training. It involves removing 15% of the characters from the original sentence and then generating the removed characters. The loss function is as follows:

$$f_{text\_loss}={\displaystyle {\sum }_{x\in D_{text}}-\log p(x^{″} | x^{\prime };{\theta }_e,{\theta }_d)}$$

(8)

The loss functions of these three pre-training tasks are summed to form the total loss function for joint training.

The fine-tuning tasks are similar to the first pre-training task in terms of input and output structure. The input data consists of SSI and the original text, and the output is structured data. The loss function is as follows:

$$f_{FT\_loss}={\displaystyle {\sum }_{(x,y)\in D_{FT}}-\log p(y | s,x;{\theta }_e,{\theta }_d)}$$

(9)

Negative samples are also introduced during fine-tuning, where information not present in the original labels is randomly inserted to improve the final training performance.

3. Power Outage Plan Information Recognition Based on the UIE Framework

3.1. Construction Approach of Power Outage Plan Information Recognition Model

This paper presents a power grid power outage plan entity and event information recognition method based on the UIE framework, which involves data preprocessing, UIE event extraction prediction model training, UIE entity recognition prediction model training, device entity recognition, and extraction of device initial and end states. The overall approach is shown in Figure 2.

The power outage plan information recognition prediction process is as follows:

(1)

Data Preparation

The original power outage plan prediction samples are organized, and event extraction samples and entity recognition samples are extracted. Based on semantic similarity, duplicate data in the samples is filtered and selected, ultimately obtaining a sample set for power outage plan event extraction and entity recognition prediction.

(2)

Model Training

The Structured Extraction Language (SEL) method is used to unify the modeling of the complex power outage plan extraction tasks, generating the structured language needed for power outage plan event extraction and entity recognition. These will serve as input variables for the power grid power outage plan entity and event information recognition prediction model.

(3)

Model Prediction

Based on the UIE framework, the event extraction model extracts the operation objects and action trigger words. The operation objects are fed into the entity recognition model to identify the specific device types within the operation objects. At the same time, regular expressions are used to extract the initial and end states of devices from the action trigger words.

3.2. Entity and Event Annotation in Power Outage Plans

Power outage plan texts are a type of power grid scheduling text that includes complex operational objects and highly specialized scheduling tasks. These plans have a critical impact on the safe and stable operation of the power grid, so accurate annotation of the involved entities and events is necessary.

Power outage plan texts can be divided into yearly, monthly, and weekly plans from a time dimension. The content mainly includes power outage plan numbers, power outage plan application units, work locations, outage work content, outage work requirements, requested start and end times, approved start and end times, approved restoration start time, first-level method opinions, second-level method opinions, and so on.

Structured data, such as the power outage plan number, application unit, work location, and related times, can be directly extracted or parsed using simple regular expressions.

In contrast, unstructured data, such as outage work content, work requirements, first-level method opinions (Operating mode opinions), and second-level method opinions (Relay protecting opinions), are more complex, varied, and lack fixed expression structures, which makes it difficult for computers to understand the content.

The unstructured data in power outage plans has the following characteristics:

(1)

A large amount of historical data exists in the power grid system but lacks label information, making it difficult to apply machine learning algorithms directly.

(2)

Each sentence corresponds to a power grid operation and involves many types of operations, which require a deep understanding of power system knowledge. Currently, there is a lack of a related knowledge base.

(3)

The content is complex, and the types of operation objects are numerous, with object names composed of many different technical terms from the power system, making extraction difficult.

Therefore, for parsing power outage plan texts, the UIE framework is used to analyze and extract the unstructured data. A typical structure of the power outage plan and extraction application is shown in the figure below (Figure 3).

Analysis of the unstructured data in the power outage plan text reveals that the content contains both entities and events, as well as action verbs. Therefore, it is proposed to design the action verbs as event trigger words in the power outage plan. First, the entity types and trigger words must be annotated. Upon identifying the action trigger word, the entity type can be extracted, and the entity name can be further obtained. The initial state and changed state of the outage equipment can also be extracted through event recognition.

Using structured unified modeling, the power outage plan entities and events are annotated, including action verbs (Action), operation devices (Indepent), and the plant stations where the operation devices are located (Indepentstart).

The power outage plan text is complex and varied, making it difficult for standard machine learning methods to achieve accurate recognition. After reasonable annotation methods are applied to the power outage plan text, the UIE framework can then be used for analysis and extraction. A typical example of TPOP annotation is shown below (Figure 4).

Subsequently, entity recognition is performed by identifying the trigger words and extracting entity names. For the trigger words in the power outage plan, due to their relatively fixed format, regular expressions are used to extract the initial and end states. For device entity extraction in the power outage plan and the synchronization with the model library’s normalization requirements, the paper summarizes the device entity types involved in the plan, including seven types of equipment: power stations, transformers, busbars, lines, beakers, disconnectors, and units. The entity components in the power outage plan are identified, and corresponding entity labels are generated.

3.3. Power Outage Schedule Information Recognition Process

The power outage plan event extraction model and entity recognition model are constructed based on the UIE framework, enabling the extraction of information from power outage plans according to event syntactic components. We use the ERNIE 3.0 framework as the core encoding module, with standardized encoded power outage plan datasets as input. The attention mechanism is employed to extract text features, and the structured action and entity information are output. During the training phase, common actions and entities related to power outage plans are introduced for fine-tuning the model parameters.

By iteratively training on the power outage plan standard dataset, the ERNIE 3.0 network’s hidden layer weights are optimized using the text data and its corresponding labels. This process allows the model to learn the distribution features of various entities and actions in the context of power outage texts. The dual-pointer decoding mechanism is then employed to predict the start and end positions of entities, combining the encoded layer’s output with the label information to achieve accurate recognition and extraction of power outage plan information.

First, based on the event syntactic logic of power outage plans, an event recognition model is built. The recognized three entity types: Action, Indepent, and Indepentstart, are composed according to the event syntactic components of the power outage plan. Action serves as the action trigger word, Start represents the plant station where the equipment is located, and Object refers to the equipment triggered by the action. A power outage event extraction model is then established. Some typical power outage plan operation event extraction patterns are shown in Figure 5.

Next, using the UIE framework for entity recognition, the operation object (Indepent) and the pre-object (Indepentstart) extracted by the event model are used to recognize the specific type of the equipment. The action (Action) extracted by the event model is then processed using regular expressions to extract the initial and end states, thus completing the power outage plan information extraction.

4. Results

All experimental results were completed on a system configured with an Intel (R) Xeon (R) CPU E5-2678, an NVIDIA RTX5000 GPU, and the Ubuntu (version 16.04.7) operating system. The Python version used was 3.8.19.

In this paper, the UIE model is adopted as the power outage plan information recognition and prediction model. The model’s maximum learning rate is 1 × 10⁻⁵, with a batch size of 32, 768 hidden layer neurons, 12 transformer layers, and a maximum sequence length of 2048. The pre-trained ERNIE 3.0 model is used for parameter initialization. This pre-trained model captures common information extraction capabilities and is fine-tuned to adapt to the tasks of power outage plan information recognition and event extraction, reducing the time spent on parameter optimization during training and improving the model’s recognition accuracy.

4.1. Experimental Datasets

The historical power outage plan texts from 2022 to 2024 of a power grid control center were used as the research subject. The proposed power outage plan entity labeling method was applied to annotate the power outage plan texts, generating a total of 3628 power outage plan entities. Among these, 3096 entities were used as training samples, while the remaining 594 entities were used as test samples.

4.2. Evaluation Indicators

Precision, recall, and F₁ score were adopted as evaluation metrics to assess the effectiveness of power outage plan event extraction and entity recognition. The formulas for these metrics are as follows:

$$P_{re}=\frac{TP}{TP+FP}$$

(10)

$$R_{ec}=\frac{TP}{TP+FN}$$

(11)

$$F_1=\frac{2\times P_{re}\times R_{ec}}{P_{re}+R_{ec}}$$

(12)

where TP (True Positive) represents the number of correct positive predictions made by the model. FP (False Positive) represents the number of negative samples incorrectly predicted as positive by the model. FN (False Negative) represents the number of positive samples incorrectly predicted as negative by the model.

4.3. Effectiveness Analysis

The UIE model was trained on the power outage plan annotated dataset, constructing a model for information extraction from power outage plan texts. The effectiveness of the model was validated through test samples. For the event extraction model, the extraction results are shown in Figure 6. In the power outage plan test set, the F₁ score for action trigger words reached 97.7%, and the F₁ score for location and location/device combinations reached 100%, demonstrating strong event extraction capability.

For the entity recognition model, the recognition results are shown in Figure 7. The figure presents the prediction accuracy, recall rate, and F₁ score for each entity type in the UIE entity recognition model. The model achieved over 90% recognition performance for all seven entity types in the power outage plan text, with some entities achieving 100% recognition accuracy, showcasing strong generalization ability.

These results indicate that the UIE model can effectively capture the specialized language features in power outage plans and accurately extract various events and entity components. The hybrid neural network model performed excellently in the event and entity information extraction tasks for power outage plan texts, validating its strong adaptability and practicality.

4.4. Model Comparison

In order to further evaluate the performance of the UIE-based power outage plan information recognition model, we compared it with several other information recognition models, including BERT-BiLSTM-CRF, word2vec-BiLSTM-CRF, and BERT-CRF models. Among them, the word2vec-BiLSTM-CRF model uses pre-trained word2vec word vectors, and the BERT-BiLSTM-CRF model uses the pre-trained BERT model as word vectors. Both of these models employ BiLSTM as the prediction encoding network. Additionally, we included the BERT-CRF model, which uses BERT as the prediction encoding network. Finally, by comparing the performance of the UIE model, we aim to validate its superiority in information recognition tasks.

We evaluated the recognition performance of the test samples using precision, recall, and F₁ score. The metrics for event extraction are presented in

Table 1. Comparison of event extraction performance of different models.

Model	Precision (%)	Recall (%)	F1 (%)
word2vec-BiLSTM-CRF	94.22	94.36	94.29
BERT-BiLSTM-CRF	94.83	94.41	94.83
BERT-CRF	97.59	95.58	96.58
UIE	97.44	98.07	97.66

. The experimental results show that the hybrid neural network model proposed in this paper significantly outperforms the comparison models in event extraction tasks for power outage plans. By using BERT word vectors combined with the BiLSTM-CRF model for event extraction, the overall recognition performance increased by 0.54% compared to using word2vec word vectors, indicating that BERT pre-trained word vectors enhance the model’s recognition robustness. After introducing BERT as the event extraction model, the performance improved by 1.75%, demonstrating that BERT better understands the semantic representation of power outage plan texts compared to BiLSTM. By adopting the UIE model proposed in this paper, the overall performance significantly improved, outperforming the best-performing BERT-CRF model by 1.08%, indicating that the UIE-based event extraction model can deeply mine the complex relationships and action trigger words in power outage plans, fully utilizing contextual information to improve recognition accuracy.

The metrics for entity recognition are shown in

Table 2. Comparison of entity recognition performance of different models.

Model	Precision (%)	Recall (%)	F1 (%)
word2vec-BiLSTM-CRF	89.10	94.73	91.83
BERT-BiLSTM-CRF	90.68	97.36	93.90
BERT-CRF	91.13	97.36	94.14
UIE	96.37	96.87	96.62

. By using BERT word vectors combined with the BiLSTM-CRF model for entity recognition, the overall recognition performance increased by 2.07% compared to using word2vec word vectors. After introducing BERT as the entity recognition model, the performance further improved by 0.24%. By adopting the UIE model proposed in this paper, the performance surpassed the BERT-CRF model by 2.48%, demonstrating that the UIE-based power outage plan entity recognition model enhances the model’s semantic understanding of entities, resulting in more accurate entity identification in entity recognition tasks.

The recognition results of the three models are shown in Figure 8. From the figure, we can see that the model proposed in this paper outperforms the comparison models in terms of entity recognition results across all categories, with most of the entity recognition results falling within the 97.3–100% range, indicating good recognition performance. The reasons for this are as follows: On one hand, the UIE model adopts unified modeling, where entity types are encoded as prompt words through structured hints, allowing the model to handle multiple entity types simultaneously. Traditional methods struggle to handle nested entities, but the UIE model can directly describe complex structures through generative output, which plays a crucial role in resolving ambiguities and increasing the accuracy of entity recognition in power outage plans. On the other hand, the ERNIE 3.0 pre-trained text encoder used in this paper, with its hierarchical attention network, can focus on local details and entity integrity within power outage plan texts. Considering these factors, the hybrid neural network model proposed in this paper demonstrates strong performance in power outage plan entity information extraction tasks. Despite the strong overall performance of the UIE model, we observed instances of entity prediction errors. A notable challenge arises from variations in terminology usage among different dispatchers. For example, the line name “Youfang Line” is sometimes abbreviated as “Youfang” in outage scheduling texts. This abbreviation bears a strong resemblance to the typical naming pattern of power substations, leading the UIE model to misclassify “Youfang” as a station entity instead of the correct line entity.

4.5. Model Application

The UIE-based framework for entity and event information recognition in power grid outage plans, proposed in this paper, has undergone preliminary research and application at a power control center of a national grid. Figure 9 demonstrates a specific example of applying the UIE framework to optimize the scheduling of power grid outage plans.

Firstly, the original outage plan text provided by the dispatch personnel is subjected to preliminary data parsing. Based on structural characteristics, the data can be roughly divided into two categories: structured data and unstructured data. Structured data, such as Application Number: ND231031XXX, Requesting Unit: XX Power Supply Company, Work Location: XX District XX Road, Outage Time: 2024-XX-XX XX:XX:XX, can be directly parsed by text processing programs into corresponding modules and stored as structured data in the database for subsequent applications. For unstructured data, such as the work requirement XX station 220 kV XX line transitioning from running state to maintenance state, a power outage plan event recognition model based on outage plan event syntax logic is first used. By identifying action trigger words, entities such as Action, Indepent, and Indepentstart are extracted.

Then, for the Indepent and Indepentstart entities, the UIE framework is applied for entity recognition, identifying the specific types of equipment involved, such as XX station and 220 kV XX line. The proposed method has successfully labeled various equipment entities involved in the outage plan, including power stations, transformers, busbars, lines, breakers, disconnectors, and units, and the recognition results effectively meet the requirements.

Finally, for the Action extracted from the event model, a regular expression is used to extract the initial state as operation and the end state as maintenance, completing the extraction of outage plan information.

The extracted data, such as outage equipment and time, is used in the outage plan optimization and scheduling application. The extracted data is processed through a power grid simulation environment, where it serves as the state space and action space for the reinforcement learning agent used to optimize the outage plan. Through training, the agent ultimately produces the optimized decision, which is the optimized scheduling result of the outage plan.

This method greatly enhances the efficiency of data collection by dispatch personnel, reduces the extraction time from a 12 h process that consumes a significant amount of manual labor time to just 1 min, improves the ability of intelligent power systems to extract outage plan information, and serves as the foundation for downstream applications such as outage plan optimization, intelligent retrieval of outage plans, and outage plan validation.

5. Conclusions

This paper addresses the challenges of recognizing complex logic and nested entities in power grid outage plan texts by proposing an entity and event information recognition method based on the UIE framework. This method introduces a Structured Extraction Language (SEL) approach to model the complex logic of outage plan extraction tasks in a unified manner, generating the structured language required for both event extraction and entity recognition of outage plans. ERNIE 3.0 is used as the pretrained text encoder to capture contextual features, enhancing the accuracy of outage plan information recognition. Ultimately, the method combines the outage plan trigger word entities and syntactic entities based on outage plan event syntax to obtain structured outage plan information.

This approach was tested on historical outage plan texts from a power control center through comparative experiments. Compared to the alternative methods, our approach achieved the best F₁ score, improving by 4.79%, 2.72%, and 2.48%, respectively, demonstrating its adaptability and practicality. This method helps improve outage plan information extraction capabilities, providing strong support for the subsequent comprehensive optimization of outage plans and related business processes. Furthermore, the proposed methodology was successfully extended to entity recognition within fault handling plans. This application also demonstrated promising performance in extracting critical entities from these complex contingency planning documents, making a significant step toward achieving intelligent scheduling. Future work will investigate the integration of encoded representations from outage scheduling texts with their corresponding relation labels to refine the calculation algorithm for entity/pair and relation label correlation scores. Additionally, addressing the lexical and structural similarities between line names and substations, we plan to study label distribution characteristics and optimize the loss function to enhance model robustness and performance.