Construction and Application of a Power Industry Knowledge Graph

1. Introduction

The power system is a large, knowledge‑intensive asset network that requires close integration of domain expertise with information technology. Traditional knowledge engineering based on expert rules is insufficient for modern electric‑grid intelligence, prompting the adoption of knowledge‑graph technology to structure concepts, entities, events, and their relationships.

2. AI Institute Overview

The Artificial Intelligence Application Institute, founded in 2018 under State Grid Corporation, focuses on five research directions: intelligent perception, big data, intelligent platforms, cognitive intelligence, and AI applications. It houses five specialized labs and a management office, employing around 70 staff with 100% master’s degrees and 34% PhDs. The institute also participates in several national AI standardization committees and provides a research platform for power‑sector AI.

3. Background of Power Knowledge Engineering

Historically, power‑sector knowledge engineering relied on expert systems and rule‑based representations. With the rise of the energy‑internet, knowledge has become more open, flat, and boundary‑blurred, increasing the complexity of intelligent cognition. Knowledge graphs are introduced to capture entities, relationships, and events in a structured manner, enabling smarter AI applications.

4. Knowledge Representation and Graph Construction

Power‑domain knowledge originates from structured sources (legacy knowledge‑engineering systems, expert databases) and semi‑/unstructured sources (standards, regulations, expert experience). It is divided into generic knowledge (e.g., equipment names, voltage levels) and specialized knowledge (e.g., customer billing). Ontology construction follows a hybrid top‑down (expert‑driven) and bottom‑up (data‑driven) approach, combining existing schemas with automatic extraction from unstructured data. The resulting ontology supports both generic and domain‑specific layers.

5. Application Practice

Four practical layers are proposed:

Data acquisition – ingesting structured, semi‑structured, and unstructured data from internal and external sources.

Graph construction – using NLP, entity/relation extraction, and graph databases (e.g., Neo4j) to store the knowledge graph.

Knowledge computation – applying representation learning, reasoning, and path‑search algorithms.

Graph application – delivering intelligent search, QA, recommendation, and decision‑support services.

Specific use cases include:

Power dispatch fault handling – combining textual regulations with D5000 system data, using Bi‑LSTM‑CRF for entity extraction and Bi‑LSTM‑Attention for relation extraction.

Power inspection work‑order processing – building a knowledge base from work orders, tickets, and inspection records, then applying Bi‑LSTM‑CRF and Bi‑GRU‑CRF models.

Power customer service intelligent QA – constructing a customer‑service ontology from 95598 call logs and knowledge bases, employing confidence‑propagation and Bi‑LSTM‑CRF for entity and relation extraction.

6. Future Challenges

Key challenges identified are:

Knowledge extraction and graph construction from heterogeneous, noisy, and redundant data.

Efficient reasoning over grid topology and real‑time decision support.

Establishing systematic quality‑evaluation metrics for domain knowledge graphs.

Developing application‑oriented tooling to rapidly generate sub‑graphs for specific business scenarios.

7. Conclusion

The power‑sector knowledge graph aims to represent entities, attributes, and relationships in a structured form, enhancing cross‑media data management and cognitive capabilities. While not a universal solution, continuous refinement and integration with existing systems are expected to meet the evolving needs of power utilities.