SUMMARY

As power systems become increasingly complex, Artificial Intelligence (AI) offers strong potential to improve decision-making, reliability, and predictive diagnostics. However, many AI applications remain limited by the fragmentation of critical engineering knowledge, which is often stored in unstructured legacy documents. This paper examines the role of structured, object-oriented engineering data as a prerequisite for scalable and reliable AI applications in the power sector. Grounded in methodologies such as Building Information Modeling (BIM) and Engineering Information Modeling (EIM), the study highlights the transition from document-centric to data-centric engineering practices. A review of the literature shows that, despite advances in standards and technologies, utilities continue to struggle with fragmented legacy data, which limits the scalable application of AI across power sector challenges. To address this gap, the paper presents an AI-based methodology for extracting structured metadata from legacy documents and integrating it into a centralized engineering database. A case study at a Brazilian Transmission System Operator demonstrates tangible benefits of the proposed approach, including a significant reduction in metadata registration time compared to manual processes, enabling large volumes of legacy documentation to be structured within realistic time and cost constraints. Additional gains include improved agility, information reliability, faster access to technical data, enhanced fault analysis consistency, and effective object-based querying of asset histories, supporting maintenance and substation modernization activities. Finally, the paper discusses the next step toward interactive AI-assisted advisors capable of reasoning directly over structured engineering databases, including the evaluation of a newly released Engineering Advisor developed in this direction. The results reinforce a fundamental AI role in power systems: advanced AI applications depend on high-quality structured data, while AI itself can be effectively used to transform legacy information into such structured data. In this context, structured data environments emerge not only as a prerequisite for human access and data governance, but also as a strategic enabler for scalable and trustworthy AI applications, accelerating digital transformation and preserving engineering knowledge across the power system lifecycle.

1 Introduction

The growing complexity of modern power systems, driven by decentralization, digitalization, and the increasing integration of renewable energy sources, has intensified the demand for advanced decision-support and diagnostic tools. In critical infrastructures such as electric power systems, rapid access to reliable and structured engineering information is essential for system resilience, safe operation, and effective decision-making, as outages can result in severe economic and societal impacts. Artificial Intelligence (AI) offers significant potential to enhance reliability, operational efficiency, and predictive capabilities; however, its effectiveness is fundamentally constrained by data availability and quality. In practice, a substantial portion of engineering knowledge remains embedded in unstructured legacy documentation or tacit human expertise, limiting its usability by data-driven algorithms.

This paper investigates structured engineering data as a foundational enabler for AI-applications in the power sector. Grounded in object-oriented methodologies such as BIM and EIM, it advocates a transition from document-centric to data-centric engineering paradigms. Despite advances in interoperability standards, including the Common Information Model (CIM – IEC 61970/61968) and IEC 61850, the literature indicates that real-world implementations remain constrained by fragmented data landscapes and the persistence of legacy documentation.

To address these challenges, the paper presents an AI-based solution for extracting and structuring legacy engineering documentation into an object-oriented engineering platform, resulting in a coherent, searchable, and machine-readable engineering knowledge base. This solution enables the systematic transformation of historical, document-centric information into structured data aligned with EIM principles. In addition, the paper explores a newly released AI-powered Engineering Advisor, evaluated as a forward-looking extension of the proposed approach. This Engineering Advisor illustrates how structured data can enable intelligent, context-aware interaction with power system technical data. Together, these contributions demonstrate a progressive digital transformation pathway, from legacy data structuring to advanced AI-assisted interaction, bridging brownfield and greenfield environments and supporting scalable, knowledge-centric power system engineering.

2 Literature Review

2.1 AI-Applications in Power Systems: Landscape and Limitations

AI has become central to research on advanced decision-support and diagnostic tools in power systems, enabling applications such as grid optimization, protection scheme analysis, and fault diagnosis [1]. However, the literature consistently shows that the effectiveness, reliability, and safety of AI-based solutions depend less on algorithmic sophistication than on the availability of curated, structured, and trustworthy data [2] [3]. This requirement is reinforced by the black-box nature of many advanced models and the high sensitivity of power systems to inaccuracies, motivating a shift from model-centric to data-centric AI development [4].

Existing studies report a wide range of AI techniques, including traditional machine learning, deep learning, reinforcement learning, and, more recently, generative AI-applied to forecasting, fault classification, stability assessment, and control [4] [5] [6] [7] [8] [9] [10]. Despite their diversity, these approaches converge on a common limitation: dependable performance requires structured, consistent, and semantically integrated data as input. Fragmented data landscapes, heterogeneous naming conventions, and the persistence of unstructured legacy documentation remain major barriers to real-world deployment [2] [9] [10]. Consequently, the literature reinforces that curated, AI-ready data is a prerequisite for scalable and trustworthy AI applications in critical power system environments [4] [11].

2.2 Engineering Information Modeling (EIM) and Object-Oriented Engineering Data Modeling

As the complexity of power systems increases, the effective management of technical knowledge becomes as critical as the deployment of new technologies. The literature shows that advances in AI and data analytics alone are insufficient to address the challenges posed by decentralized and data-intensive infrastructures when engineering information remains fragmented or poorly structured [12] [13].What is fundamentally required is structured, interoperable, and accessible information that can support both human expertise and machine-driven analytics. In this context, EIM or BIM emerges as a methodological framework for managing complexity through systematic knowledge structuring.

The EIM or BIM methodological framework represents a transition away from traditional document-centric approaches, in which engineering information is systematically registered in a centralized database and structured according to an object-oriented modeling paradigm, where assets, functions, and relationships are represented as digital objects with defined attributes and lifecycle histories. This approach enables standardization, workflow integration, traceability, and consistency across engineering phases, establishing a reliable data foundation [14] [15].

EIM extends the principles of BIM. While BIM is more commonly applied in civil and electromechanical engineering contexts and focused mainly on physical infrastructure, EIM expands these concepts to industrial and electrical engineering domains. This extension enables the application to power system data such as electrical circuits, protection settings, and logical diagrams. Although the term BIM can, in principle, be applied to any type of structured information — even when it is not associated with 3D modeling — this paper adopts the term EIM to clearly denote the broader engineering scope addressed. Although the term EIM is less widely established, its usage is consistent with terminology adopted in the literature [14] [16].

The adoption of EIM depends on the coordinated alignment of three pillars — technology, people, and processes — to ensure data consistency, traceability, and reuse throughout the asset lifecycle [16] [17]. When applied cohesively, this alignment enables engineering information to be generated in a structured manner from its inception, supporting quality assurance, lifecycle continuity, and cross-disciplinary collaboration while reducing reliance on isolated expert knowledge and manual data interpretation [15] [17] [18].

The shift to object-oriented engineering information management represents a departure from traditional Computer-Aided Design (CAD) approaches focused on geometry. Under EIM, each system component is modeled as a digital object that encapsulates functional, spatial, electrical, and performance-related attributes, rather than serving only as a graphical representation [15].

Different international standards can serve as a basis for applying EIM, depending on the engineering discipline, for maintaining semantic consistency and multi-disciplinary data interoperability. In the BIM domain, the Industry Foundation Classes (IFC), standardized as ISO 16739, provide an open, object-oriented data model for representing built assets and their lifecycle information. In power system engineering, standards such as the Common Information Model (CIM – IEC 61970/61968) and IEC 61850 offer the semantic foundations required to model electrical systems. Using EIM principles and appropriate engineering platforms, these standards can be consistently applied to structure engineering information in a way that supports interoperability and AI-based applications.

Traditional engineering workflows are document-centric, relying on isolated files such as PDFs, spreadsheets, and drawings, which leads to duplicated information, inconsistencies, and data loss during lifecycle transitions. In contrast, EIM centralizes engineering data in a structured, object-oriented database, providing a single source of truth for all stakeholders [14] [16] [17] [18]. Applying EIM significantly reduces the time to access technical information while improving operational safety through accurate and up-to-date data. By establishing a single source of truth, EIM enhances project delivery and provides a reliable foundation for AI-driven and sustainable power system operations.

The relevance of structured data environments is further illustrated by regulatory initiatives aimed at strengthening data governance at a systemic level. In Brazil, the National Electric Energy Agency (Agência Nacional de Energia Elétrica – ANEEL) has promoted the availability of structured, integrated, and machine-processable corporate data through its Open Data Plan (Plano de Dados Abertos – PDA). This initiative seeks to enhance transparency, interoperability, data quality, and accessibility, while fostering informed decision-making by public authorities, supporting social oversight, and enabling the development of new data-driven services and applications [19].

A concrete outcome of this effort is the Electric Power Transmission Facilities Database (Base de Dados das Instalações de Transmissão – BDIT), maintained by the Brazilian National Electric System Operator (Operador Nacional do Sistema Elétrico – ONS). BDIT establishes an object-based registry of transmission assets and provides a standardized framework for information exchange across institutional and organizational boundaries, fully aligned with the principles of structured and interoperable data management [20].

Despite these advances, the literature consistently reports persistent limitations related to data scarcity, confidentiality constraints, and the historical fragmentation of engineering information [21]. As a consequence, synthetic data is frequently adopted to compensate for gaps in information governance. While synthetic datasets can support hypothesis testing and methodological validation [10] [3], their widespread use ultimately highlights a deeper structural challenge: without reliable, structured, and traceable engineering data derived from real operational systems, the robustness, scalability, and operational relevance of AI-Applications remain fundamentally constrained.

2.3 Legacy Infrastructure and the Digitalization Gap in Power Systems

Although object-oriented modeling approaches and international standards establish the foundations for generating high-quality structured information from the earliest project stages, most existing power system infrastructure operates in a fundamentally different, brownfield context. Critical engineering knowledge — such as protection settings, logic diagrams, and maintenance records — remains embedded in unstructured documents and heterogeneous legacy databases. This reality confirms that interoperability is not merely a matter of protocol adoption, but a challenge of data structuring and curation at the source.

Several studies across domains ranging from chemistry and mechanical engineering to process industries highlight the substantial effort required to reconcile informational legacies dominated by large volumes of unstructured data during the digitization of existing systems. The development of digital twins and engineering databases becomes severely constrained in the absence of automated mechanisms for data extraction, organization, and linkage, as manual extraction processes are typically expensive, time-consuming, and prone to human error [22] [23] [24] [25] [26].

3 Methodology

To address the challenges associated with legacy engineering data, this paper presents a methodology based on AI-driven techniques for extracting structured metadata from legacy documents and integrating it into a centralized, object-oriented engineering database. This enables consistent data representation and lifecycle management. It is important to note that different engineering domains and heterogeneous data sources demand tailored hybrid AI models and domain-specific training strategies, depending on the target metadata and document characteristics.

3.1.1 Hybrid Approach

A hybrid AI-based approach was implemented, comprising preprocessing, processing and inference, post-processing, and validation stages. In the preprocessing phase, image processing techniques are applied to scanned documents to improve Optical Character Recognition (OCR) accuracy, including noise reduction, contrast enhancement, and layout normalization.

During processing and inference, extraction strategies are adapted according to the characteristics of each metadata field. To support this, metadata were classified based on two criteria: (i) multiplicity (single or multi-class) and (ii) knowledge of the class universe (known or unknown). For each extraction scenario, dedicated training datasets were constructed to capture the variability of real-world engineering documents, including variations in layout, templates and document quality. The overall datasets was divided into training and testing subsets, with 70% used for training and 30% for testing. The datasets were designed to reflect the distribution of metadata values, including imbalanced classes, which were addressed through proportional sampling to mitigate bias. Model robustness was assessed using k-fold cross-validation, ensuring generalization across different subsets of the data. The processing architecture combines Long Short-Term Memory (LSTM) neural networks to handle complex textual patterns with rule-based techniques for structured and deterministic cases. Hyperparameter optimization was performed to improve model performance, using the F1-score as the primary evaluation metric. In addition, LLMs, specifically GPT-4.1, were employed to support inference in scenarios involving unstructured or open-domain metadata.

Post-processing then normalizes extracted entities by resolving synonyms, abbreviations, and ambiguities, and by incorporating contextual information from file names and document metadata to ensure semantic consistency.

To evaluate the quality of the proposed architectures, performance metrics are computed by comparing the automatically extracted data with ground truth reference values. Given the high reliability requirements of the power sector, the ground truth is established through a structured human validation process consisting of two complementary stages: Review and Validation. In the Review stage, domain specialists analyze and correct the extracted metadata to ensure alignment with engineering standards and domain knowledge. In the Validation stage, a statistical quality control process is applied without allowing further data modification. This stage is based on sampling procedures defined by NBR 5426 and NBR 5427 standards, using inspection levels (typically Level II) and acceptance criteria based on the Acceptable Quality Level (AQL). The inspection evaluates metadata considering critical, major, and minor errors. If the sampled data exceeds the defined error thresholds, the batch is rejected and returned to the Review stage for reprocessing. Only data that pass both stages are considered validated and adopted as ground truth for performance evaluation.

4 Case Study

This paper presents a case study conducted at a large Brazilian Transmission System Operator (TSO), representative of the challenges faced by utilities operating extensive and long-lived power system infrastructures. The organization manages a wide portfolio of transmission assets, including hundreds of substations and tens of thousands of kilometers of transmission lines, with a significant share of its generation portfolio based on renewable energy sources. As in many mature utilities, decades of operation resulted in a vast accumulation of legacy technical documentation supporting Protection, Automation, and Control (PAC) systems.

Within this context, two complementary strategic objectives guided the case study, as described in the following sections.

4.1.1 Modernizing Legacy Technical Documentation through AI-Based Structuring

The first objective was to modernize the management of legacy technical documentation used in substation operation and maintenance by digitizing PAC projects, applying AI-based techniques to extract structured metadata, and registering this information in a commercial object-oriented engineering platform. This initiative aimed to transform fragmented, document-centric information into a continuously updated and searchable engineering knowledge base, accessible to managers, engineers, and maintenance teams through standardized workflows. In addition, coordinated mechanisms for version control, document revision, approval, and change tracking were established in the platform to ensure data integrity and traceability.

The legacy documentation presented multiple challenges related to data quality, standardization, and consistency. To address these issues, the methodology described in Section 3 was applied. The metadata categories selected for this case study include key identifiers related to substation and bay characterization, project and bay classification, voltage level, document and revision control, as well as equipment models and manufacturers. The performance metrics reported in this case study are based on a dataset comprising 1,500 document pages. Table 2 summarizes the results obtained for each metadata category.

4.1.2 Structured Data from Project Inception in New and Modernized Substations

The second objective addressed the future asset lifecycle by ensuring that new and modernized substations generate structured engineering data from their inception. Based on EIM principles, standardized database templates and modeling rules were defined so that projects natively produce consistent, traceable, and operation-ready information. In parallel, training and upskilling programs were implemented for internal teams and external partners to align people and processes with the TSO’s data standards, preventing the continued creation of unstructured legacy data.

The Engineering Platform underpins this approach by enforcing structured workflows and a single source of truth. All project information is maintained in a centralized, object-oriented database that supports real-time collaboration, version consistency, and quality control embedded directly in daily engineering activities. Engineering elements are modeled as hierarchical objects with defined attributes and relationships. Together with organizational measures, this approach enabled a sustainable transition toward data-centric, model-driven engineering practices.

4.1.3 Results and Achievements

The deployment of the solution at the Brazilian TSO resulted in significant improvements in efficiency, information governance, and operational performance. Controlled tests were conducted to compare the human effort required to manually register all relevant metadata from project sheets with the effort required when applying the proposed AI-based approach. In the fully manual scenario, the average time to register the selected metadata from a single page was approximately 96 seconds, whereas the AI-assisted workflow reduced this effort to around 10 seconds per page. Considering an estimated volume of approximately 800,000 project sheets, this reduction corresponds to an estimated saving of approximately 19,110 working hours. In contrast to a fully manual approach, which would be substantially more time- and cost-intensive, the AI-based workflow enables this transformation to be executed within realistic time and budget constraints. Over a 24-month period, PAC diagrams from more than 80 substations were successfully processed and registered, demonstrating both the scalability of the solution and the practical feasibility of large-scale structuring of legacy engineering data through AI-enabled methods.

Qualitative feedback from users indicates significant gains in agility, cost reduction, and information reliability, particularly in maintenance activities and fault analysis. Engineers reported that, prior to the implementation, accessing specific technical information could take several hours and often required physical document handling or on-site visits, whereas with the proposed solution the same information can typically be retrieved in seconds. In addition, engineers highlighted substantial improvements in information gathering for substation modernization projects. Traditionally, modernization activities require extensive access to legacy documentation, a process that could previously take weeks and involve travel-related costs. With the implemented solution, comprehensive technical information for an entire substation can now be gathered within minutes, significantly accelerating modernization workflows and reducing associated costs.

Together, these initiatives aimed to bridge the gap between brownfield and greenfield environments by establishing a unified, structured data foundation capable of supporting both current operational needs and long-term digitalization strategies. The case study demonstrates that the combined use of AI-based legacy data extraction and object-oriented engineering platforms can substantially improve access to critical technical information, enhance consistency in fault analysis and operational decision-making, enable efficient querying of device histories through object-based search mechanisms, and provide a scalable and trustworthy foundation for future AI-driven applications in power system operation and maintenance.

5 Towards Interactive AI-Assisted Engineering Advisors

The next phase of digital transformation enables direct interaction with structured engineering data through an AI-powered Engineering Advisor based on LLMs. Instead of manually searching documents, engineers query the system in natural language to obtain context-aware answers, perform semantic searches, reuse existing solutions, and receive support for calculations, consistency checks, and compliance assessments against standards.

A newly released Engineering Advisor has been analyzed in this study. Designed as a domain-specific expert, the Engineering Advisor operates directly on the engineering platform database. It reasons over relationships, attributes, and historical context, supporting procedural guidance, analytical validation, and decision support embedded in the engineering lifecycle, including multimodal interactions and automated quality checks, such as identifying modeling inconsistencies early in project development.

A representative example illustrates how automated quality-verification mechanisms enabled by the Engineering Advisor can identify modeling inconsistencies early in the engineering process. In a native project modeled within the engineering platform, an engineer selected a secondary-system device model and requested verification of whether the number of binary outputs defined during modeling was consistent with the equipment attributes stored in the database (Figure 1).

Upon receiving the request, the Engineering Advisor automatically parsed all binary-related slots associated with the device, aggregated the total number of defined binaries, and compared this value with the corresponding modeled attributes. The analysis (Figure 1) identified a discrepancy, indicating that the device had been modeled inconsistently within the platform.

The Engineering Advisor was explored across multiple representative engineering contexts, including support for querying specific engineering information, such as “Display the 50BF logic diagram for circuit breaker 14T1” and “Display the electrical diagram for the open-control function of circuit breaker 14T1,” as well as retrieval of project revision data (e.g., “What changes were made in the latest project revision?”). In addition to diagram-oriented queries, the Advisor supports semantic, asset-centered questions, such as “Which equipment models are installed in the substation?” and “How many spare binary inputs (BIs) and binary outputs (BOs) are available in each IED?”.

Across all evaluated scenarios, the Engineering Advisor consistently produced responses that were assessed as correct, based on the available structured engineering data, by dynamically combining and correlating information from multiple database layers. . These layers include digital equipment models, standardized engineering templates, and IEC 61850-aligned logical node definitions. The results indicate that, when supported by a rich and well-structured database, the Advisor enables faster access to engineering information, improves verification accuracy, reduces modeling ambiguities, and supports the adoption of AI-assisted engineering workflows across the power system lifecycle. Although quantitative performance indicators have not yet been systematically evaluated under controlled experimental conditions, these aspects are expected to be evaluated in future work based on real-world deployments.

6 Conclusion

This paper demonstrated that the effective adoption of AI in power systems is fundamentally dependent on the availability of structured, consistent, and traceable data. While AI-Applications for grid optimization, protection analysis, and fault diagnostics continue to advance, their scalability and reliability remain constrained by the predominance of unstructured legacy information and document-centric engineering practices.

By combining AI-based techniques for legacy data extraction with an object-oriented EIM approach, the proposed methodology addresses this structural limitation at its root. The case study showed that large volumes of legacy PAC documentation can be transformed into structured, machine-readable engineering data within realistic time and cost constraints — an outcome that would be economically impractical using manual approaches alone. The resulting centralized engineering database significantly improved information accessibility, traceability, and consistency, supporting both operational activities and long-term digitalization strategies.

Beyond legacy data structuring, the establishment of standardized, object-oriented engineering databases for new projects ensures that structured information is generated from the earliest stages of the asset lifecycle, preventing the continuous creation of new unstructured data. This dual strategy effectively bridges brownfield and greenfield environments, creating a unified data foundation that supports current operation and maintenance needs while enabling future digital transformation.

The paper further demonstrated that once such a structured data foundation is in place, advanced AI capabilities can evolve beyond data extraction toward intelligent, context-aware interaction with engineering information. The Engineering Advisor illustrates how an AI-based assistant can reason directly over object-oriented engineering data to support semantic search, data validation, engineering analysis, and standards compliance, embedding decision support directly into the engineering lifecycle.

Overall, the results position data structuring not as a secondary support activity, but as a strategic enabler for scalable, trustworthy, and effective AI-Applications in power systems. By aligning technology, people, and processes around structured engineering information, utilities can unlock the full potential of AI, accelerate digital transformation, and preserve critical engineering knowledge across generations supporting safer, more efficient, and more resilient power system operation.

BIBLIOGRAPHY

[1] M. Balamurugan, K. Narayanan, N. Raghu, G. B. Arjun Kumar and V. N. & Trupti, “Role of artificial intelligence in smart grid–a mini review,” Frontiers in Artificial Intelligence, 8, 1551661, 2025.

[2] M. N. Dehaghani, T. Korõtko and A. Rosin, “AI applications for power quality issues in distribution systems: A systematic review,” IEEE Access, 2025.

[3] Y. Liu, P. Li, Y. Si and L. Ma, “A Comprehensive Review of AI Integration for Fault Detection in Modern Power Systems: Data Processing, Modeling, and Optimization,” Energies, 18(18), 4983., 2025.

[4] W. Strielkowski, A. Vlasov, K. Selivanov, K. Muraviev and V. Shakhnov, “Prospects and challenges of the machine learning and data-driven methods for the predictive analysis of power systems: A review,” Energies, 16(10), 4025, 2023.

[5] J. Zhang and R. X. Gao, “Deep learning-driven data curation and model interpretation for smart manufacturing.,” Chinese journal of mechanical engineering, 34(1), 71., 2021.

[6] G. Porawagamage, K. Dharmapala, J. S. Chaves, D. Villegas and A. Rajapakse, “A review of machine learning applications in power system protection and emergency control: opportunities, challenges, and future directions.,” Frontiers in Smart Grids, 3, 1371153, 2024.

[7] Y. Chen, X. Fan, R. Huang, Q. Huang, A. Li and K. P. Guddanti, “Artificial intelligence/machine learning technology in power system applications (No. PNNL-35735).,” Pacific Northwest National Laboratory (PNNL), Richland, WA (United States)., 2024.

[8] Y. Yang, F. Tu, S. Huang, Y. Tu and T. Liu, “Research on CNN-LSTM DC power system fault diagnosis and differential protection strategy based on reinforcement learning.,” Frontiers in Energy Research, 11, 1258549., 2023.

[9] M. Hassouna, C. Holzhüter, P. Lytaev, J. Thomas, B. Sick and C. Scholz, “Graph reinforcement learning for power grids: A comprehensive survey.,” arXiv preprint arXiv:2407.04522., 2024.

[10] S. L. e. a. Choi, “Generative AI for Power Grid Operations.,” No. NREL/TP-5D00-91176. National Renewable Energy Laboratory (NREL), Golden, CO (United States), 2024.

[11] A. Aldoseri, K. N. Al-Khalifa and A. M. Hamouda, “Re-thinking data strategy and integration for artificial intelligence: concepts, opportunities, and challenges.,” Applied Sciences, 13(12), 7082., 2023.

[12] F. Henao, R. Edgell, A. Sharma and J. Olney, “AI in power systems: a systematic review of key matters of concern.,” Energy Informatics, 8(1), 76., 2025.

[13] G. L. e. a. RAJORA, “A review of asset management using artificial intelligence‐based machine learning models: Applications for the electric power and energy system,” IET Generation Transmission & Distribution v. 18, n. 12, pp. 2155-2170, 2024.