Posters

Abstract

Modern wind fleets increasingly rely on predictive analytics to raise early warning flags; however, the operational value of these alerts is often limited by the lack of contextualised maintenance knowledge across different turbine brands and models. In this study, we present an LLM-driven decision support framework built on a multi-year, multi-site database of failure, downtime, and maintenance records collected from wind farms operating multiple turbine brands and models. The records were standardised and automatically labelled using large language models according to turbine brand–model combinations, fault and downtime categories, and intervention types. The performance of downtime record labelling and LLM-based filtering for summary generation was evaluated on a representative subset of the data using standard classification evaluation metrics, demonstrating robust and consistent performance. Leveraging this validated, labelled knowledge base, the framework dynamically links each early warning flag raised for a specific turbine to historically relevant cases. For any given warning type, the system automatically filters similar past events and generates actionable insights, including expected average downtime, estimated production loss, and the most frequently applied maintenance interventions. These outputs are summarised and contextualised using LLMs and presented directly within the early warning interface, enabling operators to move beyond anomaly detection toward informed operational decision-making. Unlike traditional rule-based or purely statistical approaches, the proposed method integrates unstructured maintenance knowledge at fleet scale while remaining vendor-agnostic and adaptable to different operational practices. Field deployment across multiple wind farms demonstrates that LLM-assisted contextualisation significantly improves the interpretability and practical usefulness of early warning signals, supporting faster prioritisation, better maintenance planning, and more consistent responses across diverse turbine fleets. The results highlight the potential of large language models to bridge the gap between predictive analytics and real-world wind-farm operations by transforming historical maintenance data into directly actionable intelligence.