Posters

Abstract

We present a novel, generic, scalable and easy-to-deploy methodology for monitoring and diagnosing electromechanical assets in wind energy systems. Our approach broadly applies to sensor-equipped assets, effectively utilizing data from existing Supervisory Control and Data Acquisition (SCADA) systems. It eliminates the need for annotated historical fault datasets, which are time-consuming and costly to obtain. Functioning effectively with minimal data cleaning and not requiring precise knowledge of sensor placements—often difficult to acquire—it is non-intrusive, requires no additional hardware, and operates solely on existing SCADA data. By learning a baseline of normal operational behaviour for each asset, our methodology can detect anomalies and potential failures, including those not previously encountered. Our modelling strategy leverages artificial intelligence (AI) models inherently robust to outliers and noisy data commonly found in SCADA systems. These models identify and discard anomalous behaviour present in training datasets, ensuring they capture only normal operation patterns relevant for accurate monitoring and diagnostics. Utilizing several years of SCADA data, the models learn the unique behavioural characteristics of each turbine within a wind farm. This individualized learning is crucial since turbines of the same model and manufacturer can exhibit subtle but significant operational differences due to wear, maintenance history, settings, environmental factors, and manufacturing variations. Employing probabilistic modelling provides a range of expected values rather than single-point estimates, enhancing operators’ decision-making by quantifying inherent uncertainties for better risk assessment, resource allocation, and proactive maintenance planning. Our alarm strategy emphasizes context-aware, multi-sensor-based alerts to mitigate alarm fatigue and enhance response effectiveness. By considering each asset's specific operational conditions and context, we generate meaningful and actionable alarms that accurately reflect the system's actual state and priority. Integrating alarms across multiple components enables the identification of more complex and systemic issues that might otherwise go unnoticed. For instance, correlating an overheating component alarm with a curtailment alarm allows us to detect scenarios where a turbine deliberately limits its power output to protect itself from excessive temperatures, indicating potential underlying problems. Each alarm includes potential root causes and recommended actions, providing maintenance teams with insights that expedite troubleshooting, reduce downtime, and optimize maintenance resources. Our methodology has been successfully deployed across diverse wind farm portfolios totalling multiple gigawatts over several years, demonstrating scalability and reliability. It has consistently proven effective in identifying a wide variety of failures and performance issues, often many months before they become critical events, enabling timely interventions and significant cost savings. We will present real-world case studies highlighting the practical benefits and tangible value our approach has delivered to wind farm owners and operators across different regions and conditions. These benefits include uncovering latent health issues undetected by SCADA event logs or Condition Monitoring Systems (CMS) and proactively addressing performance challenges ranging from unexpected curtailments to assets experiencing frequent and disruptive restarts. These cases underscore the practical effectiveness and real-world impact of our AI-driven monitoring and diagnostic strategy in enhancing the reliability, availability, and overall performance of wind energy assets, contributing to more sustainable and efficient energy generation.