Introduction

The remaining useful lifetime and operational reliability of wind farms have a net impact on the current Levelized Cost Of Energy (LCOE) and influence the cost of investment for upcoming projects​​. Numerical models and installed sensors are used to monitor and predict wind turbine loads and fatigue, aiming to increase the longevity of the systems and limit operational and maintenance costs. Nevertheless, wind turbines are not easily instrumented systems with complex and moving parts, including rotating blades, generators, foundations and towers.

This study builds on previous work and introduces an improved version of the FleetLeader tool, providing a scalable virtual sensing framework that leverages limited physical sensor data, SCADA signals, and Machine Learning (ML) to estimate critical loads across entire wind farms. Earlier developments of the FleetLeader concept by ECN and later by OWI-lab have demonstrated the potential of combining SCADA data with targeted load measurements for accurate load estimation.


Method

Regression algorithms such as Neural Network (NN) and eXtreme Gradient Boosting (XGB) excel in determining often hidden correlations between features or systems. In the case of wind farms, these methods are applied to find the correlation between wind turbines, which often operate in similar conditions. While this suggests similarities between wind turbine operations and loads, wake effects, local defects or system differences can create complex correlations.

These regression models are well-suited to combine data from specialized load sensors with SCADA signals, to accurately estimate similar loads for the non-instrumented turbines inside a wind farm.

To avoid overfitting the behavior of a single turbine and the need to assume identical behavior to different turbines, at least two instrumented turbines are needed. Validation was performed using data from two Dutch wind farms (EWTW and OWEZ), each providing SCADA signals for all turbines and detailed load measurements on two turbines. The ML models were trained and tested across idle, transient, fully waked, partially waked, and free-stream conditions.

This study will also explore whether categorizing turbines by operational state improves prediction accuracy or whether data-driven categorization via ML is more effective. Additionally, it is investigated which SCADA signals are most informative for load prediction, whether load measurements enhance predictive power for other loads, and what the minimum number of sensors is for reliable predictions.


Results

Both regression models demonstrated strong capability in estimating turbine loads such as blade root bending, tower bottom bending, and main shaft torque, amongst others, using limited SCADA signals as input. Among the tested models, XGB consistently outperformed NN and reached R² scores up to 98%, indicating reliable predictions across varying environmental conditions.

It is expected that categorization and including load measurements will significantly improve the estimation of other loads.


Conclusions

The presented work demonstrates a robust approach to generalized virtual sensing in wind farms, enabling accurate load estimation of large wind farms while reducing the need for costly and impractical wind turbine sensors. By integrating targeted sensor data, SCADA signals, and ML, this method supports a transition from corrective to predictive maintenance, reduces downtime, and lowers LCOE.

Wind turbine condition monitoring and early fault detection are crucial for ensuring safe operation and reducing maintenance costs. With the rapid development of artificial intelligence, data-driven condition monitoring based on normal behavior models has advanced significantly. Among various data sources, Supervisory Control and Data Acquisition (SCADA) data has attracted widespread attention due to its easy accessibility and the absence of additional costly equipment. However, false alarms severely affect the practical effectiveness of intelligent monitoring and may even disrupt maintenance operations. To address these challenges, we propose a dynamic threshold design method for wind turbines that accounts for varying operational conditions. Firstly, the actual power curve of the wind turbine is fitted using its operational data. Based on this fitted curve, the operational conditions are classified in real time. Then, the fixed thresholds are established during the start-up and rated power operation phases to minimize the influence of isolated data points. During the maximum power point tracking phase, a dynamic threshold is adaptively determined by considering actual power fluctuations. The results demonstrate that the method can effectively adjust thresholds according to the operating conditions, thereby enabling adaptive condition monitoring and early anomaly detection. Moreover, it significantly reduces false alarms caused by power-limited operation and environmental variations.

Introduction
To reduce global warming caused by greenhouse gas emissions, the energy system needs to shift to renewable sources. In many parts of the world, including The Netherlands, offshore wind energy will play an important role in the new energy mix. The large-scale deployment of offshore wind farms will put pressure on material and human resources. Well informed asset management and automated inspection can alleviate the demand for human resources during the operational phase. This requires effective monitoring of wind turbine blades, to provide the necessary knowledge of the health of the blades. To address this challenge, World Class Maintenance (WCM), TNO and partners initiated the AIRTuB-ROMI project on Automated Inspection and Repair of wind Turbine Blades, using Resident drones for Offshore Monitoring and Inspection. This work presents the approach and first results on a method for targeting drone inspections on offshore wind turbine blades using a hybrid digital twin.
Method
The method consists of several building blocks that convert the operational (SCADA) and structural health monitoring (SHM) measurements to decision support information for the operator. First, a time-synchronized set of measurements is fed to an aero-elastic (AE) model of the offshore wind turbine to derive the instantaneous load distribution on the blade. This loading condition is then applied to a finite-element (FE) model of the blade, which provides estimates of stress and strain in the internal structure. These are condensed to two-dimensional heat maps for pressure and suction side, as overview of relative loading at each location. Together with similar maps to account for possible previously recorded damages, local SHM results and drone access limitations, the information is presented in a dashboard to the operator as individual layers and combined output. Using this overview, the operator can make an informed decision if and where to send the drone for inspection when a SHM trigger occurs. In addition to the overview, the dashboard also allows to zoom in on the time traces of all the available measurements for the selected event, to further analyze and evaluate it.
Results
The AE and FE models have been created based on design information and field observations and provide realistic load distributions and stress maps on the blade for various environmental and operating conditions. The paper will address the challenges related to dynamic loading conditions in more detail. First tests with the method indicate that the digital twin of the blade can provide additional physics-based information on stress hotspots and that the combination of the different information sources in a single dashboard aids the selection of inspection areas.
Outlook
The measurement campaign for the onshore trial is currently ongoing, with SCADA and SHM data (acoustic emission, vibration and lightning detection) being recorded. In the coming period, the method will be tested as part of the demonstration of the AIRTuB-ROMI concept. This work brings automated inspection of wind turbine blades a step closer to application in the field.

ABSTRACT
Digital twins (DTs) promise better operation and maintenance (O&M), yet most efforts within wind energy remain either conceptual reviews or narrow subsystem demonstrations, with a large gap in-between. We present a comprehensive, top-down DT framework for wind farm performance operators, filling in the holistic system-of-systems DT research gap by integrating wind farm- and component-level models. It gives a real-time, key performance indicator (KPI)-first view of farm performance and health with hierarchical drill-down capabilities (farm -> turbine -> component), accessing bottom-up models when deeper diagnosis is needed, closing the loop between overview, benchmarking and root-cause analysis.

CONTRIBUTION
- A five-module, extensible DT framework designed using systems engineering principles based on industry stakeholder requirements, with open and robust IT architecture (1) Data Processing Pipeline, 2) KPI & Metrics Engine, 3) Performance Baseline Models, 4) Benchmarking and Anomaly Detection, and 5) User Interface and Integration Layer).
- A unified performance logic that anchors KPIs derived from supervisory control and data acquisition (SCADA) data based on the state-of-the-art literature recommendations and industry implementations to baselines coming from physics-based, statistical or hybrid models (with uncertainty where possible) and contextualises them against internal expectations and external ranges.
- A deployable operator workflow --- monitor -> benchmark -> flag anomalies -> drill down for diagnostics -> generate reports.

METHOD
Operational data streams (SCADA, met-masts, grid/export meters, alarms/events, maintenance logs, asset metadata) are cleaned, harmonised and checked for reliability. Baseline models --- physics-based, statistical or hybrid --- provide expected performance and behaviour to benchmark against. The DT continuously compares measured KPIs and component signals to these baselines, highlighting deviations that warrant investigation and guiding users to the appropriate turbine/component-level models, implemented in a unified and linked user functional mock-up interface (FMI).

RESULTS
We implemented a simple proof-of-concept on real SCADA data from an industrial partner, executing the proposed pipeline. Wind farm annual energy production was calculated and benchmarked against a theoretical maximum to obtain capacity factor, then compared against common industry ranges for similar assets. This end-to-end run confirmed feasibility and user-facing clarity of the framework. Bottom-up models (going from component- to farm-level, to be used as baselines) are concurrently being developed and validated, with planned top-down/bottom-up model fusion and more concrete implementations that would make this a holistic DT for driving decisions regarding O&M and future investments.

SIGNIFICANCE
Our work addresses the research gap identified in literature on system-of-system, holistic DT framework implementations.
For operators, our DT framework delivers one entry point for performance monitoring, transparent baseline-anchored benchmarking, rapid anomaly detection and consistent reporting --- reducing cognitive load while scaling to large fleets.
For researchers, the framework provides an open, standardised integration platform that accelerates model testing and validation, provides a realistic DT context and nudges the community towards consistent KPIs, data handling practices and uncertainty treatment, improving comparability across studies and sites.
In short, our DT framework bridges concept to control-room, making DTs more actionable for daily O&M while remaining extensible for all kinds of wind farm owner needs and facilitating wind farm DT research.

Norwegian Water Resources and Energy Directorate (NVE) has recently updated the potential sea areas most suitable for offshore wind energy production. While marine icing may occur along the whole Norwegian coast, it will be most severe in the northernmost areas - the southern Barents Sea. These northernmost areas have also been identified as areas with a high probability of atmospheric icing and expected power loss, because the new generation wind farms will extend into the height of low level clouds. Hence, one can expect that these potential offshore wind areas will be exposed to both marine and atmospheric icing of considerable severity, However, experience with offshore wind farms under such icing conditions is rather limited, which affects a wide range of stakeholders that need to be involved in the development. While many icing problems have been studied for land-based wind turbines (reduced performance, vibration and fatigue, ice throw issues) icing challenges at sea are challenging due to the mixed (marine and atmospheric ) icing conditions. One also would expect challenges with saline water due to potentially high salt concentrations (under low temperature) and chemical attack (e.g., salt frost scaling). Dealing with these problems requires a detailed understanding of icing process on different surfaces for given met-ocean conditions (ice type, microstructure, growth). Here, I discuss recent results from a project on marine icing and ice adhesion (MICROSPRAY, Microstructure of sea spray ice, RCN-Norway, 2020-2025, NTNU and Equinor), in view of future challenges that offshore wind farms in waters affected by icing might face . What are the details on icing that are needed for sustainable and durable offshore wind farms in Norway?