Machine Learning or Deep Learning? The Role of Ice Sensors and Weather Forecasts in Wind Turbine Icing Prediction

Atmospheric icing on wind turbine blades remains a major barrier to reliable wind energy production in cold climates, reducing annual yields by up to 20%. In addition to production losses, ice shedding from blades creates safety hazards for personnel and infrastructure, prompting mandatory curtailments. Reliable forecasting of icing formation would allow operators to anticipate events before losses occur, preventing blade damage, reducing downtime, and improving the accuracy of day-ahead energy trading.

Predicting icing accretion is challenging because it depends on diverse atmospheric and turbine operating conditions, as well as microphysical parameters such as liquid water content (LWC) and droplet size, which are difficult to measure. These variables can fluctuate rapidly and are only partially observed by SCADA systems and conventional weather stations. Prior studies have largely been restricted to very short forecasting horizons using limited data sources, leaving a gap between research model capabilities and the operational needs of wind farm operators.

This research presents the first systematic benchmarking of machine learning (ML) and deep learning (DL) models for icing prediction across multiple horizons and input configurations. The models integrate SCADA signals, meteorological measurements, turbine-mounted icing sensors, and weather forecasts. We compare ML ensembles, including Random Forest, Extreme Gradient Boosting, and Light Gradient Boosting Machine, with a convolutional neural network–long short-term memory (CNN-LSTM) model. Models are evaluated at 1-, 6-, and 24-hour horizons using high-frequency turbine data recorded at 1-second resolution.

To assess the contribution of different data sources, three input configurations are tested: Class A (SCADA plus historical weather), Class B (Class A plus icing sensor data), and Class C (Class A plus forecasted weather including LWC). Performance is measured using the F1 score, which balances precision, avoiding false alarms that cause unnecessary shutdowns, with recall, capturing as many real icing events as possible.

Random Forest emerged as the best-performing ML model and, with Class C inputs, performed comparably to CNN-LSTM across all horizons, achieving high F1 scores at 1, 6, and 24 hours. Crucially, Random Forest executes significantly faster than CNN-LSTM, making it highly suitable for real-time deployment. Ablation analysis demonstrates that including icing sensor data significantly improves short-term, 1-hour forecasts, increasing the F1 score. At the day-ahead horizon, incorporating weather forecast data in Class C is essential for maintaining strong predictive performance.

These findings demonstrate that enriched input data matter more than model complexity. ML ensembles, when supplied with icing-relevant measurements and forecasts, can rival deep learning performance while remaining interpretable and computationally efficient. This enables proactive turbine control in the short term and more reliable energy trading strategies in cold-climate wind farms.