Transfer learning for wind speed forecasting: A scalable approach for data-scarce environments
Year: 2025
Type: article
Abstract: Accurate wind speed forecasting plays a central role in the integration of wind energy into modern power systems. However, in many regions, the deployment of data-driven models is limited by the scarcity of historical wind measurements. This study investigates the potential of transfer learning (TL) to address this issue by reusing models trained on data-rich locations to forecast wind speed in under-instrumented sites. Using a dataset from 70 meteorological stations across Spain, spanning diverse climatic conditions, we compared TL with conventional direct learning (DL) using Extreme Learning Machine (ELM) and Autoregressive (AR) models. Forecasts were performed over horizons ranging from 30 min to 6 hr and evaluated using normalized Root Mean Squared Error (nRMSE) and normalized Mean Absolute Error (nMAE). Across 4,830 TL experiments, results show that TL achieves forecast accuracy close to that of DL, with average nRMSE ranging from 0.297 (TL) to 0.292 (DL) for 30-min horizons, and from 0.674 to 0.640 for 360-min forecasts. These findings confirm that TL is a robust and scalable approach for wind speed forecasting in data-scarce environments. The methodology opens promising avenues for the deployment of forecasting tools in regions where traditional data-driven models remain inapplicable due to limited local measurements.
Authors: Milan Despotovic, Gilles Notton, Cyril Voyant
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4415106776