Evaluating Inductive Parameter-Based Transfer Learning with Deep Neural Networks for Wind Forecasting in Corsica

ICLR 2026 Conference Submission19226 Authors

Keywords: transfer learning, deep learning, fine-tuning, wind forecasting, negative transfer

Abstract: This study assesses the effectiveness of various transfer learning strategies for wind speed forecasting across meteorological stations in Corsica using deep neural networks. Leveraging inductive parameter-based transfer, models are transferred based on geographic proximity, topographic classification, dominant wind direction, and random assignment. Several architectures are evaluated, including recurrent, convolutional, attention-based, and dense networks. Results indicate that structured transfer strategies do not consistently outperform non-transfer baselines. This lack of improvement can be largely attributed to significant distributional differences in wind speed across stations, which hinder model transferability. These findings highlight the challenges posed by domain shift in a geographically heterogeneous insular context and emphasize the need for more refined similarity criteria, hybrid transfer strategies, and spatially-aware modeling, notably through graph neural networks. The results also call for a critical reassessment of commonly held assumptions about the benefits of transfer learning in complex meteorological environments.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 19226