ResNet-Based Downscaling of ERA5 Wind Data for Enhanced Wind Resource Assessment in Ambon Island, Indonesia

Abstract

Accurate wind resource assessment in complex island topographies demands high-resolution spatial datasets that conventional 31-km reanalysis products fail to capture. An advanced encoder-decoder Residual Neural Network (ResNet) framework is developed to spatially downscale ERA5 wind fields to a 1-km grid over Ambon Island, Indonesia. The proposed architecture combines geostatistical kriging with deep residual learning and multi-channel inputs of ERA5 atmospheric profiles and SRTM digital elevation data, and relies solely on independent ground observations for validation. Validation reveals a clear spatial trade-off: classical baselines tend to overfit local inland anomalies, whereas the ResNet architecture exhibits better spatial generalization over complex coastal-to-marine transition zones, with a Mean Absolute Error (MAE) of 2.17 m/s at the variable 10-m surface layer, representing a 54% error reduction with respect to raw ERA5 data. Extrapolated to a 100-m hub height, where surface friction diminishes, the architecture successfully isolates high-potential wind corridors along the southern Leitimur peninsula (4.5–7.5 m/s). While localized overestimation in built-up grids highlights the limits of pure elevation-driven downscaling without dynamic surface roughness (z_0 ) inputs, the architecture effectively captures macro-topographic wind acceleration. This framework offers a scalable, transferable methodology for wind mapping in data-sparse, topographically complex archipelagic environments.