Downscaling daily wind speed with Bayesian deep learning for climate monitoring

Wind dynamics are extremely complex and have critical impacts on the level of damage from natural hazards, such as storms and wildfires. In the wake of climate change, wind dynamics are becoming more complex, making the prediction of future wind characteristics a more challenging task. Nevertheless, having long-term projections of some wind characteristics, such as daily wind speed, is crucial for effective monitoring of climate change, and for efficient disaster risk management. Furthermore, accurate projections of wind speed result in optimized generation of wind-based electric power. General circulation models (GCMs) provide long-term simulations (often till year 2100 or more) of multiple climate variables. However, simulations from a GCM are at a grid with coarse spatial resolution, rendering them ineffective to resolve and analyze climate change at the local regional level. Spatial downscaling techniques are often used to map such global large-scale simulations to a local small-scale region. In this paper, we present a novel deep learning framework for spatial downscaling, specifically for forecasting the daily average wind speed at a local station level using GCM simulations. Our framework, named wind convolutional neural networks with transformers, or WCT for short, consists of multi-head convolutional neural networks, followed by stacked transformers, and an uncertainty quantification component based on Bayesian inference. Experimental results show the suitability of WCT when applied on four wind stations in New Jersey and Pennsylvania, USA. Moreover, we use the trained WCT on future GCM simulations to produce local-scale daily wind speed projections up to the year 2100.