Hyper-local, efficient extreme heat projection and analysis using machine learning to augment a hybrid dynamical-statistical downscaling technique

This paper describes a scalable system for quantifying hyper-local heat stress in urban environments and its expected response within the changing climate. A hybrid dynamical-statistical downscaling approach links Global Climate Models (GCMs) with dynamically downscaled extreme heat events using the Weather Research and Forecasting model (WRF). Downscaled historical simulations in WRF incorporate urban canopy physics to better describe localized land surface details in the urban environment relevant to extreme heat. This downscaled library is then used in an analog-based approach. This contribution reports a series of enhancements to existing analog-based methods which can efficiently produce more detailed results. The system here uses advanced statistical methods and simple machine learning (ML) techniques to optimize analog selection, perform spatially-consistent bias correction, and decompose patterns of extreme heat into dynamic components such as the land-sea contrast and inland sea-breeze penetration. Hindcast projections are validated against observational data from in-situ weather observing stations. The results demonstrate the scalability and efficiency of this system as it is deployed in cloud-based architectures with parallelized code. Downscaled predictions are equally applicable to heat stress at weather and climate time scales, supporting infrastructure resilience and adaptation, and emergency response.