Identifying weather regimes for regional-scale stochastic weather generators

Weather regime based stochastic weather generators (WR-SWGs) have recently been proposed as a tool to better understand multi-sector vulnerability to deeply uncertain climate change. WR-SWGs can distinguish and simulate different types of climate change that have varying degrees of uncertainty in future projections, including thermodynamic changes (e.g., rising temperatures, Clausius-Clapeyron scaling of extreme precipitation) and dynamic changes (e.g., shifting circulation and storm tracks). These models require the accurate identification of WRs that are representative of both historical and plausible future patterns of atmospheric circulation, while preserving the complex space-time variability of weather processes. This study proposes a novel framework to identify such WRs based on WR-SWG performance over a broad geographic area and applies this framework to a case study in California. We test two components of WR-SWG design, including the method used for WR identification (Hidden Markov Models (HMMs) vs. K-means clustering) and the number of WRs. For different combinations of these components, we assess performance of a multi-site WR-SWG using 14 metrics across 13 major California river basins during the cold season. Results show that performance is best using a small number of WRs (4-5) identified using an HMM. We then juxtapose the number of WRs selected based on WR-SWG performance against the number of regimes identified using metastability analysis of atmospheric fields. Results show strong agreement in the number of regimes between the two approaches, suggesting that the use of metastable regimes could inform WR-SWG design. We conclude with a discussion of the potential to expand this framework for additional WR-SWG design parameters and spatial scales.

Automated summary

Weather regime based stochastic weather generators (WR-SWGs) are proposed as a tool to better understand vulnerability to climate change, distinguishing and simulating different types of climate change with varying degrees of uncertainty. A novel framework is proposed to identify representative WRs based on performance over a broad geographic area and applied to a case study in California. Findings suggest that a small number of WRs identified using Hidden Markov Models (HMMs) perform best, with agreement between the number of WRs selected based on performance and regimes identified using metastability analysis. Future research could explore expanding this framework for additional design parameters and spatial scales.