Using information-theory to investigate climate stress test and top-down realization ensembles: an application for long-term vulnerability assessment

Abstract:

Water resources managers must plan for new and/or adaptation of existing infrastructure while the future evolution of water demand and water supply are subject to a myriad of uncertainties. The decision-making literature generally makes use of two different modeling frameworks to assess climate risk to water resource systems: i) a top-down (scenario-led) approach that uses downscaled climate projections informed by Global Circulation Models; or ii) a climate stress test (scenario-neutral) approach that uses stochastic models to generate scenarios that capture the range of plausible future climates. Both approaches, while conceptually different, lead to the generation and analysis of an ensemble of stochastic realizations that describe the future evolution of the hydro-meteorological variables that drive the water systems’ performance. The information gained changes across the ensembles and realizations through the modeling chain from stochastic generation of weather to estimated performance for the water resource system. The overarching goal of this research is to i) assess for each approach, the information redundancy in each ensemble/realization, ii) analyze the flow of information from the generation of climate scenarios to assessment of performance metrics. We use the Sierra Nevada region of the Hetch Hetchy Regional Water System (HHRWS) as case study to illustrate the proposed methodology. Using information theory-based metrics (i.e., Shannon Entropy, mutual information), we first assess the amount of information contained within temperature and precipitation realizations. Then, temperature and precipitation realizations are input to the Precipitation Runoff Modeling System (PRMS) model to simulate streamflow above Don Pedro reservoir. The information contained in each streamflow ensemble is assessed and compared with the information contained within the weather realizations. Lastly, information contained within the distribution of “water available for diversion” performance metric, is estimated from the streamflow ensemble and compared with information contained within the hydro-meteorological drivers. The use of machine learning is also assessed to maximize the information contained within the distribution of water available for diversion metric, while minimizing the number of weather realizations to be processed through the modeling chain.