Balancing Flood Risk and Water Supply in California: Policy Search Integrating Short‐Term Forecast Ensembles With Conjunctive Use

Abstract

Short‐term weather forecasts have the potential to improve reservoir operations for both flood control and water supply objectives, especially in regions currently relying on fixed seasonal flood pools to mitigate risk. The successful development of forecast‐based policies should integrate uncertainty from modern forecast products to create unambiguous rules that can be tested on out‐of‐sample periods. This study investigates the potential for such operating policies to improve water supply efficiency while maintaining flood protection, combining state‐of‐the‐art weather hindcasts with downstream conjunctive use to transfer surplus flood releases to groundwater storage. Because available weather hindcasts are relatively short (10–20 years), we propose a novel statistical framework to develop synthetic forecasts over longer periods of the historical record. Operating rules are trained with a recently developed policy search framework in which decision rules are structured as binary trees. Policies are developed for a range of scenarios with varying forecast skill and conjunctive use capacity, using Folsom Reservoir, California, as a case study. Results suggest that the combination of conjunctive use and short‐term weather forecasts can substantially improve both water supply and flood control objectives by allowing storage to remain high until forecasts trigger a release. Further, increased conjunctive use capacity reduces the importance of forecast skill, since surface storage can be moved to groundwater during the flood season without losing water supply. This analysis serves the development of forecast‐based operating policies for large, multipurpose reservoirs in California and other regions with similar flood hydroclimatology.

Renewable Energy and EU 2020 Target for Energy Efficiency in the Czech Republic and Slovakia

“…In many countries, hydro sources are often needed to generate energy for almost all fuels and technologies to generate electricity, and energy is needed to treat and transport both water and wastewater [31][32][33]. A fascinating case study on the subject is the state of California in the United States with its large water supply systems (which require a lot of energy for pumping) that moves water from the relatively humid northern areas of the state to the drier and more populated southern region (including the major metropolitan areas of Los Angeles and San Diego) [34]. Conversely, the majority of the natural gas used in the water system is used for water heating on the consumer side of the water meter.…”

Inferred inflow forecast horizons guiding reservoir release decisions across the United States

“…Forecast reliability, in turn, depends on the available predictive information. An operator might rely on upstream water storage (e.g., soil moisture, snowpack, lake levels) (Shukla and Lettenmaier, 2011), hydrological regime state (Turner and Galelli, 2016), climate indices and teleconnections (Yang et al, 2017;Libisch-Lehner et al, 2019), weather forecasts (Georgakakos et al, 2005;Shukla et al, 2012;Nayak et al, 2018), current river flow rates (Hejazi et al, 2008), knowledge of planned water releases from upstream dams, and perhaps some or all of these in combination (Denaro et al, 2017). This enormous scope for variability in forecast quality and application across dams means there is no obvious way to identify the actual operationalized forecast, or indeed the model used to assimilate it into decision making, for a given system without insight into individual agencies' models and data preferences.…”

Bootstrap Aggregation and Cross‐Validation Methods to Reduce Overfitting in Reservoir Control Policy Search

“…Bagging and calibration‐validation techniques are tested on a case study of Folsom Reservoir, California, following Nayak et al (2018). We use daily inflow data for the period of 1923–2016, split into a training set (1982–2016) and a testing set (1923–1981) based on water years beginning in October.…”

“…Several approaches have been explored to address this issue. One common approach is to optimize policies based on several random initializations (seeds) of the heuristic search and select or combine solutions from those with the highest within‐sample performance over either the historical hydrologic record (Herman & Giuliani, 2018; Nayak et al, 2018; Salazar et al, 2016) or a synthetically generated scenario ensemble (Giuliani et al, 2014, 2018; Salazar et al, 2017). Other work has reevaluated policy performance over other synthetic hydrologic sequences not used in training, but sampled from the same uncertainty characterization (Quinn et al, 2017), or over other sequences modified by additional uncertain scenario factors not considered during training (Quinn et al, 2019).…”