Search for: All records

Abstract Forecast informed reservoir operations (FIRO) is an important advance in water management, but the design and testing of FIRO policies is limited by relatively short (10–35 year) hydro‐meteorological hindcasts. We present a novel, multisite model for synthetic forecast ensembles to overcome this limitation. This model utilizes parametric and non‐parametric procedures to capture complex forecast errors and maintain correlation between variables, lead times, locations, and ensemble members. After being fit to data from the hindcast period, this model can generate synthetic forecast ensembles in any period with observations. We demonstrate the approach in a case study of the FIRO‐based Ensemble Forecast Operations (EFO) control policy for the Lake Mendocino—Russian River basin, which conditions release decisions on ensemble forecasts from the Hydrologic Ensemble Forecast System (HEFS). We explore two generation strategies: (a) simulation of synthetic forecasts of meteorology to force HEFS; and (b) simulation of synthetic HEFS streamflow forecasts directly. We evaluate the synthetic forecasts using ensemble verification techniques and event‐based validation, finding good agreement with the actual ensemble forecasts. We then evaluate EFO policy performance using synthetic and actual forecasts over the hindcast period (1985–2010) and synthetic forecasts only over the pre‐hindcast period (1948–1984). Results show that the synthetic forecasts highlight important failure modes of the EFO policy under plausible forecast ensembles, but improvements are still needed to fully capture FIRO policy behavior under the actual forecast ensembles. Overall, the methodology advances a novel way to test FIRO policy robustness, which is key to building institutional support for FIRO. 

Abstract Policy search methods provide a heuristic mapping between observations and decisions and have been widely used in reservoir control studies. However, recent studies have observed a tendency for policy search methods to overfit to the hydrologic data used in training, particularly the sequence of flood and drought events. This technical note develops an extension of bootstrap aggregation (bagging) and cross‐validation techniques, inspired by the machine learning literature, to improve reservoir control policy performance on out‐of‐sample hydrological sequences. We explore these methods using a case study of Folsom Reservoir, California, using control policies structured as binary trees, and streamflow resampling based on the paleo‐inflow record. Results show that calibration‐validation strategies for policy selection coupled with certain ensemble aggregation methods can improve out‐of‐sample performance in water supply and flood risk objectives over baseline performance given fixed computational costs. Our findings highlight the potential to improve policy search methodologies by leveraging these well‐established model training strategies from machine learning.