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.

The introduction of stochastic streamflow models by Fiering (1967), Maass et al. (1962), and others led to a revolution in water resources planning, design, and management. These models enabled hydrologists to generate representative streamflow ensembles over future planning horizons, needed to explore the consequences of future hydrologic conditions not experienced historically, and formally characterize the reliability, vulnerability, and resilience of water resource systems (Hashimoto et al., 1982;Loucks & van Beek, 2017). Traditional stochastic streamflow models are typically statistical models rather than mechanistically driven hydrologic models. Such stochastic streamflow models may be adjusted to reflect changes in seasonality or other statistical properties of flow (e.g., Quinn et al., 2018), but tying statistical hydrologic changes to climate and land use change is

The use of hydro-meteorological forecasts in water resources management holds great promise as a soft pathway to improve system performance. Methods for generating synthetic forecasts of hydro-meteorological variables are crucial for robust validation of forecast use, as numerical weather prediction hindcasts are only available for a relatively short period (10-40 years) that is insufficient for assessing risk related to forecast-informed decision-making during extreme events. We develop a generalized error model for synthetic forecast generation that is applicable to a range of forecasted variables used in water resources management. The approach samples from the distribution of forecast errors over the available hindcast period and adds them to long records of observed data to generate synthetic forecasts. The approach utilizes the Skew Generalized Error Distribution (SGED) to model marginal distributions of forecast errors that can exhibit heteroskedastic, auto-correlated, and non-Gaussian behavior. An empirical copula is used to capture covariance between variables, forecast lead times, and across space. We demonstrate the method for medium-range forecasts across Northern California in two case studies for (1) streamflow and (2) temperature and precipitation, which are based on hindcasts from the NOAA/NWS Hydrologic Ensemble Forecast System (HEFS) and the NCEP GEFS/R V2 climate model, respectively. The case studies highlight the flexibility of the model and its ability to emulate space-time structures in forecasts at scales critical for water resources management. The proposed method is generalizable to other locations and computationally efficient, enabling fast generation of long synthetic forecast ensembles that are appropriate for risk analysis.

Forecasts of heavy precipitation delivered by atmospheric rivers (ARs) are becoming increasingly important for both flood control and water supply management in reservoirs across California. This study examines the hypothesis that medium-range forecasts of heavy precipitation at the basin scale exhibit recurrent spatial biases that are driven by mesoscale and synoptic-scale features of associated AR events. This hypothesis is tested for heavy precipitation events in the Sacramento River basin using 36 years of NCEP medium-range reforecasts from 1984 to 2019. For each event we cluster precipitation forecast error across western North America for lead times ranging from 1 to 15 days. Integrated vapor transport (IVT), 500-hPa geopotential heights, and landfall characteristics of ARs are composited across clusters and lead times to diagnose the causes of precipitation forecast biases. We investigate the temporal evolution of forecast error to characterize its persistence across lead times, and explore the accuracy of forecasted IVT anomalies across different domains of the North American west coast during heavy precipitation events in the Sacramento basin. Our results identify recurrent spatial patterns of precipitation forecast error consistent with errors of forecasted synoptic-scale features, especially at long (5–15 days) leads. Moreover, we find evidence that forecasts of AR landfalls well outside of the latitudinal bounds of the Sacramento basin precede heavy precipitation events within the basin. These results suggest the potential for using medium-range forecasts of large-scale climate features across the Pacific–North American sector, rather than just local forecasts of basin-scale precipitation, when designing forecast-informed reservoir operations.

The rising frequency and intensity of heavy precipitation events due to climate change poses a significant operational challenge for water management. The use of short-term precipitation forecasts in reservoir operations has long been recognized as an opportunity to mitigate these extremes (e.g., Kelman et al., 1990;Stedinger et al., 1984;Yao & Georgakakos, 2001). Advances in forecast-informed reservoir operations (FIRO) have gained traction as a risk-based approach that conditions release decisions on hydrologic ensemble forecasts to better manage reservoir flood pool levels, optimize surface and groundwater storage, and meet environmental flow regulations (Delaney et al., 2020;Jasperse et al., 2020;Nayak et al., 2018). While many FIRO approaches directly incorporate forecast uncertainty into the decision-making process, one remaining challenge is that the forecasts and their uncertainty are generally not tailored for the decision under consideration (Turner et al., 2017). For instance, water managers who prioritize safety from flood hazards over water supply concerns may prefer forecasts that reliably capture the occurrence of heavy precipitation and floods, even if this leads to more false alarms. This study advances the development of such tailored information using deep learning techniques to address spatial error in precipitation forecasts, with the goal of producing more flexible forecasting products to incorporate within adaptive operating policies (Giuliani et al., 2021;Herman et al., 2020).The accuracy of precipitation forecasts from numerical weather prediction models has significantly improved over the last several decades (Bauer et al., 2015), but structured forecast biases are still common (Wilks, 2019). Several sources of uncertainty cause forecast bias, including model design, parameterization, and initial conditions. To address these issues, many statistical post-processing techniques have been proposed to bias-correct forecasts. These techniques, broadly termed model output statistics (MOS; Wilks, 2019), have traditionally relied on multiple regression frameworks, although machine learning (ML) methods have recently been designed for this task (Cho et al., 2020;Han et al., 2021). Traditional MOS approaches correct biases in precipitation magnitude that can result from imprecise parameterizations or coarse model resolution. However, these approaches are usually not well suited to account for spatial biases in the forecasted location of large-scale storm systems, because they are conditioned on forecasted precipitation at a specific location.