Comparing the Robustness of Forecast-Informed Reservoir Operating Policies under Forecast Uncertainty and Hydrologic Extremes

March 4, 2026

A new article, “Comparing the Robustness of Forecast-Informed Reservoir Operating Policies under Forecast Uncertainty and Hydrologic Extremes” by Will Taylor (University of California, Davis, U.S. Air Force), Zach Brodeur (Cornell University, CW3E), Jon Herman (University of California, Davis), and Scott Steinschneider (Cornell University), has been published in the Journal of Water Resources Planning and Management, a journal of the American Society of Civil Engineers.

This work represents continued progress in the ongoing research collaboration between UC Davis, CW3E, and Cornell University to research FIRO strategies through the integration of synthetic forecasts (Brodeur et al., 2025). Together, these efforts advance innovative research that test FIRO strategies over many more events and a wider range of forecast conditions than traditional hindcasts alone. It directly supports the FIRO: Resilient Water Management priority outlined in CW3E’s Strategic Plan.

The study uses Lake Oroville on California’s Feather River as a case study (Ralph, et al., 2025). Lake Oroville is the State Water Project’s largest reservoir, providing flood protection, hydropower, and water supply for millions of Californians.

This study compares three potential FIRO operating policies, all driven by 1-14 day ensemble inflow forecasts from the California-Nevada River Forecast Center:

• Ensemble Forecast Operations (EFO), which uses an optimized “risk curve” to decide when and how aggressively to draw down storage based on how many ensemble members exceed the top of conservation storage (Delaney et al., 2020).
• Model Predictive Control (MPC), which re-optimizes releases each day over a rolling forecast horizon to minimize penalties for both approaching spill and drawing the reservoir too low.
• A cumulative inflow guide-curve method, similar to approaches already being explored in Water Control Manual updates, which triggers releases based on the 70th percentile 1–5 day cumulative inflow forecast.

The policies aim to minimize the required flood pool volume needed to safely manage inflow extremes. They are benchmarked against policies using perfect 14-day forecasts and a baseline no-skill forecast.

The experimental design uses synthetic ensemble forecasts developed in Brodeur et al. (2025) for both training and testing of FIRO strategies. Synthetic ensemble forecasting enables the stochastic generation of ensemble forecast sequences against any available hydrologic sequence while preserving the skill characteristics of the original hindcast source data (CNRFC HEFS). Importantly, synthetic forecasting allows the generation of forecast sequences against not only existing historical hydrologic events at Oroville (1986, 1997, 2006, and 2017), but also scaled versions of these events (i.e. ‘100/200/300 Year Scaled’ in Figure 1) that represent plausible extremes not seen in the relatively short observed record.

Figure 1 shows the variability in FIRO strategy performance due to forecast uncertainty across many different event sizes. For instance, the MPC strategy shows high performance across scaled events for the original HEFS forecast (green stars), but a much higher degree of variability across the 100 samples of synthetic forecasts (green boxplots) as compared to other FIRO strategies (EFO, cumulative method).

Figure 1. Required flood pool by policy type and training method. Stars are simulated with HEFS forecasts, and boxplots are simulated with synthetic ensemble forecasts (100 samples). The EFO policies trained with synthetic forecasts are tested against a random synthetic ensemble different than the one used in training to determine the placement of the star. Figure 3 from Taylor et al. (2026).

For scaled events less than a 300-year event, nearly all the FIRO strategies analyzed in Figure 1 show a substantially reduced required flood pool size as compared to the current flood pool size at Oroville or the baseline strategy. The perfect forecast strategy has a required flood pool at or near zero for all event sizes, showing the value of improved skill, particularly for the 200- and 300-year extremes. The simplified cumulative method underperforms other FIRO methods, demonstrating that simplification of release rules and reduced lead-time horizons can limit performance. Lastly, the results for the EFO strategies trained with synthetic forecasts, as opposed to HEFS (i.e. EFO – Synthetic Trained), suggest that incorporating synthetic forecast uncertainty into FIRO policy training can improve out-of-sample robustness.

Figure 2. Drawdown weight effect on required flood pool. The symbols represent the individual drawdown weight value used to simulate each method. The MPC baseline (blue dot) and cumulative method (gray dot) do not have multiple series since they do not use drawdown weights in their formulation. The MPC perfect results are in orange (not shown in legend). Figure 4 from Taylor et al. (2026).

The objective function formulation uses a weighting value that reflects the degree of forecast risk aversion. Figure 2 shows that the FIRO strategies exhibit differing degrees of sensitivity to this weighting in relation to the required flood pool, especially as flood risk aversion increases (lower weight values). In many cases the choice of weight influences the performance more than the type of operating policy. This sensitivity is also highly dependent on the event and forecast sequencing, as each of the 4 events shown (1986, 1997, 2006, 2017) are scaled to the same 200-year magnitude.

All the FIRO strategies can maintain maximum drawdowns well below remaining water year inflow for Oroville (Figure 3), allowing refill even if forecasts are biased high. Nevertheless, other reservoir contexts with less reliable late-season inflows (e.g., snowmelt) may be more sensitive to these pre-event drawdowns.

Figure 3. Comparison of maximum drawdown and remaining water year inflow. Each point represents a single water year over the hindcast record (1990–2019) for each policy type and weight. Figure 6 from Taylor et al. (2026).

This study demonstrates a variety of tradeoffs for designing FIRO strategies using synthetic ensemble forecasts. MPC effectively leverages skillful forecasts, but may be more sensitive to forecast error. On the other hand, simplified release rules with shortened lead-time horizons are less sensitive to forecast uncertainty, but they are also less effective in leveraging skill and may limit FIRO benefits. Overall, the approach developed in this study offers a powerful toolkit to understand the role of hydrologic variability and forecast skill in the performance of FIRO policies and expands the role of synthetic forecasts, which are currently being integrated into ongoing FIRO viability assessments (Seven Oaks, Howard Hanson, Lake Sonoma). It also contributes to the growing evidence that FIRO can safely unlock more flexible use of existing storage, which remains a central goal of research and operations partnerships between CW3E and FIRO partners.

Citations:

Brodeur, Z. P., Taylor, W., Herman, J. D., & Steinschneider, S. (2025). Synthetic ensemble forecasts: Operations‐based evaluation and inter‐model comparison for reservoir systems across California. Water Resources Research, 61, e2024WR039324. https://doi.org/10.1029/2024WR039324

Delaney, C. J., Robert Hartman, Hartman, R., Mendoza, J., Dettinger, M. D., Delle Monache, L., Jasperse, J., Ralph, F. M., Talbot, C. A., Brown, J. D., Reynolds, D. W., & Evett, S. (2020). Forecast Informed Reservoir Operations Using Ensemble Streamflow Predictions for a Multipurpose Reservoir in Northern California. Water Resources Research, 56(9), 26604. https://doi.org/10.1029/2019wr026604

Ralph, F. M., James, J., Leahigh, J., White, M., Anderson, M., Talbot, C., Forbis, J., Fromm, J., & Haynes, A. (2025). Yuba-Feather Forecast Informed Reservoir Operations: Final Viability Assessment. University of California, San Diego. https://cw3e.ucsd.edu/FIRO_docs/Yuba-Feather_FVA/Yuba-Feather_FVA.pdf

Taylor, W., Brodeur, Z. P., Steinschneider, S., Kucharski, J., & Herman, J. D. (2024). Variability, Attributes, and Drivers of Optimal Forecast-Informed Reservoir Operating Policies for Water Supply and Flood Control in California. Journal of Water Resources Planning and Management, 150(10), 05024010. https://doi.org/10.1061/JWRMD5.WRENG-6471

Taylor, W., Brodeur, Z. P., Steinschneider, S., & Herman, J. D. (2026). Comparing the Robustness of Forecast-Informed Reservoir Operating Policies under Forecast Uncertainty and Hydrologic Extremes. Journal of Water Resources Planning and Management, 152(4), 04026005. https://doi.org/10.1061/JWRMD5.WRENG-7250