Global Parameter Sensitivity in Forecast-Informed Reservoir Operations using Model Predictive Control

August 25, 2026

A new article, “Global Parameter Sensitivity in Forecast-Informed Reservoir Operations using Model Predictive Control,” by Alex Chen (University of California, Davis, CW3E), Jon Herman (University of California, Davis), Zach Brodeur (Cornell University, CW3E), Scott Steinschneider (Cornell University) and Brett Whitin [(NOAA/NWS California-Nevada River Forecast Center (CNRFC)], has been published in the journal Water Resources Research.

The goal of the study is to understand how reservoir properties can influence forecast-informed reservoir operations (FIRO) performance. This goal aligns with one of the key objectives of the FIRO: Resilient Water Management focus area in CW3E’s Five-Year Strategic Plan (2025-2029); namely, scaling FIRO implementation throughout California and eventually nationally. To support this objective, this study presents a global sensitivity and threshold analysis framework using a Model Predictive Control (MPC) operations strategy to evaluate the sensitivity of FIRO benefits to infrastructure and forecast parameters: the maximum release, ramping rate and available forecast lead time. The study efficiently identifies key constraints on achieving desired FIRO performance.

The framework is demonstrated through six major reservoirs in California using daily ensemble hydrologic forecasts from the CNRFC Hydrologic Ensemble Forecast Service (HEFS). The MPC performance of HEFS is benchmarked against a baseline policy (no forecast) and a perfect forecast policy. The performance is measured by the required flood pool (FP), the peak accumulated storage during the flood event (Fig. 1). By comparing MPC performance across reservoirs, with parameters expressed as percentage of the peak inflow, the results are driven by two factors: the default maximum release and the size of the design flood pool. Overall, the MPC-derived release sequences, under certain parameter combinations and reservoir properties, can achieve the key objectives of reducing flood risk while maintaining high storage to support water supply reliability during the largest flood event on record (January 1997).

Figure 1. Comparison of MPC-derived required flood pool (FP) across forecast policies and reservoirs. For each reservoir, the required FP under each forecast policy (B, P and H) is shown alongside the design flood pool (gray bar), peak inflow of the 1997 flood (hatched bar), and the maximum release R_max (yellow bar). The percentage associated with R_max indicates its ratio to the peak inflow. From Figure 4 in Chen et al. (2026).

A global sensitivity analysis is then performed to evaluate the impact of the three parameters on the flood pool (FP) reduction. The pairwise relationships of all three key infrastructure and forecast parameters for reaching a certain level of FP reduction are illustrated as heatmaps, using Oroville Reservoir as an example (Figure 2). The results reveal the threshold behavior where higher values of maximum release, ramping rate, and available forecast lead time are associated with greater flood pool reduction. In addition, the results also indicate that the HEFS release sequence closely captures the performance of the perfect release sequence, showing that the MPC strategy can effectively leverage the skill in the HEFS forecasts.

Figure 2. Pairwise relationships between the parameters and the flood pool (FP) reduction when the remaining parameter is fixed at the default value: (a) R_max vs. RR at LT = 14, (b) R_max vs. LT at RR = 16%, and (c) RR vs. LT at R_max = 48%. The color scale of the heatmaps indicates the level of FP reduction for the HEFS sequence. The solid contour lines on each panel indicate the conditions under which the HEFS release sequence yields 0% FP reduction, while the dashed contour lines represent conditions under which the perfect release sequence yields 0% FP reduction. These results are based on the example of Oroville Reservoir (ORDC1). From Figure 6 in Chen et al. (2026).

Finally, Decision Tree classifiers are used to extract key parameter thresholds across reservoirs. The results show that the maximum release is the most important feature for both the ensemble and perfect forecast sequences, with additional contributions from ramping rate and lead time (Figure 3). Furthermore, the extracted thresholds from the classifiers across reservoirs identify the general parameter conditions required to achieve varying levels of flood pool reduction. This large parameter sample reflects many potential reservoir configurations and scaled events, and in practice, these parameter thresholds can support the expansion of FIRO implementation by efficiently identifying reservoir system properties where forecast-informed policies are likely to provide the greatest benefits.

Figure 3. Results from Decision Tree classifiers trained with combined samples from all six reservoirs for the HEFS and perfect forecast release sequences. (a) The parameter thresholds (R_max(%), RR(%), and LT(days)) were extracted from the first level (root), second-level, and third-level nodes of the rightmost path of the trained classifiers at different levels of flood pool reduction threshold (FPT), which isolate outcomes for the better class distribution. For example, to achieve FPT=20%, the extracted thresholds suggest that MPC with the HEFS release sequence would require R_max(%) > 23%, RR(%) > 4% and LT > 2 days. (b) Feature importances of the trained classifiers. From Figure 9 in Chen et al. (2026).

This work, led by researchers at UC Davis, strengthens the ongoing FIRO research collaboration among UC Davis, CW3E, Cornell University and CNRFC. This work also advances the understanding of HEFS forecasts’ role in FIRO’s broader applicability and provides a mechanism to help identify new sites where infrastructure and forecast characteristics can support successful FIRO implementation.

Citation:

Chen, A. B., Herman, J. D., Brodeur, Z. P., Steinschneider, S., & Whitin, B. (2026). Global parameter sensitivity in forecast‐informed reservoir operations using model predictive control. Water Resources Research, 62(4), e2025WR041515. https://doi.org/10.1029/2025WR041515