Improved reservoir sizing utilizing observed and reconstructed streamflows within a Bayesian combination framework

Abstract

Reservoir sizing is a critical task as the storage in a reservoir must be sufficient to supply water during extended droughts. Typically, sequent peak algorithm (SQP) is used with observed streamflow to obtain reservoir storage estimates. To overcome the limited sample length of observed streamflow, synthetic streamflow traces estimated from observed streamflow characteristics are provided with SQP to estimate the distribution of storage. However, the parameters in the stochastic streamflow generation model are derived from the observed record and are still unrepresentative of the long-term drought records. Paleo-streamflow time series, usually reconstructed using tree-ring chronologies, span for a longer period than the observed streamflow and provide additional insight into the preinstrumental drought record. This study investigates the capability of reconstructed streamflow records in reducing the uncertainty in reservoir storage estimation. For this purpose, we propose a Bayesian framework that combines observed and reconstructed streamflow for estimating the parameters of the stochastic streamflow generation model. By utilizing reconstructed streamflow records from two potential stations over the Southeastern U.S., the distribution of storage estimated using the combined streamflows is compared with the distribution of storage estimated using observed streamflow alone based on split-sample validation. Results show that combining observed and reconstructed streamflow yield stochastic streamflow generation parameters more representative of the longer streamflow record resulting in improved reservoir storage estimates. We also generalize the findings through a synthetic experiment by generating reconstructed streamflow records of different sample length and skill. The analysis shows that uncertainty in storage estimates reduces by incorporating reconstruction records with higher skill and longer sample lengths. Potential applications of the proposed methodology are also discussed.