Streamflow scenario tree reduction based on conditional Monte Carlo sampling and regularized optimization

• Published in: Journal of hydrology (Amsterdam) Vol. 577; p. 123943
• Main Authors: Li, Jinshu, Zhu, Feilin, Xu, Bin, Yeh, William W.-G.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2019
• Subjects: algorithms; China; covariance; Monte Carlo sampling; probability; Regularized optimization; Reservoir operation; sampling; Scenario tree reduction; stream flow; Streamflow scenario tree generation; trees; variance; China; Regularized optimization; Scenario tree reduction; Streamflow scenario tree generation; Reservoir operation; Monte Carlo sampling
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2019.123943

•Develops a conditional Monte Carlo sampling method for selecting representative scenario subsets.•Develops a regularized optimization model based on ridge regression for moment matching.•Proposes a novel method for streamflow scenario tree reduction and conducts numerical experiments using real-world data. Streamflow scenario tree reduction is essential for alleviating the computational burden of a stochastic programming with recourse model. This paper develops a new streamflow scenario tree reduction method aimed at preserving important statistical moment information and maintaining streamflow scenario probability. Specifically, we first employ a neural gas algorithm for scenario tree generation, then establish a stepwise conditional Monte Carlo sampling method for systemically reducing the number of scenarios from the full tree. We then develop a regularized optimization model based on ridge regression and moment matching to determine the posterior scenario probability. We apply the proposed method to the Qingjiang cascade reservoir system in China. The results show that the reduced tree with 35% reduction level can still maintain robust moment preservations, including the mean, variance, lag-one covariance, cross-site covariance, and scenario probability. Additionally, the stability test indicates that the proposed conditional Monte Carlo sampling method is stable and converges within a reasonable number of scenario combinations.