Bagging-based Adaptive Synthetic Oversampling Algorithm for Robust Policy Search in Reservoir Management

Abstract

Abstract Reliable policy search is essential in improving reservoir operations to satisfy multi-sectoral needs such as flood control and water supply. Given its importance, this topic has been widely explored in reservoir control studies. However, previous studies have observed that optimized policies tend to overfit to the training data, and are thus prone to be controlled mainly by infrequent extreme samples in the training data. This study proposes a bootstrap aggregation (bagging)-based Adaptive Synthetic (ADASYN) algorithm as an extension of the ADASYN and bagging techniques originated by machine learning literature. We illustrate the effectiveness of the bagging-based ADASYN algorithm using a case study of the Folsom Reservoir in Northern California with a binary tree-based control policy. The proposed algorithm variants are also developed to confirm the usefulness of the individual technique embedded in the final procedure. Results demonstrate that the proposed algorithm yields significant improvements in managing water supply and flood risks. In the proposed algorithm, the ADASYN technique facilitates creating a reliable set of policy trees while generating synthetic samples in reservoir inflow to augment infrequent extreme samples. Moreover, the bagging technique is beneficial in selecting the final policy tree while leading to improved out-of-sample performance. We conclude that this case study using the novel ADASYN algorithm highlights the potential to improve policy search algorithms by utilizing well-established training strategies from machine learning.