| SLM machine learning for the estimation of hydrological response in real time dam operation during flood events |
|---|
|
학술지명 ICOLD
저자 강부식,노준우,허영택,김성훈,강태호
발표일 2025-05-19
|
|
Real-time dam operation during flood events requires the prediction of its operation results on dam and river stage safety. During the optimization process, various operation scenarios need to be tested in advance for its possible impact on reservoir structural stability and downstream river flooding within a limited time period, and thus, one of the challenges in dam operation is to reduce model computational time and demands while keeping sufficient estimation accuracy. Recently, machine learning (ML) has been gaining certain attention in the application of the reservoir operations with the merits that it can learn and estimate efficiently the essential part of a target system and recognizably reduce the computational time and demands. Recent studies have tested and suggested that conventional Long Short Term Memory (LSTM) ML architecture is not reliable in estimating the hourly rainfall-runoff process, and more advanced architectures such as multiple timescale (MTS) LSTM are necessary. This means that feature information (instead of raw input data) extractable from current and antecedent input traces needs be used as the input to LSTM for the estimation of the rainfall-runoff response in the hourly time scale. This process follows the Atkinson-Shiffrin memory model with the three memory components of sensory, short, and long-term memory. Based on the memory model, this study suggests ML architecture called SLM (Sensory to Long-term Memory) that adds a sensory memory layer to provide feature information to LSTM layers. The applications to Namgang basins in South Korea have shown that the SLM architecture can successfully reduce the computational time while keeping sufficient accuracy in the hourly rainfall-runoff response estimation. |