Authors: Majid Hussain Shah – University of Alabama
Title: Sequence free LSTM training for long-term hydrological states
Presentation Type: Poster Presentation
Abstract: Standard Long Short-Term Memory (LSTM) networks are widely used in hydrologic modeling for learning nonlinear rainfall-runoff relationships, including in the Next Generation Water Resources Modeling Framework. However, conventional training strategies rely on shuffled mini-batches and truncated sequences that disrupt temporal continuity and limit the effective use of long-term hydrologic memory. This limitation disrupts process representation, which depends strongly on antecedent conditions (e.g., snowpack). When hidden states are repeatedly reset during training, the representation of watershed memory is weakened. We suggest a persistent LSTM framework that keeps hidden states that are specific to each basin across time segments and allows for scalable training across multiple basins. To encourage generalization, batches are shuffled across basins. To make sure that the physical consistency is kept, strict chronological ordering is maintained within each basin. Using non-overlapping time segments cuts down on redundant calculations even more by only processing each time step once. This makes training more efficient. When used on large-sample hydrology with hourly resolution, the persistent LSTM performance is acceptable while keeping long-term hydrologic history. This method offers a scalable and computationally efficient means of integrating physically significant temporal continuity into deep learning models for flood prediction, drought evaluation, and climate-resilient water resource management.