Authors: Mohammad Farmani – University of Arizona
Title: Scoring Rule-Based LSTM Model for Flow Forecasting
Presentation Type: Poster Presentation
Abstract:
Accurate representation of hydrologic signatures in data-driven streamflow models remains a key challenge, particularly in ungauged headwater basins. Here, we introduce a novel scoring rule–based loss function for a long short-term memory (LSTM) network that directly targets signature-specific behavior by incorporating probabilistic measures of the flow duration curve (FDC) and the falling and rising limb (FRL) for each signature, including baseflow index, high-flow magnitude, low-flow frequency, flow variability, and recession constant. Leveraging daily CAMELS basin attributes and accompanying meteorological forcings, a single LSTM ingests static and dynamic inputs via a spatiotemporal batching scheme. We compare model training using FDC- and FRL-based scoring rules to the traditional residual approach (Nash–Sutcliffe Efficiency) for penalizing errors. Results demonstrate that both FDC and FRL losses significantly enhance the model’s ability to reproduce key hydrologic signatures—especially in upper-catchment areas lacking gauge records—while maintaining competitive overall predictive skill. This framework offers an operationally viable pathway for calibrating and transferring LSTM-based hydrologic models across gauged and ungauged sites.