Regional vs local LSTM models for short-term streamflow forecasting under operational constraints

Abstract

Reliable short-term streamflow forecasting remains a key challenge due to data latency, uncertainty, and other real-world constraints. This study presents a regional Long Short-Term Memory (LSTM) model to predict daily mean and maximum streamflow across 340 points in Chile over a five-day horizon, explicitly accounting for operational limitations such as unavailable recent streamflow and delayed input data. Compared to locally trained models, the regional model demonstrates superior performance in temporal correlation and variance representation, with Kling–Gupta Efficiency (KGE) $\ge$ 0.6 at 156 points. Crucially, high-flow event prediction improves significantly: bias in the Fractional High-flow Volume (FHV) is reduced by $\sim$50% at the 90th percentile and $\sim$25% at the 99th percentile, demonstrating strong operational robustness with minimal degradation over the forecast horizon. These findings highlight the potential of regional deep learning models to offer scalable and resilient performance across diverse hydrological settings, supporting flood preparedness and water management.