Configuration over Architecture: Deep Learning Models for Daily Streamflow Forecasting in Snowmelt-Dominated Watersheds of Southwestern Alberta

Accurate streamflow forecasting in snowmelt-dominated headwaters is critical for flood mitigation, reservoir operations, and water resource management, yet it remains challenging due to non-linear interactions among snowmelt dynamics, meteorological forcing, and watershed hydrology. This study evaluates Long Short-Term Memory (LSTM) models for daily streamflow prediction in three mountainous watersheds of southwestern Alberta and systematically examines the sensitivity of model performance to key configuration decisions, including timestep selection, feature engineering, seasonal period, adjacent gauge inclusion, weather forecasting, and attention mechanisms. Results demonstrate high predictive accuracy of up to 0.98 NSE for typical daily streamflow when the models are optimally configured. However, all models substantially underpredicted a 1995 extreme flood event by 44–78%, reflecting the inherent challenge of predicting rare events with imbalanced training data. The extreme-event underprediction was masked by aggregate evaluation metrics over a typical flow period, which necessitated a separate evaluation for the extreme event in the LSTM performance evaluation. These findings provide guidance for configuring LSTM models for small watersheds and highlight data augmentation and uncertainty quantification as future directions for improving extreme event predictions.