Machine learning-enhanced flood forecasting using snow water equivalent: validation across multiple flood events

ABSTRACT Reliable flood forecasting in snow-fed basins requires explicit consideration of snowpack dynamics, which strongly influence streamflow variability. This study enhances flood prediction by integrating snow water equivalent from multiple mountain stations and validating across three major flood events (2008, 2012, and 2013) in the Bow River Basin, Canada. Results show that while forecasts made 7 days ahead provide early flood signals, they carry notable uncertainty, with vertical peak flow errors exceeding 7% during extreme conditions. For the 2013 extreme flood event, 5-day forecasts demonstrate markedly improved performance, achieving NSE values above 0.90, reducing root mean square error from 41.7 to 1.61 units, and lowering vertical error to below 1%. Three-day forecasts perform even better, achieving NSE values of 0.99 and minimal timing error. These results indicate that the 5-day horizon represents the optimal balance between early warning capability and actionable reliability. The findings also highlight that incorporating snow water equivalent significantly enhances prediction accuracy, improving NSE by 0.20–0.28 and reducing magnitude error by more than 50% compared to models without snowpack data. Overall, this work underscores the importance of integrating mountain snowpack information into flood forecasting frameworks to strengthen early warning systems and operational decision-making.