Improving snow water equivalent prediction with a hybrid SNOW17-LSTM model
Aamir Raj Lamichhane1 and Andrew Bennett1
1Department of Hydrology and Atmospheric Sciences
The University of Arizona, Tucson, AZ
Snowpack regulates streamflow and seasonal water availability in mountainous watersheds, making it a crucial component of the hydrologic cycle. The National Weather Service (NWS) uses SNOW17, a process-based model, to simulate snow ablation and accumulation in watersheds across the United States. However, process-based models rely on site specific parameterizations, limiting their transferability and making it difficult to capture snowpack dynamics across diverse conditions. In this work we present a novel hybrid model by integrating SNOW17 with a Long Short Term Memory (LSTM) deep learning model. We first conduct a benchmarking experiment comparing our model to standalone SNOW17 calibrated to individual sites as well as a purely LSTM based approach. Then, we will evaluate how the hybrid approach improves process representation by analysing key snowpack metrics such as timing and magnitude of peak snow water equivalent (SWE), melt timing, and responses to rain-on-snow events. Finally, we will assess the spatial transferability of the hybrid model by training it on a set of Snow Telemetry (SNOTEL) sites covering different climate types and testing its performance in unseen locations. This work provides insight into the advantages of integrating data-driven approaches with process-based models for snowpack modeling. Our hybrid model demonstrates improvement in predictions without loss in physical interpretability, which is especially important when using black-box approaches like the pure LSTM model.