Snow water equivalent forecasting in sub-arctic and arctic regions: Efficient recurrent neural networks approach

Snow water equivalent (SWE) expresses the amount of liquid water in the snow pack. Accurate SWE forecasts are essential for reliable hydrological modeling, as direct SWE measuring is labor-intensive. In this study, we systematically compared gated recurrent unit (GRU) and long short-term memory (LSTM) architectures, showing that GRU models achieve comparable accuracy with greater computational efficiency. By applying Bayesian optimization, data preprocessing, and a time-to-vector representation of temporal features, we introduce two novel GRU-based models: a lightweight model (321 parameters; average NSE = 0.91) and a more complex model (51973 parameters, average NSE = 0.95). Importantly, these models generalize effectively across geographically distant stations, demonstrating robust predictive performance under varied climatic conditions. The primary novelty of our study is identifying GRU as a computationally efficient, accurate alternative to LSTM in SWE forecasting, combined with demonstrating that compact models with smaller hidden states provide strong spatial generalization and excellent accuracy. • Recurrent neural networks show strong potential in snow water equivalent forecasting. • We compare long short term memory and gated recurrent units, detailing their design. • We propose two efficient gated recurrent unit models with state-of-the-art accuracy.