| Summary: | Snow water equivalent (SWE) expresses amount of liquid water in the snow pack. This information is crucial in many hydrological models, but the actual measurement of SWE requires a lot of labor demanding manual work. Because these in situ measurements are labor demanding and often scarce, efficient, and accurate, forecasts of SWE are crucial to get reliable estimates from current and future state of the hydrological systems. In this work, we compare two different recurrent neural network (RNN) architectures: long short-term memory (LSTM) and gated recurrent unit (GRU). We show that GRU is as accurate as LSTM, but requires less computation and is thus preferred RNN architecture for SWE forecasting. We optimise the efficiency and accuracy of these RNNs with hyper parameter Bayesian optimisation, and comprehensive data preprocessing. We furthermore improve the accuracy of the forecasts by using time to vector representation of time in the RNN architecture. We also show that the RNNs have great generalisation capabilities in SWE forecast, outperforming the commonly used degree-day physical model in generalisation capabilities and forecast accuracy. All the results are demonstrated with real in situ measurements of SWE from Finland, which were carefully selected to present great variety in natural environments. As the outcome, we propose two models which can be efficiently applied to accurate forecasts of SWE: lightweight and heavy. Lightweight model has 321 trainable parameters in total, and has average NSE of 0.91. Heavy model has 51973 trainable parameters, and average NSE of 0.95.
|
|---|