Precise prediction of the future streamflow is an important step when building an effective flood or low-flow warning system. Over the years, hydrologists have developed and used complex process-based hydrological models to solve the streamflow prediction task. However, recently a specific type of data-driven model, more precisely Recurrent Neural Networks (RNNs), have shown impressive results in many sequence predictions and time series forecasting tasks. These data-driven models are comparatively simple to build and might be able to perform equally good as well as complex process-based models. The thesis explores two RNN architectures and their applicability to the streamflow prediction problem. The two architectures are a multi-shot sequence-to-sequence and an encoder-decoder architecture. We show how to implement and apply both architectures. We present hyperparameter tuning results for the architectures and the performance results of the best-tuned models. Finally, we provide a performance comparison between the best multi-shot sequence-to-sequence model, the best encoder-decoder model, and a benchmark process-based model. The encoder-decoder architecture has several architectural advantages over the multi-shot sequence-to-sequence model. Still, the results show that both architectures can predict streamflow with an accuracy similar to the benchmark process-based model.

Keywords: RNNs, LSTMs, sequence to sequence, encoder-decoder, streamflow predictions