Abstract:
Hydrological models are one of the key challenges in hydrology. Their goal is to
understand, predict and manage water resources. Most of the hydrological models
so far were either physical or conceptual models. But in the past two decades,
fully data-driven (empirical) models started to emerge with the breakthroughs
of novel deep learning methods in runoff prediction. These breakthroughs were
mostly favored by the large volume, variety and velocity of water-related data.
Long Short-Term Memory and Gated Recurrent Unit neural networks, particularly
achieved the outstanding milestone of outperforming classic hydrological models
in less than a decade. Moreover, they have the potential to change the way hydrological modeling is performed. In this study, precipitation, minimal and maximum
temperature at the Ansongo-Niamey basin combined with the discharge at Ansongo
and Kandadji were used to predict the discharge at Niamey using artificial neural
networks. After data preprocessing and hyperparameter optimization, the deep
learning models performed well with the LSTM and GRU respectively scoring a
Nash-Sutcliffe Efficiency of 0.933 and 0.935. This performance matches those
of well-known physically-based models used to simulate Niamey’s discharge and
therefore demonstrates the efficiency of deep learning methods in a West African
context, especially in Niamey which has been facing severe floods due to climate
change.
