Abstract:
In this research, we study the ability of Long Short-Term Memory (LSTM) neu-
ral networks for forecasting karst spring discharge and compare their performance
against traditional process-based hydrological models. Two springs, the Qachqouch
and Assal spring were studied. The models were trained on five years of collected
data and tested on two years. The performance of these models is evaluated based on
standard metrics including Nash-Sutcliffe Efficiency (NSE) and Root Mean Square
Error (RMSE). The results show that LSTM models have comparable performance
relative to complex process-based models. LSTMs require less calibration param-
eters and less preprocessing of data than process-based models. Scores show that
LSTM are able to capture the non linear dynamics of Karst systems. The study
also explores the effect of changing input variables, sequence length and model ar-
chitecture to predict discharge. We notice that longer input sequences generally
enhance model efficacy, especially for capturing delayed hydrological responses typ-
ical of snow-governed systems. Additionally, results highlight that more complex
models having more layers are not better at predicting discharge.
Looking forward, a discussion on the potential of scaling these specific model to
regional ones is made. A proposed sensitivity analysis on input data aims to further
refine model performance. And finally, the potential of these models for studying
long term climate change impact is discussed.
