Henok Teklu, Matjaz Gams and Maciej Wielgosz

ABSTRACT
This study explores the use of Liquid Neural Networks (LNNs)
to predict runoff for one, three, and six days ahead, highlighting
their superior performance compared to traditional models such
as Artificial Neural Networks (ANNs), Model Trees (MTs), and
Long Short-Term Memory (LSTM) networks. LNNs leverage a
dynamic reservoir of neurons, enabling them to capture complex
temporal dependencies inherent in the rainfall-runoff process.
The study employs a case analysis of the Sieve River basin, using
historical hydrological data to train and evaluate the models. The
results demonstrate that LNNs consistently outperform other
models across all prediction horizons, achieving the lowest Root
Mean Square Error (RMSE) and Normalized Root Mean Square
Error (NRMSE) values, and the highest Coefficient of Efficiency
(COE). This indicates that LNNs are highly effective for both
short-term and long-term hydrological forecasting, offering
significant potential for enhancing water resource management
and flood prediction strategies.