Author(s): Yunus Ziya Kaya; Nigel George Wright; Xilin Xia

Linked Author(s):

Keywords: Entity aware LSTM; Caravan data set; Streamflow prediction; Static and dynamic attributes

Abstract: In recent studies, the daily discharge has been successfully predicted by using Long Short-Term Memory (LSTM). In this study, we used one of the largest hydrological datasets, namely the Caravan (Camels-GB), to perform various Long Short-Term Memory (LSTM) models for streamflow predictions of the UK. In these LSTM models, we not only tested the performance of the LSTM models countrywide, but also tested the combinations of the input parameters, which can be explained as dynamic and static attributes in the Caravan data set. As one of the goals of the study was to check the applicability of the LSTM to the large hydrological data sets, we did not perform any preprocessing on the data set, such as removing extreme records. We performed LSTM models by dividing the data set into training, validation, and test sets. We used 406 of the 408 stations available for the UK in the Caravan dataset for the training, validation, and test processes. The length of the data set was not sufficient at one of the stations, so we did not include that one. We kept 1 selected station for hold-out testing purposes. For evaluation purposes, we created spatial distribution maps for each version of LSTM based on Nash-Sutcliffe Efficiency (NSE) values.

DOI:

Year: 2026