Deciphering the Mechanism of Better Predictions of Regional LSTM Models in Ungauged Basins

Abstract

Prediction in ungauged basins (PUB) is a concerning hydrological challenge, prompting the development of various regionalization methods to improve prediction accuracy. The long short-term memory (LSTM) model has gained popularity in rainfall-runoff prediction in recent years and has proven applicable in PUB. Prior research indicates that incorporating static attributes in the training of regional LSTM models could improve performance in PUB. However, the underlying reasons for this enhancement have received limited exploration. This study aims to explore the role of static attributes in the training of the regional LSTM model. It is assumed that the regional LSTM model can induce streamflow generation mechanisms with the incorporation of static attributes and apply certain streamflow generation mechanisms to ungauged catchments based on their attributes. To this end, a grouping-based training strategy is proposed, that is, training and validating regional LSTM models on catchments with similar streamflow generation mechanisms within predefined groups. The training strategies of regional LSTM models, either incorporated with static catchment attributes or based on classification, are conducted in 363 catchments. Results demonstrate a high level of consistency in the enhancement achieved by the two training strategies. Specifically, 192 and 216 catchments exhibit enhancement compared to traditionally trained models without inclusion of attributes, with 132 catchments showing improvement under both training strategies. Furthermore, the findings indicate consistent spatial patterns and attribute distributions of enhanced catchments, as well as the notable improvement in reproducing low flow-related hydrological signatures.