Runoff predictions in ungauged basins using sequence-to-sequence models

Abstract

How to improve the performance of runoff predictions in ungauged basins (PUB) is challenging. Recently, the long short-term memory (LSTM) based models have excellent performance and receive many attentions. In this paper, to improve the performance for 1-day-ahead runoff PUB and provide good performance for multi-day-ahead runoff PUB, we propose four sequence-to-sequence (S2S) models to deal with PUB. Furthermore, we introduce two modules named attribute-weighting module and multi-head-attention module for improving the performance. To show the power of these S2S models and the advantages of those two modules, we test our four S2S models and also three benchmark models including a PUB-LSTM model and two process-driven models by using k-fold validation on 531 basins of the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) dataset. For 1-day-ahead runoff PUB, the median and the mean of Nash–Sutcliffe efficiency for the 531 basins provided by our best model achieve 0.78 and 0.70, respectively, while those provided by the PUB-LSTM model (the best one among those three benchmark models) are 0.69 and 0.54, respectively. Besides, our S2S models have promising results for multi-day-ahead runoff PUB. For 7th-day-ahead runoff PUB, the median and the mean of Nash–Sutcliffe efficiency for the 531 basins provided by our best model are 0.68 and 0.57, respectively. Furthermore, the results also show that the two modules are beneficial for improving the performance.