On the role of the architecture for spring discharge prediction with deep learning approaches

Abstract

AbstractUnderstanding karst spring flow is important to accommodate the increasing water demand caused by the population growth and manage the freshwater water resource effectively. However, due to the spatial and temporal heterogeneity and complex hydrological processes in karst systems, predicting karst spring discharge remains challenging. In this study, three deep learning‐based models, including long short‐term memory (LSTM), gated recurrent unit (GRU) and simple recurrent neural network (RNN), are framed with an encoder–decoder architecture to provide multiple‐step‐ahead spring discharge prediction. The encoder–decoder architecture includes an encoder that reads and encodes the input sequence into a vector and decoder that deciphers the vector and outputs the predicted sequence. Three hybrid models called LSTM‐ED, GRU‐ED and simple RNN‐ED are compared with single‐step models and multiple‐step models without the encoder–decoder architecture to investigate the role of the encoder–decoder architecture on multi‐step‐ahead prediction. The sensitivity of the selection of input time and lead time steps on the karst spring discharge prediction is evaluated. The predicted results are compared with the observed spring discharge. It implies that: (1) LSTM‐ED, GRU‐ED and RNN‐ED models obtain similar results on predicting karst spring discharge multiple time steps ahead; (2) three hybrid multiple‐step models outperform the single‐step models in making consistent and accurate spring discharge predictions; (3) the multiple‐step models framed with an encoder–decoder architecture obtain better spring discharge prediction results than the single‐step models and multiple‐step models without the encoder–decoder structure; (4) the LSTM‐ED, GRU‐ED and simple RNN‐ED models are sensitive to the selection of lead time and insensitive to the selection of input time step. A short lead time typically yields a more accurate spring discharge prediction.