1
Introduction
In recent years, due to the development of social economy and the impact
of human activities, the trend of global climate change has become more
apparent. Climate change may cause dramatic effects in hydrological
processes (Hoang et al., 2016). These effects may lead to more frequent
hydrological extremes such as droughts (Puri et al., 2011) and floods
(Gu et al., 2015; Ahmad and Simonovic, 2001) and pose a potential threat
to water security. IPCC reports (2001; 2007) also show that the spatial
and temporal distributions of water resources have changed and are
expected to change in the future.
In the past 100a, the temperature in China increased about 0.5-0.8℃
(Ding et al., 2006). Strong evidences show that warming climate mainly
affect some large-scale basins in China (Yang et al., 2012; Zhang et
al., 2006; He, 2017), such as Yangtze River, Yellow River, and Pearl
River. Wang et al. (2012) found the annual runoff in the whole China may
increase by about 3-10 percent by 2050 with quite uneven spatial and
temporal distributions. Meanwhile, large-scale basins usually play
significant roles in water supply, energy production, and navigation.
Climate changes also may cause severe social and economic damages to
humanity (Simonovic and Ahmad, 2005). Thus, it is particularly essential
to study and predict the change in the hydrological cycle regarding
future climate for guaranteeing water security.
Currently, one commonly-used method for projecting river discharges
under climate change scenarios is using hydrological models fed by the
outputs of climate models (Ghosh and Mujumdar, 2008; Gardner, 2009;
Shivam et al., 2019). One approach of hydrologic modeling is physical
models (e.g., SWAT, TOPMODEL and HYPE) fed with data on climate, land
use, vegetation, etc. A physical model focuses on the analysis of
hydrological processes by its abundant components. Although widely used,
it still has the limitations of too complex structures and high data
requirements (Partington et al., 2012). In addition, many model
parameters only depend on experiences (Gao et al., 2009), which may
affect accuracies of physical models. Another approach is data-driven
models which just focuses on the relationship between inputs and outputs
rather than the internal structure (Modarres, 2009). Due to the simple
focus, data-driven models always have less demand on data. In general,
precipitation and climate data are easily to obtain compared with soil,
vegetation, and groundwater data (Shoaib et al., 2014). Thus,
data-driven models have obvious advantages than physical models in terms
of predicting discharge changes regarding future warm climate
(Inmaculada et al., 2007).
Lots of statistical and autoregressive models are chosen frequently in
hydrological forecasting (Drogue et al., 2004; Cheng et al., 2016;
Okkonen and Klove, 2010). In addition to using traditional regression
analysis methods, some papers focus on machine learning recently.
However, most of them are based on “shallow” machine learning methods,
such as stepwise cluster analysis (Fan et al., 2016), back propagation
(Wang, 2010), support vector machine (Ajay et al., 2013) and artificial
neural network (Zeng et al., 2012). It can be seen that machine learning
shows better performances than traditional regression models in
hydrological forecasting. Nevertheless, the shallow learning category is
always limited by overfitting and local optimum (McInerney et al.,
2017), restricting its reliability in hydrological forecasting.
At the same time, deep learning is effective at identifying complex data
features, has relatively high accuracies, and promotes development of
data-driven models. Due to its great advantage in solving complex
problems, this approach has been widely used in many fields recently,
such as image recognition (Chen et al., 2016; Postadjian et al., 2018),
speech recognition (Graves et al., 2013), and human behaviors (Vu et
al., 2015). Its application in system simulation has significantly
improved forecasting efficiencies, beyond shallow learning methods
drastically. In this sense, deep learning provides an advanced tool for
hydrological forecasting (LeCun et al., 2015).
Comparisons of deep learning and physical models in hydrologic modeling
showed the former has better performances (Tian et al., 2018). However,
few comprehensive study of deep learning methods for hydrological
modeling was reported in literature (Shen et al., 2018), especially
those regarding watersheds in China. Existing related studies focused on
runoff series, so linking hydrology with other variables (e.g.,
meteorology) deserves more attempts (Bai et al., 2016; Cheng et al.,
2016). Furthermore, hydrological application of deep learning has often
been probed just in a single basin (Hu et al., 2018), lacking of
effective comparisons for a variety of scenarios. Generally, one study
applying deep learning into hydrological prediction under climate change
over multiple typical basins in China has not been explored.
Thus, this study aims to explore the potential advantages of deep
learning methods in hydrological prediction and to reveal the response
of hydrological systems over three typical basins in China under climate
change conditions. These explorations will provide scientific support
for flood control and reduction, and water resources planning and
management under climate change conditions. In this article, Section 2
introduces the data and methods in the study, including climate models,
hydrological models, and simulation methods. Section 3 describes the
research area. Section 4 presents the simulation and projection results,
mainly including climate simulation, hydrological simulation, climate
projection, and hydrological forecast. Section 5 concludes the
contributions, findings and limitations of this study.