6 Conclusions

The research shows that deep learning has a good application effect in runoff simulation. Compared with traditional statistical models, MLP can better learn the potential relationship between climatic factors and runoff, which is more obvious in hydrological simulations with stronger precision (e.g., daily runoff simulation). Moreover, in simulation of the peak runoff (5%), deep learning is more accurate than traditional hydrological models, showing obvious advantage.
Meanwhile, future climate and runoff is forecasted for a period of 2020-2050. The result shows that futere temperature in the three basins all showed upward trend, and the increment decrease from Jinghe, through Xiangxi to Zhongzhou River Basins. In the average annual precipitation, Xiangxi River shows growing trend, Jinghe River shows swinging change, and Zhongzhou River shows downward trend. The same trend was given by Wang et al. (2019), who found that precipitation and temperature is projected to increase in the Upper Yangze River Basin. In Yellow River Basin, Guan et al. (2019) found the temperature presented a significant rising trend, but precipitation had a decline trend.
Average annual runoff in the three basins all have upward trend. The increments ranking is consistent with the temperature increase. Runoff growth in Jinghe River and Xiangxi River is mainly in the summer, so the frequency of summer disasters may increase. There is also an upward trend in winter runoff in Jinghe River, which may improve the winter shutdown phenomenon. The runoff growth in Zhongzhou River is mainly concentrated in February to April, so the peak runoff may be advanced. Under the RCP8.5 emission scenario, the above trends are more obvious. What’s more, the four regional climate models used in this study have little difference in climate simulation accuracy in the study basins. This is not the same as trend by Chen et al. (2014), which annual runoff exhibited a decrease trend in the Yangtze, Yellow and Pearl Basins. The different results may result from the different methodology, RCMs and baseline period. However, future runoff change also is stronger under the RCP8.5 scenario than the RCP4.5 scenario in Chen’s study.
Although the results of this study are relatively abundant and reliable, there is still room for improvement. In the future research, more kinds of deep neural network methods can be applied to hydrological simulation, i.e., Convolutional Neural Network (CNN) and Cyclic Neural Network (RNN). In addition, the linear regression was used as the only bias-correction method for RCMs in this paper, so more methods can be used in the future study.