Predicting the amount of water shortage during dry seasons using deep neural network with data from RCP scenarios

The drought resulting from insufficient rainfall compared to the amount in an ordinary year can significantly impact a broad area at the same time. Another feature of this disaster is hard to recognize its onset and disappearance. Therefore, a reliable and fast way of predicting both the suffering area and the amount of water shortage from the upcoming drought is a key issue to develop a countermeasure of the disaster. However, the available drought scenarios are about 50 events that have been observed in the past. Due to the limited number of events, it is difficult to predict the water shortage in a case where the pattern of a natural disaster is different from the one in the past. To overcome the limitation, in this study, we applied the four RCP climate change scenarios to the water balance model and the annual amount of water shortage from 360 drought events was estimated. In the following chapter, the deep neural network model was trained with the SPEI values from the RCP scenarios and the amount of water shortage as the input and output, respectively. The trained model in each sub-basin enables us to easily and reliably predict the water shortage with the SPEI values in the past and the predicted meteorological conditions in the upcoming season. It can be helpful for decision-makers to respond to future droughts before their onset.