Climate change is occurring all over the world, and a lot of related research is being done. However, studies on climate change are concentrated on wealthy countries that produce a lot of pollutants but are less vulnerable to climate change. This bias is caused by limitations such as financial support and the accuracy and accessibility of the basic data and can be intensified over time. Therefore, various attempts are needed to solve these limitations. In this study, globally-available basic data and the climate change scenario were constructed for the Mekong River basin and a data-based LSTM model was applied to simulate future runoff changes by climate change scenario. The results of the future runoff change analysis using LSTM were applied to the flood inundation analysis downstream of the Mekong River and compared to the results of the SWAT model, which is a physical-based rainfall-runoff model.
The SWAT model and LSTM model were used to simulate the runoff changes, and the accuracy of each model was compared quantitatively through observation data from 2006 and 2007. The SWAT model showed NSE value of 0.84 and the LSTM model 0.99; and the runoff simulation results of each model in the verification period showed high reproducibility. The results of the runoff prediction of the future period show that the runoff results by the LSTM are similar to the runoff analysis results by the SWAT, and the reproducibility of the runoff simulation by the time series is higher than in the SWAT results. Based on the results of the LSTM model''s future runoff changes, the future seasonal flow variability of the downstream of the Mekong River is forecast to decrease, and the flood risk for the period 2056-2100 is expected to increase. This is the period during which the average annual flow rate is greatest in all four cases, and it has been analyzed that the flow rate is affected by a reduction in variability as in the flow regime analysis results.
Climate change scenarios pose very high uncertainty compared to actual climate change due to limitations such as selecting future virtual scenarios, incomplete physical understanding of various natural conditions, and computational abilities. Thus, the analysis of runoff using them is accompanied with additional errors and uncertainty. In particular, if a runoff analysis is performed through the application of a physical-based rainfall-runoff model for areas with low reliability of observations, such as this region, then the determination or estimation process of various basic data and parameters will involve a number of uncertainties. Therefore, it can be concluded that the LSTM model, which produces relatively accurate results with only a small amount of data, can be used effectively when only the time series change of the runoff amount is required.