Performance evaluation of artificial neural network models in quantile estimation model for RCP 4.5 scenario

The increase of hydrological extremes such as heavy rainfalls and floods due to climate change becomes evident. Thus, more accurate estimation of quantiles is required. To estimate future quantiles at site of interest, nonstationarity of climate change scenario must be considered. This study aims to present the performance evaluation of stationary and nonstationary models including artificial neural network (ANN) that calculate rainfall quantiles of South Korea for Representative Concentration Pathway 4.5 (RCP45). At-site frequency analysis (AFA), index flood method (IFM), quantile and parameter regression techniques (QRT and PRT) were considered for stationary quantile estimation. Nonstationary at-site frequency analysis (NS-AFA), nonstationary population index flood (NS-PIF), nonstationary index flood method (NS-IFM), nonstationary quantile and parameter regression techniques (NS-QRT and NS-PRT) were considered for nonstationary quantile estimation. After applying both the goodness-of-fit test and uncertainty measure using the rescaled Akaike information criterion (rAIC), the generalized extreme value (GEV) distribution model was finally selected as stationary and nonstationary quantile estimation model. Then, Monte Carlo simulation was conducted to estimate the bias, relative bias (Rbias), root mean square error (RMSE), and relative root mean square error (RRMSE) between rainfall quantiles from climate change scenario and those from all applied models.