Combining Long Short-Term Memory and Genetic Programming for Monthly Rainfall Downscaling in Southern Thailand's Thale Sap Songkhla River Basin
Abstract
This paper proposed a novel methodology, combining Long Short-Term Memory (LSTM) and Genetic Programming (GP), for downscaling monthly rainfall into watershed regions. The exploration of suitable downscaling models and the trend of monthly rainfall in the 2030s, 2060s, and 2080s in the Thale Sap Songkhla river basin (TSS) was investigated. The TSS is one of four major areas in Thailand’s southern basin and has a tropical monsoon climate. The monthly rainfall observed by the Royal Irrigation Department (RID) from January 1993 to December 2018 (312 months) was available at three rainfall stations. Six machine learning techniques (i.e., M5, RF, SVR, MLP, GP, and LSTM) were employed to downscale the monthly rainfall data from the General Circulation Models (GCMs) of CMIP5 (HadGEM2-ES and ACCESS1-3) and CMIP6 (HadGEM2-CGM31-LL and ACCESS-CM2) under the RCP4.5 (SSP245) and RCP8.5 (SSP585) scenarios. Since the TSS experiences significant differences in low and high rainfall for January–September and October–December, respectively, those data were analyzed separately in addition to using the whole-year data sets. This study considered six common climate variables: precipitation (pr), maximum near-surface air temperature (tasmax), minimum near-surface air temperature (tasmin), relative humidity (hur), sea level pressure (psl), and near-surface wind speed (sfcWind). These variables were chosen based on the correlation between them and the observed rainfall data. The findings of this research indicate that when LSTM and GP models are merged, they are the most efficient for downscaling monthly rainfall. The OI and r-value illustrate a highly robust relationship between the average values within the TSS watershed. These results offer valuable understandings regarding the clear strengths and limitations of every model category, which are influenced by factors such as the size of the data and the characteristics used in the model training process. Climate change is likely to have only a minor impact on rainfall patterns in TSS in the near future, both in moderate and extreme emission scenarios. However, significant changes are expected in the later stages of this century (2060s and 2080s), particularly during the monsoon season, which experiences drastic shifts.