Coupling SWAT and Bi-LSTM for improving daily-scale hydro-climatic simulation and climate change impact assessment in a tropical river basin

Abstract

Global climate change has led to an increase in both the frequency and magnitude of extreme events around the world, the risk of which is especially imminent in tropical regions. Developing hydrological models with better capabilities to simulate streamflow, especially peak flow, is urgently needed to facilitate water resource planning and management as well as climate change mitigation efforts in the tropics. In view of the need, this paper explores the feasibility of improving streamflow simulation performance in the tropical Kelantan River Basin (KRB) of Peninsular Malaysia through coupling a conceptual process-based hydrological model − Soil and Water Assessment Tool (SWAT) with a deep learning model − Bidirectional Long Short-Term Memory (Bi-LSTM) in two ways. All SWAT parameters were set as their default values in one hybrid model (SWAT-D-LSTM), whereas three most sensitive SWAT parameters were calibrated in the other hybrid model (SWAT-T-LSTM). Comparison of daily streamflow simulation results have shown that SWAT-T-LSTM consistently performs better than SWAT-D-LSTM as well as the stand-alone SWAT and Bi-LSTM model throughout the simulation period. Particularly, SWAT-T-LSTM performs considerably better than the other three models in simulating daily peak flow. Based on the latest projection results of five GCMs from the Sixth Phase of the Coupled Model Intercomparison Project (CMIP6) under three emission scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5), the best-performed SWAT-T-LSTM was run to assess the potential impacts of climate change on streamflow in the KRB. Ensemble assessment results have concluded that both average and extreme streamflow is much likely to increase considerably in the already wet northeast monsoon season from November to January, which has surely raised the alarm for more frequent flood occurrence in the KRB.