Impact of climate change on river flow, using a hybrid model of LARS_WG and LSTM: A case study in the Kashkan Basin

Abstract

Abstract

This study investigates river flow variability in the Kashkan Basin, southwest Iran, through the application of a hybrid approach involving the use of Long Ashton Research Station Weather Generator (LARS-WG) for climate downscaling and Long Short-Term Memory (LSTM) networks in the river flow forecasting. Advanced techniques like Extreme Gradient Boosting (XGBoost), Shapley Additive exPlanations (SHAP), and Analysis of Variance (ANOVA) were employed in order to decide the significant influencing factors of river flow, including rainfall, temperature, and past values of river flow. These parameters were input into the LSTM model, and TOPSIS and WASPAS were employed to choose the optimal-performing model. Historical data for the period 2001–2020 were used to estimate baseline conditions, and 2021–2040 projections were conducted under three different climate SSP (Shared Socioeconomic Pathways) scenarios: SSP1–2.6, SSP2–4.5, and SSP5–8.5.

The findings reveal a considerable rise in minimum and maximum temperatures, with the greatest increase in minimum temperatures taking place in November. Rainfall predictions have variability as some months record decreases and others show increases. River flow is anticipated to drop noticeably by 2040 compared to the historical data: from 2108 m³ to 1082 m³ (SSP1–2.6), 1070 m³ (SSP2–4.5), and 1067 m³ (SSP5–8.5) at Sarab Seydali station, and from 12,024 m³ to 9198 m³, 9057 m³, and 8858 m³ for SSP1–2.6, SSP2–4.5, and SSP5–8.5, respectively, at Poldokhtar station. This study highlights the value of machine learning models in river flow forecasting and the importance of incorporating projections of climate change into water resource management plans to address the impacts of climate change in semi-arid regions.