Precision Flood Forecasting in Dynamic Hydrological Systems: Integrating LP-III Distributions, Multilayer Neural Networks, and CMIP6 Projections for the Swat Basin

Abstract

Floods are among the most destructive natural disasters, presenting significant challenges due to their unpredictability and complex behavior. This study develops a robust flood prediction framework for the Chakdara monitoring station on the Swat River, Pakistan, by integrating traditional statistical methods with advanced machine learning (ML) models. Four statistical distributions—Log-Normal, Gumbel, General Extreme Value (GEV), and Log-Pearson Type III (LP-III)—were evaluated for flood frequency analysis. Among these, the LP-III distribution demonstrated the best performance with an R² value of 0.78. To enhance prediction accuracy, two ML models—Artificial Neural Network (ANN) and multilayer neural network (MLNN)—were employed. The MLNN model outperformed all others, achieving R² values of 0.96 for training and 0.93 for testing, confirming its high reliability for streamflow prediction. Furthermore, the trained MLNN was adapted to future climate conditions using downscaled and bias-corrected CMIP6 projections under SSP245 and SSP585 scenarios. This allowed for reliable discharge forecasting under changing precipitation and temperature trends. The proposed hybrid approach not only improves the accuracy of flood predictions but also supports long-term planning for flood risk mitigation. These findings provide essential insights for policymakers, engineers, and disaster management agencies to design adaptive infrastructure and implement proactive flood management strategies in the Swat River basin.