Artificial intelligence techniques in basin hydrology and material flow analysis and their applicability to Sri Lankan river basins

Abstract

Water resource management is a global concern due to increasing climate variability and growing water demand. Traditional physics-based models have helped understand streamflow and nutrient behavior, but they require much data, long processing times, and often do not work well in areas with missing or inconsistent data. Sri Lanka, with its diverse climate and limited hydrological data, needs better tools. Artificial Intelligence (AI) offers a promising alternative, with models like Long Short-Term Memory (LSTM) and Transformer networks showing strong results in learning from complex, time-based data. This research aims to test the ability of AI models to improve streamflow prediction, fill missing rainfall data, and model nutrient flow in Sri Lankan River basins. The study was conducted in the Kelani River Basin (wet zone) and Malwathu Oya Basin (dry zone) in Sri Lanka. LSTM and Transformer models showed strong performance in streamflow prediction in temporal context, while independent dry zone patterns enabled moderately accurate model forecasts. In contrast, rainfall imputation results were only slightly better than traditional methods, and nutrient modelling remained weak. The study concludes that AI models are promising for streamflow but recommends hybrid approaches and diverse input types for better overall performance, including material dynamics.