Rapid Prediction and Source Identification of Accidental River Pollution Using a Hybrid Machine Learning and Optimization Framework

Abstract

Rapid prediction and source identification of sudden river pollution are essential for emergency response and sustainable water-environment management, whereas conventional numerical models are often computationally intensive and too inefficient for time-critical emergency applications. This study developed an integrated framework combining process-based numerical simulation, machine-learning surrogate modeling, and intelligent optimization for the Lushui River reach in Chongyang County, Hubei Province, China. A coupled hydrodynamic–water quality model was established, with its hydrodynamic component calibrated and validated against observed water-level data. Latin hypercube sampling (LHS) was then used to generate a database of sudden pollution scenarios. An HGS-optimized kernel extreme learning machine (HGS-KELM) was developed as a rapid surrogate for nonlinear source–response relationships, and its predictive performance was compared with that of GPR and XGBoost. HGS-KELM achieved the best predictive performance, with RMSE = 0.0523, MAE = 0.0413, and R2 = 0.9944. HGS-KELM was further coupled with GA, WOA, and EEFO for source inversion. GA achieved the highest source-parameter recovery accuracy, with relative errors of 6.49% for source strength and 8.24% for source distance, while maintaining the lowest parameter errors across different observational noise levels. The framework provides technical support for rapid pollution prediction, source identification, and sustainable watershed water-environment management.