Feature-enhanced hybrid-optimized convolutional neural network-long short-term memory framework for real-time early...

In sudden river pollution incidents, rapid prediction of pollutant migration is critical for emergency response. This study proposes a seconds-level prediction model based on a feature-enhanced convolutional neural network-long short-term memory (CNN-LSTM) framework to forecast suspended-solids pollution in the Shizi River reach affected by combined sewer overflow (CSO) discharges. Training and validation datasets were generated using a coupled one-dimensional-two-dimensional hydrodynamic-water-quality model driven by overflow hydrographs derived from rainfall events with different return periods. The framework integrates multi-scale convolutional feature extraction with recursive temporal modeling to predict spatial suspended-solids concentration fields, affected areas, and event-scale pollutant loads. The model captures the nonlinear dependence of pollutant load on flow and source concentration across a wide concentration range, achieving coefficient of determination (R²) values consistently above 0.90. Spatial agreement remains high under low-to-moderate concentrations, and intersection over union (IoU) exceeds 0.98 in the late stage of high-concentration cases. Upon training completion, the model generates event-level predictions within seconds, providing rapid early warning and decision support for emergency management of sudden river suspended solids pollution.