Enhanced total phosphorus prediction and explanation for the Minjiang upper reaches using bidirectional long short-term memory, attention mechanism, and particle swarm optimization

Study region Minjiang River basin, situated in the upper reaches of the Yangtze River, China. Study focus Total phosphorus (TP) is a key contributor to water eutrophication, necessitating accurate forecasting for effective control and management. This study addresses the challenge by developing and evaluating the PSO-BiLSTM-Attn (Particle Swarm Optimization - Bidirectional Long Short-Term Memory - Attention) framework for high-frequency TP prediction in the Minjiang River basin. The proposed framework is benchmarked against BP, LSTM, Transformer, and BiLSTM-Attn models using 8-hourly data from three monitoring stations across multiple forecast horizons. This study then further investigates how prediction performance varies across stations with different TP fluctuation regimes and applies explainability analysis to identify dominant drivers of TP prediction. New hydrological insights for the region Performance analysis displayed that PSO-BiLSTM-Attn achieved superior TP-level prediction compared with the baseline model, with R 2 values of 0.8963 (station S1), 0.9412 (station S2), and 0.9138 (station S3). Stations with larger TP variability exhibited lower predictive performance, indicating that disturbances or stronger anthropogenic perturbations reduce model predictability. In the prediction step, single-step predictions achieved optimal accuracy for TP concentration, suggesting that TP dynamics were dominated by short-term memory and rapidly changing external inputs. The explainability analysis further indicates that TP predictions correlated positively with TN and EC, reflecting synergistic nutrient co-accumulation and pollutant-driven TP elevation, respectively. These findings provide insight into how TP dynamics respond to spatial heterogeneity and short-term disturbances in human-impacted river systems.