LSTM-Transformer hybrid model predicts and unveils total phosphorus dynamics and pollutions in Poyang Lake

Addressing the persistent issue of Total Phosphorus (TP) exceedance in Poyang Lake, this study integrates spatiotemporal analysis with deep learning techniques using data from 2013 to 2022 to identify heterogeneity patterns, driving mechanisms, and predict future trends (2023-2026). After evaluating univariate models, an LSTM-Transformer hybrid model was developed by incorporating key influencing factors-such as water level, rainfall, and TP load-through a multi-head self-attention mechanism, significantly improving prediction accuracy (test set R² = 0.6428, RMSE <0.017). Attribution analysis revealed that external phosphorus inputs constitute the primary driver (56.7 %), while rainfall and water level exert indirect effects through runoff and sediment resuspension. Model projections indicate that the western and northern regions of the lake will remain pollution hotspots, particularly during autumn and winter, while the southern region exhibits a rising TP trend mainly due to internal loading. We recommend region-specific and seasonal management strategies-including enhanced wet-season pollution interception, optimized aquaculture feed input, and real-time monitoring at key cross-sections-to support targeted eutrophication control.