可解释性长短期记忆模型用于预测湖泊总磷浓度变化OA北大核心CSTPCD

The prediction and source identification of total phosphorus(TP)in lakes is critical for the management of water re-source and watershed ecology.However,non-stationarity caused by inconstant hydrodynamic conditions and the complex biochemi-cal reactions pose significant challenges in accurate forecast of lake TP concentrations.To address this challenge,this study intro-duced the Seasonal and Trend decomposition using Loess(STL)technique and SHapley additive exPlanations(SHAP),and com-bined them with Long Short-Term Memory(LSTM)and Gated Recurrent Unit(GRU)to develop an interpretable prediction frame-work.The framework was applied to enhance the prediction of lake TP concentrations and improving their interpretability.The study achieved the following results.(1)In the prediction of TP concentrations in Lake Luoma,this framework achieved a higher model fit with an R2 value of 0.878,outperforming LSTM and CNN-LSTM.By increasing the prediction time step to 8 hours,the frame-work achieved a better model fit with a decrease of MRE and RMSE by 47.1％and 33.3％,respectively.An analysis of the predic-tion trend for Lake Luoma revealed that the average TP concentration is 0.158 mg/L during the flood season,202.1％higher than that during non-flood seasons.(2)Canal inflow was the most influential factor on TP concentrations,with a contribution of 60％.Different sections(Sanwan and Sanchang)had large spatiotemporal variations in phosphorus sources influenced by hydrodynamics and meteorological factors.This study highlighted the potential of neural network models in predicting water pollution,and offered valuable insights into enhancing the learning capabilities and interpretability of traditional neural networks.