基于时序大模型与数据同化的府河水质实时模拟预测框架OACHSSCD

This study proposed a novel modeling framework that integrated a time series model with a data assimilation technique for river water quality prediction at the watershed scale to enhance the accuracy of real-time prediction and forewarning capability,and to overcome the limitations of traditional data-driven models-such as difficulties in integrating real-time observations and susceptibility to error accumulation.The study area was the Fu River,a major inflow river into Baiyangdian Lake,and total nitrogen(TN)was used as the water quality parameter for testing the model.The modeling framework employed the pre-trained temporal foundation model Chronos as the core predictor and incorporated the Ensemble Kalman Filter(EnKF)to construct a dynamic data assimilation module.This enabled the continuous integration of real-time monitoring data and model state correction during dynamic prediction.Long Short-Term Memory(LSTM)and Transformer models were selected as benchmarks for comparing model performance.The prediction performance of the proposed framework was systematically evaluated at two monitoring sections of the Fu River with distinct water quality fluctuation patterns(the AnZhou and NanLiuZhuang sections),and the influence of assimilation frequency on monitoring data was investigated.The results demonstrated that:(1)The developed Filter-Chronos framework significantly improved the real-time simulation accuracy of hourly TN concentrations.At the NanLiuZhuang section,it reduced the prediction RMSE by approximately 73.4%and increased R2 to 0.998.At the AnZhou section,which exhibited more concentration fluctuations,RMSE decreased from 0.532 mg/L to 0.374 mg/L,confirming the effectiveness of data assimilation in mitigating model bias and error propagation.(2)Prediction performance was sensitive to data assimilation frequency,and this sensitivity increased with the intensity of water quality fluctuations.At the highly variable AnZhou section,maintaining a high assimilation frequency(e.g.,every 4 hours)was essential for ensuring prediction reliability,while the impact of assimilation frequency was less pronounced at the relatively stable NanLiuZhuang section.(3)In both overall predictions and peak concentration event predictions,the Filter-Chronos framework outperformed the LSTM and Transformer models coupled with EnKF.This highlights the advantage of Chronos,which provided superior prior estimates for the assimilation system by leveraging its strong generalization capacity for temporal patterns acquired through pretraining.This study presented an innovative methodology and a practical case for building a high-accuracy,adaptive real-time prediction and dynamic forewarning framework for river water quality,offering valuable insights for safeguarding water quality in sensitive aquatic environments and supporting refined water environmental risk management.