ADAPT: A novel IoT-driven analytical data assimilation method based on phase-space tuning for long-sequence water quality forecasting

Abstract

Water quality models often suffer from performance degradation due to parameter obsolescence caused by external environmental changes. To address this, this study proposes a novel framework named Analytical Data Assimilation via Phase-space Tuning (ADAPT). Unlike traditional data assimilation methods that directly update state variables, ADAPT dynamically calibrates model parameters by establishing a robust link between parameters and water quality dynamics using Aquaformer, a Transformer-based deep learning model driven by phase-space reconstruction. The method was validated through digital-twin and real-world experiments on the Diannong River, China. Results demonstrate that ADAPT significantly outperforms the Ensemble Kalman Filter, reducing prediction errors by 36.26 % at monitored sites and 54.66 % at unmonitored sites. ADAPT exhibits superior transferability and stable error control, effectively overcoming the limitations of traditional methods in spatial generalization. This study provides a reliable, physics-informed solution for high-frequency auto-calibration in smart water management systems.