Long-term forecasting of water quality and algal dynamics in riverine systems using advanced physicochemical-informed machine learning models.

Monitoring water quality in riverine ecosystems is vital for safeguarding public health and maintaining environmental sustainability. Although data-driven modeling approaches have enhanced the ability to forecast chlorophyll-a (Chl-a) and orthophosphate (PO43-), achieving high forecast accuracy across diverse environmental conditions remains a significant challenge. This study introduces two novel hybrid models, the Wavelet-Physicochemical Enhanced Kinetic Neural Network (WT-PhyEKNN) and the Wavelet-Physicochemical Regularized Kinetic XGBoost Tree (WT-PhyRKXT), which integrate biogeochemical kinetics governing nutrient cycling, algal growth, and heat‑oxygen dynamics. WT-PhyRKXT surpassed alternative models by incorporating biogeochemical process logic into its architecture, anchoring predictions in ecological theory instead of relying on statistical correlations. In 360-day ahead forecasts for the Des Plaines River, WT-PhyRKXT-5 emerged as the superior framework, achieving R2 values of 0.880 for Chl-a and 0.839 for PO43-, with relative errors (RE) of 4.38% and 17.80%, respectively. It should be noted that Config 5 requires future exogenous driver data (e.g., surface heat fluxes, radiation) for operational forecasting, which may be obtained from meteorological forecasts or reanalysis products. Beyond its statistical performance, the model demonstrated capacity to support ecological diagnostic applications. The model forecasted 95.09% of days as poor or very poor for PO43- in close agreement with the observed value of 94.15%, and captured critical seasonal patterns, including the July peak in Chl-a and PO43-, increase from August to December. The model's categorical fidelity provided information for risk assessments, with results consistent with a persistent water quality decline in the study area potentially driven by internal loading. The framework is applicable to riverine systems with pronounced seasonal dynamics and moderate data availability, though transferability to other hydroclimatic settings requires further evaluation. WT-PhyRKXT-5 thus established a new paradigm, a scalable, interpretable, and robust tool that advances forecasting from statistical correlation to physicochemical-informed ecological analysis, providing support for adaptive management of impaired freshwater systems.