Reimagining coastal water quality forecasting with hydrodynamic simulation and advanced machine learning

Effective and sustainable management of coastal ecosystems relies on accurate water quality monitoring and forecasting, particularly in regions facing heavy pollutant loads. Phosphate and nitrate play crucial roles in coastal water quality and ecosystem health. However, predicting their concentrations remains challenging due to complex spatiotemporally varying environmental interactions. Leveraging advanced numerical modelling and artificial intelligence (AI), this study introduces a novel hybrid machine learning (ML) framework based on least-squares support vector regression (LSSVR) to simulate coastal water quality using physically grounded simulation data generated by the TELEMAC-WAQTEL numerical framework, calibrated against field-measured boundary and initial conditions. The proposed model is applied to Ha Long Bay, Vietnam, utilizing a dataset comprising 8760 hourly samples collected over one year (2021-2022). The data was generated through a two-dimensional TELEMAC hydrodynamic model integrated with a water quality module. Key hydrodynamic variables, including free surface elevation (FS), velocity (U), and flow direction (α), serve as inputs to predict nitrate (NO₃) and phosphate (PO₄) concentrations. To optimize the LSSVR hyperparameters, we introduced the Improved Dragonfly Algorithm (IDA), which integrates Cauchy walk mutation and adaptive inertia weighting to enhance convergence, global search efficiency, and robustness. Benchmarking against Gaussian Process Regression (GPR), ε- LSSVR, ANN, Random Forest, and Model Tree demonstrate the superior predictive capability of IDA-LSSVR. On an independent testing dataset comprising 20% of the total samples (1752 observations), IDA-LSSVR achieved the lowest RMSE (NO₃ = 1.55E-02 mg/L, PO₄ = 5.88E-04 mg/L), MAE (NO₃ = 3.45E-07 mg/L, PO₄ = 2.39E-04 mg/L), and MAPE (NO₃ = 3.41, PO₄ = 4.49E-04), coupled with the highest R² (NO₃ = 0.95, PO₄ = 0.97). By coupling high-resolution physics-based hydrodynamic simulations with advanced machine learning and optimisation techniques, the proposed framework provides an accurate, computationally efficient, and reliable tool for coastal water quality forecasting. The integration of physically consistent simulation data with robust uncertainty-aware prediction offers valuable decision-support capabilities for nutrient management, eutrophication risk assessment, and environmental monitoring in dynamic coastal environments. The framework therefore presents a promising approach for supporting sustainable coastal ecosystem management and protecting vulnerable marine systems from nutrient-driven degradation.