Machine learning for water quality prediction and uncertainty assessment

Effective water quality monitoring requires predictive models that combine high accuracy, interpretability, and credible uncertainty quantification. Machine learning (ML) techniques have emerged as powerful tools for predicting water quality and quantifying associated uncertainties. Similarly, Bayesian deep learning (BDL) frameworks enable probabilistic predictions that quantify uncertainties. These approaches can capture nonlinear interactions and provide robust predictions in diverse environmental conditions. This study integrated ensemble ML and BDL to assess the complex relationships between physicochemical parameters and the Water Quality Index (WQI). Six supervised ensemble ML algorithms, namely Decision Tree (DT), Random Forest (RF), Extra Trees (ERT), XGBoost, CatBoost, and LightGBM were evaluated using Bayesian optimisation to identify the optimal hyperparameter configurations. DT achieved the highest predictive accuracy with MAE = 0.657 and 0.428, RMSE = 1.181 and 0.747, MAPE = 10.561 and 7.155, R 2 = 0.960 and 0.987, and nRMSE = 0.065 and 0.042 for the training and test sets, respectively. The DT outperformed more complex ensemble models, and SHapley Additive exPlanations (SHAP)-based eXplainable Artificial Intelligence (XAI) identified the most influential predictors, aligning model predictions with underlying hydrochemical processes. To capture predictive uncertainty, a probabilistic BDL was developed using variational inference, yielding probabilistic outputs and explicit epistemic uncertainty estimates. ROC analysis confirmed strong performance across WQI classes, with AUC scores of up to 0.90 for WQI classes. The probabilistic approach provides actionable insights for adaptive water quality management, enabling targeted monitoring in areas of high uncertainty and supporting transparent, evidence-based decision-making. These results underscore the value of integrating machine learning, and Bayesian optimisation to advance robust and adaptive water quality assessment. The proposed workflow provides a scalable framework to enhance monitoring, optimize resources, and advance sustainable water management aligned with the SDGs. • Bayesian optimisation of Decision Tree achieved highest accuracy and interpretability among six ML models. • SHAP-based XAI revealed dominant hydrogeochemical drivers of WQI variation. • Bayesian deep learning quantified epistemic and aleatoric predictive uncertainties. • Probabilistic framework supports adaptive, risk-aware water-quality monitoring. • The framework is scalable, energy-efficient AI workflow for advancing adaptive water management.