Artificial Intelligence for Predicting and Optimizing Water Quality in Aquatic Systems

Maintaining optimal water quality is essential for the sustainability of large aquarium systems, yet conventional
treatment technologies such as reverse osmosis and membrane filtration remain energy intensive and expensive.
Although artificial intelligence has been widely applied for water quality prediction, most existing studies focus
mainly on predictive modeling and rarely integrate optimization, inverse prediction, interpretability, and
comparative evaluation within a unified decision support framework. This study develops an interpretable
artificial intelligence framework for proactive water quality management by predicting, optimizing, and
inversely estimating two key indicators: electrical conductivity (EC) and total dissolved solids (TDS). A database
of 110 raw water samples containing 21 routinely measured physicochemical parameters was used to develop
decision tree ensemble models. RUSBoost and AdaBoostM5 models were optimized using a genetic algorithm and
evaluated against optimized baseline models, including Regression Tree, Support Vector Regression, and
Gaussian Process Regression, under the same validation conditions. The proposed boosting models achieved
superior predictive performance compared with the baseline approaches. The RUSBoost model provided the best
performance for conductivity prediction, achieving an overall correlation coefficient of R = 0.9996 (R² =
0.9993), whereas AdaBoostM5 achieved the highest accuracy for TDS prediction with R = 0.9995 (R² = 0.9990).
Both models exhibited very low prediction errors, with overall RMSE values of 2.99 μS cm⁻¹ for conductivity and
2.62 mg L⁻¹ for TDS. Cross-validation and external validation using an independent water source confirmed the
robustness and stability of the developed framework. Predictor importance analysis, representative decision tree
structures, and constrained optimization enabled the identification of dominant physicochemical factors and
feasible water quality configurations. The developed MATLAB based decision support application integrates
prediction, optimization, and inverse prediction functionalities. The proposed framework provides an inter-
pretable and practical tool for proactive aquatic water quality management, supporting more efficient control of
physicochemical parameters and reducing unnecessary dependence on costly treatment processes.