The role of optimizers in developing data-driven model for predicting lake water quality incorporating advanced water quality model

Abstract

Lake water is considered as one of the most precious freshwater resources for human consumption and thus, it is required to monitor the lake water quality (WQ) to maintain the good water state. For these purposes, the practice of integrating data-driven models in assessing and predicting lake WQ has received wider application. Nevertheless, the optimal performance of the data-driven models depends on the model hyperparameters and hyperparameter optimization techniques have crucial role in this regard. Therefore, the research aimed to assess the impact of various optimization techniques on these prediction models in order to predict the lake WQ. For the purposes of reliable lake water state assessment, this study presents a comprehensive analysis of lake WQ prediction, integrating sophisticated Root Mean Squared (RMS)-Water Quality Index (WQI) model with eight machine learning algorithms (ML), two Artificial Intelligence (AI) techniques, and five optimization methods. Comparing five optimization methods—Grid Search (GS), Random Search (RS), Bayesian Optimization (BO), Optuna (OPT), and Tree-based Pipeline Optimization Tool (TPOT)—revealed their significant influence on model performance and computational efficiency. Based on the analysis, random forest (RF) integrated with TPOT consistently demonstrated robust performance compared to other models, achieving an impressive R² value of 0.94 during training (Root mean squared error-RMSE= 0.85, Mean squared error-MSE= 0.73, Mean absolute error-MAE= 0.48, and Percentage of absolute bias error-PABE= 2.35), testing (RMSE= 0.43, MSE= 0.18, MAE= 0.15, and PABE= 0.19), and validation (RMSE= 0.96, MSE= 0.96, MAE= 0.823, and PABE= 1.08), indicating its precision in predicting lake WQ. The results reveal that TPOT and OPT show remarkable effectiveness in optimizing the hyperparameter space to enhance model accuracy across various ML/AI model combinations. The research findings suggest that RMS-WQI with RF-TPOT integration model could be effective to predict lake WQ that can be helpful approach for sustainable water resource management in terms of computational-cost efficiency and reliable WQ assessment. However, by addressing these complexities and highlighting the impact of various optimization techniques, the research provides crucial guidance for researchers, advancing the field of lake WQ prediction and ensuring effective utilization of ML/AI integration models. • Integrated the RMS-WQI with diverse ML/AI optimizers for precise lake water predictions. • Evaluated fifty ML/AI-optimizer combinations, highlighting RF_TPOT's accuracy. • RF with TPOT optimizer excelled in accurate lake water quality predictions. • Both RMS-WQI and RF_TPOT exhibit a 2 % expected uncertainty, showcasing resilience of the models. • TPOT demonstrated superior performance in extracting optimal hyperparameters.

Topics & Concepts