Comparing neural network architectures for simulating pollutant loads and first flush events in urban watersheds: Balancing specialization and generalization.
Journal:
Chemosphere
PMID:
40279933
Abstract
This study investigates the effectiveness of artificial neural networks (ANNs) models in predicting urban water quality, specifically focusing on first flush (FF) event classification and pollutant event mean load (EML) predictions for total suspended solids (TSS), total nitrogen (TN), and total phosphorus (TP) . The importance of accurate water quality models in urban areas is underscored by their role in addressing pollution challenges, ensuring public health, and maintaining ecosystem integrity. This research compares single-output, double-output, and multi-output feedforward neural network models to assess their performance in varying levels of specialization and generalization. The models were trained and evaluated using a comprehensive dataset, that consists of 577 data points, including observed events, simulated values, and generated data. Their performance achieved average F1 scores of 0.70 and accuracy of 0.77 for TSS, TN, and TP classification tasks. For EML predictions, Nash-Sutcliffe efficiency (NSE) values averaged 0.85 for TSS, 0.77 for TN, and 0.83 for TP, indicating strong predictive capabilities. Results reveal that while single-output models excel in precision for specific tasks, multi-output models offer improved efficiency and adaptability, completing training in about 6.75% of the total time required for the single-output models. Multi-output models generalize across multiple pollutants, making them more suitable for integrated urban water quality management, though with a slight reduction in precision. These findings highlight the potential of ANN-based machine learning approaches to enhance urban water quality management, supporting regulatory compliance and environmental sustainability.