Knowledge-Based Feature Selection Substantially Enhances Data-Driven Wastewater Treatment Modeling

Data-driven modeling in wastewater treatment is increasingly constrained by the reality of small, high-dimensional data, where the abundant monitoring parameters in small-sized data sets obscure fundamental mechanistic understandings. This study proposes a knowledge-driven feature selection framework that integrates mechanistic insights with statistical correlations to identify the most informative predictive features. Using nitrous oxide (N₂O) emission prediction at a full-scale plant as a case study, we compared classic deep-learning feature selection algorithms using attention mechanisms against two new knowledge-based approaches: (i) expert-guided feature selection and (ii) large language model (LLM)-augmented feature selection. Expert-knowledge-guided feature selection substantially enhances predictive accuracy, achieving a mean R² of 0.723 and an MAE of 0.033, compared to R² = 0.712 and MAE = 0.033 for the best-performing attention-based architecture. More importantly, the proposed framework markedly improves model generalizability: under out-of-distribution high-flow conditions where the attention-based model fails to capture N₂O emission patterns, the expert-selected model continues to reproduce the dominant temporal dynamics of N₂O emissions. The LLM-assisted approach also delivers competitive accuracy (mean R² = 0.596, MAE = 0.041) and similarly preserves generalizability under an input distributional shift. By introducing mechanistic understanding into the feature selection process, this framework offers a generalizable pathway for addressing complex wastewater treatment challenges while maintaining a computational efficiency.