Clustering as a Prerequisite for Reliable Machine Learning Prediction of Multi-Odor Systems in Wastewater Treatment

Complex odor emissions from wastewater treatment plants consist of multiple volatile compounds that exhibit heterogeneous temporal dynamics and low linear correlations, making accurate prediction and interpretation difficult when analyzed on a single-compound basis. This study investigates whether clustering can serve not only as an exploratory tool but as an essential preprocessing step to enhance machine-learning performance in multi-odor prediction systems. A total of 22 designated odorants were continuously monitored, and their pairwise dependencies were evaluated using Pearson correlation and mutual information. Data-driven clustering was performed through K-means, hierarchical linkage, and principal-component–based latent grouping, and the resulting structures were quantitatively compared with functional-group-based chemical classifications using the consistency ratio and Jaccard similarity index. Cluster validity was further examined using the Silhouette Coefficient, Davies–Bouldin Index, and Calinski–Harabasz Index. The predictive contribution of clustering was verified by training XGBoost regression models on both raw and cluster-structured datasets. The clustered dataset yielded higher predictive accuracy, with increased R2 and reduced MAE and RMSE across most odorants. SHAP analysis further confirmed that clustering improved model interpretability by stabilizing feature contributions and reducing noise-driven importance shifts. The findings demonstrate that clustering is not a supplementary diagnostic tool, but a prerequisite for building reliable, high-performance machine-learning models in complex odor systems. This integrative framework offers a methodological foundation for multi-odor forecasting, source tracking, and next-generation odor management platforms.