Research on CatBoost model based on AutoEncoder dimensionality reduction in pollution source apportionment

Water pollution monitoring data typically exhibit characteristics of complexity, high dimensionality, and non-linearity. However, traditional receptor models-including PCA, PMF, and APCS-MLR-struggle to handle nonlinear high-dimensional water quality data and are susceptible to interference from outliers and missing values, resulting in inaccurate pollution source apportionment. To address this issue, this study aimed to establish a precise pollution source apportionment model to determine the quantity of potential pollution sources, identify their types, and quantify their contribution rates. The proposed method follows three key steps: first, PCA is applied to determine the number of potential pollution sources based on the cumulative variance contribution rate of principal components; second, an AE model is used for dimensionality reduction and pollution source identification; third, a CatBoost model is employed to quantify the contribution rate of each identified source. Taking the Qinhuai New River as a case study, four types of pollution sources were identified: organic pollution or domestic sewage discharge sources, industrial pollution sources, urban runoff or soil erosion sources, and agricultural pollution sources, with their respective contribution rates being 31.1%, 21.5%, 21.7%, and 25.7%. Model evaluation results demonstrate that the AE model achieves a reconstruction $R^2$ $>$ 0.95 and a MSE $<$ 0.05; Meanwhile, the CatBoost model for contribution rate quantification also yields an $R^2$ $>$ 0.95 and MSE $<$ 0.05, indicating high fitting accuracy. Overall, the PCA-AE-CatBoost model outperforms the PCA-APCS-MLR and PCA-CatBoost models, providing a more accurate technical basis for pollution control.