The global attention to the pollution of nutrients and per-and polyfluoroalkyl substances (PFAS) is increasing, which requires the development of more efficient and low-cost remediation strategies. This study focuses on the remediation of nitrogen, phosphorus, and various PFAS components by aquatic plants, representing a sustainable alternative to traditional physical and chemical methods. In order to overcome the limitations of traditional experimental methods, we constructed a comprehensive dataset by systematically searching for literature in databases such as CNKI, Science Direct, and Web of Science from 2015 to 2025. The final database contains 128 independent experimental records, covering pollutant remediation data under different experimental conditions, ensuring the universality of research conclusions. The input features cover 14 dimensions, including aquatic plant types (submerged, emergent, and floating), planting density, environmental temperature, and initial concentrations of nitrogen, phosphorus, and eight specific PFAS components (such as PFOS, PFOA, PFBA). In order to mitigate inherent systematic biases in multi-source literature data, this study implemented strict quality control protocols. For missing values of secondary environmental parameters, RF-based imputation is used for processing, which can better preserve the nonlinear structure of the data than simple mean replacement. All numerical features are standardized using StandardScaler to eliminate dimensional deviations. In addition, data heterogeneity was quantitatively evaluated using one-way analysis of variance (ANOVA) and effect measures (η2). The results indicate that planting density and temperature are the main sources of statistical heterogeneity, explaining 41.67% and 41.47% of the total variation, respectively (P<0.001). We comprehensively evaluated the impact of various factors on the removal rate of PFAS using Multi-objective Random Forest (MTRF), Multilayer Perceptron (MLP), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) algorithms.?Use R2 and Root Mean Square Err (RMSE) as evaluation metrics for different models.The MTRF model achieved average R2 values of 0.68 and 0.72 in nitrogen removal and PFNA prediction, respectively, demonstrating substantial predictive ability significantly better than the MLP model with negative R2 values and inability to capture complex patterns. The in-depth mechanism explanation using the SHAP framework reveals that initial phosphorus concentration has a positive effect on nitrogen absorption, and phosphorus can promote nitrogen absorption by improving plant metabolic capacity. For PFAS, a threshold suppression effect was observed; Due to oxidative stress and enzyme activity inhibition caused by reactive oxygen species (ROS), high initial concentrations and extreme temperatures are negatively correlated with removal efficiency. Through virtual screening, Vallisneria natans was identified as the dominant species, possibly due to its strong root to stem transport ability and high bioaccumulation factor for PFAS. By optimizing the operating conditions, the optimal planting density for removing nutrients from Vallisneria natans is 19 plants/m2, and the optimal restoration time is 69 days; The optimal planting density for removing PFAS is 59 plants/m2, and the optimal restoration time is 34 days. The machine learning model established in this study not only provides accurate predictions for plant remediation efficiency under complex water quality conditions, but also provides theoretical basis and engineering decision support for low-cost and systematic treatment of water composite pollution.