Hydraulic infiltration and decontamination performance of bioretention systems are influenced by plant species and planting layers. However, traditional evaluation methods are limited by their high subjectivity and inability to capture complex relationships among multidimensional data accurately. This study developed a coupled PCA-PP-GA model, integrating principal component analysis (PCA) for dimensionality reduction, projection pursuit (PP) for comprehensive evaluation, and genetic algorithm (GA) for optimization. Through analyzing Pearson correlation coefficients and principal component loadings, which showed strong multicollinearity and significant weights, the rationality of employing PCA for dimensionality reduction was validated. In evaluating five plant species (Cynodon dactylon, Hemarthria sibirica, Paspalum wettsteinii, Lolium perenne, and Festuca elata) across growth stages and different planting layers for stormwater runoff control, results indicated that Cynodon dactylon exhibited the highest score of 1.22, and L4 planting layer composition (10.0% loamy sandy soil + 90.0% fine sand) scored 1.47. Furthermore, compared to the analytic hierarchy process (AHP) and traditional projection pursuit model (PP-GA), the data-driven PCA-PP-GA offers a more comprehensive consideration of both cost and pollutant removal efficiency, demonstrating advantages in reducing subjective bias and enhancing information screening efficiency. This study provides a reference for evaluating the effectiveness and implementation of ecological engineering in stormwater runoff control.