X-ray fluorescence analysis (XRF) is a remarkably effective analytical technique for quantitatively studying heavy metal elements in soils. Due to matrix effects and elemental interferences, existing machine-learning methods suffer from inadequate performance and instability in predicting lead (Pb) and cadmium (Cd) concentrations using soil XRF spectra. Therefore, this paper proposes a PCA-BLS method for the XRF quantitative analysis of heavy metals in soil based on principal component analysis (PCA) combined with the broad learning system (BLS). It can accurately, efficiently, and stably determine concentrations of Pb and Cd in soil. First, the 56 standard soil data are feature-reduced using PCA. The first three principal components of Pb and Cd are selected as features. Then, the optimal principal component features are fed into the width learning system for calibration and testing. Using the grid search determine the optimal network structure. The three optimum parameters for the BLS corresponding to the Pb and Cd elements are 2,11,11 and 3,19,15, respectively. Using support vector regression (SVR), BP neural network, and the original BLS compared with the PCA-BLS. PCA-BLS achieved performances of 0.954, 1.433, and 1.014 in the R2, RMSE, and MAPE corresponding to Pb. In the quantitative Cd, PCA-BLS obtains the R2 of 0.982, RMSE of 1.215, and MAPE of 1.059. Grid search visualization demonstrates the stable performance of PCA-BLS in predicting two heavy metal elements. The experimental results show that PCA-BLS can effectively correct for matrix effects and interferences in soil XRF. The PCA-BLS is a promising method for quantitative XRF spectroscopy that accurately predicts Pb and Cd elemental concentrations while maintaining model stability.