<p>Partial least squares (PLS) regression is widely used in high-dimensional data modeling, but adaptively selecting the optimal number of latent components remains a critical challenge for improving model performance. To address limitations in existing methods, such as poor distribution alignment in distance-based approaches and instability caused by sampling randomness in resampling-based techniques, we propose a novel component number selection framework for PLS, termed SS-PLS. The proposed method integrates the SPlit algorithm, which optimizes training-test set partitioning via an energy distance objective function, with cross-validation to identify representative support points in a continuous space and map them to original data samples. By ensuring distribution consistency between the test set and the full dataset, SS-PLS mitigates biases arising from distribution deviations in resampling methods. Experimental comparisons on four real-world high-dimensional datasets demonstrate that SS-PLS consistently achieves superior predictive accuracy and stability, outperforming other methods while selecting fewer components on average. This work pioneers the application of SPlit for component selection in PLS regression, offering a deterministic and efficient solution for high-throughput data modeling in chemometrics, genomics, and beyond.</p>