Power grid marketing inspection relies on work order data to identify rule violations. However, manual labeling of audit results is costly and often infeasible at scale. This study proposes a self-training partial least squares (STPLS) model to address this challenge by leveraging both labeled and unlabeled data for semi-supervised modeling. The method iteratively augments labeled samples using high-confidence predictions, improving calibration accuracy under limited supervision. Experiments on over 10,000 real-world grid work orders demonstrate the model’s effectiveness in semantic audit topic alignment and time-series prediction of work order volumes. Additional validation on a spectroscopic benchmark dataset confirms the generalizability of the approach. The proposed method supports efficient and scalable intelligent inspection in the power grid domain.