High impact icing events pose a threat to the stable operation of power systems. However, under small-sample scenarios, traditional physics-knowledge guidance and data-driven prediction methods often have poor performance. Therefore, a physics knowledge guided iterative self-transfer learning (PKG-ISTL) model is proposed, integrating physics informed neural network (PINN) with self-transfer learning (STL). First, the data is divided into source and target domains, and a three-dimensional tensor containing spatial, feature, and temporal dimensions is constructed to achieve sliding-window-based icing prediction. Second, the model is built by combining PINN and STL. In the source domain branch, a PINN is trained to guide the model using physics knowledge. In the self-migration branch, multi-kernel maximum mean discrepancy is applied for domain adaptation. In the target domain branch, knowledge distillation is used to transfer physics knowledge from the expert model to the trained model. Finally, historical icing data from transmission lines in Guangxi province are used for simulation. Through interpretability analysis, the influence of meteorological, mechanical, line, and icing factors on line icing is revealed. Results show that compared to traditional data-driven models, the PKG-ISTL model improves prediction accuracy by 47.69%, verifying its effectiveness in small-sample scenarios.