Knowledge Distillation TabPFN for Predicting Dismantled Material Weight of Retired Oil-Immersed Transformers

The dismantled components of retired oil-immersed power transformers, particularly the windings and core, have high recycling economic value. However, accurately estimating their weight based solely on nameplate information prior to disassembly remains a key challenge for recycling enterprises in formulating rational plans and enhancing bidding competitiveness. This paper proposes a knowledge distillation-based prediction method, where a categorical boosting algorithm (CatBoost) serves as the teacher model, and a tabular prior-data fitted network (TabPFN) acts as the student model for knowledge distillation. The nameplate information of transformers is used as initial input features, from which derived features are constructed to enrich the information representation and enhance the model’s predictive capability. Process-level weighing data (e.g., total and de-oiled weight) are used as privileged information during training. The distillation targets are constructed in a proportional space to account for physical differences between windings and cores, with sample-level transfer controlled to suppress negative transfer. The proposed method achieved R2 values of 0.9593 and 0.9813, with RMSE of 11.49 kg and 16.56 kg, and MAPE of 8.50% and 5.46%, respectively. The proposed distilled TabPFN model enhances prediction accuracy, providing quantitative support for recycling pricing and dismantling decisions, and contributing to improved waste management and resource recovery.