Abstract: Due to its strong nonlinear modeling ability, deep learning has a good application prospect in the assessment of total transfer capacity (TTC) of transmission interfaces. However, considering the time variability and uncertainty of the power system, it is necessary to update data and models quickly to satisfy the needs of online applications. To take full advantage of the historical scenario data and reduce the computational cost of model updating, a TTC assessment method based on active transfer learning is proposed, which includes two stages. In the first stage, the transfer learning pre-training is introduced and the generalization bound as well as the optimal empirical error combination weight is derived to guide the pre-training of the new model with the minimum generalization error. In the second stage, the active learning and model fine-tuning method are introduced, and the important samples are actively queried based on the TTC assessment network sensitivity, which greatly reduces the number of required new labeled samples for model updating, and the model fine-tuning further improves the model performance. Case study shows that, compared with the traditional training method, the proposed method greatly reduces the number of required labeled samples as well as the time cost for model updating, improving the efficiency of model transfer.