Predictive maintenance is essential in prognostics and health management, where prediction of remaining useful life (RUL) plays a key role. However, various challenges exist in deep learning-based RUL prediction models, including hyperparameter tuning, uncertainty, and applying RUL for maintenance. To address such issues, this paper proposes a novel PdM framework by combining interval RUL prediction with maintenance policy optimization. A bidirectional temporal convolutional network with multi-head attention method is adopted for RUL prediction, and a physical model can integrate the predicted RUL intervals for maintenance decision-making. Moreover, the differential creative search algorithm is introduced to optimize hyperparameters and decision maintenance variables. A case study is conducted with the C-MPASS aero engine dataset. The results show that the proposed model can reduce RMSE by over 3.20% and 3.68% on the FD002 and FD004 datasets, respectively. Sensitivity analysis also confirms its robust performance despite the variations in maintenance costs or times.