Engine performance variations and traction fluctuations have a great impact on the adaptability of shift strategies for high-power power shift tractors (PSTs). In order to construct a dynamic accurate model and deal with traction fluctuation to improve the adaptability of PST shift strategy, an adaptive shift strategy development method was proposed based on digital twins. On one hand, the engine state change was regarded as an internal disturbance, and the virtual PST engine was calibrated in real time based on the deep deterministic strategy gradient algorithm, which was combined with the PST mechanism model to realize the real-time dynamic and accurate modeling of the PST. On the other hand, the traction fluctuation was treated as an external disturbance, and a deep Q-network was used to automatically generate the shift strategy. Finally, the virtual PST training simulation under plowing conditions and the speed tracking comparison test between the proposed method and the fuzzy adaptive method were carried out. The results showed that the average tracking errors of engine torque and fuel consumption rate did not exceed 7.28N·m and 1.55g/(kW·h), and dynamic and accurate modeling of physical PST was achieved. After using for a long time, the changes of engine and traction force caused that the shift points and fuzzy rules of the fuzzy adaptive method were no longer fully applicable, and the shift performance was gradually deteriorated. In contrast, the shift performance and speed tracking effect of the proposed method were good throughout, and the mean value of speed tracking error, mean value of fuel consumption rate, and total number of shifts were 0.0125m/s, 229.76g/(kW·h),and 42, respectively, which were reduced by 0.91%, 11.14%, and 34.38% compared with those of the fuzzy adaptive method. The adaptability and superiority of the proposed method were verified.