Digital twin (DT) describes the application of holistic simulations to virtually mirror the physical systems. In chemical applications, DT can be used as an operator training system, process simulation, and process control. DT can be constructed by first principle models or data-driven models. However, the development of the first principle models is time-consuming and requires a lot of prior knowledge, which cannot provide immediacy. We present a state-of-the-art data-driven digital twin based on the sequence-to-sequence (StS) model. The model is validated using the dynamic simulation of the vapor-recompression C3 process. The rolling training algorithm is presented as a plug-in in StS training to achieve auto-error-calibration for rolling prediction. Therefore, the proposed model achieves long-term horizontal predictions using rolling prediction based on initial observed data and various manipulation sequence conditions without any new historical inputs. It is revealed that the proposed StS model is a true data-driven digital twin of a physical simulator for implementing infinity rolling predictions.