Abstract: Accurate prediction of well trajectory is fundamental to well trajectory control, and therefore, extremely important for improving drilling efficiency. However, there are many factors that may change well trajectory, and the downhole mechanical behavior is complex, which leads to high difficulties in accurately predicting well trajectory. This presents a dual-input sequence-to-sequence (Di-S2S) model. The model considers time series features, such as WOB and ROP, and non-time series features, such as drilling mode, formation stratigraphy and BHA structure. The non-time series features were numerically characterized with dimensionality reduction via a natural language processing process, and a dynamic updating mechanism based on incremental training was built for the model. The data of 12 wells were analyzed with the Di-S2S model, and the results were compared with those of the LSTM and BP models. The results show that the average absolute error of well inclination angles is reduced by 49% and 8% respectively, and the average absolute error of azimuths is reduced by 49% and 24%, respectively, compared with the LSTM and BP models. Moreover, compared with the offline model, the average absolute errors of well inclination and azimuth of the dynamic updating model, both lower than 0.2°, are reduced by 61% and 67% respectively. The presented Di-S2S model has high accuracy and enables real-time prediction. This research provides technical support for steerable drilling.