Abstract: Accurate prediction of natural gas well production is of great significance to optimize development decisions. Existing prediction methods mostly focus on overall modeling, which is difficult to adapt to the life cycle evolution of natural gas wells, which is characterized by “increasing production, stabilizing production, and decreasing production”; moreover, most of the existing data-driven models ignore the physical law of reservoir seepage field evolution over time, which is difficult to accurately reflect the temporal changes of physical properties. In this paper, a gas well production prediction method (FGPM) is proposed from the perspective of life cycle. Firstly, the life cycle of gas wells is divided by the breakpoint detection algorithm, and the relative fluctuation rate of production is combined to identify the production stage of gas wells; subsequently, a prediction model based on the coding-decoding structure is constructed, with feature matching training conducted for each production stage and integrated into the full-life-cycle model; finally, the model optimization is carried out from the optimization of the model hyper-parameters and the integration of seepage flow laws. The experiments proved that: (1) the prediction accuracy of the full life cycle model is higher compared with that of the single cycle model. Compared with the classical methods such as TCN, LSTM and GRU, the prediction accuracy of FGPM is improved by 42.10%, 39.40% and 36.16%, respectively. (2) Optimization measures for FGPM design play a positive role in model performance enhancement: (a) the model with optimized hyper-parameters has improved its MAPE, MAE and RMSE by 9.9%, 16.2% and 19.4%, respectively; (b) the FGPM incorporating the constraints of seepage law has a MAPE of only 4.874479%, and the model performance has been further improved.