Abstract: The accurate prediction of natural gas well production is of paramount significance for optimizing development and operational decisions. Most existing forecasting methodologies predominantly emphasize holistic modeling approaches, which struggle to adapt to the characteristic lifecycle evolution pattern of gas wells—typically encompassing the "production increase", "stable production", and "production decline" phases. Furthermore, prevalent data-driven models often overlook the physical principles governing the temporal evolution of reservoir seepage fields, thereby failing to accurately capture time-dependent variations in reservoir properties. To address these limitations, this paper proposes a gas well production prediction method (Full life cycle gas production forecast model, FGPM) from a lifecycle perspective. The proposed methodology begins by segmenting the gas well lifecycle using a breakpoint detection algorithm (Pruned exact linear time, PELT), and subsequently identifies the specific production phase based on the relative fluctuation rate of production data (PELT-V). Next, a prediction model based on an encoder–decoder (ED) architecture is constructed. This model is trained using phase-specific feature matching and ultimately integrated into a comprehensive full-lifecycle forecasting framework. Finally, model optimization is conducted from two key aspects: hyperparameter tuning (Improved Artificial Fish Swarming Algorithm, IAFSA) and integration of seepage flow principles. The experimental outcomes are as follows: (1) The PELT algorithm, combined with the relative volatility of daily gas production, can effectively divide the production data into three stages, namely, increase, stable, and decline. (2) To verify the accuracy and reliability of FGPM, it was comprehensively compared with long short-term memory (LSTM), gated recurrent unit (GRU), and temporal convolutional network (TCN). The experimental results show that the FGPM performs particularly well in the task of forecasting daily natural gas production, achieving an MAPE of 5.82489%, an MAE of 0.552519, and an RMSE of 0.685578, all of which outperform the three classical network models. Compared with the multiscale fusion-based time series forecasting model TimeMixer, FGPM also demonstrates superior predictive performance, which fully demonstrates the superior performance of FGPM in the prediction of natural gas daily production. (3) To evaluate the contribution of each module in FGPM, multiple ablation experiments were designed. In Experiment 1, the PELT-V algorithm was removed, and the IAFSA-ED model was used for prediction to verify the effectiveness of the lifecycle division strategy. In Experiment 2, both IAFSA and PELT-V were removed, and only the ED model was used to validate the effectiveness of PELT-V and IAFSA. In Experiment 3, the proposed FGPM model was applied for prediction. The ablation results show that the prediction performance of the IAFSA-ED model is significantly better than that of the original ED model, demonstrating the advantage of the proposed IAFSA. Furthermore, compared with the IAFSA-ED model, FGPM achieves reductions of 6.79%, 16.06%, and 9.37% in MAPE, MAE, and RMSE, respectively, which clearly confirms the positive impact of the proposed PELT-V lifecycle division strategy on prediction performance. (4) Finally, the effect of introducing physical parameters on further improving prediction performance was investigated. Compared with the purely data-driven FGPM, the FGPM incorporating flow-mechanism constraints achieves better predictive performance, with MAPE, MAE, and RMSE reaching 4.874479%, 0.63981, and 0.797082, respectively. In summary, accurate life-cycle division of gas wells has a significant impact on improving the performance of downstream prediction tasks. Compared with using a single model to represent the entire life cycle, a stagewise ensemble modeling approach can comprehensively exploit the differences in production characteristics across different periods, thereby achieving lower prediction errors. Moreover, hyperparameter optimization algorithms offer clear advantages in enhancing the model predictive performance, and the integration of physical information with data-driven methods can effectively improve the prediction accuracy.