川南中深层页岩气开发过程中，常规油气藏工程方法，如裂缝扩展、应力诱导分析和数值模拟等研究过程使得加密井的预测工作繁重，且无法有效应对不同生产阶段的产能差异性，应用条件苛刻。为了快速且准确预测加密井产能，根据老井生产压力曲线呈趋势性“三段式”递减的特征，将剧烈下降期作为前期产水期，快速下降和缓慢下降期作为后期产气期两部分，采用优化速度快、具有自适应性和信息反馈机制的灰狼优化算法（GWO）对长短期记忆（LSTM）神经网络模型进行超参数择优，分别构建由GWO计算最优解确定隐含层神经元个数、丢包率和批次数的前、后期模型，通过损失曲线和性能指标曲线确定迭代次数，采用线性学习率热身的方法动态调整学习率，实现高速训练过程，形成分阶段的产量预测模型。实例研究表明：GWO优化的LSTM神经网络模型在预设学习率为0.002、迭代450次的条件下，短时间内能够快速实现收敛，最终性能指标达到0.923。GWO优化的LSTM神经网络模型与传统LSTM神经网络模型预测结果相比，前、后期平均绝对误差分别降低了1.290 m³/d和0.213×10⁴ m³/d；与数值模拟拟合结果相比，产气量预测的平均绝对误差降低了0.24×10⁴ m³/d。因此，改进后的LSTM神经网络模型在不同生产阶段的产能预测中表现出色，且对应阶段模型能够准确预测川南中深层页岩气加密井的产能变化，为加密井产能预测方法提供理论依据。

During the development of middle and deep gas reservoirs in South Sichuan, conventional reservoir engineering methods—such as fracture propagation, stress-induced analysis, and numerical simulation—render productivity prediction of infilling wells laborious and ineffective in addressing variations in production capacity across different production stages, with stringent application conditions. In order to quickly and accurately predict the production capacity of infilling wells, this study classifies the “three-stage” declining trend observed in the production pressure curves of existing wells into: (1) A drastic decline period, regarded as the initial water production stage; (2) a rapid decline period; and (3) a slow decline period, both considered part of the later gas production stage. The Grey Wolf Optimizer(GWO) algorithm, a fast optimization algorithm with adaptive capabilities and an information feedback mechanism, is applied for hyperparameter optimization of the Long Short-term Memory (LSTM) neural network. Two stage-specific models were constructed, with the number of hidden layer neurons, dropout rate, and batch size determined by the optimal solutions obtained via GWO. The number of iterations was selected based on the loss curve and performance metric curve, while a linear warm-up strategy was used to dynamically adjust the learning rate, facilitating high-speed training and the formation of a staged productivity prediction model. Example studies show that the GWO-optimised LSTM neural network model achieves rapid convergence with a preset learning rate of 0.002 and 450 iterations, ultimately reaching a performance index of 0.923. Compared to the conventional LSTM neural network model, the average absolute errors during the early and later stages are reduced by 1.290 m³/d and 0.213 × 10⁴ m³/d, respectively. Compared with numerical simulation fitting results, the average absolute error in gas production prediction is reduced by 0.24 × 10⁴ m³/d. Therefore, the improved LSTM neural network model demonstrates excellent performance in capacity prediction across different production stages, and the stage-specific productivity variations in infilling wells within middle and deep shale gas reservoirs in South Sichuan. This provides a theoretical foundation for productivity prediction methods of infilling wells.