Abstract
<jats:p>Underground gas storage (UGS) plays a critical role in seasonal peak shaving and supply security. Multi-well production allocation—determining optimal injection/production rates for each well under safety constraints—remains a challenging sequential decision problem due to the strong coupling between reservoir dynamics, wellbore hydraulics, and surface pipeline constraints. This paper presents the first comprehensive DRL framework for UGS multi-well production allocation, evaluating four algorithms (DDPG, SAC, PPO, TD3) against three baselines (rule-based heuristic, GA, NSGA-II) on a realistic 15-well UGS dataset. Experiments conducted on a Hygon K100_AI DCU accelerator (65 GB VRAM) with 500,000 training steps per algorithm and 5 random seeds reveal that SAC achieves the best DRL performance (mean reward −177.75 ± 3.47), while PPO converges fastest (~7,000 episodes). A genetic algorithm baseline initially appears to outperform all DRL methods, but this advantage does not survive controlled evaluation: under a single-surrogate normalized protocol, SAC (−113.41 ± 3.64) is statistically indistinguishable from GA (−115.24 ± 0.54), demonstrating that the apparent gap is an artifact of differing evaluation conditions rather than a genuine algorithmic advantage. Constraint-handling experiments show that fixed-penalty SAC maintains low violation rates (1.9–2.5%) across normal, high-demand, and sensor-noise scenarios. Ablation studies on reward weight sensitivity and planning horizon further characterize algorithm behavior. This work establishes the first benchmark for DRL-based UGS production allocation and provides practical guidelines for algorithm and constraint-handling strategy selection.</jats:p>