The frequency and intensity of localized torrential rains caused by climate change are increasing, exacerbating flood damage in urban areas. Particularly in urban areas near rivers, they have limitations in effectively responding to short-duration heavy rainfall. Therefore, this study developed a Bi-LSTM-based inflow prediction model and a PPO reinforcement learning-based pump operation decision-making system for the drainage pump station in Samcheok City, Gangwon Province. The inflow prediction model was trained using Huff rainfall distributions and 898 SWMM simulation datasets to predict 8-hour inflows from 6-hour input data. It demonstrated an NSE of 0.98 for Huff 1st percentile rainfall and an NSE of 0.90 for Huff 4th percentile rainfall. The pump operation system was trained using the PPO algorithm, designed with five objective functions: flood prevention, safe water level maintenance, energy efficiency, and proactive response. Compared to rule-based operation, it reduced the maximum water level by an average of 0.61m (16.8%) and achieved a maximum reduction rate of 25.6% under 100% drainage capacity extreme conditions. The AI-based pump operation system developed in this study is expected to overcome the limitations of existing reactive responses, enable proactive responses, and contribute to reducing urban flood damage.