This study proposes a method to optimize parking lot revenue while ensuring that occupancy does not exceed 100% in an O2O (online-to-offline) platform-based parking management system. To address the issue of lead time resulting from the difference between the time of sale and the time of entry, we introduce a dynamic parking permit sales model using Deep Q-Networks (DQN), a type of reinforcement learning. The proposed model determines whether to sell parking permits every hour by observing the parking environment, occupancy, and the number of permits sold. The reward structure includes both immediate and delayed rewards, and a backtracking mechanism is implemented to impose penalties by tracking when the permit that caused the occupancy to exceed the limit was sold. Numerical analysis shows that the proposed model is effective in reducing congestion and enhancing profitability compared to the continuous sales method.