With the continuous growth of urban electricity demand

direct expansion variable air volume (DXVAV) air conditioning systems are investigated as flexible adjustable loads to participate in power system demand response regulation. However

it remains a challenge to optimize power control while satisfying thermal comfort and safety constraints

particularly in avoiding power peak rebound. This paper proposes a model-free control strategy based on deep reinforcement learning(DRL)

leveraging its self-learning and adaptive capabilities to achieve intelligent regulation of DXVAV systems. This study incorporates a safety layer based on nonlinear mapping to ensure component safety

designs an adaptive reward mechanism to address power peak rebound during recovery phases and maintain grid stability and develops a cumulative reward optimization strategy to track and optimize control policies

thereby resolving control delay issues. Simulation results demonstrate that the proposed strategy effectively reduces real-time power consumption to demand response targets while ensuring smooth power recovery

significantly enhancing system flexibility and safety.