间冷塔喷雾参数设计和优化往往依赖经验和不断试错，无法充分发挥系统的动态特性来应对复杂性。本文提出了一种基于深度确定性策略梯度（DDPG）和物理约束神经网络（PINN）的间冷塔喷雾优化，旨在降低机组背压和间冷塔进风口温度。通过将DDPG算法应用于深度神经网络，将间冷塔进风口温度、机组背压等参数输入到价值网络和策略网络，得到包含阀门开启程度以及喷雾压力调节的优化方案。此外，在奖励函数中引入 PINN 的物理方程对智能体进行物理约束，从而使智能体在环境中表现出更合理更可靠的物理行为。通过实验，验证了该优化能够有效降低进风口温度以及机组背压，进风口温度下降约4℃，机组背压下降约2.5kPa且达到目标值速度增加12.5%。

The design and optimization of intercooler tower spray parameters often rely on experience and constant trial and error, which cannot fully utilize the dynamic characteristics of the system to cope with the complexity. In this paper, an intercooler tower spray optimization based on deep deterministic policy gradient (DDPG) and physically-constrained neural network (PINN) is proposed, aiming to reduce the unit backpressure and intercooler inlet temperature. By applying the DDPG algorithm to the deep neural network, parameters such as intercooler tower inlet temperature and unit back pressure are input to the value network and the policy network to obtain an optimization scheme that includes the degree of valve opening as well as spray pressure adjustment. In addition, the reward function board is redesigned, and the physical equations of PINN are introduced into the reward function to physically constrain the intelligences, so as to make the intelligences exhibit more reasonable and reliable physical behaviors in the environment. Through experiments, it is verified that the optimization can effectively reduce the inlet temperature and the unit back pressure, with the inlet temperature decreasing by about 4°C, the unit back pressure decreasing by about 2.5 kPa and the speed of reaching the target value increasing by 12.5%.

国家重点研发计划资助（2023YFC3803900）