基于强化学习的核电数字孪生模型自动化同步研究

AUTOMATED SYNCHRONOUS RESEARCH ON NUCLEAR POWER DIGITAL TWIN MODEL BASED ON REINFORCEMENT LEARNING

核电运行数字孪生系统与机组的高效同步是电厂运行优化的基础,是核电安全的重要保障。以核电蒸汽系统数字孪生模型为研究对象,提出了一种基于孪生延迟深度确定性策略梯度的自适应智能优化算法(TD3PSO),通过构建动作网络与动作价值网络,利用神经网络动态生成基本粒子群优化算法运行过程中所需要的超参数,实现超参数的自我探索,降低人工对算法的干预,解决基本粒子群优化算法容易陷入局部最优的问题。实验表明,TD3PSO算法优于基本粒子群优化算法,相比于人工调试的结果,在优化精度上提高了93.39%,自动化同步效果显著。

The efficient synchronization of the digital twin system and the unit of nuclear power operation is the basis for the optimization of power plant operation and an important guarantee for nuclear power safety.Taking the digital twin model of nuclear steam system as the research object,this paper proposes an adaptive intelligent optimization algorithm based on the deep deterministic policy gradient of twin delay(TD3PSO).By constructing the action network and action value network,the neural network is used to dynamically generate the hyperparameters required during the operation of the elementary particle swarm optimization algorithm,so as to realize the self-exploration of the hyperparameters,reduce the manual intervention in the algorithm,and solve the problem that the elementary particle swarm optimization algorithm is easy to fall into local optimum.Experiments show that the TD3PSO algorithm is superior to the elementary particle swarm optimization algorithm,and compared with the results of manual debugging,the optimization accuracy is improved by 93.39%,and the automatic synchronization effect is remarkable.