【摘要】 针对辐射屏蔽优化设计中存在的消耗时间长、优化效率低的问题，提出一种基于粒子飞行样本更新策略的动态径向基代理模型。首先采用径向基神经网络建立真实目标函数的初始代理模型，然后通过差分进化算法对代理模型进行全局寻优，然后基于代理模型寻优结果和粒子飞行样本更新策略产生新样本点，最后将新样本点加入原有样本点后重新更新代理模型并循环迭代，直至满足收敛条件。该方法以代理模型拟合精度为依据控制原有样本点向随机样本点和最优预测样本点的飞行速度，可以实现动态代理模型全局探索与局部探索的自适应平衡。为验证方法的有效性，将所提方法应用于12个数值测试函数和船用反应堆辐射屏蔽优化设计工程实例，并与其他优化方法计算结果进行对比。结果表明：对于数值测试函数，所提方法在寻优结果、样本点数量和算法鲁棒性方面均具有显著优势，对于辐射屏蔽优化设计实例，所提方法得到的中子透射率为另外两种方法的48%和8%，所需样本点数量为静态代理模型的25%，证明该方法是求解辐射屏蔽优化等昂贵优化问题的有效方法。更多还原

【Abstract】 [Background] Radiation shielding is crucial for ensuring the environmental safety of personnel and nuclear facilities in the nuclear industry. As it usually consumes a long time in single simulation calculation, the optimization design of radiation shielding is a classical expensive optimization problem. [Purpose] This study aims to reduce the number of sampling points required in the radiation shield optimization design and improve the efficiency of intelligent optimization algorithms. [Methods] A dynamic radial basis function based on particle flying(PF-DRBF) surrogate model was proposed in this study for radiation shielding optimization. Firstly, a radial basis neural network was used to build the initial surrogate model of the actual objective function, and the surrogate model was globally searched for optimality by a differential evolutionary(DE) algorithm. Thereafter new sample points were selected to join the original sample points based on the result of the surrogate model search and the particle flight sample update strategy, and the surrogate model was updated based on the new set of sample points and iterated until the convergence condition was satisfied. Since the flight speed of the original sample point to the random sample point and the optimal predicted sample point based on the fitting accuracy of the surrogate model were controlled by the model, the adaptive balance between the global exploration and the local exploration of the dynamic surrogate model was achieved. Finally, in order to verify the effectiveness of the method, the proposed method was applied to 12 numerical test functions and the optimization design for radiation shielding of marine reactors, and the calculation results of other optimization methods, i. e., mode pursuing sampling(MPS) method and dynamic radial basis function based on trust region(TR-DRBF) method, were compared. [Results] The comparative results show that for numerical test functions, the proposed PF-DRBF method has significant advantages in search accuracy,search efficiency, and algorithm robustness. For the radiation shielding optimization, the neutron transmittance obtained by the proposed method is 48% and 8% of MPS and TR-DRBF methods, and the number of required sample points is 25% of the static surrogate model. [Conclusions] The results of this study indicate that by using the dynamic surrogate model based on the particle flying algorithm, the number of sample points needed to solve the expensive optimization problem is greatly reduced. The dynamic radial basis surrogate model with particle flight is an effective method for radiation shielding optimization.更多还原