Optimizing and evaluating multiple kernel density estimators for local-scale atmospheric dispersion modeling at a representative AP1000 nuclear power plant site in China

This study proposes a particle-age-based bandwidth calculation method for the Uniform kernel, an efficient kernel density estimator (KDE) in Lagrangian Particle Dispersion Model (LPDM), to mitigate excessive smoothing of concentrations near release points. The proposed method, along with Parabolic and Gaussian kernels, was jointly validated against wind tunnel experiments of a representative AP1000 nuclear power plant site with complex topography and irregular building layouts. Sensitivity analyses were also performed to quantify the performance of three KDEs with different LPDM parameters. Results show that proposed method improves Uniform kernel’s accuracy and meets acceptable criteria, achieving similar performances as the other two KDEs. Sensitivity analyses illustrate that three KDEs can suppress increasing stochastic errors with high-resolution grid settings. Uniform kernel’s bandwidths near release point should vary sharply horizontally and smoothly vertically, with appropriate bandwidth multiplication factors from 0.3 to 0.5. Parabolic and Gaussian kernels show less sensitivity to maximum horizontal bandwidth. The integer governing the Gaussian kernel’s cut-off mechanism should be at least 2. Uniform and Parabolic kernels may lead to partial loss of particle mass contribution with coarser grid sizes. Gaussian kernel demonstrates more robust performance with lower numbers of released particles.