Study On Back Calculation Of Airborne Radionuclides Dispersion With Genetic Algorithm

In a nuclear emergency response support system, the difusion simulation of at-mospheric radionuclides is a critical component.A predication of the nuclides difu-sion made by the dispersion model can provide information for the guidance of publicevacuation and escape, and a technology for the evaluation of the consequences ofa nuclear accident. Radionuclide difusion model requires the input source term in-formation and meteorological information, which is not easy to obtain under accidentsituations.To get a better source term estimation,a technology was developed, in whichthe source term and difusion coefcients is adjusted to minimize the gap of the out-put of dispersion model and the observation data.As the technical detail suggests, asource term back calculation problem can always be converted to a minimal optimiza-tion problem.and genetic algorithm is a good candidate for solving such a optimizationproblem with high degree of nonlinearity, many parameters, and many local optima.For the study on the application of genetic algorithm in radionuclide difusion sourceterm back calculation, we did some research as the following list:1. Developed a radionuclide difusion simulation library, named pysjtupuf.2. Did a deep research of genetic algorithm’s mechanism, and the current status inthis area.3. Did a research on the ftness functions, which is part of genetic algorithms usedin the procedure of difusion coefcients correlation.As the result shows, to get a better estimation of the diffusion coefficients, observation error should be taken into consideration, the greater a observed data has, the smaller the weight it should be given in the fitness function.4. The Kincaid experimental data was converted into an appropriate format for pysj-tupuff, on which the coefficients correction procedure was performed, and a boost of the dispersion model’s foreign ability was seen with the corrected diffusion coefficient. By comparing the model’s performance in different experimen, we saw a positive correlation between the number of observation stations and the model’s forecast ability.