关键词: 核事故, 源项估算, 反向传播神经网络(BPNN), 维数约简, 主成分分析(PCA), 随机森林(RF)

Abstract: In order to accurately estimate the radioactive source terms in the nuclear accident consequence assessment, the BPNN inversion model of nuclear accident source terms was optimized. A PCA-BPNN assessment model was built after reducing the number of factors influencing the terms from 10 to 6. The cumulative contribution rate of the 6 factors was greater than 85%. Random forest algorithm was used to calculate values of importance of the influence factors to remove the wind direction and the height of the mixed layer, and to build an RF-BPNN assessment model. The estimation results by the above three models were analyzed. The results show that compared with BPNN model, PCA-BPNN model and RF-BPNN model have shorter estimation time and smaller error, the two can reflect the source information of the accident truthfully, and that RF-BPNN model has better accuracy and stability than PCA-BPNN model.

Key words: nuclear accident, source term assessment, back propagation neural network(BPNN), dimensionality reduction, principal component analysis(PCA), random forests(RF)