Research on predicting the diffusion of radionuclide in complex building scenarios based on a hybrid deep learning model

In the event of an extreme nuclear accident, predicting the dispersion of radionuclide is critical for nuclear emergency response. Traditional atmospheric dispersion models struggle to balance accuracy and time efficiency, failing to meet the demands of nuclear emergency situations. Therefore, this study proposes an innovative hybrid deep learning model—Dual CNN-LSTM. On the Indianapolis dataset, the model demonstrates favorable predictive performance, with a coefficient of determination(R² = 0.6351, RMSE = 0.0495, training time = 1919.00 s, prediction time = 1.11 s).The study also found that incorporating Gaussian plume results into the input features reduced the model’s performance in complex scenarios. Through simulation validation, the plume shapes produced by the model were found to be highly consistent with experimental data, indicating that atmospheric stability significantly affects concentration peaks and providing a scientific basis for relevant decision-making.