During the transportation of components related to nuclear materials, accidental chemical explosions may occur, resulting in the release of radionuclides. Effective decision-making during nuclear transport accidents, especially in cases with incomplete source information and a complex terrain, requires the rapid prediction of changes in radionuclide concentration. This paper proposes a method for predicting the concentration of radionuclides resulting from nuclear transport explosion accidents based on stacked long short-term memory (LSTM) networks. Specifically, this study considered plutonium-containing explosive transport and chemical explosion accidents under the pad surface of a hill as a research scenario. The diffusion data of radionuclide Pu-239 were simulated using the computational fluid dynamics (CFD) software OpenFOAM. Nuclide concentration and meteorological time series data of a specific area were selected for stacked LSTM network training and prediction based on geographical characteristics and population density. The proposed model, optimized using grid search, can stably achieve a mean absolute percentage error (MAPE) of less than 5% within 150 iterations for Pu-239 nuclide concentration prediction. The model is highly efficient and has significant practical value for use in nuclear emergencies.