Machine learning-enabled weather forecasting for real-time radioactive transport and contamination prediction

Abstract

Effective responses to nuclear accidents require accurate and fast predictions of radionuclide transport in the atmosphere. However, real-time forecasting has been challenging, since predicting the transport requires meteorological forecasting, which relies on computationally intensive numerical weather simulations. To address this challenge, we have developed statistical emulators for the Weather Research and Forecasting model (WRF) to forecast wind and temperature fields around the source location. The emulator is constructed based on pre-computed WRF simulation results and their boundary conditions from global circulation models (GCM), enabling the rapid prediction of local-scale wind/temperature fields by ingesting the GCM-predicted boundary conditions available online for the next several days. We have explored two methods: (1) deep neural network and (2) a vector autoregressive model with exogenous variables (VARX). A Gaussian plume model is then used to take these field as input and to predict the distribution of radioactive material/dose in the environment. The results show that both emulators are capable of forecasting the wind field for several days in the future in real-time, while not compromising accuracy. The autoregressive model has shown a strong performance of predicting the temperature field, owing to the strong temporal autocorrelations in the temperature data. This framework offers a powerful tool for informed decision-making during emergencies such as enabling meteorology-informed timing of containment venting, and real-time guidance of evacuation strategies.