Abstract:During severe nuclear accidents at Nuclear Power Plants (NPPs), it is essential to predict the dispersion of radioactive plumes using atmospheric models for informed decision-making and protection of people near affected areas. However, forecasting Atmospheric Radionuclide Dispersion (ADR) requires intricate and time-consuming physical simulations. Therefore, this study delves into the exploration of Long Short-Term Memory (LSTM), a Deep Learning model renowned for its proficiency in handling sequence and time series data, to anticipate the trajectory of the maximum whole-body dose rate coordinates over time. The investigation utilizes data derived from the ADR simulator of a Brazilian Pressurized Water Reactor (PWR), incorporating genuine meteorological data records from the NPP's vicinity to simulate the radioactive plume dispersion under hypothetical severe accident scenarios. The investigation yielded an LSTM model capable of forecasting the movement of the maximum whole-body dose rate coordinates at (t + 1) time steps. This developed LSTM model presented a Mean Absolute Error (MAE) of 6.331 when evaluated against a simulated test dataset.

Exploring LSTM-based prediction for radioactive plume atmospheric dispersion in nuclear power plant emergencies: A preliminary study