Hazardous Materials Prediction Using an Artificial Neural Network and Meteorological FASDAS Data Assimilation

Estimating and quantifying in an accurate way, the distribution of pollutant concentrations plays a crucial role in the emergency response management in major nuclear accidents. As in many areas of the environmental studies, air quality models are used to simulate atmospheric dispersion phenomena, although they lack accuracy due to the uncertainty in source term and meteorological fields. Even though the ability of a mesoscale meteorological model to assimilate observational data is an efficient way to improve operational air quality model forecasts, errors still occur in the flux-adjusting surface data assimilation system (FASDAS) algorithm when predicting the wind components, a factor that influences directly the plume shape, concentration distribution, and ground deposition of radionuclides. Thus, operational models should carefully handle this meteorological parameter with both accuracy and efficiency. In response to this problem, our newly proposed combined data assimilation scheme is used to combine deep learning and meteorological data assimilation. In particular, a correction of FASDAS real-time wind predictions method is proposed using an artificial neural network (ANN). This alternative wind forecast method aims to train the ANN in a manner that it can predict wind speed and direction efficiently, correcting, therefore, the errors introduced with the relaxation assimilation methods. The combined system rapidly corrects the errors in the wind data, thereby significantly improves the Gaussian dispersion model results.