Dynamic prediction of spatial dose rate distribution during nuclear severe accidents by means of active machine learning and mobile sensors

Abstract

The prediction of atmospheric dispersion of nuclear material is crucial to support emergency response during nuclear accidents, due to the potential severity and far-reaching consequences of such accidents. This importance is particularly evident in the context of severe accidents, like the incident that occurred at the Fukushima Daiichi Nuclear Power Plant (NPP). NPP are commonly supported by atmospheric dispersion systems (ADS), which estimate the spatial dose rates distribution, referred to as the dose rate map (DRM), by means of physical model simulations, including source term (ST), wind fields, atmospheric transport and diffusion, and dose calculations. However, it is well-known that a real accident scenario may differ from those modeled especially under severe conditions, and this fact may lead to erroneous predictions resulting in poor or catastrophic decision making. Aiming to minimize this problem, several efforts have been made to improve predictions based on field measurements, such as inverse problems and other methods to correct/adjust ST parameters. Such methods, however, complement the physical model calculations. This paper proposes a novel approach to predict DRM at ground level, based only upon field measurements (avoiding execution of physical model simulations). The idea is to dynamically adapt a set of mobile sensors (for example monitoring drones swarm) to follow the radioactive plume. To achieve this, an active machine learning (AML) approach was developed based on Gaussian process regression (GRP) and swarm intelligence. To test the approach, computational experiments using realistic simulated data (generated on the ADS simulator of a Brazilian NPP) has been used. As result, DRMs have been reproduced with good qualitative (visual) representation and satisfactory quantitative metrics. The average correlation coefficient between predicted and real DRMs was approximately 0.72 (ranging from 0.64 to 0.97 during plume evolution), demonstrating to be a promising approach to predict DRMs.