Surrogate modelling for the forecast of Seveso-type atmospheric pollutant dispersion

Abstract

This paper presents a framework for the development of a computationally-efficient surrogate model for air pollution dispersion. Numerical simulation of air pollution dispersion is of fundamental importance for the mitigation of pollution in Seveso-type accidents, and, in extreme cases, for the design of evacuation scenarios for which long-range forecasting is necessary. Due to the high computational load, sophisticated simulation programs are not always useful for prompt computational studies and experimentation in real time. Surrogate models are data-driven models that mimic the behaviour of more accurate and more complex models in limited conditions. These models are computationally fast and enable efficient computer experimentation with them. We propose two methods. The first method develops a grid of independent dynamic models of the air pollution dispersion. The second method develops a reduced grid with interpolation of outputs. Both are demonstrated in an example of a realistic, controlled experiment with limited complexity based on an approximately 7 km radius around the thermal power plant in Šoštanj, Slovenia. The results show acceptable matching of behaviour between the surrogate and original model and noticeable improvement in the computational load. This makes the obtained surrogate models appropriate for further experimentation and confirms the feasibility of the proposed method.