This paper presents a method for developing computationally-efficient surrogate models for the spread of air pollution. Mitigating the pollution of Seveso-type accidents and designing evacuation scenarios require long-term prediction, which is obtained with numerical simulations of the spread of air pollution. Sophisticated simulation programs frequently possess high computational load and are not suitable for real-time computational studies and experiments. Data-driven surrogate models that are computationally fast are used for such investigations. We propose a grid of independent dynamical Gaussian-process models (GP-GIM) to simulate the spread of atmospheric pollution. This is demonstrated using a realistic example of limited complexity based on a thermal power plant in Šoštanj, Slovenia. The results show an acceptable behaviour match between the surrogate and original models, with a tenfold decrease in computational load. This confirms the feasibility of the proposed method and makes the resulting surrogate model suitable for further experiments.