Use of neural network to improve the dispersion models performances: proposal of an advanced methodology

Using the results of a dispersion model, the reproduced concentration levels were evaluated for two typical calibration data sets. The model reproduces concentration levels with a precision that is not always optimal. Given this fact, a neural net was trained on the basis of the meteorological situation and the values for the pollutant levels reproduced by the model. Applying this methodology, the ground level concentrations indicate an improvement in the predictive capacity of dispersion models, when appropriately filtered by a neural net. Notwithstanding the limited nature of the example dealt with here, the proposed methodology can be generalised, thus opening new perspectives for integrated models in the simulation of complex situations