BOOSTING METHOD OF HETEROSKEDASTIC MODELS FOR THE PREDICTION OF THE SAHARA DUST CONCENTRATION IN THE ATMOSPERIC AIR OF UKRAINE

Abstract

The paper presents new method of heteroskedastic models boosting and its applied usage on the example of the Sahara dust concentration in the atmospheric air of Ukraine. Recently the increased frequency of fine dust transfer from the Sahara Desert across the Mediterranean sea on the territory of Europe, in particular – in Ukraine, is observed. This phenomenon complicates the prediction of atmospheric air quality as a result of destruction of the stable regularities of pollution, as new factors are added, they need special models for adequate description. Special meteorological mode of the Sahara Desert dust spread enables to assume that the dispersion of the remains of ARIMA model may be random process and for its description it is expedient to use heteroskedastic models, such as GARCH. However, the conventional GARCH-models are efficient only if one dominating random process is available. If several such processes are added, conventional models loose their efficiency.

In the given study the application of boosting approach for the construction of the assembly models is suggested, these models, unlike the available, comprise the cooperation of the decision trees and heteroskedastic models for modeling complex heteroskedastic processes. The suggested method, as it is accepted for the boosting models, is based on the iterative process of models selection, where the next model takes into account the errors of the previous model. For the verification of the efficiency of the method the data of civil monitoring of atmospheric air EcoCity were used, in particular, data regarding Vinnytsia region by PM1 index, which indicate the periods, when the concentration of the Sahara fine dust in the atmospheric air of the region reached especially abnormal values.

It was proved that the process of the Saharan dust spreading in Vinnytsia region is heteroskedastic. SARIMAX models and typical GARCH-models using Python-libraries statsmodels and arch are constructed. It is revealed that the model ARIMA demonstrates far better results as compared with classic
GARCH-models with different parameters, this shows insufficient efficiency of these GARCH-models. The suggested method of boosting heteroskedastic models allows to reach far greater accuracy than all these models in the whole range of values, except the value of the largest anomaly, this value is impossible to foresee. Thus, the forecasting method, developed in this research is an efficient approach for the solution of complex forecasting problems, the example of which is forecasting of the atmospheric air quality during the Sahara dust spreading in Ukraine.