Being able to predict high concentrations of tropospheric ozone is important because of its negative impact on human health. In this paper eight regressor-selection methods are utilised in a case study for ozone prediction in the city of Nova Gorica, Slovenia. The comparison of the selected methods proved to be useful for building models that successfully predict the ozone concentrations for the treated case. Different regressors are selected for different models, with different methods based on the validation procedure’s cost functions. Namely, for the model to predict the maximum daily ozone concentration, ten regressors are selected; for the average concentration of ozone between 8.00 and 20.00 h, fifteen regressors are selected; and for the average daily concentration, ten regressors are selected. The result of the study is a regressor selection that is specific for a particular geographical location. Moreover, the study reveals that regressor selection, as well as the obtained models, differ depending on the kind of averaging interval of the ozone concentration.
COBISS.SI-ID: 28481319
Forecasting the ozone concentration and informing populations about times when air-quality standards are not being met is an important task. One of the possibilities for carrying out such forecasting is Gaussian-process (GP) models and artificial neural networks, both of which can make a forecast using available present-time or historical measurements of air-pollution or meteorological parameters at the location of automatic air-quality measuring stations. In this paper an on-line updating, or evolving, GP model is evaluated. Its main advantage is an ability to learn with almost no prior knowledge or data. This means that it can be used for modelling a range of variables shortly after the measurement station is established. The evolving GP model for ozone forecasting is compared to the full GP model and the multilayer-perceptron neural networks model. The investigation shows that the evolving GP model performs sufficiently well for it to be used for informing citizens when alarm-level concentrations occur.
COBISS.SI-ID: 29253671
Many meteorological parameters present a natural diurnal cycle because they are directly or indirectly dependent on sunshine exposure. The solar radiation diurnal pattern is important to energy production, agriculture, prognostic models, health and general climatology. This article aims at introducing a new type of radial frequency diagram – hereafter called sunflower – for the analysis of solar radiation data in connection with energy production and also for climatological studies. The diagram is based on two-dimensional data sorting. Firstly data are sorted into classes representing hours in a day. Then the data in each hourly class is sorted into classes of the observed variable values. The relative frequencies of the value classes are shown as sections on each hour’s segment in a radial diagram. The radial diagram forms a unique pattern for each analysed dataset. Therefore it enables the quick detection of features and the comparison of several such patterns belonging to the different datasets being analysed. The sunflower diagram enables a quick and comprehensive understanding of the information about diurnal cycle of the solar radiation data. It enables in a graphical form, quick screening and long-term statistics of huge data quantities when searching for their diurnal features and finding the differences between the data for several locations. The results of the data analysis using the sunflower diagram show how daily or monthly-based patterns are identified within small or huge data sets. The paper demonstrates the sunflower diagram usefulness over a wide range of applications from green energy production to weather analysis.
COBISS.SI-ID: 28621607