Predicted Ozone Concentrations Research Articles

Prediction of hourly ozone concentrations with multiple regression and multilayer perceptron models

In this work ozone observations of an urban area of the east coast of the Iberian Peninsula, are analyzed. The data set contains measurements from five automatic air pollution monitoring stations (background suburban or traffic urban). The application of multiple linear regression and neural networks models is considered. These models forecast hourly ozone levels for short-term prediction intervals (1, 8, and 24 h in advance). The study period is 2010–2012. The input variables are meteorological observations, ozone and nitrogen oxides concentrations, and daily and weekly seasonal cycles. The performance criteria to evaluate the computations accuracy are the residual mean square error, the mean absolute error, and the correlation coefficient between observations and predictions. These criteria have better results for the 1-h and 24-h predictions in all the locations. The comparison of multiple linear regressions and multilayer perceptron networks indicates that the second approach allows to obtain more accurate forecast for the three prediction intervals.

Forecasting ozone concentrations in the east of Croatia using nonparametric Neural Network Models

Ozone is one of the most significant secondary pollutants with numerous negative effects on human health and environment including plants and vegetation. Therefore, more effort is made recently by governments and associations to predict ozone concentrations which could help in establishing better plans and regulation for environment protection. In this study, we use two Artificial Neural Network based approaches (MPL and RBF) to develop, for the first time, accurate ozone prediction models, one for urban and another one for rural area in the eastern part of Croatia. The evaluation of actual against the predicted ozone concentrations revealed that MLP and RBF models are very competitive for the training and testing data in the case of Kopački Rit area whereas in the case of Osijek city, MLP shows better evaluation results with 9% improvement in the correlation coefficient. Furthermore, subsequent feature selection process has improved the prediction power of RBF network.

Prediction of column ozone concentrations using multiple regression analysis and principal component analysis techniques: A case study in peninsular Malaysia

The aim of this study is to develop new algorithms of the column ozone (O3) in Peninsular Malaysia using statistical methods. Four regression equations, denoted as O3 NEM, O3 SWM, (PCA1) O3 NEM season, and (PCA2) O3 SWM season, were developed. Multiple regression analysis (MRA) and principal component analysis (PCA) methods were utilized to achieve the objectives of the study. MRA was used to generate regression equations for O3 NEM and O3 SWM, whereas a combination of the MRA and PCA methods were used to generate regression equations for PCA1 and PCA2. The results of the best regression equations for the column O3 through MRA by using four of the independent variables were highly correlated (R = 0.811 for SWM, R = 0.803 for NEM) for the six-year (2003–2008) data. However, the result of fitting the best equations for the O3 data using four of the independent variables gave approximately the same R values (≈0.83) for both the NEM and SWM seasons using the combined MRA and PCA methods. The common variables that appeared in both regression equations were H2O vapor and NO2. This result was expected because NO2 is a precursor of O3. The correlation coefficients (R) of the validation for the NEM and SWM seasons were 0.877–0.888 and 0.837–0.896, respectively. These statistical values indicated a very good agreement between the monthly predicted and observed O3 for Peninsular Malaysia.

Development of Long-term Spatiotemporal Models for Ambient Ozone in Six Metropolitan regions of the United States: The MESA Air Study.

Current epidemiologic studies rely on simple ozone metrics which may not appropriately capture population ozone exposure. For understanding health effects of long-term ozone exposure in population studies, it is advantageous for exposure estimation to incorporate the complex spatiotemporal pattern of ozone concentrations at fine scales. To develop a geo-statistical exposure prediction model that predicts fine scale spatiotemporal variations of ambient ozone in six United States metropolitan regions. We developed a modeling framework that estimates temporal trends from regulatory agency and cohort-specific monitoring data from MESA Air measurement campaigns and incorporates land use regression with universal kriging using predictor variables from a large geographic database. The cohort-specific data were measured at home and community locations. The framework was applied in estimating two-week average ozone concentrations from 1999 to 2013 in models of each of the six MESA Air metropolitan regions. Ozone models perform well in both spatial and temporal dimensions at the agency monitoring sites in terms of prediction accuracy. City-specific leave-one (site)-out cross-validation R2 accounting for temporal and spatial variability ranged from 0.65 to 0.88 in the six regions. For predictions at the home sites, the R2 is between 0.60 and 0.91 for cross-validation that left out 10% of home sites in turn. The predicted ozone concentrations vary substantially over space and time in all the metropolitan regions. Using the available data, our spatiotemporal models are able to accurately predict long-term ozone concentrations at fine spatial scales in multiple regions. The model predictions will allow for investigation of the long-term health effects of ambient ozone concentrations in future epidemiological studies.

Comparison of regional and global land cover products and the implications for biogenic emission modeling

Accurate estimates of biogenic emissions are required for air quality models that support the development of air quality management plans and attainment demonstrations. Land cover characterization is an essential driving input for most biogenic emissions models. This work contrasted the global Moderate Resolution Imaging Spectroradiometer (MODIS) land cover product against a regional land cover product developed for the Texas Commissions on Environmental Quality (TCEQ) over four climate regions in eastern Texas, where biogenic emissions comprise a large fraction of the total inventory of volatile organic compounds (VOCs) and land cover is highly diverse. The Model of Emissions of Gases and Aerosols from Nature (MEGAN) was utilized to investigate the influences of land cover characterization on modeled isoprene and monoterpene emissions through changes in the standard emission potential and emission activity factor, both separately and simultaneously. In Central Texas, forest coverage was significantly lower in the MODIS land cover product relative to the TCEQ data, which resulted in substantially lower estimates of isoprene and monoterpene emissions by as much as 90%. Differences in predicted isoprene and monoterpene emissions associated with variability in land cover characterization were primarily caused by differences in the standard emission potential, which is dependent on plant functional type. Photochemical modeling was conducted to investigate the effects of differences in estimated biogenic emissions associated with land cover characterization on predicted ozone concentrations using the Comprehensive Air Quality Model with Extensions (CAMx). Mean differences in maximum daily average 8-hour (MDA8) ozone concentrations were 2 to 6 ppb with maximum differences exceeding 20 ppb. Continued focus should be on reducing uncertainties in the representation of land cover through field validation.Implications: Uncertainties in the estimation of biogenic emissions associated with the characterization of land cover in global and regional data products were examined in eastern Texas. Misclassification between trees and low-growing vegetation in central Texas resulted in substantial differences in isoprene and monoterpene emission estimates and predicted ground-level ozone concentrations. Results from this study indicate the importance of land cover validation at regional scales.

Evaluation of MEGAN predicted biogenic isoprene emissions at urban locations in Southeast Texas

Summertime isoprene emissions in the Houston area predicted by the Model of Emissions of Gases and Aerosol from Nature (MEGAN) version 2.1 during the 2006 TexAQS study were evaluated using a source-oriented Community Multiscale Air Quality (CMAQ) Model. Predicted daytime isoprene concentrations at nine surface sites operated by the Texas Commission of Environmental Quality (TCEQ) were significantly higher than local observations when biogenic emissions dominate the total isoprene concentrations, with mean normalized bias (MNB) ranges from 2.0 to 7.7 and mean normalized error (MNE) ranges from 2.2 to 7.7. Predicted upper air isoprene and its first generation oxidation products of methacrolein (MACR) and methyl vinyl ketone (MVK) were also significantly higher (MNB = 8.6, MNE = 9.1) than observations made onboard of NOAA's WP-3 airplane, which flew over the urban area. Over-prediction of isoprene and its oxidation products both at the surface and the upper air strongly suggests that biogenic isoprene emissions in the Houston area are significantly overestimated. Reducing the emission rates by approximately 3/4 was necessary to reduce the error between predictions and observations. Comparison of gridded leaf area index (LAI), plant functional type (PFT) and gridded isoprene emission factor (EF) used in MEGAN modeling with estimates of the same factors from a field survey north of downtown Houston showed that the isoprene over-prediction is likely caused by the combined effects of a large overestimation of the gridded EF in urban Houston and an underestimation of urban LAI. Nevertheless, predicted ozone concentrations in this region were not significantly affected by the isoprene over-predictions, while predicted isoprene SOA and total SOA concentrations can be higher by as much as 50% and 13% using the higher isoprene emission rates, respectively.

Evaluating methods for spatial mapping: Applications for estimating ozone concentrations across the contiguous United States

Abstract Understanding spatial variability of air pollutant concentrations is critical for public health assessments. Our goal is to examine ground-level ozone and comparatively evaluate method performance for predicting and mapping national concentrations across the United States, while assessing the importance of accounting for spatial variability. Cross-sectional US EPA ozone monitoring data was acquired for three days in 2006, plus environmental covariates of land use, traffic, temperature, elevation, and population. Evaluation of ozone variability was assessed with land use regression (LUR) and spatially explicit kriging models. Ozone concentration was predicted with four approaches, including LUR, inverse distance weighting (IDW), ordinary kriging, and universal kriging, and evaluated with a Monte Carlo cross-validation simulation. Results were mapped for the continental United States. Temperature, elevation, and distance to major roads were significantly related to ozone concentrations and examination of spatial dependence on LUR models revealed the presence of residual spatial variation. Cross-validation results found kriging outperformed both LUR and IDW in terms of root mean squared prediction error. We demonstrate that national-level ozone is best evaluated using the statistically optimal kriging models, which account for residual spatial variation. Universal kriging was preferred over ordinary kriging by allowing us to assess the significance of environmental covariates both for inference and prediction of ozone concentrations.

Prediction of ozone concentration in tropospheric levels using artificial neural networks and support vector machine at Rio de Janeiro, Brazil

It is well known that air quality is a complex function of emissions, meteorology and topography, and statistical tools provide a sound framework for relating these variables. The observed data were contents of nitrogen dioxide (NO2), nitrogen monoxide (NO), nitrogen oxides (NOx), carbon monoxide (CO), ozone (O3), scalar wind speed (SWS), global solar radiation (GSR), temperature (TEM), moisture content in the air (HUM), collected by a mobile automatic monitoring station at Rio de Janeiro City in two places of the metropolitan area during 2011 and 2012. The aims of this study were: (1) to analyze the behavior of the variables, using the method of PCA for exploratory data analysis; (2) to propose forecasts of O3 levels from primary pollutants and meteorological factors, using nonlinear regression methods like ANN and SVM, from primary pollutants and meteorological factors. The PCA technique showed that for first dataset, variables NO, NOx and SWS have a greater impact on the concentration of O3 and the other data set had the TEM and GSR as the most influential variables. The obtained results from the nonlinear regression techniques ANN and SVM were remarkably closely and acceptable to one dataset presenting coefficient of determination for validation respectively 0.9122 and 0.9152, and root mean square error of 7.66 and 7.85, respectively. For these datasets, the PCA, SVM and ANN had demonstrated their robustness as useful tools for evaluation, and forecast scenarios for air quality.

Prediction of Maximum Ground Ozone Levels using Neural Network

Ozone is one of the most effective pollutants in lower atmosphere. Concentration of ozone in atmosphere reveals its impact on plants, human and on other organic materials. Many techniques had been used in past to calculate the concentration of ozone with the help of other environmental factors like wind, humidity, temperature and etc. Prediction models like Artificial Neural Network (ANN) have gained much reputation in calculating accurate results with learning data. This paper shows a study of integration of predicted ozone concentration by two ANN proposed models. The study initiated with data collection from the study area. The collected data is then fed to the proposed ANN models as training data to get the concentrations of ozone with many input variables temperature, humidity, wind speed, incoming solar radiation, sulfur dioxide, nitrogen dioxide, and previous ozone data as predictor. The study shows the great dependence of ozone concentration upon environmental factors. The two proposed Back Propagation (BP) models clearly gave good results according to statistical indicators. In terms of the gradient, mean error and the standard deviation values, the proposed two BP models perform well for both data sets.

Cell structural changes in the mesophyll of Norway spruce needles by elevated ozone and elevated temperature in open-field exposure during cold acclimation

The effects of elevated ozone (1.4× ambient) and temperature (ambient +1.3 °C) alone and in combination were studied on the needle cell structure of soil-grown Norway spruce seedlings in the late growing season and winter. Temperature treatment continued over winter and lengthened the snow-free period. Elevated temperature caused microscopic changes related to photosynthesis (decreased chloroplast size and increased number), carbon storage (reduced starch and increased cytoplasmic lipids) and defence (decreased mitochondrial size and proportion per cytoplasm, increased peroxisomes and plastoglobuli, altered appearance of tannins). The results suggest increased oxidative stress by elevated temperature and altered allocation of limited carbon reserve to defence. The number of peroxisomes and plastoglobuli remained high in the outer cells of needles of ozone-exposed seedlings but decreased in the inner cells. This may indicate defence allocation to cells close to the stomata and surface, which are experiencing more oxidative stress. Ozone reduced winter hardiness based on seasonal changes in chloroplast shape and location in the cells. The effects of ozone became evident at the end of the growing season, indicating the effect of cumulative ozone dose or that the seedlings were vulnerable to ozone at the later phases of winter hardening. Elevated temperature increased cellular damage in early winter and visible damage in spring, and the damage was enhanced by ozone. In conclusion, the study suggests that modest air temperature elevation increases stress at the cell structural level in spruce seedlings and is enhanced by low ozone elevation. Future climatic conditions where snow cover is formed later or is lacking but temperatures are low can increase the risk of severe seedling damage, and current and future predicted ozone concentrations increase this risk.

Urban Ozone Concentration Forecasting with Artificial Neural Network in Corsica

Abstract Atmospheric pollutants concentration forecasting is an important issue in air quality monitoring. Qualitair Corse, the organization responsible for monitoring air quality in Corsica (France), needs to develop a short-term prediction model to lead its mission of information towards the public. Various deterministic models exist for local forecasting, but need important computing resources, a good knowledge of atmospheric processes and can be inaccurate because of local climatical or geographical particularities, as observed in Corsica, a mountainous island located in the Mediterranean Sea. As a result, we focus in this study on statistical models, and particularly Artificial Neural Networks (ANNs) that have shown good results in the prediction of ozone concentration one hour ahead with data measured locally. The purpose of this study is to build a predictor realizing predictions of ozone 24 hours ahead in Corsica in order to be able to anticipate pollution peaks formation and to take appropriate preventive measures. Specific meteorological conditions are known to lead to particular pollution event in Corsica (e.g. Saharan dust events). Therefore, an ANN model will be used with pollutant and meteorological data for operational forecasting. Index of agreement of this model was calculated with a one year test dataset and reached 0.88.

Evaluating hydraulic and disinfection efficiencies of a full-scale ozone contactor using a RANS-based modeling framework

The capability of predicting hydraulic and disinfection efficiencies of ozone disinfection contactors is essential for evaluating existing contactors and improving future designs. Previous attempts based on ideal and non-ideal models for the hydraulics and simplified mechanisms for chemical reaction modeling have resulted in low accuracy and are restricted to contactors with simple geometries. This manuscript develops a modeling framework for the ozonation process by combining computational fluid dynamics (CFD) with a kinetics-based reaction modeling for the first time. This computational framework has been applied to the full-scale ozone contactor operated by the City of Tampa Water Department. Flow fields, residence time distribution, ozone concentration distribution, and concentration-contact time (CT) distribution within the contactor have been predicted via the computational framework. The predictions of ozone and bromate concentrations at sample points agree well with physical experimental data measured in the contactor. The predicted CT values at the contactor outlet demonstrate that the disinfection performance of the ozone contactor operated by the City of Tampa Water Department is sufficient to meet regulation requirements. The impact of seasonal flow rate change on disinfection performance is found to be significant and deserves attention during the management and operation of a water treatment plant.

<b>Surface ozone measurements and meteorological influences in the urban atmosphere of Campo Grande</b> - doi: 10.4025/actascitechnol.v36i1.18379

The objective of this research was to develop a model to predict ozone concentrations on the meteorological variables. Concentration measurements of ozone were conducted at the Federal University of Mato Grosso do Sul, in the period 2007-2011. Predictor variables related to climate (relative humidity, wind speed, precipitation and temperature) have been provided by Embrapa-Gado de Corte, Campo Grande, Mato Grosso do Sul State. The model is able to explain over 70% of the variability concentrations of surface ozone.

Prediction of Ozone Concentration in Semi-Arid Areas of China Using a Novel Hybrid Model

The semi-arid regions of China support delicately balanced ecosystems that are likely to be significantly affected by small changes in the level of atmospheric ozone. To predict ozone concentrations in such areas, this paper proposes a new hybrid forecasting model called an ARMT-CPSO-BP neural network, which is based on the association rule mining technique (ARMT), a chaotic particle swarm optimization algorithm (CPSO) and a back-propagation (BP) neural network (CPSO-BP neural network). This paper first uses the ARMT to find correlations between meteorological variables and ozone concentrations. The significant correlation coefficients are then fed into the CPSO-BP neural network to obtain a prediction. Finally, the predicted results are compared with the results obtained from the BP neural network and the regression model. The comparison shows that the proposed hybrid model is superior to both the BP neural network and the regression model. The hybrid model reduces the square root mean square error (RMSE) by 31.6% compared to the BP model and by 23.8% compared to the regression model. The hybrid model is promising for forecasting ozone concentration in arid and semi-arid areas.

Neural network-based meta-modelling approach for estimating spatial distribution of air pollutant levels

Continuous measurements of the air pollutant concentrations at monitoring stations serve as a reliable basis for air quality regulations. Their availability is however limited only at locations of interest. In most situations, the spatial distribution beyond these locations still remains uncertain as it is highly influenced by other factors such as emission sources, meteorological effects, dispersion and topographical conditions. To overcome this issue, a larger number of monitoring stations could be installed, but it would involve a high investment cost. An alternative solution is via the use of a deterministic air quality model (DAQM), which is mostly adopted by regulatory authorities for prediction in the temporal and spatial domain as well as for policy scenario development. Nevertheless, the results obtained from a model are subject to some uncertainties and it requires, in general, a significant computation time. In this work, a meta-modelling approach based on neural network evaluation is proposed to improve the estimated spatial distribution of the pollutant concentrations. From a dispersion model, it is suggested that the spatially-distributed pollutant levels (i.e. ozone, in this study) across a region under consideration is a function of the grid coordinates, topographical information, solar radiation and the pollutant's precursor emission. Initially, for training the model, the input–output relationship is extracted from a photochemical dispersion model called The Air Pollution Model and Chemical Transport Model (TAPM–CTM), and some of those input–output data are correlated with the ambient measurements collected at monitoring stations. Here, improved radial basis function networks, incorporating a proposed technique for selection of the network centres, will be developed and trained by using the data obtained and the forward selection approach. The methodology is then applied to estimate the ozone concentrations in the Sydney basin, Australia. Once executed, apart from the advantage of inexpensive computation, it provides more reliable results of the estimation and offers better predictions of ozone concentrations than those obtained by using the TAPM–CTM model only, when compared to the measurement data collected at monitoring stations.

Forecasting the Tropospheric Ozone using Artificial Neural Network Modelling Approach: A Case Study of Megacity Madras, India

Ozone, which is a secondary pollutant in the troposphere, is very injurious to human health causing irritation of respiratory system, reducing lung capacity, etc. Adversely affecting the plant growth and deteriorating the materials. Thus it is of prime importance to predict the ozone concentration so that effective mitigation strategies can be adopted. As the formation of the tropospheric ozone is dependent on various meteorological parameters and the concentration of various other air pollutants, it is necessary to consider this dependence aspect in the modelling approach. In this background, the present paper puts forward the study of short-term prediction of tropospheric ozone concentration using Artificial Neural Network (ANN) modelling approach for a busy traffic junction of Madras city, one of the four megacities of India. 8-hourly averaged values of 11 air pollutants concentrations and 6 meteorological parameters were used for the study. The respective data was collected at a busy traffic junction of the city for a period of 19 months i.e. during September 2008–March 2010. 70% of the data was used for training the ANN models while the remaining of 30% data was used for validating them. By changing the neural architecture, 34 ANN models were formulated which were statistically analyzed. Based on the encouraging results (d=0.80, r=0.69, etc.), the paper puts forward the suitability of ANN modelling approach for the short-term prediction of tropospheric ozone concentration.

A comparative analysis of methodologies of daily metroplex ozone concentration prediction

This paper compares three methods of predicting the changes in ozone concentration: linear regression, classification and regression tree (CART) analysis, and the T-method. Using linear regression on these results, a linear equation defining the change of the independent variable versus the dependent variables is created. The strength of the relationship is assessed using the R-squared value and adjusted R-squared value. Classification and regression tree analysis uses a tree-building methodology to generate decision rules, using patterns from historical data obtained on both the dependent variable and the independent or ‘predictor’ variables to create a prediction model. The T-method is used to calculate an overall prediction based on the dynamic signal-to-noise ratio to obtain an overall estimate of the true value of the output for each signal member. It was found that for this nearly directly correlated dataset the T-method performed comparably to linear regression and was a better predictor than the CART method.

Development and initial application of a sub-grid scale plume treatment in a state-of-the-art online Multi-scale Air Quality and Weather Prediction Model

Traditional Eulerian air quality models are unable to accurately simulate sub-grid scale processes, such as the near-source transport and chemistry of point source plumes, because they assume instantaneous mixing of the emitted pollutants within the grid cell containing the release, and neglect the turbulent segregation effects that limit the near-source mixing of emitted pollutants with the background atmosphere (e.g., Kramm and Meixner, 2000). Observations by Dlugi et al. (2010) show that the segregation of chemically reactive species can slow effective second-order reaction rates by as much as 15%, due to inhomogeneous mixing of the reactants. This limitation of traditional grid models applies to both “off-line” models, in which externally derived meteorology is used to drive the chemistry model, and newer “on-line” models, such as the Weather Research and Forecasting model with Chemistry (WRF/Chem), that simulate the emissions, transport, mixing, and chemical transformation of trace gases and aerosols simultaneously with the meteorology. While a number of approaches have been used in the past to address this limitation, the approach that has been most effectively used in operational models is the plume-in-grid (PinG) approach, in which a reactive plume model is embedded within the grid model to resolve sub-grid scale plumes. This paper describes the implementation of such a PinG treatment in WRF/Chem, based on a similar extension to the U.S. EPA Community Multi-scale Air Quality (CMAQ) model. The treatment, referred to as Advanced Plume Treatment, has been tested in CMAQ over more than a decade and has been used successfully in both episodic and long-term applications for assessing point source contributions to ozone and particulate matter. This paper presents the application of the PinG version of WRF/Chem for a three-day episode in July 2001, including a model performance evaluation and comparison of model results with and without PinG treatment. The results from the model application show that overall model performance is only slightly affected when the PinG treatment is used, although there are some generally small improvements, with the PinG treatment showing a 5% lower bias in predicting ozone concentrations, and 3% lower bias in sulfate predictions. However, the predicted spatial patterns of ozone and PM2.5 concentrations from the two simulations show both large decreases of up to 40 ppb ozone and 14 μg m−3 PM2.5, and increases of up to 80 ppb ozone and 33 μg m−3 PM2.5 as a result of using the PinG treatment. These differences are attributed to both direct effects of the PinG treatment (i.e., differences in dispersion, transport and chemistry of point source emissions) and indirect effects (i.e., impacts of air quality changes on meteorology).

Indirect prediction of surface ozone concentration by plant growth responses in East Asia using mini-open top chambers

We developed small and mobile open top chambers (mini-OTC) measuring 0.6m (W) × 0.6m (D) × 1.2m (H) with an air duct of 0.6m (W) × 0.23m (D) × 1.2m (H). The air duct can be filled with activated charcoal to blow charcoal filtered air (CF) into the chamber, as opposed to non-filtered ambient air (NF). Ozone sensitive radish Raphanus sativus cv. Red Chime and rosette pakchoi Brassica campestris var. rosularis cv. ATU171 were exposed to NF and CF in mini-OTCs at different locations in East Asia. A total of 29 exposure experiments were conducted at nine locations, Shanghai, China, Ha Noi, Vietnam, Lampang, Phitsanulok and Pathumtani, Thailand, and Hiratsuka, Kisai, Abiko and Akagi, Japan. Although no significant relationships between the mean concentrations of ambient O(3) during the experimental period and the growth responses were observed for either species, multiple linear regression analysis suggested a good relationship between the biomass responses in each species and the O(3) concentration, temperature, and relative humidity. The cumulative daily mean O(3) (ppb/day) could be indirectly predicted by NF/CF based on the dry weight ratio of biomass, mean air temperature, and relative air humidity.

Evolving Gaussian process models for prediction of ozone concentration in the air

Ozone is one of the main air pollutants with harmful influence to human health. Therefore, predicting the ozone concentration and informing the population when the air-quality standards are not being met is an important task. In this paper, various first- and high-order Gaussian process models for prediction of the ozone concentration in the air of Bourgas, Bulgaria are identified off-line based on the hourly measurements of the concentrations of ozone, sulphur dioxide, nitrogen dioxide, phenol and benzene in the air and the meteorological parameters, collected at the automatic measurement stations in Bourgas. Further, as an alternative approach an on-line updating (evolving) Gaussian process model is proposed and evaluated. Such an approach is needed when the training data is not available through the whole period of interest and consequently not all characteristics of the period can be trained or when the environment, that is to be modelled, is constantly changing.