Optimization of Neural Network by using cluster analysis for Ozone

Introduction As is well known, air quality problems, due to the ozone (O3) could produce effects on human health related to respiratory problems, damage to ecosystems, agricultural crops and materials (World Health Organisation –WHO-, 2003). The ozone is classified as a secondary pollutant and its levels are determined by complex photochemical reactions with primary pollutants (EPA, 2006). The concentrations are strongly dependent both on micro-meteorological conditions linked to turbulence and to the effects related to the seasons. As described in different works, the prediction of ozone levels is very complex to obtain by mathematical models (Comrie, 1997). In this work, NNs have been developed to forecast ozone levels in complex systems using hourly pollutant and meteorological data as input parameters. In our approach, in order to improve the NN ability to “capture” the true hidden relation inside the data set by means of the pre-selection of information, we applied the cluster analysis techniques. Description of used data set The ozone concentration follows the maximum of the temperature, especially during the daytime, when the highest values of ozone in urban areas are related to the high values of global solar radiation and pollutants. The analysis of the summer daytime reveals that the ozone peak is at 17.00, the temperature peak is at 15.00 (the temperature values varied in the range of 18° to 26°C) and the GSR peak is at 12.00 for the summer. The seasonal variation of O3 shows low concentrations in late autumn and winter and high concentrations in late spring and early summer (average daily 146.94 and 132.3 mg/m3 respectively) Usually, the distributions of observed pollutants in urban sites are skew, because low values are more frequent than the higher values. In our case, ozone distribution is highly skewed. In fact, about 97% for 2007 and 96% for 2006 of patterns belong to the class 0-120µg/m3, whereas less than 0.1% for 2007 and 0.2% for 2006 is above the information threshold (180µg/m3). Methodologies The main target of the study is to suggest a way to implement the learning of neural network models for the ozone forecasting in urban areas. In general, the best practice before running the NN model consists in the pre-processing techniques related to the selection of the best information, linked to the pattern selection, to the physics inside your system or to pollutant trends. The pattern selection is a very important task that should be solved in order to achieve a good generalisation of the net, above all if the net is used to simulate chemical reactions in the atmosphere. This task ensures the quality of the data and it optimises the efficiency of the computation time during the elaboration phase and improves the NN learning process. We tested the NN performance by means of two ways of pattern selection. The first one is the random pattern selection and the second one is the cluster analysis pattern selection. Results and Discussion We applied NN to the results coming from the pattern selection process to forecast ozone concentrations using as input data, meteorology, as well as primary and secondary pollutants (CO, NO, NO2). We carried out 27 simulations using different percentages of input patterns for the training. All the results are referred to the generalization phase, where the patterns are never seen by the NN (NG in the tables). The results obtained by Cluster Analysis applied to the NN (named CANN) are compared to the Conventional Random Pattern Selection applied to the NN (named CRPSNN), our benchmark, with different percentages of input patterns from 0.01% to 100% (from S1 to S27). In general, we observed that the NN performance shows different values for the ozone predictions. In particular shows that CANN is performed and predicted better than CRPSNN (Table 1). In terms of global fit, CANN has a better performance (R˛ from 0.59 to 0.89) than CRPSNN (R˛ from 0.05 to 0.97). These results show a meaningful difference and demonstrate that the pre-selection by cluster can simulate in best way the physics of ozone (i.e. can reproduce the observed skew distribution for ozone). Moreover, we observed that in S10 (with only 1% of total data) we obtain R2=0.56 utilizing CANN, whereas we obtain R2=0.41 for CRPSNN (see Tables 1-2). The first important result of our elaboration shows that CA sampling is, on average, more efficient than CRPSNN when using small amounts of patterns during the training and, consequently, could be adapted to simulate rare events, such as what happens during the extreme ozone episodes. Table 1. Cluster Patterns Selection (CANN): Generalisation phase (N=14324) Conclusions Our research shows a good capacity of the NN to analyse the large complex data sets and to model the ozone levels using the clustering approach during pattern pre-processing phase. The capability of the Neural Network technique, applied to multivariate and non-linear problems, to capture the environmental information inside the data depended not only on the learning methods used, but also on the preliminary study of patterns. This study is related to the quality of the data used to train the neural network. The problem of pre-processing and of a proper sampling plan for the input data is essential to obtain a good forecasting performance of the NN. We observed that the neural classifier trained after the random pattern choice, is able to distinguish only average/stable situations. On the contrary, the NN, after cluster pattern choice, is able to distinguish outlier situations too. In conclusion, the clustering approach, adopted as a pattern selection approach, obtains better predictions of pollutant phenomena. The results are very encouraging and our simulations based on cluster analysis demonstrated that this method is feasible and effective, resulting in a substantial reduction of data input requirement and outperform other techniques applied in this context. The determination coefficient RD substantially shows better performance in the combined forecast procedure. References Comrie R.S, 1997. Comparing neural network and regression models for ozone forecasting. Journal of the Air and Waste Management Association 47, 653-663. EPA, 2006. Air Pollutants. Epa.gov. 2006-06-28. http://www.epa.gov/ebtpages/airairpollutants.html. Retrieved 2010-08-29.