Article Preview
TopIntroduction
World Health Organization under their air pollution control program recently revealed that 9 out of 10 people in the world is inhaling polluted air and around seven million deaths happen every year due to poor air quality. Many of the studies concluded that it causes around one third deaths due to lung cancer, respiratory diseases, heart problems and other implications. According to IQAir (2019) report, India is playing major contribution in depiction of global air pollution. Ten Indian cities on the scale of worst PM2.5 were highlighted by WHO in its Global Urban Ambient Air Pollution Database (WHO, 2019). The capital of India, Delhi, is one of those cities. Due to urbanization of Indian people, the quality of air in Delhi, turn out to be worsen since last 30 years and now it becomes the global concern. There are several reasons which impact the quality of air like unrestrained emission sources, adverse meteorological conditions, agricultural pollutants (Guttikunda & Gurjar, 2012) ozone from background, and transport emissions due to traffic flow etc. (West et al., 2009).
Presence of different pollutants makes the prediction of air pollution as a complex task. Several studies available in literature considered factors such as transport emissions and meteorological conditions. Many of the existing air quality models works either on the deterministic analysis of factors or on the statistical analysis that comprises the semiempirical statistical relationship among different factors. The incompetence to determine the atmospheric and boundary conditions, limits us to consider the problem as fully deterministic. Such studies examined number of factors which affects the quality of air, though, influenced by numerous uncertainties among inputs (Mallet & Sportisse, 2008). On the contrary work, the stochastic methods involve many assumptions related to distribution and randomness. In other study, authors pointed that statistical models could work on the nonlinearity linked with the data but with an assumption that data must follow normal probability distribution (Bengio et al., 1994).
Recently, machine learning approaches have been gaining success in each field. Like statistical models, these approaches did not essentially involve prior assumptions. However, these approaches can execute self-formulation of the most efficient model by minimizing the cost function. This feature makes it preferred approach for the prediction of air quality. In one of the studies, authors discussed a lazy learning method which attempts to reduce the impact of parameters namely humidity, temperature, wind, and solar radiation (Kalapanidas & Avouris, 2001).
Athanasiadis et al. (2003) simulated and classified the ozone levels based on metrological data and pollutants like NO2 using σ-fuzzy lattice neurocomputing classifier. One of the authors presented the prediction of CO using adaptive Neuro-fuzzy model (Jain & Khare, 2010). In the same year, Kurt and Oktay (2010) applied the artificial neural network (ANN) to forecast the daily concentrations of CO, SO2, and PM10 in air. Fu et al. (2015) introduced a concept of the ray model over the conventional FFNN models. In another study, authors analysed the concentration of PM2.5 and concluded that in some of the cases regression models can be the better option over classifiers (Ni et al., 2017).
Recent studies presented RNNs as a better approach over ANNs while predicting air quality due to its capability of modelling the interrelated sequence of time series (Kim et al., 2010). In another study, authors (Fan et al. (2017) developed a model using deep RNNs. They used the concept of deep feed forward neural networks. According to open literature, RNNs have the ability to handle long sequences of data in time series; however, sometimes this time gap can be very large (Bengio et al., 1994), which makes researchers to think of new approach over RNNs.