Bi-Directional Long Short-Term Memory Network (BiDLSTM) Model Based Air Pollution Prediction

Air pollution is a serious problem over the globe. It could put human life in peril by bringing on illness and occasionally even death. The purpose of air quality forecasting is to provide early warnings and minimize the consequences of air pollution. An important gauge of air quality is the concentration of small particulate matter (PM2.5), which are inhalable particles with a diameter of around 2.5 micrometers. These particles pose major health risks because they can deeply penetrate the lungs. This research investigates precise prediction of PM2.5, which may help to mitigate detrimental outcomes. Authors have proposed stacked bi-directional long short-term memory (BiDLSTM) model for forecasting. This will capture complicated temporal correlations in air quality data, this cutting-edge model makes use of LSTM neural networks' bidirectional sequence processing capabilities. The BiDLSTM model can make extremely precise forecasts about the amount of pollution in the future by examining historical air quality measurements taken in various directions in time. In this paper, air quality dataset of India is used for model performance evaluation. Along with this, LSTM model is used for comparison purpose. Different evaluation parameters like average absolute deviation, root mean square deviation and determination coefficient are calculated to quantify the model's performance. The evaluation findings suggest that BiDLSTM model performs excellently.

I. Introduction

With rapid economic expansion and industrialization in recent decades, the problem of environmental pollution has become more serious. As a result of industrialization, factories and automobiles continue to emit more exhaust fumes, raising the level of air pollution. Pollution in the air occurs when dangerous compounds, such as particles (PM2.5 and PM10), gases, and other pollutants, are discharged in substantial number into the atmosphere [1]. The lives of people are significantly impacted by the quality of the air. Particularly for people residing in high-pollution regions like India and China, air pollution has significant adverse effect [2]. According to a World Health Organization (WHO) estimate, roughly seven million people die each year as a result of the effects of air pollution. According to the article of New York Times, air pollution in India has reached a critical level. An atmospheric particulate with dimension of 2.5 micrometers or smaller is denoted as PM2.5, whereas one with dimension of 10 micrometers or smaller is labeled as PM10. These particles are conveniently inhaled and quickly taken up by the alveoli. The cardiovascular and respiratory systems may be harmed, and asthma attacks may be triggered [3]. Consequently, research on accurate forecasting of quality of air is extremely significant and are regarded as a critical aspect in environmental preservation. Air quality tracking, modelling, and accurate forecasts are also critical for determining upcoming levels of pollution as well as associated health hazards. The cause and method of PM2.5 production is extremely complex due to various complex aspects, like nonlinear properties in both space and time, that have a significant impact on accuracy of prediction. For the purpose of air quality forecasting, time series data-based methodologies such as conventional machine learning algorithms are often used. Many classical time series prediction methods [4], such as Box-Jenkins models, Exponential Smoothing methods, State Space Models, and so on. However, classical methods are incapable of accurately capturing the complex nonlinear character of air quality, such as PM2.5 levels.