Multi-Step API Prediction in Malaysia: Evaluating Recursive NARX and SVM Models

The increasing problem of air pollution has led to the advancement and improvement of air quality prediction studies. Predicting air quality in advance is crucial for mitigating the detrimental effects of air pollution on public health and economic activities. This study focuses on the development and evaluation of the Nonlinear Autoregressive Exogenous Neural Network (NARX) and Support Vector Machine (SVM) models for multi-step prediction of Malaysia's Air Pollutant Index (API). The models were constructed using a dataset from air quality monitoring stations in Malaysia's three prominent industrial areas: Pasir Gudang, Larkin, and TTDI Jaya. The model development process began by constructing a single-step API predictor, then developing and analyzing a multi-step API predictor using the recursive approach. The prediction performance was assessed using the Root Mean Square Error (RMSE) and Coefficient of Determination values $(\mathrm{R}^{2})$. The results indicate that the recursive NARX demonstrates promising performance compared to the recursive SVM in multi-step API prediction. However, additional analysis shows that the outliers strongly affected multi-step NARX's prediction, suggesting that the recursive NARX model cannot predict sudden fluctuations not seen in training.

I. Introduction

Air pollution poses a pressing global concern, posing significant threats to public health and economic well-being worldwide [1]–[2]. Like many other nations, Malaysia has also faced detrimental health consequences and economic losses due to air pollution issues [3]–[4]. For instance, a study identified a clear link between increased hospital admissions for cardiovascular diseases in Kuala Lumpur and two prominent air pollutants, namely Nitrogen Dioxide (NO2) and Ozone (03) [5].