PM2.5 Forecasting Model based on Linear and Non-linear Hybrid Algorithm

Abstract

Air pollution is one of the harmful problems that the world has focused on and needs to be solved urgently because air pollution has a direct impact on humans leading to premature death caused by various diseases such as asthma inflammatory respiratory disease, lung cancer, and so on. The air pollutants, especially tiny particulate matter (PM), are currently receiving attention because they are a major problem in many large and populated cities around the world. This paper proposed a time-series model for forecasting PM2.5 in advance through a machine learning process with a linear and non-linear hybrid algorithm. A hybrid algorithm that brings together the capabilities of autoregressive integrated moving average (ARIMA) and the adaptive-neuro fuzzy inference system (ANFIS) is used to find the linear and nonlinear correlation of the PM2.5 time-series data. The proposed model is called ARIMA-FIS which uses the gradient descent (GD) method in the learning process. The dataset used in this research is the daily recorded of PM2.5 values in Rayong province, which is the industrial city in Thailand. The results showed that the ARIMA-FIS model had the best performance in forecasting PM2.5 in advance with the least error at 3.46 of mean absolute error (MAE) and 5.11 of root mean square error (RMSE). The proposed model gave the percentage of RMSE almost 3% better than the other standard time-series models.