Fine particulate matter concentrations forecasting using long short-term memory network and meteorological inputs

BACKGROUND AND OBJECTIVES: In metropolitan settings, the requirement to travel and participate in everyday tasks exposes numerous individuals to the harmful effects of air pollutants, specifically particulate matter ۲.۵, which has the potential to impact their well-being. Developing precise forecasting models is crucial in mitigating air pollution and providing accurate predictions for the people.Nonetheless, the deficiency in acquiring observable data can frequently lead to unsatisfactory performance of forecasting models in various scenarios. The objective of this study is to address the issue by examining the most effective approaches for predicting the non-linear time-series data of daily particulate matter ۲.۵ concentration using meteorological inputs.METHODS: The concentration data of particulate matter ۲.۵ at Central Jakarta and South Jakarta were collected using sensors from the United States of America Embassy in Indonesia and Indonesia’s Meteorological, Climatological, and Geophysical Agency. Conversely, the meteorological information was collected through the Merra-۲ satellite. This study introduces the long short-term memory deep learning model and contrasts it with the one-dimensional convolution neural network as well as their hybrid counterpart. The dataset is split into ۸۰ percent training and ۲۰ percent testing data. The root mean square and mean absolute error values are then calculated to determine the performance of the models.FINDINGS: A combination of long short-term memory and fully connected layers using dropouts and early stopping patience techniques has been successfully developed to model the non-linear time-series data of daily particulate matter ۲.۵ concentration. The model effectively captured the patterns present in the historical data, resulting in outcomes that exhibited similar patterns. The long short-term memory model demonstrates an overall root mean square error and mean absolute error values of ۱۸.۵۳ micrograms per cubic meter and ۱۴.۹۲ micrograms per cubic meter in Central Jakarta and ۱۹.۴ micrograms per cubic meter and ۱۵.۶۱ micrograms per cubic meter in South Jakarta, where the best seasonal data were found to be in the June-July-August and December-January-February seasons respectively.CONCLUSION: The air pollution forecasting models, which were created using both seasonal and overall time-series data, have the ability to predict air pollution levels by utilizing historical pollution data and meteorological inputs. The proposed long short-term memory model outperforms the one-dimensional convolution network and their hybrid combination. It has effectively surpassed the constraint of collecting observable data, attaining minimal error values on both sensors and satellite data, signifying a noteworthy progression compared to previous studies. Therefore, it might benefit areas lacking sufficient data, providing a valuable tool for air pollution mitigation.

Keywords:

Air pollution , Hybrid deep learning models , Long short term memory (LSTM) , Fine particulate matter (PM۲.۵) forecasting