Prediction of PM2.5 and CO2 concentrations using the PCA-LightGBM method in Bandung, Indonesia

Abstract

Poor air quality due to large amounts of human activity shows the need to increase public awareness and alertness by building a system predicting future pollutant concentrations. This research creates a prediction system using the LightGBM algorithm for PM_2.5 and CO₂ pollutant parameters with an additional parameter reduction method using PCA to increase prediction accuracy. The number of valid datasets is 918 for each of the five parameters at each measurement station, with data gaps filled using median values so that they can be used for predictions. The prediction results show that the best accuracy for PM_2.5 is at the Deli station, which uses PCA with a MAPE of 21.5%, and for CO₂, it is achieved at the Deli station without PCA with a MAPE of 4.8%. Based on its accuracy, PCA is less suitable if there are outliers in the dataset, but PCA is ideal for homogeneous datasets. Overall, the prediction results based on accuracy for PM_2.5 are in the feasible category, and for CO₂, they are in the accurate and very accurate category. To optimize prediction results, especially in the long term, it is necessary to retrain with a complete and up-to-date dataset to better suit air conditions.