The Impact of Data Imputation and Feature Extraction on PM2.5 Forecasting Performance in Bangkok Using Long Short-Term Memory Neural Networks

Abstract

The last few years have seen a dramatic increase in PM2.5 air pollution in Thailand's major cities. Various works have tried to develop efficient Long Short-Term Memory (LSTM) deep neural network models for PM2.5 concentration forecasting. However, little has been studied about the impact of data imputation and feature extraction on the model performance in this context. In this paper, we imputed missing values using Kalman Smoothing and Linearly Weighted Moving Average. We utilized the LSTM Autoencoder (LSTM AE) for feature extraction. Using the Chokchai Police station in Bangkok as a case study to predict PM2.5 in the next 24 hours, we demonstrated that the performance gain from training LSTM models with imputed data is more than 7 percent overall with respect to the root mean square error (RMSE) and more than 10 percent overall with respect to the mean absolute error (MAE). Improvement with LSTM AE varies according to time steps. Forecasting 22 to 24 hours ahead tends to favor the use of LSTM AE. © 2020 ACM.