Imputation of Missing Data in Time Series for Air Pollutants Using Long Short-Term Memory Recurrent Neural Networks

Abstract

Long Short Term Memory (LSTM) Recurrent Neural Networks has been shown to be capable of learning long time dependencies, and has been successfully applied to many studies, such as machine translation, speech recognition and air pollution concentration prediction. The present research has shown that the presence of missing data could dramatically degrade the results of data mining and categorical predictions with the aid of the machine learning technique including LSTM networks. Therefore, this paper focuses on imputation of missing data in the time series of air pollutants using LSTM networks to improve the PM2.5 concentration prediction accuracy. Experimental result shows that the proposed LSTM-based imputation method presents better PM2.5 concentration prediction accuracy than mean-imputation method and moving average imputation method.