BiLSTM-A: A missing value imputation method for PM2.5 prediction

With the frequent occurrence of haze in recent years, air quality has attracted more and more public attention. Realtime air quality acquisition can help people to obtain environmental status in time and avoid environmental pollution exposure. However, confusion in sensor system behavior or errors in the transmission process can result in missing data. Missing data interpolation (MDI) is the task of predicting missing parts in the dataset. The problem of missing data draws difficulties in decision-making process, and there is an urgent need for accurate and effective estimation methods to impute the data. The existing imputation algorithms are usually assume that the data have obvious statistical regularity, which limit the interpolation model to the mathematical model with a clear and simple description, ignoring a large number of hidden key factors. Our goal is to address fewer data missing situations. In this study, a novel method based on bidirectional Long Short-Term Memory (LSTM) named BiLSTM-A is proposed for missing value imputation. We evaluated our method using a dataset containing daily PM2.5 measurements from the beijing region in 2015. The results show that our model has excellent performance. In all time intervals, our model achieves the minimum error.