PM2.5 prediction with Recurrent Neural Networks and Data Augmentation

Abstract

This paper presents three novel models based on recurrent neural networks (RNN) such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) for PM2.5 prediction with data augmentation (DA). The data augmentation technique is based on linear interpolation, it allows to find a linear function with each pair of items from the original time series. A space parameter allows to define the number of synthetic items to be generated, with this is possible to enlarge the original time series and improve the precision of the regression models. The baseline models as GRU, LSTM and GRU+LSTM got regular and bad prediction results, while the same ones with data augmentation as DA+GRU, DA+LSTM and DA+GRU+LSTM got excellent predictions showing the superiority of the proposals models. Likewise, according to the Mean Absolute Percentage Error (MAPE), the data augmentation allows to improve a regular GRU model by 18.6288% and bad models as LSTM and GRU+LSTM by 21.7683% and 31.0092% respectively.