Deep Learning Implementation Using Long Short Term Memory Architecture for PM_2.5Concentration Prediction: a Review

PM2.5 is a respirable fine particle with harmful effects. PM2.5 prediction research plays a role to contribute scientific recommendations in actions for controlling air pollution. Nowadays, PM2.5 prediction using deep learning is the most chosen since it is an efficient and cost-effective method for air quality modeling. As a subset of deep learning, LSTM is strong in remembering past experiences and able to identify highly complex relationships of sequential data. LSTM has higher accuracy in time series prediction compared to other deep neural network models. On the other hand hybrid CNN-LSTM, a method that can produce good predictive models since CNN's ability to extract spatial-temporal features might collaborate with LSTM, it is possible to identify stationary and non-stationary data. Meteorological parameters and other gaseous pollutants affect the concentration of PM2.5 proven by correlation analysis results. Therefore, it is appropriate to use as input for air quality modeling using either LSTM or the hybrid CNN-LSTM method.