Development of a CNN+LSTM Hybrid Neural Network for Daily PM2.5 Prediction

Abstract

A CNN+LSTM (Convolutional Neural Network + Long Short-Term Memory) based deep hybrid neural network was established for the citywide daily PM2.5 prediction in South Korea. The structural hyperparameters of the CNN+LSTM model were determined through comprehensive sensitivity tests. The input features were obtained from the ground observations and GFS forecast. The performance of CNN+LSTM was evaluated by comparison with PM2.5 observations and with the 3-D CTM (three-dimensional chemistry transport model)-predicted PM2.5. The newly developed hybrid model estimated more accurate ambient levels of PM2.5 compared to the 3-D CTM. For example, the error and bias of the CNN+LSTM prediction were 1.51 and 6.46 times smaller than those by 3D-CTM simulation. In addition, based on IOA (Index of Agreement), the accuracy of CNN+LSTM prediction was 1.10–1.18 times higher than the 3-D CTM-based prediction. The importance of input features was indirectly investigated by sequential perturbing input variables. The most important meteorological and atmospheric environmental features were geopotential height and previous day PM2.5. The obstacles of the current CNN+LSTM-based PM2.5 prediction were also discussed. The promising result of this study indicates that DNN-based models can be utilized as an effective tool for air quality prediction.