A deep learning approach via multifractal detrended fluctuation analysis for PM2.5 prediction

Abstract

Accurate prediction of PM 2.5 concentrations in Asian regions with air pollution as a hot topic is still considered as a strong premise for the public to take proactive measures for effective control of the adverse effects of PM 2.5. In this study, a hybrid model consisting of convolutional neural network (CNN)-long short-term (LSTM) with multifractal detrended fluctuation analysis (MF-DFA) for air pollution and meteorological time series data was proposed, and it was used to predict the next day's 24-h average PM 2.5 concentration in Beijing City. MF-DFA was employed to identify the fractal characteristics of the measured data series; it has been considered for the historical data of the target monitoring station and, as a result, enabling CNN-LSTM model to adopt valuable input data. CNN of the hybrid model efficiently extracted the inherent features relevant to PM 2.5 in the input data, and LSTM fully conveyed the historical information of the time series data. Air quality and meteorological data from January 1st, 2015 to February 24th, 2018 were adopted. Through the comparison of the performance indices of the proposed model with MLP and LSTM models, the proposed model showed better prediction accuracy. When the integrated data (RMSE 11.732, MAE 5.18, MAPE 0.131) were used as input of the model, it is guaranteed that the performance of the model is improved compared to the cases of using air quality data (RMSE 11.886, MAE 6.17, MAPE 0.272) and meteorological data (RMSE 29.673, MAE 25.171, MAPE 0.425) as input, respectively.