Haze Forecasting via Deep LSTM

Abstract

PM$_{2.5}$ is a crucial indicator of haze pollution, which can cause problems in respiratory systems. Accurate PM$_{2.5}$ concentration forecasting systems are essential for human beings to take precautions. State-of-the-art methods including support vector regression (SVR), artificial neural network (ANN) and Bayesian, try to forecast PM$_{2.5}$ concentrations of the following 3 days via building an approximation from weather features to PM$_{2.5}$ concentration. However, the performances of these methods are poor because they ignore the essence of the problem: PM$_{2.5}$ concentration is the product of a time series.

This paper aims to propose more accurate forecasting algorithms to forecast PM$_{2.5}$ concentration. First, we employ the recurrent neural network with Long Short Term Memory kernel to handle the time series forecasting. Secondly, in order to further improve the performance, a convolutional neural network (CNN) is utilized as feature extractor to generate input for LSTM. Two models are proposed to handle the forecast for the following 3 and 7 days: (i) based on 2 days’ weather features and PM$_{2.5}$ concentrations; (ii) based on 4 days’ (including 2 days of this year, the day of last year, and the day two years ago) weather features and PM$_{2.5}$ concentrations. Finally, all experiments are compared with the root of mean squared errors (RMSE) for each city and averaged root of mean squared errors (ARMSE) of all cities. Experiments are tested on two datasets: one with hourly meteorological data and daily air-pollution data of 104 cities in east China from 2013 to 2017, the other with both hourly meteorological and air-pollution data in 5 cities from 2010 to 2015. Experimental results show that the proposed methods significantly outperform the state-of-the-art.