Modelling Missing Data for PM2.5 Time Series Forecasting with Computational Intelligence

Abstract

The paper presents two missing data filling methods which can be applied to time series forecasting. The basic idea of the proposed methods is that usually, the forecasted parameter (in this case PM2.5 air pollutant concentration) is dependent on some related parameters that influence its value. When the parameter time series have missing data due to various reasons (e.g. faulty measurement instruments), the time series of other parameters (if available) can be used to fill in the missing values. One method is based on an artificial neural network that has as input the values of the other related parameters measured at time t and as output the value of the missing value of the forecasted parameter at time t. The other method is Holt-Winters which uses as inputs previous values of the forecasted parameter. These methods are proper for cases with larger gaps in the time series (more than several days). These filling methods are compared in terms of statistical indicators (e.g. RMSE). Also, a comparative study was performed for PM2.5 forecasting accuracy analysis with two forecasting methods: a feed forward artificial neural network and Holt-Winters.