Evaluating hourly air quality forecasting in Canada with nonlinear updatable machine learning methods

Abstract

Air quality data (observational and numerical) were used to produce hourly spot concentration forecasts of ozone (O3), particulate matter 2.5 μm (PM2.5), and nitrogen dioxide (NO2), up to 48 h for six stations across Canada—Vancouver, Edmonton, Winnipeg, Toronto, Montreal, and Halifax. Using numerical data from an air quality model (GEM-MACH15) as predictors, forecast models for pollutant concentrations were built using multiple linear regression (MLR) and multi-layer perceptron neural networks (MLPNN). A relatively new method, the extreme learning machine (ELM), was also used to overcome the limitation of linear methods as well as the large computational demand of MLPNN. In operational forecasting, the continual arrival of new data necessitates frequent model updating. This type of learning (online sequential learning) is straightforward for MLR and ELM but not for MLPNN. Forecast performance of the online sequential MLR (OSMLR) and online sequential ELM (OSELM), together with stepwise MLR, all updated daily, were compared with MLPNN updated seasonally and the benchmark climatology model. OSELM, combining relatively inexpensive frequent model updating with nonlinear modeling capability, tended to outperform the other models in mean absolute error and correlation. Compared to the linear models, the nonlinear models (OSELM and MLPNN) often had worse bias (mean error) and more severe underprediction of extreme events.