Monday, 7 January 2019: 3:15 PM
North 125AB (Phoenix Convention Center - West and North Buildings)
Of all environmental threats, air pollution is responsible for the most deaths globally, specifically being responsible for one out of every eight deaths and a total of approximately seven million deaths according to World Health Organization data available for 2012. For reducing the adverse health effects of PM2.5 on humans, the precise prediction of PM2.5 concentration associated with unhealthy values of air quality index (AQI > 101) is very important. Most learning algorithms do not exhibit good performance for these scenarios, especially extreme events. The objective of this study is to evaluate the performance of different machine learning algorithms for short-term forecasting of PM2.5 such as Gaussian Process Regression, Support Vector Machine, Random Forest (RF), and Long Short-Term Memory (LSTM). We also introduce a new hybrid framework based on transfer learning to improve the prediction of PM2.5 concentration for extreme events. Our hybrid model is comprised of two major components: 1) A RF model that classifies the future concentrations into extreme and normal, and 2) Two LSTM models, one for regular forecasting and another (which is trained using transfer learning mechanism) for extreme concentration forecasting. The models were trained and tested using collected PM2.5 from the Airnow database and the corresponding synchronized climatological data of nearby stations, available in NOAA’s National Climatic Data Center. Our results suggest that our proposed hybrid transfer learning model outperforms the single-based models, while all models underperform between November and January owing to the important effect of reduced boundary layer height on PM levels. The better performance of the hybrid model can be attributed to the transfer learning mechanism, in which large scale time series data of other regions helps tune the model for the target dataset. A quantitative report of model comparisons will be presented, which is based on their performance in each month for two cities differing in landscape, emissions, and meteorology (i.e., Los Angeles and Phoenix).
![](images/flag.png)
![](images/blue_ribbon.png)