Short-Term Forecasting of PM2.5 in High-Density City: A Spatio-Temporal Deep Learning Model Framework in Hong Kong

This presentation would introduce a hybrid model framework to predict air quality at Environment Protection Department (EPD) air quality monitoring stations in Hong Kong over the next 48 hours. This hybrid predictive model uses historical air monitoring records (SO₂, O₃, CO, NO₂, PM_2.5, PM₁₀), current meteorology data, weather forecasts in predicted time period at the target station and stations within a few kilometers. This predictive model consists of four parts: 1) a linear-regression-based temporal predictor to predict following hours concentration for each station, 2) a neural network-based spatial predictor to model the impact from nearby regions, 3) a regression-tree-based dynamic aggregator combining the predictions of the temporal and spatial predictors according to the situation of meteorological and air quality observations, 4) an inflection predictor to capture sudden changes in air quality. To evaluate this hybrid model, first thing is to compare the traditional regression methods with neural network by feed all the features into one single model directly. Secondly, examine each part of model’s necessity by controlling the components of the model to see the difference of results. Thirdly compare this hybrid model with some baseline methods, like Auto-Regression-Moving-Average, traditional air quality forecast model (CMAQ). After training this model with more data, this model is hoped to contribute in the platform of the HSBC project: Personalized Real-time Air-Quality Informatics System for Exposure- Hong Kong.