In this paper, we forecast the reading of an air quality monitoring station in the next 48 hours, using a data-driven method that considers the current meteorological data, weather forecasts, and the air quality data of the station and that of other stations within a few hundred kilometers to the station. Our predictive model is comprised of four major components: 1) a linear regression-based temporal predictor to model the local factor of air quality, 2) a neural network-based spatial predictor modeling the global factors, 3) a dynamic aggregator combining the predictions of the spatial and temporal predictors according to the meteorological data, and 4) an inflection predictor to capture the sudden changes of air quality. We evaluate our model with the data of 43 cities in China, surpassing the results of multiple baseline methods. We have deployed a system in Chinese Ministry of Environmental Protection, providing 48-hour fine-grained air quality forecasts for four major Chinese cities every hour. The forecast function is also enabled on Microsoft Bing Map and MS cloud platform Azure. Our technology is general and can be applied globally for other cities.