Generalizable Deep Learning for Urban Air Pollution Forecasting: AirNet-Adapt with Cross-Domain Alignment on Beijing and UCI Air Quality Data

The problem of urban air pollution prediction is still an urgent issue since cities become more exposed to environmental threats and human health problems. This paper presents a domain- adaptive deep learning framework, called AirNet-Adapt, that can improve generalization to a wide range of urban settings. AirNet-Adapt uses the concept of crossdomain alignment to combine spatial-temporal patterns of pollution with the use of the dataset on Beijing and sensor- based atmospheric measurements on the UCI Air Quality dataset. The model learns variations in the behaviour of pollutants in different cities using spatial feature extraction and adaptive representation learning. Through experimental analysis, it is proven that composite feature matching across multiple domains can enhance accuracy of prediction much higher than traditional single-city based forecasting models. AirNet-Adapt offers a scalable and robust system that can be used to forecast urban air pollution in a robust and generalizable way.