Shift-Aware Dual-Encoder Transfer Learning Improves PM2.5 Forecasting in Data-Limited Settings

A recent arXiv preprint (2608.14456) presents a dual-encoder transfer learning framework that is shift-aware, aimed at forecasting PM2.5 levels over short horizons in scenarios with limited target-domain data and differing statistical properties between source and target domains. This research tackles the issue where models relying solely on local data struggle with intricate temporal dynamics, while basic transfer learning may result in adverse transfer effects. The innovative framework integrates knowledge from the source domain with learning tailored to the target. The source encoder utilized hourly data from 10 U.S. monitoring sites. It was subsequently tested with two years of hourly data from 77 Taiwanese stations, adhering to a chronological train-validation-test approach. Among four key baselines, the frozen-source dual-encoder model demonstrated superior performance, achieving MSE = 21.8960, MAE = 3.1597, and R^2 = 0.8725, showcasing the method's efficacy in data-scarce settings. This research advances air quality forecasting by providing a solid transfer learning strategy in the presence of domain shifts.