Data package for Multi-Sequence Retrieval-Augmented PM2.5 Forecasting: Physical Predictability and Auxiliary Information Gains

This data package accompanies the study "Multi-Sequence Retrieval-Augmented PM2.5 Forecasting: Physical Predictability and Auxiliary Information Gains". It provides meteorological inputs, ordered station metadata, static geographic features, and numerical analysis outputs for PM2.5 forecasting at 1,029 monitoring stations in China. PM25_weather_and_station_data_v1.zip contains nine hourly meteorological variables from ERA5 obtained through the Open-Meteo Historical Weather API, together with station metadata and derived geographic features. The meteorological matrix has shape (1029, 9, 43848) and covers 1 October 2017, 00:00, to 1 October 2022, 23:00, in China Standard Time (UTC+8). Variable definitions, source attribution, preprocessing details, and file checksums are provided in the archive. Analysis_outputs.zip contains numerical result summaries and plotting scripts for the model comparisons, meteorological-noise experiments, stratified analyses, and zero-shot transfer analyses. Hourly PM2.5 observations are available from the original CNEMC/QuotSoft archive and are not included in this deposit. Acquisition and preprocessing scripts are available in the associated versioned software repository: https://github.com/TLHeVRT/PM25_prediction_with_AAT/