Beyond Traditional Models: Foundation Models for Accurate Particulate Matter Prediction

Accurate PM 2.5 forecasting is critical for public health, yet traditional deep learning models trained on location-specific data often lack generalizability across geographic contexts. Time Series Foundation Models (TSFMs) offer zero-shot forecasting capabilities through large-scale pre-training, but their efficacy for air quality prediction using low-cost sensor (LCS) data remains unexplored. This study presents an empirical benchmarking study of TSFMs for PM 2.5 prediction using real-world LCS data evaluating Google’s TimesFM and IBM’s Granite against three deep learning architectures (CNN, LSTM, Transformer) across 34 datasets from 10 cities spanning 6 countries. We evaluate two forecasting strategies: direct prediction of reference station-equivalent data and prediction of LCS data with subsequent calibration. Our results show that TSFMs consistently outperform traditional deep learning models, particularly for short-term forecasts, while retaining an stable advantage over longer horizons. Zero-shot TSFMs, applied without any task-specific training, perform competitively across all evaluated sites, providing empirical evidence of transferability to unseen sensor deployments. Fine-tuning offers limited additional benefit over the zero-shot configuration. Direct reference station prediction using fine-tuned Granite emerges as the most effective strategy, achieving the lowest errors and highest R 2 values. No clear performance decline is observed in datasets with substantial missing data compared to those with few missing values when linear interpolation is applied to fill gaps. Input window length has limited impact on forecasting accuracy. These findings establish TSFMs as a promising and practically viable direction for air quality monitoring with LCS networks.