Related Experiment Video
Updated: Jul 27, 2026
10:29
Calibrated Passive Sampling - Multi-plot Field Measurements of NH3 Emissions with a Combination of Dynamic Tube Method and Passive Samplers
Published on: March 21, 2016
Evaluating deep learning time series models for PM2.5 forecasting across diverse horizons
Ling Zeng1, Runan Dong1,2, Meng Yuan1,2
1Geomathematics Key Laboratory of Sichuan Province, Chengdu Technological University, Chengdu 610059, China.
Iscience
|February 18, 2026
Summary
Transformer-LSTM models excel at forecasting PM2.5 (particulate matter) in Chengdu. Integrating meteorological data and ensuring complete seasonal training significantly enhances prediction accuracy for air quality management.
Area of Science:
- Environmental Science
- Data Science
- Atmospheric Chemistry
Background:
- Urban air pollution, especially PM2.5, presents a significant public health risk in cities like Chengdu, China.
- Geographical factors (basin topography) and high emission levels intensify PM2.5 concentrations, necessitating accurate forecasting.
- Existing forecasting methods may not fully capture the complex dynamics of PM2.5 influenced by various environmental factors.
Purpose of the Study:
- To evaluate the performance of four deep learning time series algorithms (LSTM, CNN-LSTM, Transformer, Transformer-LSTM) for PM2.5 forecasting.
- To compare univariate and multivariate model configurations using auxiliary pollutants and meteorological data.
- To assess the impact of data completeness and prediction horizons on forecasting accuracy.
Main Methods:
- Utilized two years of daily PM2.5 data from Chengdu, China (November 2022-October 2024).
- Compared LSTM, CNN-LSTM, Transformer, and Transformer-LSTM models in univariate and multivariate settings.
- Incorporated auxiliary pollutants (CO, NO2, O3, SO2) and meteorological factors (temperature, pressure, precipitation, wind speed).
- Evaluated models across different forecasting horizons (monthly, seasonal, half-year, annual) and data completeness scenarios.
Main Results:
- Transformer-LSTM demonstrated superior performance, indicated by higher R-squared and lower MAE% and RMSE% values.
- Augmenting models with meteorological factors yielded better results than using only auxiliary pollutants.
- Complete seasonal training datasets significantly improved model performance.
- Gaps exceeding three months in training data reduced prediction reliability due to evolving PM2.5 dynamics.
Conclusions:
- Deep learning models, particularly Transformer-LSTM, are effective for PM2.5 forecasting in complex urban environments.
- Meteorological data integration is crucial for enhancing PM2.5 prediction accuracy.
- Maintaining seasonal data completeness and timely predictions are vital for reliable air quality management strategies.
