Forecasting air quality time series using deep learning

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Air quality management relies on accurate air pollution time series data for exposure assessment and regulatory compliance.
• Predicting air pollution, specifically ozone (O3) concentrations, is crucial for public health and environmental protection.

Purpose of the Study:

• To apply deep learning (DL) techniques, specifically Long Short-Term Memory (LSTM) recurrent neural networks (RNNs), for predicting 8-hour averaged surface ozone (O3) concentrations.
• To develop a forecasting model capable of predicting air pollution up to 72 hours in advance with low error rates.
• To assess the model's ability to forecast the duration of continuous O3 exceedances.

Main Methods:

• Utilized hourly air quality and meteorological data to train an LSTM-based deep learning model.
• Implemented a novel imputation technique for handling missing data and outliers.
• Employed decision trees to identify and reduce the number of input features from 25 to 5, enhancing model accuracy.
• Conducted parameter sensitivity analysis to optimize RNN look-back nodes.

Main Results:

• Achieved low error rates, with Mean Absolute Errors (MAE) less than 2 for predictions up to 72 hours.
• Successfully forecasted the duration of continuous O3 exceedances.
• Reduced feature set from 25 to 5, leading to improved prediction accuracy.
• Identified optimal RNN parameter settings for accurate forecasting.

Conclusions:

• Deep learning, particularly LSTM networks, offers a powerful tool for accurate air pollution time series forecasting.
• The developed model enables air managers to predict long-range air pollution using key parameters, facilitating real-time monitoring and continuous prediction.
• The novel data imputation and feature selection methods enhance the efficiency and accuracy of air quality forecasting models.