Explainable sequence-to-sequence GRU neural network for pollution forecasting.

Area of Science:

• Environmental Science
• Artificial Intelligence
• Data Science

Background:

• Deep neural networks (DNNs) show promise in air quality forecasting, outperforming traditional methods.
• DNNs have historically been treated as 'black boxes,' limiting understanding of their predictions.
• Explainable AI (XAI) techniques offer insights into DNN decision-making processes.

Purpose of the Study:

• To adapt and apply Layer-wise Relevance Propagation (LRP), an XAI technique, to a sequence-to-sequence neural network model.
• To identify key meteorological and temporal features influencing the accumulation of major air pollutants ([Formula: see text], [Formula: see text], [Formula: see text], [Formula: see text]).
• To demonstrate the utility of XAI in understanding and potentially controlling air pollution.

Main Methods:

• Utilized a sequence-to-sequence neural network architecture incorporating Gated Recurrent Unit (GRU) layers for air pollution forecasting.
• Extended the Layer-wise Relevance Propagation (LRP) technique to analyze the developed neural network model.
• Generated explanation heatmaps to visualize feature importance for pollutant accumulation predictions.

Main Results:

• LRP successfully identified significant meteorological and temporal input features contributing to the prediction of four major air pollutants.
• The identified features align with established knowledge in environmental science and pollution research.
• Explanation heatmaps provided by LRP offer interpretable insights into the model's forecasting behavior.

Conclusions:

• XAI, specifically LRP, is appropriate for understanding complex air pollution forecasting models.
• The findings open new possibilities for targeted air pollution control and mitigation strategies.
• Applying XAI to deep learning models enhances transparency and trust in environmental predictions.