Revealing Influence of Meteorological Conditions on Air Quality Prediction Using Explainable Deep Learning

Abstract

Meteorological conditions have a strong influence on air quality and can play an important role in air quality prediction. However, due to the “black-box” nature of deep learning, it is difficult to obtain trustworthy deep learning models when considering meteorological conditions in air quality prediction. To address the above problem, in this paper, we reveal the influence of meteorological conditions on air quality prediction by utilizing explainable deep learning. In this paper, (1) the source data from air pollutant datasets, including PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> , PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sub> , SO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> hourly concentration, and the meteorological condition datasets measuring the temperature, humidity, and atmospheric pressure are obtained; (2) the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models are established for air quality prediction in 4 conditions; (3) the SHapley Additive exPlanation (SHAP) method is employed to analyze the explainability of the air quality prediction models. We find that the prediction accuracy is not improved by considering only meteorological conditions. However, when combining meteorological conditions with other air pollutants, the prediction accuracy is higher than considering other air pollutants. In addition, the largest contribution to air quality prediction is atmospheric pressure, followed by humidity and temperature. The reason for the different accuracies of the prediction may because of the interaction between meteorological conditions and other air pollutants. The investigated results in this paper can help improve the prediction accuracy of air quality and achieve trusted air quality predictions.