Explainable Deep Learning for Air Quality Index Prediction: Integrating Meteorological and Traffic Data for Anomaly Identification
TL;DR
A explainable deep learning model taking into account meteorological parameters, traffic-flow factors and historical pollutant information to predict the AQI and detect abnormal pollution events is designed and shows that the prediction accuracy and interpretability can be achieved concurrently.
Abstract
Air quality forecasting has become an important tool for the management of public health and urban planning with the rapid increase of the urban population and the vehicular traffic. Although deep learning models have been able to predict the Air Quality Index (AQI) with a high accuracy, their black-box nature limits their use by regulators and city planners who need transparent and trustworthy support for their decision-making. This paper aims to design a explainable deep learning model taking into account meteorological parameters, traffic-flow factors and historical pollutant information to predict the AQI and detect abnormal pollution events. The combination of CNN-BiLSTM with an attention mechanism is used to model the spatial correlation between pollutants and long-term temporal correlations. To understand the model, SHapley Additive exPlanations (SHAP) and Integrated Gradients, which quantify the contribution of different meteorological and traffic features to each prediction, are used. An unsupervised residual-thresholding module also identifies abnormal AQI episodes that do not match the normal pattern, e.g. due to an increase in traffic congestion or an atmospheric temperature inversion. The experimental results using a multi-source dataset that is a combination of air-quality monitoring records, meteorological observations, and traffic-sensor data demonstrate that the proposed model achieves a coefficient of determination (R²) of 0.943 and a root-mean-square error (RMSE) of 9.82, surpassing the baseline LSTM, GRU, and CNN-LSTM models. The lagged PM2.5 concentration, traffic volume and relative humidity are the top three factors, which agree with the knowledge of the atmospheric chemistry, and thus confirms the validity of the model's reasoning process. The proposed framework shows that the prediction accuracy and interpretability can be achieved concurrently, and can provide a practical and transparent decision support tool for environmental agencies and smart-city traffic management systems.