Deep learning-based AQI forecasting: a CNN-LSTM model with visual insights from SHAP-LIME and PDP

Precise air quality forecasting is essential for assessing environmental health risks and facilitating timely intervention measures. This paper introduces a hybrid deep learning framework for multi-step Air Quality Index (AQI) forecasting based on multivariate time series data of Anugul, Odisha, and Cities of India. The proposed architecture incorporates convolutional neural networks (CNN) to identify spatial correlations among contaminants and long-short-term memory (LSTM) networks to identify long-term trends. The model exhibited robust predictive capability for Anugul city, Odisha, and cities of India with a Mean Squared Error (MSE) of 130.66, 1192, 130.66, Root Mean Squared Error (RMSE) of 11.40, 34.53, 37.75, and Mean Absolute Error (MAE) of 8.38, 27.65, 23.81, respectively, which is better than the accuracy of customary statistical and isolated deep learning models. To improve interpretability, Shapley Additive Explanations (SHAP), LIME (Local Interpretable Model-Agnostic Explanations), and Partial Dependence Plots (PDPs) were integrated, allowing for analysis of the contribution of each pollutant to the AQI predictions. The tools of interpretability aid in identifying environmental and industrial variables impacting air quality, enabling greater insights and giving more confidence in model outcomes. One-way ANOVA suggests that CNN-LSTM is a powerful model for AQI prediction when trained and validated separately for each data set. The results of p-values for all three metrics (MSE, RMSE, MAE) are below 0.5, which suggests the model’s performance is highly region-dependent. The proposed hybrid CNN-LSTM method provides stable AQI prediction and enables interpretable decision-making via model explainability, providing application value for environmental monitoring and policymaking.