Integration of Explainable Artificial Intelligence into Hybrid Long Short-Term Memory and Adaptive Kalman Filter for Sulfur Dioxide (SO2) Prediction in Kimberley, South Africa

Air pollution remains one of the environmental issues affecting some countries, which leads to health issues globally. Though several machine learning and deep learning models are used to analyze air pollutants, model interpretability is a challenge. Also, the dynamic and time-varying nature of air pollutants often creates noise in measurements, making air pollutant prediction (e.g., Sulfur Dioxide (SO2) concentration) inaccurate, which influences a model’s performance. Recent advancements in artificial intelligence (AI), particularly explainable AI, offer transparency and trust in the deep learning models. In this regard, organizations using traditional machine and deep learning models are confronted with how to integrate explainable AI into air pollutant prediction systems. In this paper, we propose a novel approach that integrates explainable AI (xAI) into long short-term memory (LSTM) models and attempts to address the noise by Adaptive Kalman Filters (AKFs) and also includes causal inference analysis. By utilizing the LSTM, the long-term dependencies in daily air pollutant concentration and meteorological datasets (between 2008 and 2024) for the City of Kimberley, South Africa, are captured and analyzed in multi-time steps. The proposed model (AKF_LSTM_xAI) was compared with LSTM, the Gate Recurrent Unit (GRU), and LSTM-multilayer perceptron (LSTM-MLP) at different time steps. The performance evaluation results based on the root mean square error (RMSE) for the one-day time step suggest that AKF_LSTM_xAI guaranteed 0.382, LSTM (2.122), LSTM_MLP (3.602), and GRU (2.309). The SHapley Additive exPlanations (SHAP) value reveals “Relative_humidity_t0” as the most influential variable in predicting the SO2 concentration, whereas LIME values suggest that high “wind_speed_t0” reduces the predicted SO2 concentration.