A hybrid long short-term memory with generalized additive model and post-hoc explainable artificial intelligence with causal inference for air pollutants prediction in Kimberley, South Africa

The study addresses the problem of nonlinear characteristics of common air pollutants by proposing a deep learning time-series model based on the long short-term memory (LSTM) integrated with a generalized additive model (GAM). LSTM model captures both nonlinear relationships and temporal long-term dependencies in time-series data, and GAM provides insight into the statistical relationship between selected features and the target pollutant. The post-hoc eXplainable artificial intelligence (xAI) technique, local interpretable model-agnostic explanation (LIME), further explains the nonlinearity. Finally, causal inference was determined on the impact of the air pollutants relationship, thereby offering further interpretability in which deep learning models are deficient. Meteorological and air pollutant statistical records were leveraged from a Hantam (Karoo) air monitoring station in South Africa, and through a random sampling approach, synthetic data were generated for the city of Kimberley. The model was evaluated with the mean squared error (MSE), root mean squared error (RMSE) and mean absolute error (MAE) for different time-steps. The proposed referred to as long short-term memory generalized additive model based post-hoc eXplainable Artificial Intelligence (LSTM-GAM_xAI) model with a 10-day time-step and 5-day time-step for multiple pollutants prediction guaranteed least MSE. Though the causal effect analysis show no <i>p</i>-values (>0.88) for variables, the experiment results show that LSTM-GAM-xAI guaranteed the lowest MSE values across different time-steps.