Optimized Deep Learning Model for PM2.5 and PM10 Concentration Prediction in Formal ELV Recycling Zone: A Stacked LSTM-GRU Approach with Explainable AI

OpenAlex İngilizce

This paper presents an optimized and explainable stacked LSTM-GRU model for predicting PM2.5 and PM10 concentrations in a formal End-of-Life Vehicle (ELV) recycling zone using multivariate and multioutput dataset. The model was developed using sensor data of weather and pollution parameters collected through an in-house developed device, consisting of a set of highly precise sensors and enhanced the interpretability of the model with Explainable AI (XAI) techniques. The Bayesian optimization technique was employed to accurately fine-tune the hyperparameters of the model for greater accuracy. The results show that the model significantly outperforms baseline models. For PM2.5, the model shows performance improvements of 66.75% for MAE, 81.59% for MSE, and 66.68% for MAPE. For PM10, the model demonstrates performance improvements of 64.45% in MAE, 81.64% in MSE, and 63.65% in MAPE. XAI reveals the dominance of past values and decomposition components, along with meteorological parameters for PM2.5 and PM10 predictions. The results demonstrate the model’s superior accuracy and interpretability, offering valuable insights for air quality management. The integration of Explainable AI further contributes to the model’s practical utility by providing transparent insights into feature importance. This research contributes a novel approach to air quality prediction in industrial settings, offering both improved accuracy and actionable insights.