A Machine Learning and Explainable AI (XAI) Framework with LLM Fine-Tuning for PM2.5 Prediction and EV Impact Analysis

Abstract

Prediction and mitigation are important challenges in order to adequately address the environmental and public health threat of air pollution PM2.5. This study presents a new “Machine Learning (ML)” and “Explainable Artificial Intelligence (XAI)” method that combines with fine-tuned Large Language Model (LLM) for forecasting PM2.5 atmospheric pollution and analyzing the effects of Electric Vehicles (EVs) on the environment. Data from Air Quality Data in India from Kaggle were preprocessed by removing missing values, duplicate records, detecting outliers, scaling features and splitting it into 80:20 train test sets. The ensemble learning models such as Random Forest, XGBoost, LightGBM and LSTM were tested using MAE, RMSE and R2 metrics. LightGBM achieved the best performance with MAE = 4.08, RMSE = 5.02, and R² = 0.97. The feature importance analysis indicates that PM10 (0.284), NO₂ (0.226) and temperature (0.173) are the most important features. The SHAP indicates that PM10 (0.312) is the most important feature. Seasonal analysis revealed that the season with the highest PM2.5 was December (165.8 µg/m³) and that of the lowest was July (36.9 µg/m³). The proposed framework offers reliable, understandable and policy informed air quality forecasts for sustainable environmental management.