A Risk-Aware Explainable Temporal Fusion Transformer Framework with Dynamic Pollution Risk Index for Multi-Class Air Quality Forecasting

Forecasting the air quality has gained more importance owing to the growing concern of impact of air pollutants on human health and environmental sustainability. On the other hand, available models can be limiting in commodity temporal dependencies, lack interpretability and do not yield actionable risk. As a way to overcome these limitations, this paper presents a new risk-aware and explainable deep learning architecture for multi-class air quality forecasting using the Temporal Fusion Transformer (TFT). The suggested method includes a powerful preprocessing pipeline, with K-Nearest Neighbor (KNN) imputation for missing data, temporal feature engineering (lag, rolling and ratio based features), as well as balanced class distribution to improve authenticity of the data and generalization ability of the model. This work uses a TFT model to extract underlying long-term temporal patterns and intervariable relationship functionalities for multiple city-based multivariate air quality data. Moreover, a hybrid mechanism of explainability is proposed (mainly SHAP-based feature attribution and attention-driven importance analysis), to produce comprehensible interpretations for the pollutants influence on prediction breadth. A new Dynamic Pollution Risk Index (DPRI) is also introduced, which quantifies the severity of pollution risk using predicted class probabilities and feature importance weights to enable visual, intuitive categorization into low, moderate and high levels. Through a wide range of experiments, the proposed framework shows significant improvement over state-of-the-art such as traditional deep learning models (LSTM, CNN-LSTM and FGCE-BiLSTM) both in terms of accuracy, precision, recall and F1-score. By integrating explainability and risk-aware modelling, both predictive performances improve as well as decision-making for environmental monitoring and policy planning. The framework is scalable and interpretable, which makes it an ideal approach for developing next-generation air quality intelligence systems.