Abstract
Air Quality Index (AQI) forecasting is challenging because atmospheric pollution is governed by nonlinear pollutant interactions, temporal dependencies, meteorological conditions, missing observations, and abrupt emission events. This paper presents a Digital Twin (DT)-enabled deep learning framework for AQI prediction in which a virtual environmental state is continuously updated from real-world observations and supplied to a temporal prediction model. The proposed predictive engine combines a Temporal Convolutional Network (TCN) and RWKV-style sequence modeling with cross-attention to capture local temporal patterns, long-range dependencies, and feature interactions. The experimental configuration uses a 24-step sequence, an 80-epoch training schedule, and a DT-enhanced data generation/update pipeline. On the user's experimental results, the proposed model achieved MAE = 1.48095, RMSE = 1.57756 and R² = 0.995628 with the Digital Twin configuration, while the corresponding configuration without DT achieved MAE = 1.40353, RMSE = 1.55070 and R² = 0.999681. These results show that the present DT configuration does not improve the three reported numerical metrics, although it provides a richer virtual-state and scenario-analysis capability. The paper therefore treats Digital Twin integration as an architectural contribution rather than claiming unsupported predictive superiority. The framework is designed for real-time AQI monitoring, forecasting, pollution-event analysis, and future scenario simulation.
