DynLink-AQ: Adaptive inter-station connectivity for multi-pollutant, multi-horizon air quality forecasting across
Arun Raj Velraj1, Senthil Kumar Jagatheesaperumal1
1Department of Electronics and Communication Engineering, Mepco Schlenk Engineering College, Sivakasi, India.
None:
Air quality in Delhi has deteriorated significantly over the past decade, yet accurate high-resolution forecasting across multiple pollutants remains a major challenge due to heterogeneous monitoring networks, missing data, and complex spatial - temporal interactions. Motivated by the need for reliable early-warning systems, this study proposes DynLink-AQ, an end-to-end framework for multi-pollutant forecasting using data from 39 CPCB stations over 2009-2023. The system integrates rigorous data quality control and robust spatio-temporal imputation, followed by feature engineering enriched with meteorological drivers and temporal encodings. Unlike static distance-based station graphs, DynLink-AQ learns time-varying inter-station connectivity by inferring adaptive graph attention weights from station embeddings, spatial proximity, and temporal similarity, enabling event-driven and meteorology-linked coupling to be captured. Built upon this structure, the model alternates temporal attention blocks with spatial adaptive graph-attention layers to capture deep spatial - temporal dependencies. The framework supports multi-task prediction for 1-24-hour pollutant horizons with optional uncertainty quantification, and hyperparameters are tuned using the Enzyme Action Optimizer Algorithm (EAOA) under rolling-window training. Extensive walk-forward and spatial generalization experiments demonstrate strong predictive skill across pollutants, with ablation studies confirming the importance of dynamic connectivity learning and meteorological features. The study additionally provides GIS-ready outputs for seamless visualization and operational use in pollution management.Implications: DynLink-AQ enables operational, network-wide forecasting of PM2.5, PM10, NO2, and O3 across Delhi using Central Pollution Control Board monitoring data. By learning time-varying inter-station connectivity and combining spatial graph attention with temporal attention, the model improves 1-24h predictions and provides uncertainty bounds for risk-aware alerts. Agencies can use these forecasts to issue timely hotspot formation, alerts, and plan short-term mitigation (traffic control, construction restrictions, industrial scheduling) during unexpected times. The framework is transferable to other cities with dense station networks and can integrate meteorological drivers already available hourly.
