NEXUS: A compact neural architecture for high-resolution spatiotemporal air quality forecasting in Delhi National Capital Region

Rampunit Kumar; Aditya Maheshwari

cs.LGFeb 23, 2026

NEXUS: A compact neural architecture for high-resolution spatiotemporal air quality forecasting in Delhi National Capital Region

Authors: Rampunit Kumar, Aditya Maheshwari

Abstract

Urban air pollution in megacities poses critical public health challenges, particularly in Delhi National Capital Region (NCR) where severe degradation affects millions. We present NEXUS (Neural Extraction and Unified Spatiotemporal) architecture for forecasting carbon monoxide, nitrogen oxide, and sulfur dioxide. Working with four years (2018--2021) of atmospheric data across sixteen spatial grids, NEXUS achieves R $^2$ exceeding 0.94 for CO, 0.91 for NO, and 0.95 for SO $_2$ using merely 18,748 parameters -- substantially fewer than SCINet (35,552), Autoformer (68,704), and FEDformer (298,080). The architecture integrates patch embedding, low-rank projections, and adaptive fusion mechanisms to decode complex atmospheric chemistry patterns. Our investigation uncovers distinct diurnal rhythms and pronounced seasonal variations, with winter months experiencing severe pollution episodes driven by temperature inversions and agricultural biomass burning. Analysis identifies critical meteorological thresholds, quantifies wind field impacts on pollutant dispersion, and maps spatial heterogeneity across the region. Extensive ablation experiments demonstrate each architectural component's role. NEXUS delivers superior predictive performance with remarkable computational efficiency, enabling real-time deployment for air quality monitoring systems.

Figures & tables

Figure 1. NEXUS architecture showing the complete processing pipeline from raw spatiotemporal inputs through patch embedding, low-rank projection, stacked NanoBlocks with parallel pathways (CompactKernel, MicroConv, FusionGate), weighted spatial pooling, and prediction head. The compact design achieves superior forecasting performance through strategic exploitation of atmospheric data structure—local temporal coherence enables patch-based dimensionality reduction, dominant meteorological modes justify low-rank projections, and multi-scale parallel processing captures concurrent atmospheric processes across temporal scales.

Carbon Monoxide (CO)
Metric	SCINet	Autoformer	FEDformer	NEXUS
R 2	0.7537	0.8840	0.8890	0.9404
RMSE	0.5527	0.3794	0.3711	0.2718
MAE	0.3893	0.2516	0.2538	0.1831
sMAPE (%)	65.63	45.07	47.15	34.90
IoA	0.9270	0.9697	0.9691	0.9847

Table 1. Comprehensive performance comparison across models and pollutants. Best results highlighted in bold.

Improvements versus Baselines
Model	Parameters	Avg R 2	Avg RMSE	Avg MAE
SCINet	35,552	0.7531	0.5828	0.4017
Autoformer	68,704	0.8804	0.4053	0.2540
FEDformer	298,080	0.8747	0.4152	0.2660
NEXUS	18,748	0.9355	0.2967	0.1900
vs SCINet	-47.27%	+24.23%	-49.08%	-52.69%

Table 2. Parameter efficiency, computational requirements, and average performance across all pollutants.

Figure 2. Temporal evolution of predicted versus observed pollutant concentrations over year-long test period (July–December 2021). NEXUS captures sustained low-concentration baseline during summer monsoon months and dramatic concentration spikes during winter pollution episodes.

Figure 3. Diagnostic plots for model residuals including scatter plots, quantile-quantile comparisons, and scale-location relationships. Residuals show approximately zero mean, near-normal distribution, and relatively constant variance.

Figure 4. Relationships between weather variables and pollutant concentrations. Strong negative correlations with temperature and wind speed confirm their roles in controlling dispersion.

Figure 5. Mean diurnal variation showing systematic 24-hour patterns. Morning peaks coincide with traffic emissions and shallow boundary layers, afternoon minima reflect enhanced mixing.

Figure 6. Spatial distribution during winter months showing elevated concentrations in northwestern industrial regions and systematic gradients reflecting transport patterns.

Figure 7. Mean concentrations stratified by temperature and wind speed quartiles. Cold and calm conditions produce most severe pollution, exceeding six times warm and windy regime levels.

Figure 8. Seasonal patterns in monthly means showing pronounced November-December peaks contrasted against March-September baseline values. NEXUS accurately tracks these seasonal variations.

Figure 9. Scatter plots comparing predicted and observed concentrations across test samples, showing tight clustering around perfect prediction line with minimal systematic bias.

Configuration	Avg R 2	$\Delta$ R 2	Parameters	Training Time
NEXUS	0.9355	baseline	18,748	25 min
w/o Patch Embedding	0.8535	-8.77%	24,320	38 min
w/o Low-rank Projection	0.8820	-5.72%	22,144	31 min
w/o NanoBlock Pathways	0.8768	-6.28%	14,592	19 min
w/o Weighted Pooling	0.8972	-4.09%	18,320	24 min
Single NanoBlock	0.8654	-7.49%	12,416	18 min

Table 3. Ablation study showing contribution of each architectural component.

Lightweight Neural Networks

See all 18 papers

GraphSVR: A Graph Convolutional Support Vector Regression Framework for Robust Spatiotemporal Air Pollution Forecasting
May 5, 2026Nourin Jahan, Muhammed Navas T, Tanujit Chakraborty +1Air Quality Prediction Support Vector Machines
Department of Mathematics, School of Advanced Sciences, Vellore Institute of Technology, Vellore, Tamil Nadu, India · SAFIR, Sorbonne University, Abu Dhabi, UAE · SCAI - Sorbonne Cluster for Artificial Intelligence, Paris, France +1
AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting
Oct 19, 2026Yiheng Wang, Kai Zheng, Yuetan Lin +5Air Quality Prediction Spatiotemporal Forecasting
Fudan University · Shanghai Academy of AI for Science · University of Shanghai for Science and Technology
NeuroDDAF: Neural Dynamic Diffusion-Advection Fields with Evidential Fusion for Air Quality Forecasting
Apr 1, 2026Prasanjit Dey, Soumyabrata Dev, Angela Meyer +1Air Quality Prediction Spatiotemporal Forecasting

May 5, 2026J/K move · Enter open · S save

GraphSVR: A Graph Convolutional Support Vector Regression Framework for Robust Spatiotemporal Air Pollution Forecasting

Nourin Jahan, Muhammed Navas T, Tanujit Chakraborty, Madhurima Panja

Department of Mathematics, School of Advanced Sciences, Vellore Institute of Technology, Vellore, Tamil Nadu, India · SAFIR, Sorbonne University, Abu Dhabi, UAE · SCAI - Sorbonne Cluster for Artificial Intelligence, Paris, France+1

Air Quality Prediction Support Vector Machines Spatiotemporal Forecasting Spatiotemporal GNNs

Open paper AskPDF arXiv

Urban air quality forecasting is challenging because pollutant concentrations are nonlinear, nonstationary, spatiotemporally dependent, and often affected by anomalous observations caused by traffic congestion, industrial emissions, and seasonal meteorological variability. This study proposes a Graph Convolutional Support Vector Regression (GraphSVR) framework for robust spatiotemporal forecasting of urban air pollution. The model combines graph convolutional learning to capture inter-station spatial dependence with support vector regression to model nonlinear temporal dynamics while reducing sensitivity to outlier observations. The proposed framework is evaluated using air quality records from 37 monitoring stations in Delhi and 18 stations in Mumbai, representing inland and coastal metropolitan environments in India. Forecasting performance is assessed across multiple horizons and compared with established temporal and spatiotemporal benchmarks. The results show that GraphSVR consistently improves predictive accuracy and maintains stable performance across seasons and outlier-prone pollution episodes. Statistical test further confirms the reliability of the proposed approach across the datasets. Furthermore, the conformal prediction approach is integrated with GraphSVR to generate calibrated prediction intervals, enhancing its practical value for uncertainty-aware air quality monitoring and public health decision-making.

Abstract

Lightweight Neural Networks

Explore similar work

GraphSVR: A Graph Convolutional Support Vector Regression Framework for Robust Spatiotemporal Air Pollution Forecasting

AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting

NeuroDDAF: Neural Dynamic Diffusion-Advection Fields with Evidential Fusion for Air Quality Forecasting