Graph Convolutional Support Vector Regression for Robust Spatiotemporal Forecasting of Urban Air Pollution

Technical Summary: Graph Convolutional Support Vector Regression for Robust Spatiotemporal Forecasting of Urban Air Pollution

Problem Statement
Urban air quality forecasting faces significant challenges due to the nonlinear, non-stationary, and spatiotemporally dependent nature of pollutant concentrations. These dynamics are frequently disrupted by anomalous observations arising from traffic congestion, industrial emissions, and seasonal meteorological variability. Traditional statistical models (e.g., ARIMA, STARMA) often fail to capture complex nonlinearities and spatial dependencies, while standard deep learning approaches (e.g., LSTM, Transformer) may lack robustness against outliers and do not inherently provide calibrated probabilistic forecasts. This study addresses the need for a framework capable of robust spatiotemporal forecasting in environments characterized by high pollution variability and outlier-prone data, specifically within the Indian urban contexts of Delhi and Mumbai.

Methodology
The authors propose a Graph Convolutional Support Vector Regression (GCSVR) framework, which integrates spatial graph learning with robust temporal regression. The methodology consists of two primary modules:

1. Spatial Module (Graph Convolutional Network):

   • Monitoring stations are modeled as nodes in an undirected graph $G=(V, E)$, where edges represent spatial relationships.
   • The adjacency matrix is constructed using weighted Haversine distances between station coordinates, encoded via a Gaussian kernel to define edge weights.
   • Graph Convolutional Layers (GCN) aggregate information from neighboring stations to generate spatial embeddings. The model employs a first-order approximation of Chebyshev polynomial filters to ensure computational efficiency.
   • This module produces graph-enhanced node embeddings that capture inter-station spatial dependence.

2. Temporal Module (Support Vector Regression):

   • For each station, the input vector is formed by concatenating the local lagged pollutant time series with the graph-enhanced spatial embedding derived from the GCN.
   • A separate SVR model is trained for each station using a Radial Basis Function (RBF) kernel.
   • Crucially, the SVR utilizes an $\epsilon$-insensitive loss function. This loss function ignores errors within a specified tolerance margin ($\epsilon$), thereby reducing sensitivity to isolated noisy or extreme observations (outliers) compared to squared-error-based methods.
   • Multi-step forecasting is achieved through a recursive strategy, where predicted values are appended to the input sequence for subsequent time steps.

3. Uncertainty Quantification:

   • To generate calibrated prediction intervals, the framework integrates Conformal Prediction.
   • This distribution-free approach computes conformity scores based on forecast errors within a rolling calibration window, producing adaptive prediction intervals that widen during volatile episodes and narrow during stable periods.

Experimental Evaluation
The framework was evaluated using daily air quality records for PM2.5 and PM10 from:

• Delhi: 37 monitoring stations (2019–2023 training, 2023 testing).
• Mumbai: 18 monitoring stations (2021–2023 training, 2024 testing).

Performance was assessed across three forecasting horizons (30, 60, and 90 days) using six metrics: Mean Absolute Error (MAE), Mean Absolute Scaled Error (MASE), Root Mean Squared Error (RMSE), Symmetric Mean Absolute Percentage Error (SMAPE), Pinball Loss, and Continuous Ranked Probability Score (CRPS). The GCSVR model was compared against nine benchmarks, including temporal models (ARIMA, LSTM, DeepAR, Transformer, N-BEATS) and spatiotemporal models (STARMA, GSTAR, GpGp, STGCN).

Key Results

• Predictive Accuracy: GCSVR consistently achieved the lowest or near-lowest error rates across most metrics and horizons for both cities. It demonstrated particular robustness during high-pollution episodes and winter seasons in Delhi, where outlier-prone data typically degrade other models.
• Stability: Unlike deep learning benchmarks (e.g., Transformer, DeepAR) which showed increased variability and error magnitudes at longer horizons, GCSVR maintained stable performance as the forecasting horizon extended from 30 to 90 days.
• Statistical Significance: Multiple Comparison with the Best (MCB) tests confirmed that GCSVR's superior performance was statistically significant, with the model achieving the lowest mean ranks across the majority of evaluation metrics in both cities.
• Uncertainty Calibration: The integration of conformal prediction successfully generated adaptive intervals that captured the volatility of pollution episodes, providing reliable uncertainty bounds for decision-making.

Significance and Claims
The paper claims that the GCSVR framework offers a robust solution for urban air pollution forecasting by effectively combining the spatial dependency modeling of graph neural networks with the outlier resilience of Support Vector Regression. The authors assert that this hybrid approach is particularly valuable for:

• Handling Outliers: The $\epsilon$-insensitive loss function allows the model to maintain accuracy in the presence of anomalous pollution spikes common in cities like Delhi.
• Generalizability: The framework demonstrated consistent performance across two distinct climatic and geographic environments (inland Delhi and coastal Mumbai).
• Policy Utility: By providing both accurate point forecasts and calibrated uncertainty intervals, the model supports evidence-based public health planning, early warning systems, and mitigation strategies.

The study concludes that while the current framework relies on historical concentrations and distance-based graphs, it establishes a scalable, statistically grounded tool for spatiotemporal risk assessment that can be extended in future work to include meteorological and emission covariates.