OmniST: Topology-Invariant Spatiotemporal Forecasting with Guaranteed-Coverage Uncertainty at Unseen Sensor Locations
Organizations: Department of Computer Science and Engineering, Indian Institute of Technology Patna, India · Sopra Banking Software India as Senior Software Development Engineer
Abstract
Standard spatiotemporal graph convolutional networks need a fixed-size adjacency matrix, coupling model weights to a specific sensor count N; they cannot produce any output at a sensor absent from training. We present OmniST, a topology-invariant forecaster whose computation has zero structural dependency on the number of nodes, and which, composed with inductive split-conformal calibration, gives distribution-free, coverage-guaranteed prediction intervals at arbitrary (including never-seen) sensor locations, preserving that guarantee as sensors are added or removed. Topology invariance follows from coordinate-based tokenization, k-nearest-neighbor attention from geographic coordinates at each forward pass, and per-node temporal processing via a bidirectional GRU. On real Delhi air-quality data (37 stations, 2019-2023), OmniST attains a 30-day PM2.5 MAE of 31.0 ug/m3 and degrades at most 1.1 percent when 40 percent of stations are removed. Its inductive forecasts at unseen stations beat nearest-neighbor interpolation (about 12 percent) and a learned kriging GNN (IGNNK, about 21 percent), and, unlike either, carry a finite-sample coverage guarantee. The guarantee holds at 90 percent or above under 30-50 percent node dropout across four real datasets over three domains (Delhi and Beijing air quality, NOAA GHCN weather, Caltrans PeMSD7-M traffic), and still holds at 325 real sensors (full PeMS-BAY). We also fix a temporal-drift failure: static conformal under-covers over time (about 0.50 late), but a rolling adaptive-conformal estimator restores target coverage (0.89 overall, 0.96 late) at the cost of wider intervals.
Figures & tables
| Month | OmniST | E-STGCN ∗ | Note |
| Jan | 57.3 | 56.1 | Comparable |
| Mar | 24.2 | 25.8 | Comparable |
| Apr | 22.0 | 25.4 | OmniST better |
| May | 30.9 | 29.0 | Comparable |
| Jun | 18.1 | 22.0 | OmniST better |
| Jul | 11.3 | 15.7 | OmniST better |
| Active stations | MAE ( g/m 3 ) | Degradation |
|---|---|---|
| 37 (full network) | 35.33 | — |
| 32 (5 failures) | 35.72 | |
| 27 (10 failures) | 35.72 | |
| 22 (15 failures) | 35.58 |
| Method | MAE ( g/m 3 ) |
|---|---|
| Nearest-neighbor (copy closest training station) | 30.3 |
| IGNNK (learned inductive kriging [ 9 ] ) | 34.0 1.3 |
| OmniST + node-dropout training augmentation | 27.1 1.1 |
| OmniST (coordinate-based) | 26.7 1.2 |
| Setting | Dataset | OmniST | E-STGCN |
| Transductive MAE (seen) | Delhi AQ | 28.6 1.2 | 33.5 1.9 |
| Weather | 5.55 0.14 | 5.30 0.29 | |
| Inductive (unseen nodes) | Delhi AQ | MAE 26.7, cov 0.91 | cannot predict |
| Weather | MAE 5.0, cov 0.92 | cannot predict | |
| Transductive = train and test on the same stations. On Delhi OmniST is better (paired ); on weather E-STGCN is marginally better ( , n.s.) — reported both ways. Inductive = predict at held-out stations: the fixed training adjacency of E-STGCN ( / ) has no row/column for an unseen station, so it structurally produces no output; OmniST does, with a coverage guarantee. | |||
| Method | Coverage | Mean width | Guarantee? |
|---|---|---|---|
| Nearest-neighbor + residual | 0.925 | 149.7 | No |
| OmniST + conformal | 0.915 0.006 | 128.5 3.7 | Yes |
| Coverage under training-node dropout (OmniST + conformal): | |||
| 30% stations dropped | 0.914 | — | Yes |
| 50% stations dropped | 0.913 | — | Yes |
| This is the original 3-seed single-domain run. The 10-seed re-run (Table VI ) gives a consistent Delhi coverage of 0.914 (width 127.0); the small differences are seed/run variation, not a discrepancy. | |||
| Dataset (type) | OmniST | NN | drop 30% | drop 50% |
|---|---|---|---|---|
| Delhi AQ (environmental) | 0.914 .006 | 0.925 | 0.912 | 0.916 |
| Beijing AQ (environmental) | 0.928 .003 | 0.869 | 0.928 | 0.930 |
| Weather (meteorology) | 0.919 .007 | 0.896 | 0.919 | 0.920 |
| PeMSD7-M (transportation) | 0.915 .008 | 0.887 | 0.916 | 0.917 |
| OmniST = conformal coverage at unseen nodes; NN = nearest-neighbor residual baseline. | ||||
| 10 seeds (air quality, weather); 6 seeds (228-node traffic). Fig. 1 . | ||||
| Dataset | OmniST | NN | IDW | mean- | GP |
|---|---|---|---|---|---|
| Delhi AQ | 0.914 | 0.925 | 0.871 | 0.813 | 0.845 |
| Beijing AQ | 0.928 | 0.869 | 0.818 | 0.789 | 0.789 |
| Weather | 0.919 | 0.896 | 0.927 | 0.945 | 0.893 |
| Traffic | 0.911 | 0.887 | 0.905 | 0.906 | 0.922 |
| All methods use identical split-conformal calibration. OmniST is the only one | |||||
| on every dataset; interpolators under-cover (tighter but invalid) on several. | |||||
| Dataset | metric | full | no-GAT | k=4 / k=16 |
|---|---|---|---|---|
| Delhi AQ | coverage | 0.915 | 0.912 | 0.914 / 0.913 |
| ind-MAE | 26.7 | 26.4 | 26.2 / 26.1 | |
| Weather | coverage | 0.917 | 0.925 | 0.920 / 0.922 |
| ind-MAE | 4.99 | 4.80 | 4.92 / 4.97 | |
| Traffic | coverage | 0.919 | 0.919 | 0.919 / 0.919 |
| ind-MAE | 9.55 | 9.54 | 9.63 / 9.63 |
| Method | Overall cov. | Late-window cov. | Width |
|---|---|---|---|
| Static split-CP | 0.758 | 0.498 | 95.4 |
| Rolling window | 0.840 | 0.799 | 127.4 |
| Adaptive CP (ACI) | 0.847 | 0.905 | 139.0 |
| Rolling-ACI | 0.890 | 0.962 | 170.1 |
| Late-window = final third of the test year (worst drift). | |||
| Width in g/m 3 . Restoring drift coverage costs width (honest trade-off). | |||
Spatiotemporal Forecasting
Spatiotemporal forecasting predicts how quantities change across locations and time, from traffic and weather to disease and environmental systems.
See all 448 papers