AI forecasting model adapts to sensor changes, guarantees uncertainty at new locations

The authors introduce a spatiotemporal architecture that works with any number of sensor nodes and gives statistically valid prediction intervals at never-seen coordinates.

Academic

Praveg Vashishtha

Indian Institute of Technology Patna · Sopra Banking Software India

Research Digest·4 November 2026·2 min read

Illustration: AI forecasting model adapts to sensor changes, guarantees uncertainty at new locations

In one line

OmniST forecasts at unseen sensor locations with guaranteed uncertainty coverage via topology-invariant design and conformal calibration.

Before you read this

Two ideas this paper builds on

conformal predictionA method that produces prediction sets with guaranteed coverage, enabling risk control without distributional assumptions.6 papers →sparse attentionSelectively attending to only the most relevant positions or tokens to reduce computational overhead.46 papers →

Read it as

Sensor networks change: stations are added, removed, or fail. Standard forecasting models break when the number of sensors changes, requiring retraining. They also cannot predict at a location they never saw during training.

OmniST is a forecaster built so its internal computations do not depend on the total number of sensors. It uses sensor coordinates to find nearby neighbors and processes each sensor independently. This lets it work with any number of sensors, old or new.

The model predicts air pollution levels at never-seen stations about 12 percent more accurately than simple interpolation. More importantly, it provides prediction intervals that, with 90 percent or higher probability, contain the true value, even when up to half the training sensors are removed. This guarantee holds without retraining.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

What it released

Code

Concepts

topology-invariant forecasting conformal prediction sparse attention adaptive conformal inference inductive prediction