More Sensors Only One Field: Rethinking Continual Spatio-Temporal Forecasting

2026-09-25

AI Summary

This study addresses the failure of conventional graph-based spatial representations in sensor networks undergoing expansion, which hinders continuous spatiotemporal forecasting. Motivated by the insight that sensor addition alters observational evidence rather than underlying dynamics, this work proposes the STFO operator and introduces a unified field evolution representation framework. The method parameterizes knowledge as a shared operator, decoupling layout variations through distinct observation and query interfaces. By integrating normalized coordinate aggregation, spectral descriptors, Fourier propagation, and attention mechanisms, it enables spatial mapping reuse while adaptively handling process drift. Experimental results demonstrate that STFO achieves state-of-the-art performance across multiple datasets, reducing the mean absolute error by 8.4% and 4.7% on the PEMS-Stream and CA-Stream benchmarks, respectively.

Abstract

Continual spatio-temporal forecasting supports traffic management and environmental monitoring under evolving dynamics and expanding sensor networks. However, conventional graph-based continual learning methods tie forecasting representations to the current sensor layout, so sensor expansion can alter the representation of learned spatial relationships. Our key insight is that sensor expansion changes the evidence available about a process without necessarily changing the dynamics to be learned. We propose STFO (Spatio-Temporal Field Operator), which parameterizes forecasting knowledge as a shared field-evolution operator and handles changing sensor layouts through observation and query interfaces. Normalized coordinate-based aggregation lifts irregular sensor histories onto a fixed latent grid, enabling reuse of learned spatial maps across observation sets without sensor-specific parameters. To accommodate process drift, a spectral descriptor summarizes variation across spatial scales and conditions Fourier propagation and attention to adapt operator responses to the current spatial regime. Coordinate-based decoding queries the evolved field at sensor locations and combines spatial corrections with local-history predictions. Experiments on PEMS-Stream, CA-Stream, and AIR-Stream demonstrate state-of-the-art average forecasting performance. STFO-Large reduces average MAE over DOL by 8.4% on PEMS-Stream and 4.7% on CA-Stream. Our code is available at https://github.com/Xielewei/Spatio-Temporal-Field-Operator.

Problem

• Continual Spatio-Temporal Forecasting
• Sensor Expansion
• Graph-based Continual Learning
• Spatial Representation

Innovation

• Continual Spatio-Temporal Forecasting
• Spatio-Temporal Field Operator
• Normalized Coordinate-based Aggregation
• Spectral Descriptor
• Coordinate-based Decoding