Problem
Research questions and friction points this paper is trying to address.
Time-series Foundation Models
Model Predictive Control
Control Excitation
Zero-shot Forecasting
Input-response Relationship
Time-series Foundation Models for Predictive Control: The Role of Excitation
🤖 AI Summary
This study addresses the frequent failure of time-series foundation models in predictive control due to their inability to accurately model responses to intervention actions. Using heat pump control as a representative scenario, we conduct closed-loop experiments integrating model predictive control with zero-shot forecasting. Our findings reveal a critical insight: low prediction error does not necessarily translate to effective control performance. Furthermore, we establish contextual excitation as a fundamental prerequisite for achieving model controllability. Specifically, sufficient contextual excitation is essential for recovering input-response relationships and ensuring system controllability. Preliminary closed-loop evaluations further demonstrate the practical potential of short context windows in real-world control applications.
📝 Abstract
Deploying model predictive control (MPC) requires constructing or identifying a predictive model for each target system. Time-series foundation models (TSFMs) offer an attractive option thanks to strong zero-shot forecasting capabilities across systems. However, low forecast error does not guarantee that a TSFM captures the system's response to the alternative actions considered by the controller. We study this gap using residential heat-pump control as a test bed, measuring the agreement between predicted and ground-truth effects of control interventions. Importantly, we find that TSFMs can recover the system's input-response relationship when the context contains sufficient independent control excitation. Common fine-tuning pipelines and feature smoothing reduce, but do not eliminate, the need for in-context excitation. Our results indicate that current TSFMs used for predictive control require sufficiently informative control variation in the inference context. Initial closed-loop results show promise for shorter context windows.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Time-series Foundation Models
Model Predictive Control
Control Excitation
Zero-shot Forecasting
Closed-loop Control
🔎 Similar Papers
No similar papers found.
