Action-Conditioned Chronos-2 Hybrid for Sample-Efficient MPC Dynamics Modeling: Industrial SO2 Validation

Abstract

Time-series foundation models are designed to continue observed trajectories, whereas model predictive control (MPC) must compare counterfactual rollouts under candidate future actions. A forecaster whose predictions are invariant to those actions cannot rank controls by predicted plant response, irrespective of forecast error. We formalize this forecasting-to-control gap and convert pretrained Chronos-2 into an action-conditioned dynamics model. Chronos-2 Hybrid combines low-rank target-domain adaptation, separate process-history and future-decision pathways, and gated residual fusion. Removing future-decision conditioning degraded action-response fidelity in all ten Quadrotor seeds. In bidirectional transfer between two 660 MW thermal power units, the unit-balanced area under the sulfur dioxide trajectory root-mean-squared error (RMSE) learning curve was 14.95% lower than that of an input-matched scratch model; at 6 and 24 h of target labels, the RMSEs were 29.0% and 18.5% lower. A separate Benchmark Simulation Model No. 1 study showed the same low-budget pattern under long process memory. Across eight outcome independently matched desulfurization system episodes, the normalized emission–effort performance index was 29.8–69.6% lower during strategy-assisted operation. These results identify the action-conditioned interface required to use a pretrained forecaster in MPC and show why forecast error alone is insufficient to assess control readiness.