Disturbance-Adaptive Data-Driven Predictive Control: Trading Comfort Violations for Savings in Building Climate Control

A team from EPFL has published a breakthrough paper in IEEE Transactions on Control Systems Technology introducing a Disturbance-Adaptive Data-Driven Predictive Control (DAD-DPC) framework for building climate systems. This AI-driven approach fundamentally improves upon traditional Model Predictive Control (MPC) by directly learning building thermal dynamics from operational data, eliminating complex manual modeling. Crucially, DAD-DPC employs conformal prediction to mathematically guarantee that occupant comfort violations stay within user-defined bounds, even when facing inevitable real-world uncertainties like sensor noise or unexpected occupancy. This allows building operators to explicitly trade marginal, bounded comfort deviations for significant energy savings, moving beyond rigid setpoints.

The technical core combines Willems' Fundamental Lemma for data-driven modeling with robust conformal prediction techniques to handle disturbances without needing their statistical distribution. Validated on the high-fidelity BOPTEST simulation platform and a real occupied campus building (Polydome), the system demonstrated remarkable performance. A setup allowing a 5% comfort violation bound delivered energy savings ranging from 11.2% to 30.1% across four different building cases compared to standard controllers. This represents a major step toward practical, scalable AI for building management, offering a plug-and-play solution that reduces commissioning effort while providing predictable performance and substantial operational cost reduction.