Abstract
Learning-based model predictive control (MPC) is a promising approach for HVAC optimization, especially where slow thermal dynamics limit conventional control, yet end-to-end deployment on real-building IoT platforms remains rare. This paper presents a modular AI-enabled MPC framework comprising multi-step neural state prediction, forecast-aware constrained optimization, building management system integration, and periodic model updating from operational data. The framework is deployed on a cloud-based building IoT platform and validated on passive displacement cooling (PDC), where PI control is prone to overcooling. In paired-comparison experiments, the controller reduced mean maximum overcooling from \(1.10 \,\mathrm{ \mathrm{^{\circ }\mathrm{\mathrm{C}}}}\) to \(0.20 \,\mathrm{ \mathrm{^{\circ }\mathrm{\mathrm{C}}}}\) and mean settling time from 175 min to 40 min relative to PI control. Retraining with operational data further reduced the 60 min-ahead median prediction error from \(0.33 \,\mathrm{ \mathrm{^{\circ }\mathrm{\mathrm{C}}}}\) to \(0.14 \,\mathrm{ \mathrm{^{\circ }\mathrm{\mathrm{C}}}}\) while reducing closed-loop oscillation. Because the framework’s architecture is not specific to PDC, the same pipeline can be instantiated for other HVAC subsystems on the same platform by replacing the predictor training data and adjusting the MPC objective.

This publication has 11 references indexed in Scilit: