From Model-Centric to Data-Centric: A Practical MPC Implementation Framework for Buildings

The potential of using Model Predictive Control (MPC) to improve building operation has been shown in many studies. Unfortunately, real-world applications are still restricted by the high implementation cost and the unguaranteed profitability. In the traditional paradigm of "model-centric" MPC, most effort is devoted to constructing the control-oriented model given specific building properties and data availability. Due to the significant heterogeneity among buildings, the results are hardly reproducible, and a high level of customization is required for each new building. To address this issue, we propose a new "data-centric" approach for MPC, which starts with control-oriented data curation that acquires the necessary and cost-effective data concerning the intended control purpose and the building characteristics. The foundation of data-centric MPC is a standardized framework to quantify the data requirements and the established relationships between data usage and control performance. Such an end-to-end framework promotes actual MPC applications with controllable costs and reliable outcomes. We use tropical office buildings as an example to consolidate the data-centric MPC framework. Two use cases are provided to demonstrate its benefits. Over 10% of energy saving was achieved without excessive occupant-related data, and occupant-centric control significantly improved the thermal comfort only with proper data acquisition.