Lawrence Berkeley National Laboratory

Model Predictive Control (MPC) has demonstrated potential for reducing building energy costs and integrating buildings into the electric grid, yet adoption in large commercial buildings remains limited. We developed, deployed, and demonstrated an MPC in a large office building from 2020 to 2025, an uncommonly long duration for MPC field research that has enabled us to gain valuable insights. The MPC achieved 45% energy savings in efficiency mode and an estimated 61% annual cost reduction under experimental dynamic prices compared to baseline rule-based control, leveraging thermal mass for load shifting across all four seasons. However, the demonstration revealed critical barriers to broader adoption: uncertain cost-to-benefit ratios, including dependence on specialized expertise for deployment, integration hurdles due to proprietary and many different data streams (APIs/protocols), gaining facility staff trust, and updating the Building Management System (BMS). There were also operational challenges from unexpected events such as wildfire protocols, equipment failures, and inconsistent data. Initial use of semantic data models for data acquisition is promising, but encountered performance issues with large queries. This manuscript summarizes the long-term deployment evolution and performance over time, details operational challenges and solutions, and proposes scalability pathways, including the opportunities and challenges with leveraging the use of Generative AI.