Nearly 40% of all U.S. energy consumption occurs in residential and commercial buildings, and about 40% of that use is associated with HVAC systems. An opportunity to lower consumption in these sectors through the application of deep learning technology is being advanced by U.S. Pacific Northwest National Laboratory (PNNL) researchers.

Model predictive control (MPC) has emerged as a building efficiency optimization method with potential to reduce HVAC energy use by up to 50% without impacting occupant comfort, but high costs associated with software and computing have constrained its application. The researchers devised a new deep learning approach that uses building data and physics-informed control policy architecture to train the MPC. The learning approach allows non-experts to optimize control of a building's energy systems without the need for additional computing power and proprietary software.

The new method starts with a building model represented as a physics-based deep neural network developed to learn faster and with less data than traditional off-the-shelf neural networks. Traditional online MPCs continually solve optimization problems while a building is operating and requires installation of computing resources inside the building and proprietary optimization solvers. These requirements are eliminated with the deep MPC method, which trains a neural control policy that can be installed on low-cost embedded hardware.

Deep MPC learns to make optimal decisions and can adapt by learning from new data. The method is generic enough that it can be used with any type of building, offering it solves the cost-effective deployment in every building.

The prototype code is available on GitHub along with example applications to facilitate adoption in the building controls community and to stimulate continuing improvements.

Structural equivalence of implicit MPC in a dense form with the proposed deep learning-based reformulation of MPC. Source: Ján Drgona et al./PNNL