Skip to content
Home » Deep learning trains buildings to optimise efficiency

Deep learning trains buildings to optimise efficiency

The Pacific Northwest National Laboratory (PNNL) has released a prototype code to speed up implementation of controls for up to 50% more efficient buildings.

Around the world, residential and commercial buildings account for nearly 40% of all energy consumption. About 40% of that energy use is from heating, ventilation, and air-conditioning (HVAC) systems.

An emerging efficiency method for buildings called model predictive control (MPC) could reduce HVAC energy use by up to 50% without impacting occupant comfort. But the method has been hampered by high costs associated primarily with software and computing.

Now, a team of scientists at Pacific Northwest National Laboratory (PNNL) has developed a new deep learning approach that uses building data and physics knowledge to train the MPC. The learning approach allows non-experts to optimise control of a building’s energy systems without the need for additional computing power and proprietary software.

“The deep MPC method lowers costs and leads to faster design and installation of the control systems compared to traditional methods,” said Draguna Vrabie, a data scientist in PNNL’s Advanced Computing, Mathematics, and Data Division.

Vrabie teamed with PNNL data scientists Aaron Tuor and Ján Drgoňa to develop the deep MPC method through funding from the US Department of Energy’s Building Technologies Office. The team recently released the prototype code on GitHub, along with example applications, to facilitate adoption in the building controls community and to stimulate new research and development.

The researchers’ technical paper describing the deep MPC methodology is under review by Institute of Electrical and Electronics Engineers (IEEE).

Overcoming MPC obstacles with deep learning methods

PNNL’s new approach starts with a building model represented as a deep neural network. Neural networks, a form of artificial intelligence, are mathematical models that can be trained to recognise patterns and relationships in connected sets of data.

Because the model incorporates physics knowledge, it learns faster and from much less data than traditional off-the-shelf neural networks. Once installed, traditional online MPC continually solves optimisation problems while the building is operating.

This requires installing computing resources inside the building and proprietary optimisation solvers. In contrast, the deep MPC method trains a neural control policy that can be installed on low-cost embedded hardware.

A flexible, adaptive generalist with the expertise of a specialist

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

“The deep learning MPC is a generalist that can quickly adapt to behave like an expensive physics-based MPC, a highly trained specialist which understands all the details of a single building,” said Drgoňa.

Although still in the prototype stage, the flexibility and low cost of the deep learning MPC method makes it possible to transition to practice and apply at a large scale. The team continues to increase the methodology’s range of practical applications.

Source: Pacific Northwest National Laboratory