Neural differential equations for temperature control in buildings under demand response programs

Heating Ventilation and Air Conditioning (HVAC) are energy-intensive systems that greatly contribute to peak demand, which can cause stability and reliability issues in the grid. The use of adaptive smart temperature controllers combined with demand response programs play an important role in addressing this challenge. However, deploying such controllers is difficult, mainly due to the need for detailed models of buildings which can be expensive to acquire. This work proposes a general purpose approach to alleviate this problem through the use of continuous-time neural differential equations models to predict the temperature and HVAC power usage. A fully data-driven approach is used to train the models, thus requiring no prior knowledge about buildings apart from metering data. Therefore, this approach is easily adaptable to different buildings. Furthermore, we show that we can adapt such models to incorporate high-level prior knowledge about building physics to further enhance their efficiency and interpretability. A planning algorithm embedded in an MPC framework is designed to control the temperature setpoint to limit the HVAC power consumption and increase eligibility to a demand response program. Extensive empirical tests are conducted on simulation and real data to assess each model’s performance. The experiments show that continuous-time models are more sample-efficient and robust to missing and irregular observations. Our experiments also reveal that for the same level of planning accuracy, continuous-time models require fewer samples than their discrete-time counterparts when adjusting temperature setpoints to lower power consumption during demand response events.