Learning indoor temperature predictions for optimal load ensemble control

Aggregation of electrical appliances in residential households is a potent source for harnessing demand-side flexibility that can be leveraged by utilities or demand response aggregators for various transmission- and distribution-level services. However, the aggregated flexibility of these resources depends on such external factors as behavioral preferences of electricity consumers and temperature. More importantly, these external factors can be interdependent, e.g. ensuring the comfort of electricity consumers requires maintaining in-door temperatures within a certain range. This paper develops a deep learning approach for in-door temperature predictions and then integrates it with optimal load ensemble control. To improve the accuracy of deep learning, which is notorious for a lack of physical interpretability and performance guarantees, we employ the concept of physics-informed neural networks, which allows for incorporating a physical (thermal) building model. We use a real-world National Institute of Standards and Technology (NIST) data set to demonstrate the usefulness of temperature learning for such demand response application.