Database-Driven Iterative Learning for Building Temperature Control

Building interior temperatures are affected by the outdoor air temperature. Noting that outdoor weather patterns are somewhat repetitive in nature and historical records of outdoor temperature are readily accessible, we present a datadriven iterative learning approach to improve room temperature tracking performance over time. By comparing short term temperature forecast with the past data, chains of (non-consecutive) days exhibiting similar outside temperature patterns can be identified. The corresponding building operation record (heat input and temperature output trajectories) may then be used in iterative learning control (ILC) to update the input based on the past temperature tracking error. Multi-zone buildings are strictly passive from the heat input to temperature output in all zones. This property assures the convergence of ILC iteration if the update gain is suitably bounded, without the need of an accurate model. This means that for each chain, the zone temperature deviation from the specified profile will converge to zero as the number of days in the chain grows (i.e., as more iterations of ILC are performed). Using a six-zone physical testbed with programmable ambient temperatures, we demonstrate the practicality of the proposed approach in multiple experimental trials. Additional longer-duration simulations are performed based on the actual temperature recorded in Orlando, FL and New York, NY over a two-year period. In all cases, ILC is shown to improve tracking error in the presence of ambient temperature fluctuations.