Demand response (DR) is a large source of flexibility in power systems. DR refers to moderating the power
consumption of electric loads like air conditioners, electric vehicles, and water heaters to provide power system
services such as load-shifting and regulation. Loads that participate in DR programs are rewarded with
decreased electricity prices, rebates, or direct payments. Water heaters account for a substantial fraction of
winter power consumption in Quebec. Water heaters have flexible power consumption because water may be
heated hours before it is used. Therefore, water heaters are a natural candidate for DR. A key challenge in DR
is load uncertainty. Conventional power system components like transmission lines, transformers, and energy
storage have low uncertainty because they have precise physical models, high-bandwidth communication, and
precise actuation. None of these are true for loads. For example, a water heater depends on its human users, is
hard to measure the power usage of if it is inside of a metered building, and its power consumption may be a
complicated function of temperature setpoint. Online convex optimization is a subfield of machine learning
that manages uncertainty by leveraging information as it becomes available. We will apply online convex
optimization to load-shifting with water heaters in Quebec.