Large neighborhood search for temperature control with demand response

Demand response is a control problem that optimizes the operation of electrical loads subject to limits on power consumption during times of low power supply or extreme power demand. This paper studies the demand response problem for centrally controlling the space conditioning systems of several buildings connected to a microgrid. The paper develops a mixed integer quadratic programming model that encodes trained deep neural networks that approximate the temperature transition functions. The model is solved using standard branch-and-bound and a large neighborhood search within a mathematical programming solver and a constraint programming solver. Empirical results demonstrate that the large neighborhood search coupled to a constraint programming solver scales substantially better than the other methods.