Deep Learning Explicit Differentiable Predictive Control Laws for Buildings

We present a differentiable predictive control (DPC) methodology for learning constrained laws unknown nonlinear systems. DPC poses an approximate solution to multiparametric programming problems emerging from explicit model (MPC). Contrary MPC, does not require supervision by expert controller. Instead, system dynamics is learned the observed system’s dynamics, and neural law optimized offline leveraging closed-loop model. The combination of penalty methods constraint handling outputs inputs allows us optimize law’s parameters directly backpropagating economic MPC loss through performance proposed method demonstrated in simulation using multi-zone building thermal dynamics.