Adaptive control and identification for heating demand-response in buildings

Abstract

An adaptive control and system identification framework is presented. This framework is applied for closed-loop system identification and adaptive heating demand-response control for residential buildings. The building envelope, considered for control purposes within the context of a Model Predictive Control (MPC) strategy, is based on the simplified Resistive-Capacitive analogy. These models characterized with parametric and state uncertainty, are adapted by incorporating, on-line, learning-based Moving Horizon Estimation (MHE) for the joint state-parameter based estimation problem. Learning-based Moving Horizon Estimation, in this context, incorporates a reinforcement learning (RL) strategy. The latter entails a Q-factor parametrized value function approximation of the MHE arrival cost which is adapted (learned) on-line using a temporal difference algorithm based on the MPC control policy.