Approximate Model Predictive Control for Heating of Arbitrary Buildings Without Thermal Modeling

Model Predictive Control (MPC) is a promising method for enabling grid-interactive heating in buildings. However, deriving an accurate building model, which is a prerequisite for a successful implementation of MPC, hinders its widespread adoption. The present paper eliminates the need for buildingspecific modeling by developing an approximate MPC (AMPC) based on a neural network that emulates MPC behavior without requiring a building model. The AMPC is trained on data created by controlling the heating system of 16,000 building models with MPC in response to dynamic pricing. Utilizing data from diverse buildings forces the AMPC to learn building-agnostic decision-making, thereby enabling its application to buildings with unknown dynamics. The AMPC’s control performance is validated on 687 unseen building models for 25 weeks each across different climate zones. Compared to the MPC, 50% of the buildings, the AMPC achieves less than 7% excess cost and less than 2 Kh/d increase in thermal discomfort, while simultaneously reducing the computation time by 95%. For very well-insulated buildings that barely require heating, the AMPC incurs substantial relative cost increases, even though the absolute cost increases are most probably negligible.