When deterministic models fall short: A motivation for enriching building simulations with data-driven stochasticity

In the past decade(s) there has been a lot of studies on how advanced control methodologies, such as model predictive control (MPC) and reinforcement learning (RL), can improve the performance of heating, ventilation and air conditioning (HVAC) systems in buildings. Most studies are based on simulations. Simulation environments are necessary for comparison of different control algorithms, as it is challenging to perform experimental studies with reproducible conditions in real buildings. However, the validity of the simulation results compared to implementation in a real building is seldom documented.

The approaches for creating building simulation environments can range from highly detailed physical models (white box) to purely data-driven models with no prior knowledge embedded (black box).

While white-box models are able to include the complicated internal dynamics resulting from the interaction between the building envelope, the HVAC system and the control system, data-driven approaches (at least if based on real measurement data) are typically better suited to represent the stochastic effects typically present in buildings (e.g. driven by user presence and behaviour), as it is challenging to implement noise with unknown properties and state connection in complex dynamical models.

In this study, we compare the performance of an MPC in a real-life implementation in a building with its performance in simulation environments. A grey-box MPC algorithm has been implemented and tested in an office building with a waterborne heating system. A white-box emulator model of the same building has been developed using Modelica and implemented in the BOPTEST framework. Our results show that the MPC prediction performance is significantly better in the simulation environment than in the real-life implementation.

We aim to contrast the physics-based model used in this study with a data-driven model that does not rely on a state-space representation of the building. Instead, it propagates measurements of input-output data directly using a linear combination of past measurements found via subspace identification. Although this approach is not able to include nonlinear dynamic effects to the same extent as state-space descriptions, it alleviates the need for detailed modeling efforts which facilitates a large-scale adoption for varying houses and buildings.

In the paper, we will evaluate the benefits and disadvantages of the different approaches for creation of simulation environments, as a tool to infer performance of the MPC in the real life application, and come with recommendations for future improvements.