Context-based Thermodynamic Modeling of Buildings Spaces
Introduction
This article is a follow-up of our previous research, where we proposed using a reinforcement learning (RL) approach to optimized the operation of the HVAC system. Context-based models where the result of our struggle to find suitable representations for the environment while organizing ideas that were being proposed in different research fields (hybrid systems, lumped RC models and context-based reasoning). In the discussion section of our RL article we argued that an intelligent system should be able to explain learned policies - a requirement that our proposal was not able to accomplish. With RL, there is no insight on the internal structure of the building environment or its governing principles.
We have assumed that one of the desired requirements for ambient intelligence is to have occupants informed about important aspects on their environment. These requirements include giving information on how energy is being used, notifications about improper use of operable windows when the HVAC is on, or when there are other more economical conditions that can guarantee the same comfort levels by taking advantage of e.g. solar gains and/or natural ventilation. As we proposed for future work in the article, by using context-based models we envision an application output where it is possible to show, for example, that by leaving a door open a thermal zone can be heated by taking advantage of solar gains in another thermal zone; or a smart recommendation showing that a certain window should be closed, for energy savings. To accomplish this level of reasoning and planning the building automation system should be able to represent knowledge and reason at a symbolic level while creating (or including) meaningful models of the dynamical building environment suitable for describing the non-trivial interaction of both continuous and discrete variables.
Context-based Thermodynamic Modeling
Thermodynamic models are frequently used for modeling the thermal behavior of building spaces. However, the occurrence of events such as, for example, doors, windows and blinds being opened or closed, can drastically affect the underlying processes that govern the dynamics of temperature evolution of building spaces, rendering current thermodynamic models less effective for control and prediction. This article presents a framework for appropriate model structure and parameter selection that accounts for such discrete disturbances based on the notion of context. Contexts are modeled as discrete states, represented by a set of discrete variables, associated with the operation of the building. Depending on how context changes our thermodynamic model transitions through a set of different linear time-invariant sub-models. Each sub-model is effective in representing the thermal behavior of a building space under a given context and the result is a hybrid automaton that effectively adjusts to the discrete and continuous dynamics of the building environment. We present an application example and use the outputs of EnergyPlus as reference for model performance evaluation. We show, through different context changes, how a context-based model can be used to represent, with reasonable accuracy, the evolution of temperatures in a simulated thermal zone.