Overview / Usage

Existing technologies for environmental and sustainable building architects is time-consuming and rely on several assumptions to be made concerning building envelope and typical weather conditions. For instance, a popular open-source tool used by Archiects is BeOPT (https://www.beopt.net/features), which relies on another open source tool, EnergyPlus, to perform various simulations of building envelope.

While BeOPT has no doubt greatly contribute to the activity of architects in building design, it still relies on several older models and hard-coded assumptions [1] to be chosen a-priori by architects, requiring several iterations to realistically, at best case scenario, the envelope of buildings to achieve the desired thermal comfort outcome, while constraining energy consumption.

For model and simulation tunning, models such as BeOPT are often used in combination with thermal comfort adaptive methods, such as the PMV and Thermal Comfort Methods, which in turn, also rely on outdated regressive models under not typical conditions. For example, the PMV still relies on assumptions where participants remain in a sealed chamber with air conditioner and have no access to windows throughout their day, which is often not the case in a normal house envelope or office job these days, while the Thermal Comfort Method relies on the opposite, where no Air Conditioner is available, but only windows and the thermal comfort indoors is solely dictated by the outdoor environment.

[1] https://www.beopt.net/features

Methodology / Approach

We propose the use of deep learning to model the various components required by building architects based on sensor available data to construct mixed mode models.

To see how deep learning can be leveraged to address the limitation of existing models, I reproduced the previously available code for PMV and Adaptive Method as R Notebooks, which are available here:

Thermal Comfort: http://rpubs.com/carlosandrade/pmv_ppd

Adaptive Method: http://rpubs.com/carlosandrade/adaptive_method

As it can be observed from the commentary in the R Notebooks, the process relies on the usage of data sensors for temperature, humidity, (as one would expect to affect thermal comfort), wind speed (as one would expect for usage of fans), and indoor and outdoor temperatures. However, several assumptions are performed on existing methods, as it can be observed in the R notebooks, such as the air adjustment speed for the Adaptive Method, where a simple regression model is used to observe the relationship between thermal comfort and air speed.

With the growing usage of IoT and the increasing plethora of sensors to monitor different parts of the house, different learning architectures and layers can be used to learn realistic parameters of BeOpt models, instead of relying on apriori and imprecise assumptions. Considering deep learning can augment in several layers at it's simplest form regression functions such as used by the coded models still in use today, and the availability in data, the different dimensions it could capture offer a promising alternative to more classical approaches.