UOW SMART building research revealed
As reported in FM Media earlier this year, building efficiency is set to be transformed with the release of new research undertaken by academics at the SMART Infrastructure Facility at the University of Wollongong (UOW), supported by industry partners and a government agency.
Funded by Grosvenor Engineering Group (Grosvenor), Enviro Building Services (Enviro) and the NSW Government (through the Department of Industry), the project was part of SMART’s Digital Living Lab, which provides an Internet of Things (IoT) network to create smarter living in buildings.
The research achieved three main goals – improve the accuracy of near real-time counting of people in a room or building; accurate forecasting of indoor temperature to help assess the impact of different load regimes and model of operations for HVAC systems; and building a smart sensor that can predict problems in rotating equipment such as fans.
“The outcomes were two-fold and are changing how the heating, ventilation and cooling of buildings can be used more efficiently in the built environment. An accurate real-time people counting solution for indoor environments that respects privacy is now available for building managers to utilise,” says senior research fellow Dr Rohan Wickramasuriya.
“A deep learning-based indoor temperature forecasting algorithm has been developed which provides a great alternative to the traditional approach that requires detailed information about a building’s construction, fit-out and modelled occupancy. Training this algorithm for a new building is straightforward, hence it will cut time and costs when a forecasting model is required to predict indoor temperatures.”
“Both Grosvenor and Enviro are leading building services providers that were open to innovate and collaborate with the university’s research facilities to deliver the latest project. The technical skills and domain knowledge of the research team, as well as strong engagement from industry partners helped achieve the project’s goals,” he added.
Anonymised real building data was collected from equipment maintained by Grosvenor, while Enviro’s office spaces provided image data. This data was used to train deep neural networks to predict the outcomes.
“The accurate detection of building occupation is paramount to keeping indoor environments within a set comfort zone. Temperatures can vary dramatically depending on the number of people in a given room. This research produced a solution that has a much higher accuracy (compared to the existing solution), allowing building managers to better respond to cooling and heating demand. Another opportunity includes empowering tenants with occupancy data to support better space utilisation leasing decisions,” says Rod Kington, national sustainability manager for Grosvenor Engineering Group.
“The temperature forecasting model is unique and can predict the temperature of a room or building within the next 24 hours at 15-minute intervals and takes into consideration the building occupancy, weather forecast and historical room temperature. This scenario analysis tool will significantly reduce overall running costs of a building,” he added.
Products from the research that are now being used by the industry include a cost-effective accurate people counter, combining off the shelf components with powerful machine learning software. The camera includes a Raspberry Pi unit and a Raspberry Pi camera module V2. This unit runs a custom-trained Yolov3 deep learning algorithm. The images captured by the Raspberry Pi camera are analysed locally using the Yolvo3 algorithm to count the number of people in the image, even partially obscured occupants. The processed image is discarded and only the count is transferred to a database server.
The installed software package includes all data analysis, modelling and visualisation tasks, including neural networks, all implemented in Python language.
“The research undertaken into deep neural networks is a future driver of artificial intelligence in smart buildings. We started off with a deep neural network architecture called YOLO (You Only Look Once) that can detect objects in real time scenes,” says Wickramasuriya.
“Using transfer learning, we custom trained the YOLO algorithm to detect people in the images, reviewed annotated images and then evaluated the accuracy of the algorithm. The test set accuracy of the algorithm was 92 percent which is an excellent result, given the existing solution’s accuracy is often only 65 percent.”
“In addition, to forecast room temperature, we used another deep network architecture called Long Short Term Memory (LSTM). LSTM models were compared against classical time series forecasting models and vastly outperformed the classical models from an accuracy perspective.”
Image courtesy of UOW.