[by Umair Khan, Anas Battah and Ali Moro]
This article will shed light on numerous applications of data from smart building and how occupancy and temperature prediction app helped SYK understand their buildings and the ecosystem around them better, as well as how the prototype further validated the value of data in the current economy.
Data driven applications are fairly novel and the idea is still in its infancy. With the advent of building automation spawned the concept of smart buildings powered by hundreds of sensors, and these sensors continuously recording data from the building environment. Just like any other domain data accumulation has enabled us to dig out edifice insights never thought of before.
Data from smart buildings is used in various applications. Since a gargantuan amount of data is collected smart ways are required to visualise it, as too much data looks amazingly similar to no data at all. One of the most popular data driven applications is Analytics Dashboard which converts the plethora of data into intelligible information in form of charts, reports, flags and alarms.
A derivative/extension of dashboard is Building Management System (BMS). It allows users to control the electrical and mechanical equipment of the building such as ventilation, lighting, power systems, fire systems, and security systems. Predictive maintenance is another application of smart buildings data: machine learning algorithms consume data to train complex models for maintenance prediction. IBM is one of the pioneers in this domain. It has paved the way for companies with large infrastructure to dispel uncertainty when handling its maintenance.
A short video in which team Smart Energy demonstrate the results of their project
SYK, being one of the largest university building owners, hosts some state-of-the-art smart buildings. These buildings are powered by sensors primarily provided by Siemens and Helvar. Both vendors store sensor data and provide access through web-services. The data is usually in a raw form with timestamps which require some pre-processing to make it application ready. During this exploratory phase of 3 months, we at Floworks/SCIL learned about smart building data, experimented with it and developed two different applications. One shows the Occupancy Status of a room and Occupancy Rate (based on historical data) to learn more about room usage. The other one is the indoor temperature forecast application which uses historical data to train deep neural network for temperature predictions. The applications are discussed in detail in the following sections.
Occupancy Application: The application uses a basic 3D model of one of the floors (KampusKlubi) of the building Kampusareena at Tampere University Of Technology, colour codes different types of rooms of the floor and then uses another set of colour codes to display occupancy status. Three different statuses define occupancy, namely, occupied, occupied with no recent activity, and vacant. The application also shows the occupancy rate of rooms for the last 3 months. This is only to show the potential that this sort of information can be used make critical decisions such as converting a room to storage as a result of low usage due to any implicit reason. Moreover, the stakeholder benefitting the most from this application is a university student who can instantly locate vacant rooms in the university.
The application is scalable and open to inclusion of advanced features such as suggestion of the closest vacant room. Using historical occupancy data, Machine Learning models can be trained to predict occupancy of different rooms. The information could help in scheduling meetings etc.
Temperature Forecast Application: The application is built on top of a recurrent neural network – a Long-Short-Term-Memory network. LSTM networks are special kinds of recurrent neural networks capable of retaining and losing information at will. These are one of the most popular choices for sequence prediction. Temperature data from different rooms – after necessary preprocessing – was used to train models consisting of stacked LSTM networks. Despite the lack of training data the models achieved an impressive Root Mean Squared Error (RMSE an accuracy metric) score of ~0.32. The best model was then connected to a web interface through web-services for prototyping purpose.