Time series data
The Modulai team used a combination of temperature, CO2, outdoor temperature, climate system and meeting room booking data, gathered over a year at the company’s head offices to predict how the temperature would be in a meeting room 25 minutes in the future, to enable energy savings and as well as increased comfort.
Bespoke sources of data, controlled and gathered by the client, where used to build machine learning models that predict the future temperature in a meeting room.
The prediction could then be used to control the output effect of the relevant climate control systems. While traditional HVAC control systems only have access to data for current and recent temperature variations, an intelligent controller unit could utilize a far broader set of data (such as if an area is or is likely to be occupied) when setting the output effect.
The company had two of its offices fitted with nearly 1000 temperature, CO2-level, sound and other sensors. Modulai also had access to climate system data and data from the room booking system.
How Modulai did it
Various convolutional and recurrent neural network architectures were tested out and benchmarked against more traditional algorithms such as linear regression and tree ensemble methods. The final architecture showed a significant increase in predictive power compared to existing systems. The project used a stack of relational and time series databases for data storage and retrieval as well as Python, Tensorflow and Scikit-learn for modeling.