Using interpretable machine learning for indoor CO₂ level prediction and occupancy estimation

Student thesis: Doctoral Thesis

Abstract

Management and monitoring of rooms' environmental conditions is a good step towards achieving energy efficiency and a healthy indoor environment. However, studies indicate that some of the current methods used in environmental room monitoring are faced with some challenges such as high cost and lack of privacy. As a result, there is need to use a method that is simpler, reliable, affordable and without any privacy issues. Therefore, the aims of this thesis were: (i) to predict future CO₂ levels using environmental sensor data, (ii) to determine room occupancy using environmental sensor data and (iii) to create a prototype dashboard for possible future room management based on the models developed for room occupancy and CO₂ prediction. Machine learning methods were used and these included: Gradient Boosting ensemble model (GB), Long Short-Term Memory recurrent neural network model (LSTM) and Facebook Prophet model for time series (Prophet). The sensor data were recorded from three different office locations (two test sites at a university and a real-world commercial office in Glasgow, Scotland, UK). The results of the analysis show that with LSTM method, a Root Mean Square Error (RMSE) (absolute fit of the model results to the observed data) of 0.0682 could be achieved for two-hour time interval CO₂ prediction and with GB, of 82% accuracy could be achieved for proposed room occupancy estimation. Furthermore, as the model understanding was raised as a key issue, interpretable machine learning methods (SHapley Additive exPlanation. (SHAP) and Local Model-agnostic explanations. (LIME)) were used to interpret room occupancy results obtained by GB model. In addition a dashboard was designed and prototyped to show room environmental data, predicted CO₂ levels and estimated room occupancy based on what the sensor data and models might provide for people managing rooms in different settings. The proposed dashboard that was designed in this research was evaluated by interested participants and their responses show that the proposed dashboard could potentially offer inputs to building management towards the control of heating, ventilation and air-conditioning systems. This in turn could lead to improved energy efficiency, better planning of shared spaces in buildings, potentially reducing energy and operational costs, improved environmental conditions for room occupants; potentially leading to improved health, reduced risks, enhanced comfort and improved productivity. It is advised that further studies should be conducted at multiple locations to demonstrate generalisation of the results of the proposed model. In addition, the end benefits of the model could be assessed through applying its outputs to enhance the control of HVAC systems, room management systems and safety systems. The health and productivity of the occupants could be monitored in detail to identify whether resulting environmental improvements deliver improvements in health and productivity. The findings of this research contribute new knowledge that could be used to achieve reliable results in room occupancy estimation using machine learning approach.
Date of Award11 Nov 2021
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde
SupervisorMarilyn Lennon (Supervisor) & Richard Bellingham (Supervisor)

Cite this

'