Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Comparison of machine learning models for occupancy prediction in residential buildings using connected thermostat data
https://orcid.org/0000-0002-7720-7830
Abstract Occupancy detection capabilities provided by modern connected thermostats enable adaptive thermal control of residential buildings. While this adaptation might simply consider the current occupancy state, a more proactive optimized system could also consider the probability of future occupancy in order to balance comfort and energy savings. Because such proactive control relies on accurate occupancy prediction, we comparatively evaluate a number of machine learning models for predicting measurements of the future occupancy state of homes that is critically enabled by thermostat data from real households in ecobee's Donate Your Data program. We consider a variety of models including simple heuristic and historical average baselines, traditional machine learning classifiers, and sequential models commonly used for time series prediction. We evaluate the performance of each model according to temporal, behavioural, and computational efficiency characteristics. Our key overall finding is that the random forest algorithm matched or outperformed the other candidate models, had consistently high accuracy predicting over a range of time horizons, and is relatively efficient to train for individual “edge” devices.
Highlights Leveraged actual consumer longitudinal data from connected thermostat devices to build occupancy prediction models. Compared various models from the built environment, machine learning, and deep learning communities. Highlighted aspects of the models including predictive accuracy, temporal, behavioural, and computational efficiency. Quantified the effects of additional PIR sensors on sensing motion in the home and found evidence of an upper sensing limit.
Comparison of machine learning models for occupancy prediction in residential buildings using connected thermostat data
https://orcid.org/0000-0002-7720-7830
Abstract Occupancy detection capabilities provided by modern connected thermostats enable adaptive thermal control of residential buildings. While this adaptation might simply consider the current occupancy state, a more proactive optimized system could also consider the probability of future occupancy in order to balance comfort and energy savings. Because such proactive control relies on accurate occupancy prediction, we comparatively evaluate a number of machine learning models for predicting measurements of the future occupancy state of homes that is critically enabled by thermostat data from real households in ecobee's Donate Your Data program. We consider a variety of models including simple heuristic and historical average baselines, traditional machine learning classifiers, and sequential models commonly used for time series prediction. We evaluate the performance of each model according to temporal, behavioural, and computational efficiency characteristics. Our key overall finding is that the random forest algorithm matched or outperformed the other candidate models, had consistently high accuracy predicting over a range of time horizons, and is relatively efficient to train for individual “edge” devices.
Highlights Leveraged actual consumer longitudinal data from connected thermostat devices to build occupancy prediction models. Compared various models from the built environment, machine learning, and deep learning communities. Highlighted aspects of the models including predictive accuracy, temporal, behavioural, and computational efficiency. Quantified the effects of additional PIR sensors on sensing motion in the home and found evidence of an upper sensing limit.
Comparison of machine learning models for occupancy prediction in residential buildings using connected thermostat data
https://orcid.org/0000-0002-7720-7830
Abstract Occupancy detection capabilities provided by modern connected thermostats enable adaptive thermal control of residential buildings. While this adaptation might simply consider the current occupancy state, a more proactive optimized system could also consider the probability of future occupancy in order to balance comfort and energy savings. Because such proactive control relies on accurate occupancy prediction, we comparatively evaluate a number of machine learning models for predicting measurements of the future occupancy state of homes that is critically enabled by thermostat data from real households in ecobee's Donate Your Data program. We consider a variety of models including simple heuristic and historical average baselines, traditional machine learning classifiers, and sequential models commonly used for time series prediction. We evaluate the performance of each model according to temporal, behavioural, and computational efficiency characteristics. Our key overall finding is that the random forest algorithm matched or outperformed the other candidate models, had consistently high accuracy predicting over a range of time horizons, and is relatively efficient to train for individual “edge” devices.
Highlights Leveraged actual consumer longitudinal data from connected thermostat devices to build occupancy prediction models. Compared various models from the built environment, machine learning, and deep learning communities. Highlighted aspects of the models including predictive accuracy, temporal, behavioural, and computational efficiency. Quantified the effects of additional PIR sensors on sensing motion in the home and found evidence of an upper sensing limit.
Comparison of machine learning models for occupancy prediction in residential buildings using connected thermostat data
Huchuk, Brent (Autor:in) / Sanner, Scott (Autor:in) / O'Brien, William (Autor:in)
Building and Environment ; 160
04.06.2019
Aufsatz (Zeitschrift)
Elektronische Ressource
Englisch
Deep Learning Models for Future Occupancy Prediction in Residential Buildings
| Springer Verlag | 2023
Developing a residential occupancy schedule generator based on smart thermostat data
| Elsevier | 2024