Publication Year: 2018 Publication Type : Thesis

Abstract:

With advancement in sensors and the Internet of Things, gathering spatiotemporal information from one’s surroundings has become more convenient. There are multiple phenomenological behaviours, such as indoor comfort and occupancy trends, that can be inferred from this information. There are multiple advantages to having an accurate indoor occupancy prediction, including better understanding of space-room utilisation, which can be used to further inform energy consumption reduction, human indoor comfort optimisation and security enhancement.

We use non-intrusive ambient sensors to infer indoor occupancy patterns. Non-intrusive ambient sensors are utilised because they are commonly available in building management systems (BMSs). Machine learning techniques are applied and data-driven approaches are implemented to identify indoor human occupancy and predict comfort.These facilitate the decision-making tasks for building management professionals and are used in real-time monitoring. Our preliminary study with multiple ambient sensors reveals that carbon dioxide is one of the best predictors of indoor human occupancy.

We design a seasonal trend decomposition algorithm by implementing pervasive sensing and leveraging carbon dioxide data from BMS sensors. The first model is seasonal decomposition for human occupancy counting (SD-HOC), a customised feature transformation decomposition prediction model. This provides a novel way to estimate the number of people within a closed space, using one carbon dioxide sensor. SD-HOC integrates a time lag and line of best fit model in the preprocessing algorithms and customises different regression algorithms for each subcomponent, to predict each respective human occupancy component value. Utilising several machine learning techniques, a set of prediction values for each component is obtained. Finally, additive decomposition is used to reconstruct the prediction value for human indoor occupancy.

We improve the algorithm to cover multiple buildings with different contexts and locations and develop a large Room Utilisation Prediction with carbon dioxide sensor (RUP). RUP improves SD-HOC and is able to predict a larger number of occupants, up to three hundred, using data from a single carbon dioxide sensor. RUP de-noises and pre-processes the carbon dioxide data. We use multiple variants of seasonal decomposition techniques and feature factorisation for both occupant and carbon dioxide datasets, and develop a zero pattern adjustment model to increase the accuracy. We run our model in two different locations that have different contexts.The prediction accuracy results outweigh the state-of-the-art techniques for time series decomposition and regression.

RUP is a reliable model for any building with adequate historical data. In the real world, this condition is not always feasible, due to several limitations such as a new building only having limited historical data, or government/military buildings that have strictly controlled access to historical ambient sensor data. One way to solve this problem is by implementing a transfer learning technique with SD-HOC. We design a semi-supervised domain adaptation method for carbon dioxide - human occupancy counter (DA-HOC) to estimate the number of people within one room, by using a carbon dioxide sensor with a limited number of training labels (as little as one day of historical data). The DA-HOC model is trained using data from
a source domain that has a more complete set of training labels, and transferred to predict the occupancy of a much larger room of the target domain, with very little training data. We enhance DA-HOC into DA-HOC++ and successfully experiment with the model to transfer the knowledge from one room to five different rooms in different countries.

Moving beyond indoor human occupancy, each occupant’s comfort is also a crucial problem that needs to be considered. Indoor comfort prediction is crucial for energy efficiency cost adjustment, human productivity and non-wastage of resources. Maintaining human indoor comfort levels at acceptable values is one of the primary goals in any building and room utilisation. The main problem is that everybody has a different level of acceptance of what is comfortable. We implement a machine learning algorithm to predict the thermal comfort for each occupant. Our model successfully achieves a respectable accuracy of comfort prediction to help the BMS adjust the temperature.

This thesis presents several contributions in machine learning for indoor human occupancy and comfort prediction. This research implements and extends existing data mining techniques to solve problems on time series prediction. The solutions are scalable and can also work with minimal sets of historical training data with a transfer learning method. The research contributions in this thesis present multiple occupancy algorithms for both indoor human occupancy and thermal comfort. We believe that this research provides a big step towards building a robust solution for smart homes and smart buildings, in which the buildings are more aware of their occupants and can adapt to their needs.

BibTex:

Cite: