Development of the Machine Learning-Based Approach to Access Occupancy Information Through Indoor CO2 Data

Development of the Machine Learning-Based Approach to Access Occupancy Information Through Indoor CO2 Data: for Demand-Driven Building Heating Energy Management

Wang, T.

Keyson, D.V. (mentor)
De Hoogh, M.P.A.J. (mentor)

Electrical Engineering, Mathematics and Computer Science

Electrical Sustainable Energy Technology

Master of Sustainable Energy Technology

2017-07-19

Demand driven building energy management system has gained significant interest in recent years. In order to provide insights for the demand driven building heating energy management at IDE building, TU Delft, this research develops an approach to access real time occupancy information by applying supervised machine learning (ML) classification algorithms on the indoor CO2 data collected from a wireless sensor network (WSN) in the IDE building. Three types of ML classification algorithm including K-nearest neighbors (KNN), support vector machine(SVM) and decision tree (DT) have been applied to build the models, then the models are evaluated and compared with the aim to find the best ones for this case. Data collected from one single occupancy room and one double occupancy room based on the WSN are used to train different classification models, then the models are tested not only on these two rooms (local occupancy modeling), but also on another test room (global occupancy modeling). Results show that, as long as the CO2 sensors are located in the similar distance with occupants, the models trained at one room can be applied for another room with slightly lower accuracy. At single occupancy room, the binary occupancy detection (two classes: 0 or at least 1 occupant) models can reach the accuracy range from 90.6% to 93.8% for local occupancy modeling and from 86% to 89.1% for global occupancy modeling. The SVM (linear) model yields the highest accuracy. At double occupancy room, the binary occupancy detection (two classes: 0 or at least 1 occupant) models can reach the accuracy range from 96.2% to 97.2% for local occupancy modeling and from 89.1% to 91.6% for global occupancy modeling. The KNN (K=10) model yields the highest accuracy. The occupants number estimation (three classes: 0, 1, or 2 occupants) models can reach the accuracy range from 90.2% to 93.7% for local occupancy modeling and from 83.5% to 90.5% for global occupancy modeling. The SVM (linear) mode yields the highest accuracy. It is also found that SVM (linear) model yields the most stable prediction results with less overfitting in the time series, thus it is concluded as the most fit model for both binary occupancy detection and occupants number estimation with the aim to apply for demand-driven heating energy management.

Machine learning
occupancy detection
demand-driven control
heating energy

http://resolver.tudelft.nl/uuid:98a8d01e-6259-4fed-8704-5af0615304d2

Student theses

master thesis

(c) 2017 Wang, T.