Predicting Carbon Dioxide Levels for Built Environment of Flexible and Hot-Desking Offices in Different Scenarios

This paper proposes a novel featured neural network of predicting carbon dioxide levels for built environment of flexible offices in Singapore via data-driven approach. In this work, different versions, scenarios and training algorithms are evaluated. Key findings are as follows: (1) Both version 1.1 and 1.2 are able to perform well on predicting CO
2
levels at accuracies of 90%+ and 74%+, respectively. (2) Predictive results are not significantly impacted with consolidated remarks between scenarios. (3) Experimental results show that the neural networks with batch gradient descent (BGD) outperforms stochastic gradient descent (SGD) in all versions and scenarios. The advantage differences between BGD and SGD are about 3% for version 1.1, and 13% for version 1.2.