Modeling and Prediction of Occupancy in Buildings Based on Sensor Data Using Deep Learning Methods

Accurate modelling and prediction of indoor occupancy can lead to efficient optimization and control of building energy consumption. This research uses indirect ambient sensor measurements and heterogeneous data types, together with state of the art techniques for data-driven modelling based on deep neural networks architectures, for estimating building occupancy. The methodology steps include input variable selection, comprehensive data pre-processing, implementation of several models using convolutional neural networks, fully connected neural networks and long short-term memory models, and evaluation on a reference public occupancy dataset. Various design and parametrisation options are investigated in a dual formulation, as both classification and regression problem. An application of the work consists of accurate building occupancy estimations, measured using standardised metrics, that can be subsequently used in a predictive building energy control framework. One main finding of the study shows that the classification approach, which categorizes occupancy in coarse-grained occupancy levels, performed better than the fine-grained regression approach in terms of accuracy and robustness. A classification accuracy for the five-sensor occupancy model of 94% is reported, while the regression equivalent accuracy value stands at 80% with a Mean Squared Error (MSE) indicator of 0.1934.