Evaluation of Deep Learning and Machine Learning Algorithms for Building Occupancy Classification on Open Datasets

Abstract

Accurately estimating and forecasting building occupancy represents an important tasks for higher level indoor energy management and control routines. Extended availability of public and open datasets reflecting indoor conditions through various sensor measurement and indirect proxies of human activity enable reliable benchmarking of new techniques for pre-processing and learning of occupancy patterns. In this work we present a comparative study between deep learning, such as convolutional neural networks, and conventional machine learning approaches, such as decision trees and random forests, on an a reference occupancy dataset. The various design decision and parametrisation options are discussed. The building occupancy classification task involves generating model outputs for various discrete occupancy categories. Standardised metrics such as accuracy, precision, recall and the F1-score are used for replicable benchmarking of the results. Main finding of the study is that, though generally the deep learning methods offer better overall results, the addition of relevant features (sensors) to the input dataset can yield better results for the conventional machine learning models with significantly lower training time and model size. This results in suitable, fast-inference, models for embedded deployment in physical proximity to the process.