Bayesian filtering for building occupancy estimation from carbon dioxide concentration

Chaoyang Jiang, Zhenghua Chen, Rong Su, Mustafa K. Masood, Yeng Chai SohEnergy 2019 · DOI

TL;DR

This paper introduces a novel Bayesian filtering framework to accurately estimate building occupancy levels using carbon dioxide (CO2) concentration. By fusing a statistical model that captures temporal occupancy patterns with an observation model that links CO2 levels to occupancy, the method significantly improves estimation accuracy and reduces time delays compared to existing approaches. This advancement is crucial for optimizing building energy consumption and improving occupant comfort.

What Problem Does This Paper Address?

Efficiently managing energy consumption in buildings is a major global challenge, with HVAC (heating, ventilation, air conditioning, and lighting) systems being a significant contributor to energy usage. The key to optimizing these systems lies in understanding and predicting building occupancy. Traditional methods for occupancy estimation have limitations: vision-based systems raise privacy concerns, terminal-based methods require active user participation, and simple environmental sensors often lack accuracy or can be costly. While CO2 concentration is a promising indicator of occupancy due to its direct correlation with human respiration, accurately translating CO2 levels into precise occupancy estimates is difficult. This is due to factors like ventilation rates, building air tightness, and the inherent noise and delays in CO2 measurements. This paper addresses the challenge of developing a robust, non-intrusive, and low-cost method for real-time occupancy estimation, specifically focusing on leveraging CO2 data.

Methods

The core of this paper is a Bayesian filtering framework that combines two distinct models:

  • Statistical Model: This model captures the temporal dynamics of occupancy. Instead of a simple homogeneous Markov model, the authors employ a first-order inhomogeneous Markov chain. This allows for time-varying transition probabilities, reflecting how occupancy patterns can change throughout the day or week. This model is trained on historical occupancy data to learn these temporal dependencies.
  • Observation Model: This model establishes the relationship between CO2 concentration and occupancy. The authors propose an ensemble Extreme Learning Machine (ELM) to identify this model. ELMs are known for their fast learning capabilities. The "ensemble" aspect is crucial: multiple ELMs are trained, and their results are combined through a voting mechanism to mitigate the issue of diverse solutions arising from randomly initialized ELM weights. This ensemble ELM is trained to predict the occupant number from CO2 data, venting status, and past occupancy information.

The Bayes filter then acts as the fusion engine. It recursively combines the predictions from the statistical model (transition probabilities) and the observation model (likelihood of CO2 measurements given occupancy) to produce a refined, real-time posterior probability distribution of the occupancy level. This approach leverages both historical patterns and current sensor readings for more accurate and timely estimations.

Datasets

The study utilizes a test-bed in an office space where CO2 concentration was measured at a 1-minute interval. Occupancy data was collected over 30 weekdays using a semi-automatic approach involving video recordings and manual verification to establish a highly accurate "ground truth." This dataset was used for both training the statistical model (transition probabilities) and the observation model (ensemble ELM). For comparative analysis, data from a previous study ([16]) was also used, which included 31 days of data.

Key Insights and Ideas

  • The Power of Fusion: The paper strongly advocates for combining different information sources. Fusing a temporal statistical model with a CO2-based observation model proves significantly more effective than using either model in isolation.
  • Inhomogeneous Markov Chains Matter: Recognizing that occupancy patterns aren't static and vary over time (e.g., morning vs. evening, weekday vs. weekend) is key. Using an inhomogeneous Markov chain allows the model to adapt to these dynamic temporal dependencies.
  • Ensemble ELM for Robustness: The random nature of ELM weights can lead to varied results. Employing an ensemble approach with voting not only improves the accuracy of the observation model but also makes it more robust to these variations.
  • Bayes Filter as a Corrector: The Bayes filter acts as a sophisticated post-processor, correcting errors and reducing time delays introduced by the observation model by incorporating prior knowledge from the statistical model. This is particularly effective during periods of low activity or when ventilation systems are off.

Key Results

  • Improved Occupancy Estimation: The proposed Bayes filter, using the ensemble ELM as its observation model, achieved an estimation accuracy of 85.57%, outperforming the ensemble ELM alone (83.11%) and a single ELM (77.29%).
  • Reduced Errors: The Normalized Root-Mean-Squared Error (NRMSE) was reduced to 0.1270 with the Bayes filter, compared to 0.1377 for the ensemble ELM.
  • Enhanced Detection Performance: Detection accuracy, precision, and recall were also significantly improved by the Bayes filter, reaching 97.69%, 95.65%, and 97.77%, respectively.
  • State-of-the-Art Comparison: When compared to other state-of-the-art methods using a different dataset and occupancy level definition, the proposed Bayes filter still demonstrated superior performance in both estimation accuracy (77.29%) and detection accuracy (97.29%), often outperforming methods that used more environmental parameters than just CO2.
  • Time-Delay Reduction: The Bayes filter effectively reduced time delays in occupancy detection, especially noticeable around midnight when ventilation systems are off, by leveraging historical occupancy data.

Limitations and Future Work

  • Data Dependency: The performance of the Bayes filter is contingent on the accuracy of the estimated transition probability matrix and the observation model. The authors note that more historical occupancy data would improve the transition probability estimation.
  • Future Data Acquisition: If additional real occupancy data is scarce, the authors suggest using generative adversarial networks (GANs) to simulate data for enhancing the transition probability matrix.
  • Observation Model Improvement: While the ensemble ELM is effective, further research could focus on developing even better observation models, potentially using semi-supervised learning techniques given the ease of collecting CO2 data.
  • Generalizability: The paper demonstrates effectiveness in an office setting. Future work could explore its applicability to other building types and occupancy scenarios.

Summary updated

BibTeX
@inproceedings{jiang2019bayesian,
  author = {Chaoyang Jiang and Zhenghua Chen and Rong Su and Mustafa K. Masood and Yeng Chai Soh},
  title = {{Bayesian filtering for building occupancy estimation from carbon dioxide concentration}},
  booktitle = {Energy},
  year = {2019},
  doi = {10.1016/j.enbuild.2019.109566},
}
Figure