PresenceSense: Zero-training Algorithm for Individual Presence Detection based on Power Monitoring

Ming Jin, Ruoxi Jia, Zhaoyi Kang, Ioannis C. Konstantakopoulos, Costas J. Spanos

Abstract

Non-intrusive presence detection of individuals in commercial buildings is much easier to implement than intrusive methods such as passive infrared, acoustic sensors, and camera. Individual power consumption, while providing useful feedback and motivation for energy saving, can be used as a valuable source for presence detection. We conduct pilot experiments in an office setting to collect individual presence data by ultrasonic sensors, acceleration sensors, and WiFi access points, in addition to the individual power monitoring data. PresenceSense (PS), a semi-supervised learning algorithm based on power measurement that trains itself with only unlabeled data, is proposed, analyzed and evaluated in the study. Without any labeling efforts, which are usually tedious and time consuming, PresenceSense outperforms popular models whose parameters are optimized over a large training set. The results are interpreted and potential applications of PresenceSense on other data sources are discussed. The significance of this study attaches to space security, occupancy behavior modeling, and energy saving of plug loads.

Introduction

Optimizing HVAC and lighting systems in commercial buildings requires accurate, granular occupancy data. However, traditional sensing methods face significant hurdles. Passive Infrared (PIR) sensors are often too coarse to detect stationary occupants, while cameras and acoustic sensors raise privacy concerns. Furthermore, high-performing occupancy models typically require extensive "ground truth" labels—manual logs of arrival and departure times—which are labor-intensive and difficult to scale. PresenceSense (PS) addresses these challenges by leveraging existing "smart plug" infrastructure to infer presence from power consumption data without requiring manual training labels.

Methods

PresenceSense treats occupancy detection as a semi-supervised learning problem using a "zero-training" approach.

Feature Engineering

The algorithm analyzes power data through three distinct "views":

  1. Power Level: Baseline demand (e.g., a charging laptop).
  2. Edge Effects: Sudden changes in power signaling state transitions (e.g., turning on a monitor).
  3. Rippling Effects: High-frequency fluctuations indicative of active device usage, captured via metrics like Standard Deviation (SD) and Mean Absolute Difference (MAD).

The PresenceSense Algorithm

PS utilizes a variation of co-training. It begins with a rough "prior" based on a typical work schedule. Multiple classifiers are trained on different views of the unlabeled data using this prior. These classifiers then iteratively "vote" on labels, refining the model's understanding of the user's specific patterns. To prevent the model from learning noise, an early stopping rule based on the Probably Approximately Correct (PAC) learning framework is applied.

Experimental Setup

The study was conducted at the UC Berkeley CREST graduate office testbed. Primary data consisted of individual plug-load measurements (desktops, monitors, lamps) collected via ACme sensors at 1-second resolution. To establish ground truth and provide benchmarks, the researchers deployed additional hardware.

Experimental setup showing the ultrasonic sensor and chair-mounted acceleration sensor used for ground truth collection.

Ground truth was established through 5-minute resolution logs from four users, supplemented by ultrasonic sensors measuring distance to the user and acceleration sensors attached to office chairs.

Results and Discussion

PresenceSense achieved high accuracy, ranging from 87% to 93%. Notably, this zero-training algorithm outperformed traditional supervised models that were optimized using full training sets.

The Significance of Rippling Effects

A key insight of the study is that simple power thresholds are often insufficient because many devices remain powered on when a user is absent. However, the "rippling effects"—the fluctuations in power consumption—are highly correlated with active human presence.

Observation of rippling effects in power consumption during periods of presence vs. absence.

The algorithm proved robust to inaccurate initial priors, as the iterative voting process allowed it to converge toward true occupancy patterns. While dedicated hardware like ultrasonic sensors achieved slightly higher accuracy (93.7%), PS offers a compelling software-only alternative for environments where power monitoring is already present. In contrast, WiFi proximity logs were found to be poor indicators of desk presence (47% accuracy), as mobile devices often remain connected to access points even when the user has left the immediate area.

Conclusion

PresenceSense demonstrates that individual occupancy can be accurately detected using non-intrusive power monitoring without the need for manual labeling. While the system currently relies on active device usage, future work could explore integrating PS with other data sources or social incentives to further enhance building energy efficiency and occupancy modeling.

Summary updated

Abstract

Non-intrusive presence detection of individuals in commercial buildings is much easier to implement than intrusive methods such as passive infrared, acoustic sensors, and camera. Individual power consumption, while providing useful feedback and motivation for energy saving, can be used as a valuable source for presence detection. We conduct pilot experiments in an office setting to collect individual presence data by ultrasonic sensors, acceleration sensors, and WiFi access points, in addition to the individual power monitoring data. PresenceSense (PS), a semi-supervised learning algorithm based on power measurement that trains itself with only unlabeled data, is proposed, analyzed and evaluated in the study. Without any labeling efforts, which are usually tedious and time consuming, PresenceSense outperforms popular models whose parameters are optimized over a large training set. The results are interpreted and potential applications of PresenceSense on other data sources are discussed. The significance of this study attaches to space security, occupancy behavior modeling, and energy saving of plug loads.
BibTeX
@inproceedings{jin2014presencesense,
  author = {Ming Jin and Ruoxi Jia and Zhaoyi Kang and Ioannis C. Konstantakopoulos and Costas J. Spanos},
  title = {{PresenceSense: Zero-training Algorithm for Individual Presence Detection based on Power Monitoring}},
  booktitle = {BuildSys},
  year = {2014},
  doi = {10.1145/2674061.2674073},
}
Figure