TL;DR
This paper tackles the challenge of collecting sufficient labeled data for occupancy prediction in smart buildings by introducing a "weakly supervised" framework. It leverages an interactive learning approach to gather high-quality, albeit limited, training data, which is then used to train models for predicting both discrete occupancy states and continuous occupant counts. This method offers a privacy-preserving and cost-effective alternative to traditional data collection for occupancy prediction.
What Problem Does This Paper Address?
Accurate occupancy prediction is crucial for optimizing energy consumption in buildings, particularly for HVAC and lighting systems. However, a significant hurdle in developing effective occupancy prediction models is the difficulty and cost associated with collecting large, labeled datasets. Traditional methods often rely on extensive sensor networks or manual labeling, which can be intrusive, time-consuming, and expensive. This paper addresses this gap by proposing a novel framework that uses interactive learning to collect a smaller, yet high-quality, dataset, enabling weakly supervised occupancy prediction. This approach aims to overcome the data scarcity problem while respecting occupant privacy.
Methods
The core of the proposed methodology lies in its two-stage process:
- Data Preparation (Occupancy Estimation): This stage focuses on generating reliable occupancy labels. It starts with feature selection from various sensor readings (CO2, motion, power, humidity, temperature, acoustic pressure, door position). Then, an interactive learning (IL) approach is employed. IL involves a user interacting with the system to provide labels, but in a way that minimizes user effort and maximizes data usefulness. The paper references a previously developed IL method that uses a "spread rate" concept to determine when to ask for user input, ensuring data quality and reducing the number of interactions needed. This process generates weakly labeled data.
- Main Pipeline (Occupancy Prediction): Once the weakly labeled data is prepared, it's fed into various machine learning models for occupancy prediction. The paper explores two main prediction tasks:
- Discrete Occupancy Prediction: This involves predicting distinct states: absence of occupants, presence of one occupant, and presence of more than one occupant.
- Continuous Occupancy Prediction: This aims to predict the exact number of occupants.
Several machine learning algorithms were evaluated, including:
- Multi-Layer Perceptron (MLP)
- Recurrent Neural Networks (RNNs), specifically LSTM, GRU, and bi-LSTM.
- LightGBM, a gradient boosting decision tree algorithm, which is highlighted as a key model due to its efficiency and scalability.
The study also incorporates additional features beyond raw sensor data, such as time-based features (hour, day, month) and calendar features (office events, meetings), to enhance prediction accuracy.
Datasets
The study utilizes data collected from a real-world case study: an office at the Grenoble Institute of Technology accommodating a professor and three Ph.D. students. This office is equipped with various sensors monitoring:
- Temperature (indoor and outdoor)
- Humidity
- Motion detection
- CO2 concentration
- Acoustic pressure
- Power consumption
- Door position
The data was collected over a period of one year, with raw sensor readings sampled every 5 minutes. For prediction tasks, this data was often resampled to an hourly rate. The paper does not introduce a new benchmark dataset but rather uses data from a specific, real-world office environment.
Key Insights and Ideas
- Weakly Supervised Learning for Occupancy: The core idea of using a limited, high-quality, interactively labeled dataset for occupancy prediction is a significant contribution, addressing the practical challenges of data acquisition.
- Interactive Learning's Value: The paper demonstrates that interactive learning can effectively generate reliable occupancy labels with minimal user intervention, making it a viable approach for privacy-conscious environments.
- LightGBM's Performance: LightGBM emerges as a strong performer, particularly for short-term predictions and when dealing with limited datasets, often outperforming more complex deep learning models like LSTMs.
- Impact of Additional Features: Incorporating calendar information and time-based features significantly improves prediction accuracy, highlighting the importance of contextual data beyond raw sensor readings.
- Discrete vs. Continuous Prediction Challenges: The study reveals that predicting discrete occupancy states (especially binary absence/presence) is generally more accurate than predicting the exact continuous number of occupants, likely due to the inherent noisiness and complexity of continuous values.
Key Results
- Discrete Prediction:
- For binary occupancy prediction (absence/presence), the LightGBM model achieved an overall precision of 94% for 1-hour ahead predictions.
- For multi-level occupancy prediction (absence, 1 occupant, >1 occupant), the model could distinguish absence from presence effectively, but struggled with differentiating between one and more than one occupant, with F1-scores around 50% for the latter categories in 24-hour predictions.
- Continuous Prediction:
- Random Forest outperformed Linear Regression and Gradient Boosting for continuous occupancy estimation, achieving an R2 score of 0.91.
- For continuous occupancy prediction, LightGBM also showed superior performance compared to RNN-based models (LSTM, GRU) and MLPs, achieving an R2 score of 0.502 for 24-hour ahead predictions.
- Feature Impact: Adding calendar and time features improved LightGBM's prediction performance.
Limitations and Future Work
- Data Imbalance: The study acknowledges data imbalance issues, particularly in binary prediction where the "absence" class dominates, requiring techniques like class weighting to mitigate bias.
- Continuous Prediction Accuracy: While Random Forest performed well in continuous estimation, the prediction accuracy for continuous occupancy, especially for longer horizons (12h/24h), was less satisfactory compared to discrete predictions. This is attributed to the inherent fluctuation in continuous data.
- Future Work Suggestions:
- Investigate the association between occupancy prediction and load forecasting.
- Explore the use of future load prediction as an additional feature.
- Further research into using an interactive approach directly in the prediction phase.
- The authors provide a GitHub repository for the source code to ensure reproducibility and facilitate future research.
Summary updated