Weakly Supervised Occupancy Prediction Using Training Data Collected via Interactive Learning.

Area of Science:

• Building Science
• Artificial Intelligence
• Machine Learning

Background:

• Accurate occupancy prediction is crucial for smart building energy efficiency.
• Collecting sufficient historical occupancy data for training prediction models is a significant challenge.

Purpose of the Study:

• To propose a weakly supervised occupancy prediction framework using interactive learning for data collection.
• To predict discrete occupancy states (absence, single occupant, multiple occupants) and continuous occupant numbers.
• To evaluate the performance of the proposed framework against traditional methods.

Main Methods:

• Utilized an interactive learning approach to estimate and collect occupancy data.
• Developed a weakly supervised prediction framework incorporating sensor readings and interactive estimations.
• Trained and compared various machine learning algorithms, including LightGBM and recurrent neural networks, on the collected dataset.

Main Results:

• The interactive learning approach effectively facilitated the collection of necessary training data.
• LightGBM demonstrated superior performance for short-term occupancy prediction compared to recurrent neural networks, especially with limited data.
• For a 24-hour forecast, LightGBM improved prediction accuracy from 38% to 50% for non-aggregated, single-office data.

Conclusions:

• The proposed interactive learning-based framework offers a viable solution for data collection challenges in occupancy prediction.
• Weakly supervised learning combined with LightGBM provides an effective strategy for accurate short-term occupancy forecasting in smart buildings.
• This approach enhances the potential for improved energy management systems through reliable occupancy prediction.