Occupancy State Prediction by Recurrent Neural Network (LSTM): Multi-Room Context

Area of Science:

• Building Science
• Artificial Intelligence
• Environmental Monitoring

Background:

• Building energy consumption is heavily influenced by occupant behavior and occupancy states.
• Accurate monitoring of occupant habits is crucial for energy efficiency, but non-intrusive methods are preferred to avoid disrupting natural behaviors.

Purpose of the Study:

• To develop a non-intrusive model for ascertaining building occupancy states.
• To utilize environmental data and artificial neural networks (ANNs) for predicting occupant behavior patterns.

Main Methods:

• Employed artificial neural networks (ANNs), specifically a long short-term memory (LSTM) network.
• Utilized non-intrusive sensors to collect environmental data, including CO2 concentration and noise levels.
• Implemented rule-based a priori labeling for training the predictive model.

Main Results:

• The model successfully predicts four distinct occupancy states in a residential building.
• Achieved promising prediction accuracy ranging from 78% to 92%, despite lacking direct occupancy data.
• Demonstrated the feasibility of using environmental data for inferring occupant presence and behavior.

Conclusions:

• Non-intrusive environmental sensing combined with ANNs offers a viable approach to monitoring building occupancy states.
• The developed LSTM model shows significant potential for improving building energy management through behavior-based predictions.
• Further research can refine the model for enhanced accuracy and broader application in smart building systems.