Deep Learning Based Timeseries Modeling For Room Occupancy Estimation

Sammanfattning

Accurate estimation of occupant presence in indoor environments is crucial for optimizing energy use, improving comfort, and ensuring safety. Traditional deep learning approaches often struggle to capture the complex and long-range temporal dependencies inherent in occupancy data. To address this challenge, we investigate a transformer-based model for analyzing time series data from non-intrusive environmental sensors, including light,sound, temperature, and CO2 concentration. Using the self-attention mechanism, the transformer architecture effectively models temporal patterns and outperforms conventional CNN and LSTM baselines. We systematically evaluate model performance across different sensor combinations, temporal feature configurations, and windowsizes. A key finding is that a minimal sensor setup using light and sound achieved superior performance compared to larger sensor arrays, demonstrating that strategic feature selection can enhance system accuracy while reducing complexity. The transformer model achieved the highest overall accuracy of 98.8 percent, confirming its robustness for occupancy estimation tasks. These results highlight the potential of transformer architectures to enable more intelligent, efficient, and scalable building management systems through accurate occupancy estimation.