Predicting Indoor PM2.5 Levels Using Deep Learning for Enhanced Digital Twin Applications

Abstract: Monitoring and predicting indoor PM2.5 levels is critical for ensuring healthy indoor environments. However, the integration of real-time data acquisition and instant PM2.5 forecasting within digital twin systems remains underdeveloped in practical applications. This study presents a CNN-BiLSTM hybrid model for forecasting indoor PM2.5 concentrations over a 72-hour horizon. In a real-world case study, the model achieved a root mean square error (RMSE) of 4.884 μg/m3 and a Mean Absolute Error (MAE) of 4.092 μg/m3. The integration of the model into a digital twin platform demonstrates its potential to enhance indoor air quality management through real-time, data-driven interventions.