Predicting PM_2.5 in Well-Mixed Indoor Air for a Large Office Building Using Regression and Artificial Neural Network...

Area of Science:

• Environmental Science
• Building Science
• Data Science

Background:

• Exposure to fine particulate matter (PM2.5) poses significant health risks.
• Indoor PM2.5 monitoring and regulatory standards are underdeveloped, especially for large commercial office buildings.
• Predictive models for indoor PM2.5 are scarce in these settings.

Purpose of the Study:

• To develop and compare statistical models for predicting indoor PM2.5 concentrations in a commercial office building.
• To evaluate the performance of various modeling techniques, including LSTM, MLR, PLS, DLM, and LASSO.
• To assess the models' accuracy with and without ambient PM2.5 data.

Main Methods:

• Statistical modeling approaches were employed to predict hourly indoor PM2.5 concentrations.
• Multiple linear regression (MLR), partial least squares regression (PLS), distributed lag model (DLM), least absolute shrinkage selector operator (LASSO), artificial neural networks (ANN), and long-short term memory (LSTM) were compared.
• Environmental and meteorological parameters were utilized as predictors, with varying combinations tested.

Main Results:

• The Long Short-Term Memory (LSTM) model achieved the lowest root-mean-square error (RMSE) of 1.73 μg/m³ for predicting hourly indoor PM2.5.
• Other models showed higher RMSEs, ranging from 2.20 to 4.71 μg/m³, when using ambient PM2.5 data.
• The developed models demonstrated good predictive skill even when ambient PM2.5 information was excluded.

Conclusions:

• LSTM models significantly outperform traditional statistical and machine learning methods for indoor PM2.5 prediction in office buildings.
• The models effectively capture temporal patterns in predictor variables, leading to improved accuracy.
• Accurate indoor PM2.5 prediction is achievable using environmental and meteorological data, even without direct outdoor PM2.5 measurements.