Predicting fine particulate matter levels in Finnish buildings
della Vecchia, Salvatore (2019)
della Vecchia, Salvatore
2019
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2019060414581
https://urn.fi/URN:NBN:fi:amk-2019060414581
Tiivistelmä
Fine particulate matter (PM 2.5 ) is considered one of the most harmful air pollutants. While
a large proportion of the particles is originating from outdoor sources, people are mostly
exposed while indoors. Predicting future trends of PM 2.5 concentrations could help build-
ings owners and operators developing better control strategies, and minimizing delays in
responding to potential indoor air quality (IAQ) issues. Machine Learning and Deep
Learning methods, in particular Long-Short Term Memory Neural Networks (LSTM),
have shown good results in predicting sequential data. In this study, PM 2.5 data from 260
sensors in 119 Finnish buildings were collected during the period 2014/09 - 2019/01.
The mean PM 2.5 concentration observed was 1.01 μg/m 3 (SD2.41 μg/m 3 ). Different meth-
ods were compared to predict from one hour up to 8 hours lead times. Three methods
were tested for short term predictions (+1 hr): Autoregression, Random forest for regres-
sion, and LSTM, while the latter two were tested for long-term predictions (+8 hr). For
short term prediction, all methods used a univariate time series (historical hourly indoor
PM 2.5 average). The best prediction was obtained using LSTM with only one lag variable
(mean absolute error 0.21 μg/m 3 , mean squared error 0.85 μg/m 3 ). For long term predic-
tion, both methods were first tested using a multivariate time series (historical indoor and
outdoor PM 2.5 ). An additional time series containing the outdoor PM 2.5 forecasts for the
next eight hours was then added to the models, which significantly improved the model
accuracy. The lowest Mean Absolute Error (0.49 μg/m 3 )and Mean Squared Error (1.83
μg/m 3 ) were obtained using LSTM with eight lag variables and eight forecasts. In conclu-
sion, long-term predictions are more challenging, but the predictions can be improved by
multivariate methods.
a large proportion of the particles is originating from outdoor sources, people are mostly
exposed while indoors. Predicting future trends of PM 2.5 concentrations could help build-
ings owners and operators developing better control strategies, and minimizing delays in
responding to potential indoor air quality (IAQ) issues. Machine Learning and Deep
Learning methods, in particular Long-Short Term Memory Neural Networks (LSTM),
have shown good results in predicting sequential data. In this study, PM 2.5 data from 260
sensors in 119 Finnish buildings were collected during the period 2014/09 - 2019/01.
The mean PM 2.5 concentration observed was 1.01 μg/m 3 (SD2.41 μg/m 3 ). Different meth-
ods were compared to predict from one hour up to 8 hours lead times. Three methods
were tested for short term predictions (+1 hr): Autoregression, Random forest for regres-
sion, and LSTM, while the latter two were tested for long-term predictions (+8 hr). For
short term prediction, all methods used a univariate time series (historical hourly indoor
PM 2.5 average). The best prediction was obtained using LSTM with only one lag variable
(mean absolute error 0.21 μg/m 3 , mean squared error 0.85 μg/m 3 ). For long term predic-
tion, both methods were first tested using a multivariate time series (historical indoor and
outdoor PM 2.5 ). An additional time series containing the outdoor PM 2.5 forecasts for the
next eight hours was then added to the models, which significantly improved the model
accuracy. The lowest Mean Absolute Error (0.49 μg/m 3 )and Mean Squared Error (1.83
μg/m 3 ) were obtained using LSTM with eight lag variables and eight forecasts. In conclu-
sion, long-term predictions are more challenging, but the predictions can be improved by
multivariate methods.