The Influence of Outdoor Conditions on Indoor Air Quality in Norwegian Schools

Abstract: This study examines the influence of indoor and outdoor environmental conditions on indoor particulate matter (PM10) levels in Norwegian school classrooms, aiming to identify significant factors affecting indoor air quality (IAQ) and develop predictive machine learning (ML) models. Data were collected from sensors in classrooms and nearby weather stations, capturing parameters such as CO₂, temperature, relative humidity (RH%), and PM levels. Random Forest (RF) models were employed to predict indoor PM10 levels, with SHAP analysis used to interpret the relative importance of input parameters. Results showed that outdoor temperature and PM2.5 were the most influential factors in winter, while indoor RH% and CO₂ were more significant during other seasons, reflecting the combined effects of environmental and human activities. RF models achieved high prediction accuracy (up to 90% of R² values), offering ML feasibility to be implemented as predictive tool. Practical recommendations include enhancing outdoor air filtration during winter and optimizing indoor humidity control year-round. This article is limited in specific restricted sample sizes and geographical focus, future research should explore broader datasets and implement real-time IAQ interventions to validate these findings.

Keywords: Indoor Environment, Outdoor Conditions, Machine Learning, Particulate Matters, Random Forest, Feature Importance

Background

Poor indoor air quality (IAQ) in classrooms can lead to respiratory issues, allergic reactions, and other health problems, making optimal IAQ maintenance crucial [1, 2]. IAQ now can be effectively monitored using sensors that measure air temperature, carbon dioxide (CO₂), volatile organic compounds (VOCs), and relative humidity (RH%) [1]. Particulate matter (PM) is a significant component of indoor pollution, categorized by particle size into PM10, PM2.5, and PM1.0, with fine and ultra-fine particles posing the most severe health risks due to their ability to penetrate deep into the lungs and even cellular membranes. Meanwhile, outdoor pollutants can infiltrate buildings through ventilation systems, significantly influencing indoor PM levels [2]. Indoor climate conditions, such as RH and human activities, also affect indoor PM levels [3-5]. Fortunately, machine learning (ML) has proven effective in IAQ prediction, ML method is advantageous for predicting indoor PM levels, with promising applications, such as using Artificial Neural Networks (ANN) to predict CO₂ and indoor PM data [6] and using Random Forest (RF) algorithm [7] in schools and university already.

This research addresses a gap in the study of environmental parameters relationship by specifically examining PM levels in school classrooms, using ML as predictive tools. This study aims to investigate the influence of both indoor and outdoor conditions on indoor PM values, uncovering the causes behind their significant impact on IAQ. Additionally, it seeks to develop a ML model for predicting indoor PM values, contributing to the body of knowledge on effective IAQ management and offering actionable insights for improving air quality in schools.

Methods

The indoor data for the analysis was sourced from the free-standing sensors (N3smart) installed within the selected rooms. Outdoor conditions mean outdoor data originated from various weather stations located near the schools provided publicly by Norwegian Center for Climate Services. Specifically, the outdoor PM10 and PM2.5 values were gathered publicly from weather stations affiliated with Norwegian Institute for Air Research.

The sensors that collected data inside the schools were provided by N3. These sensors gather data concerning CO₂, air temperature, RH%, and PM10. These sensors are positioned in the middle of the classrooms on the wall opposite the door. Several different schools in Norway were chosen for the analysis, partly due to their proximity to weather stations. Ventilation system in schools mostly apply ePM1.0 intake air filter (efficiency 85%). Therefore, the indoor PM10 level read by the sensors is likely dominated by fine particles. Various weather stations were used to source data for analysis to support availability of outdoor data and outdoor PM levels. The selected schools, rooms, weather station and PM level data source are presented in Table 1.

Table 1. Selected schools and rooms.

In this study, supervised regression ML method with RF algorithm was utilized to predict indoor PM10 values based on eight parameters. Recorded indoor PM10 values were used as training data for the model. These input-target parameters for ML were employed as shown in Table 2. Additionally, decision tree and linear regression models were employed, but RF algorithm results consistently yielded more accurate and were therefore used for the article. Later, Sharpley Additive Explanation (SHAP) analysis will contribute to explain the significance of influence by all input parameters through variation of mean-relative importance coefficients.

Table 2. Input & Target for ML Modelling.

Results & Discussions

The outdoor and indoor PM10 values data are presented for all schools in January and September in Table 3. Regarding Percentile.75% (P.75) and Percentile.95% (P.95), indoor PM10 in all schools are considered exceeding 15 µg/m³ set by national health guidelines for a few moments [8]. criteria for zoning in the planning of activities or construction. September exhibits lower values of PM10 than January. However, there is a noticeable increase of PM levels for all schools between the 6th and 12th of September where Sahara dust intrusion came to Norway [9], occurred high correlation between indoor – outdoor PM, as presented in Figure 1. The correlation is calculated for the entire day and separately for workdays when ventilation is active.

Table 3. Descriptive Statistic of PM levels (µg/m³) in all schools – year: 2023.

Figure 1. Pearson correlation coefficient between indoor PM10 levels with indoor - outdoor parameters.

Figure 2 shows relative importance by SHAP analysis had the largest impact on the indoor PM10 values for January 2023. The most influential feature in January was the outdoor air temperature, closely followed by outdoor PM2.5 values and indoor relative humidity. In January, the most influential feature was outdoor air temperature, closely followed by outdoor PM2.5 values and indoor relative humidity. The outdoor air temperature had the biggest impact on the indoor PM10 values followed by the PM2.5 values. These two parameters can be related. The low temperatures make the pollutants accumulate at lower altitudes. When higher outdoor PM2.5 concentration occurs, they will naturally have a higher impact on the indoor values by infiltrating ventilation system through indoors. Indoor air temperature also exhibited significance, while indoor CO₂ values, outdoor PM10, and relative humidity values played minor roles.

Figure 2. Accuracies of supervised ML models with RF algorithm in all schools with mean relative importance using SHAP analysis by all input parameters: using Data in January 2023.

Impact on the PM10 values has different patterns for September 2023, as shown in Figure 3. It can be observed that indoor relative humidity had the most significant overall impact, with some variations across different rooms. The CO₂ values exhibit a strong influence on indoor PM10 values in September. This influence is likely due to consistent student activity, which remains constant regardless of the seasons. Besides, outdoor PM2.5 also played a major role, while indoor RH% and air temperature had supportive influence.

Figure 3. Accuracies of supervised ML models with RF algorithm in all schools with mean relative importance using SHAP analysis by all input parameters: using Data in September 2023.

Conclusion

This study investigated the influence of indoor and outdoor environmental conditions on indoor PM10 levels in Norwegian school classrooms, aiming to understand the significant factors by applying ML method. The findings revealed that outdoor air temperature and particulate matters (PM2.5) significantly impact indoor PM levels in winter (January), while indoor RH% and CO₂ play a dominant role during other seasons (September), influenced by seasonal and occupants’ activity-related variations. If the ventilation system runs routinely, ML models based on indoor-outdoor data, especially by Random Forest algorithm predict PM10 levels effectively with reliable accuracy supported by SHAP analysis, which enhanced interpretability by quantifying parameter significance. These insights provide actionable recommendations for managing IAQ in schools, such as prioritizing air filtration during winter and controlling indoor humidity levels during warmer months. The study focused on a limited number of classrooms and regions, which may limit generalizability; future research should expand to include diverse schools and optimize the ML prediction results.

References

1.      Grimsrud, D., B. Bridges, and R. Schulte, Continuous measurements of air quality parameters in schools. Building research and information, 2006. 34(5): p. 447-458.

2.      Braniš, M., P. Řezáčová, and M. Domasová, The effect of outdoor air and indoor human activity on mass concentrations of PM10, PM2.5, and PM1 in a classroom. Environmental research, 2005. 99(2): p. 143-149.

3.      Chao, C.Y., T.C. Tung, and J. Burnett, Influence of different indoor activities on the indoor particulate levels in residential buildings. Indoor and Built Environment, 1998. 7(2): p. 110-121.

4.      Jeong, S.-G., et al., Application of pre-filter system for reducing indoor PM2.5 concentrations under different relative humidity levels. Building and Environment, 2021. 192: p. 107631.

5.      Thatcher, T.L. and D.W. Layton. Deposition, resuspension, and penetration of particles within a residence. Atmospheric environment, 1995. 29(13): p. 1487-1497.

6.      Hatta, M. and H. Han. Predicting indoor PM2.5/PM10 concentrations using simplified neural network models. Journal of Mechanical Science and Technology, 2021. 35: p. 3249-3257.

7.      Yuchi, W., et al., Evaluation of random forest regression and multiple linear regression for predicting indoor fine particulate matter concentrations in a highly polluted city. Environmental pollution, 2019. 245: p. 746-753.

8.      Folkehelseinstitutt, N., Anbefalte faglige normer for inneklima, Revisjon av kunnskapsgrunnlag og normer - 2015. 2015: Oslo. p. 147.

9.      Cogorno, I.R., Sand fra Sahara – men ingen sommer i sikte, in avisa OSLO. 2023.