Monitoring Air Quality and Estimation of Personal Exposure to Particulate Matter Using an Indoor Model and Artiﬁcial Neural Network

: Exposure to particulate materials (PM) is known to cause respiratory and cardiovascular diseases. Respirable particles generated in closed spaces, such as underground parking garages (UPGs), have been reported to be a potential threat to respiratory health. This study reports the concentration of pollutants (PM, TVOC, CO) in UPGs under various operating conditions of heating, ventilation and air-conditioning (HVAC) systems using a real-time monitoring system with a prototype made up of integrated sensors. In addition, prediction of the PM concentration was implemented using modeling from vehicle tra ﬃ c volumes and an artiﬁcial neural network (ANN), based on environmental factors. The predicted PM concentrations were compared with the level acquired from the real-time monitoring. The measured PM 10 concentrations of UPGs were higher than the modeled PM 10 due to short-term sources induced by vehicles. The average inhalable and respirable dosage for adult was calculated for the evaluation of health e ﬀ ects. The ANN predicted PM concentration showed a close correlation with measurements resulting in R 2 ranging from 0.69 to 0.87. This study demonstrates the feasibility of the use of the air quality monitoring system for personal-exposure to vehicle-induced pollutant in UPGs and the potential application of modeling and ANN for the evaluation of the indoor air quality. H.-J.O. and J.K.; methodology, H.-J.O. and J.K; investigation, H.-J.O.; data curation, H.-J.O. and J.K.; modelling, H.-J.O. and J.K.; writing—original draft preparation, H.-J.O.; writing—review and editing, H.-J.O. and J.K. All authors


Introduction
The development of underground parking garages (UPGs) has become widespread in cities due to the increasing number of automobiles [1][2][3][4][5]. Nowadays, newly developed commercial and/or residential areas are mostly designed with UPGs equipped with ventilation systems (including natural ventilation); however, the air quality in UPGs has become an issue due to the impact of the volume of traffic on indoor air quality [1,2,[5][6][7]. Pollutants such as carbon monoxide (CO), carbon dioxide (CO 2 ), nitrogen dioxide (NO 2 ), particulate matter (PM), and volatile carbon (quantified as total volatile carbon; TVOC) are released from automobiles in these parking garages [8][9][10][11].
Indoor air quality is affected by a constantly changing interaction of indoor and outdoor factors. Pollutants generated from automobiles may accumulate in UPGs, and/or penetrate into the indoor environment [1,8,12]. Particulate materials (PM 10 , PM 2.5 ) within buildings are mostly derived from outdoor particles drawn in through HVAC systems [1,[13][14][15][16][17][18]. If the ventilation system is

Monitoring Locations
This study was conducted in spring ( Figure 1a) were approximately 15-20 years old and located in a commercial area close to heavy-traffic roads (4000-9000 cars/day): 250-500 cars/day for UPG1 and 200-450 cars/day for UPG2. Figure 1b shows the number of automobiles based on business-related trips obtained from the Korea Transportation Safety Authority, which was used to determine future background traffic volumes for the modelling in this study. The characteristics of the different locations selected for this study are presented in Table 1.

Monitoring Locations
This study was conducted in spring ( (Figure 1a) were approximately 15-20 years old and located in a commercial area close to heavy-traffic roads (4000-9000 cars/day): 250-500 cars/day for UPG1 and 200-450 cars/day for UPG2. Figure 1b shows the number of automobiles based on business-related trips obtained from the Korea Transportation Safety Authority, which was used to determine future background traffic volumes for the modelling in this study. The characteristics of the different locations selected for this study are presented in Table 1.

Pollutants Measurements and Instrumentation
During the sampling period, the inlets of the indoor sampler were located at about 1.2-1.5 m above ground level at mid-points in the sampled indoor areas. Indoor and outdoor particulate matter (PM10, PM2.5) were determined by Grimm 1.109 OPC (Grimm Technologies Inc., Douglasville, GA) every 5 mins, both indoors and outdoors.
For TVOC and CO, the concentrations in both UPGs and outdoors were determined by Graywolf's portable semi-permanent VOC meter (Graywolf Sensing Solutions, LLC., Shelton, CT, USA) and CO meter (Testo SE & Co., West Chester, PA, USA), respectively. Additionally, temperature (Temp.) and relative humidity (RH) were investigated using a thermo-hygrometer (Testo SE & Co.) in indoor and outdoor locations, respectively.
The integrated sensor platform consisted of sensor modules (see Supplementary Materials Figure  To evaluate the efficiency of pollutant reduction, variations of PM were investigated as a function of operation conditions based on the air volume in the HVAC system: NV (natural ventilation), 20% (OP1), 50% (OP2), and 100% (OP3) for operation conditions.

Pollutants Measurements and Instrumentation
During the sampling period, the inlets of the indoor sampler were located at about 1.2-1.5 m above ground level at mid-points in the sampled indoor areas. Indoor and outdoor particulate matter (PM 10 , PM 2.5 ) were determined by Grimm 1.109 OPC (Grimm Technologies Inc., Douglasville, GA, USA) every 5 min, both indoors and outdoors.
For TVOC and CO, the concentrations in both UPGs and outdoors were determined by Graywolf's portable semi-permanent VOC meter (Graywolf Sensing Solutions, LLC., Shelton, CT, USA) and CO meter (Testo SE & Co., West Chester, PA, USA), respectively. Additionally, temperature (Temp.) and relative humidity (RH) were investigated using a thermo-hygrometer (Testo SE & Co.) in indoor and outdoor locations, respectively.
The integrated sensor platform consisted of sensor modules (see Supplementary Materials Figure   To evaluate the efficiency of pollutant reduction, variations of PM were investigated as a function of operation conditions based on the air volume in the HVAC system: NV (natural ventilation), 20% (OP1), 50% (OP2), and 100% (OP3) for operation conditions.

Estimation of Personal Exposure Using an Indoor Model
To estimate the indoor PM 10 concentrations, we used given knowledge of activities (based on indoor pollutant sources and outdoor automobiles) in IE1 (adjacent to UPG1, without a window), and an indoor model from Koutrakis et al. [40].
where C in [µg/m 3 ] is the indoor PM 10 concentration, C out [µg/m 3 ] is the outdoor PM 10 concentration, α [h −1 ] is the air exchange rate, p is the dimensionless penetration efficiency factor, E i [µg/h] is the emission rate for the i th indoor particle source, V [m 3 ] represents the room volume, and D [h −1 ] is the particle deposition factor. The parameters (α, D, p, E, applied in office) for indoor environments were found from the literature [40,41]. We performed preliminary tests of particles in indoor environments based on the indoor particle model [42]. The calculated p and D ranged from 0.8 to 1.0 and from 0.35 to 0.40 h −1 , respectively, and these values were similar to those reported in a past study [41]. We have applied the values measured in office environments: 1.2 h −1 for α, 0.39 h −1 for D, 1.0 for p and 1.1 µg/h for E. For the room volume, each area and height in the indoor environment were used in Equation (1).
To estimate health effects related to the PM exposure, we considered monthly input data: the physical attributes of the building (air exchange rate, penetration efficiency factor, room volume, particle deposition factor) and human activity induced indoor sources (short-term activity, automobile frequency). The PM dosage for adult was considered for output data as below:  [43]. Human exposure to PM could include human factors, especially adult such as physiological (tidal volume, breathing frequency) and behavioral factors (exposure time): 0.5 L/breath for TV, 16 breath/min for f and 540 min (as working time).

Prediction of Indoor PM Concentration Using an ANN
The indoor PM concentration was predicted using an ANN, which consisted of an input layer, multi-layers (including a hidden layer), and an output layer [44]. Layers composed of multiple working neurons were connected to all neurons in a neighboring layer through adaptive synaptic weights, and they summed up all input results received from the previous layer and formed a neuron output according to the transmission function. For the transfer function in the hidden layer, the logistic sigmoid function was used and the Levenberg-Marquart (LM) [44,45] algorithm was applied to the early stopping [46].
Learning is determined by a network to change weights using a backpropagation algorithm through two phases. For the forward phase, the training data set is propagated through the hidden layer and comes out the neural network through the output layer, and the output value is compared to the actual target output value. The error between the output layer and the actual values is calculated and propagated back toward the hidden layer [44,47], while in the backward phase, for the adjustment of weight, the derivative of the network error is fed back to the network and weighted to reduce the error for each iteration, requiring the neural model to produce the desired output. The prediction of PM concentrations was conducted for two days in advance, and the predicted PM concentration for one day in advance was also used as input data for the next day's prediction, together with other prediction data [48].
To reduce the learning error, the forward and reverse propagation of the weight were repeated, and the calculation of the error and the total error were calculated. This method continued until a predetermined threshold error. The errors of the neural network generally continued until convergence was minimized using a root mean squared error (RMSE) [35], prediction accuracy (PA), normalized absolute error (NAE), and the index of agreement (IA) [35]. To verify the accuracy of the data obtained from the ANN model, the coefficient of determination (R 2 ) was calculated; the closer this value is to 1, the higher the accuracy. We used 80% of the input data for the learning algorithm, and the remaining 20% of data were applied to a cross-validation method that attempted to validate the indoor air quality prediction data.
The equations of performance indicators are from Karppinen et al. [49]: where N is the number of samples, O i is the concentrations of indoor PM 10 , P i is the predicted indoor PM (PM 10 , PM 2.5 ) and O is the mean value of measured PM (PM 10 , PM 2.5 ).
In this study, one-way ANOVA was used for the PM concentrations as a function of the investigated locations (i.e., UPG1, UPG2, IE1 and IE2) and ventilation operating (NV, OP1, OP2, OP3). For significant effects, all multiple pairwise comparisons were performed using the Holm-Sidak method with Sigma Plot 2011 (version 12.3, Systat Software Inc., San Jose, CA, USA). Data normality was assessed across all replicates using the Shapiro-Wilk test, and, if needed, a nonparametric Mann-Whitney rank-sum test was used to determine significantly different pairs.

Analysis of Variations in the Measured Particulate Matter
To investigate the variations in PM concentrations (PM 10 , PM 2.5 ) as a function of the effect of the ventilation in an HVAC system, the conditions of ventilation were determined as follows: NV for natural ventilation; 20% of air volume for OP1; 50% of air volume for OP2; 100% air volume for OP3 ( Figure 2).
The ventilation system mostly operated with a limited air volume (OP1: 20% and OP2: 50%), but maximum air volume (OP3: 100%) was applied in the morning and afternoon. Normally, it is operated differently for the underground parking garages' HVAC systems. For the indoor environment, the HVAC system was operated regularly, once (for 1 h) in the morning and once (for 1 h) in the afternoon, every day, whereas the HVAC system in the underground parking garages operated continuously for 2 h at night due to heavy and/or accumulated pollutants.

Analysis of Variations in the Measured Particulate Matter
To investigate the variations in PM concentrations (PM10, PM2.5) as a function of the effect of the ventilation in an HVAC system, the conditions of ventilation were determined as follows: NV for natural ventilation; 20% of air volume for OP1; 50% of air volume for OP2; 100% air volume for OP3 ( Figure 2). The ventilation system mostly operated with a limited air volume (OP1: 20% and OP2: 50%), but maximum air volume (OP3: 100%) was applied in the morning and afternoon. Normally, it is operated differently for the underground parking garages' HVAC systems. For the indoor environment, the HVAC system was operated regularly, once (for 1 h) in the morning and once (for 1 h) in the afternoon, every day, whereas the HVAC system in the underground parking garages operated continuously for 2 h at night due to heavy and/or accumulated pollutants. Table 2 shows the concentrations of PM as a function of ventilation system during spring (SP), summer (SU) and winter (WI) seasons. When the HVAC system was operated at an air volume of 100% in the underground parking garage, the concentrations for PM10 and PM2.5 achieved with natural ventilation (NV, 0% ventilation operation) were 199.5 μg/m 3 (SP) to 386 μg/m 3 (SU) and 71.1 μg/m 3 (SP) to 137.4 μg/m 3 (SU), respectively. When the different ventilations were applied, the PM10 concentrations were 127.0 μg/m 3  Table 2 shows the concentrations of PM as a function of ventilation system during spring (SP), summer (SU) and winter (WI) seasons. When the HVAC system was operated at an air volume of 100% in the underground parking garage, the concentrations for PM 10  in summer, while when the operation rates of the HVAC system were changed, the highest PM 10 concentration was found in the spring season, and for PM 2.5 the value measured during the winter season showed the highest value. Table 3 shows pairwise comparisons of PM 10 concentrations by different seasons. There was a significant difference for PM 10 between SP and WI (p < 0.001) for OP1 operation of the HVAC system, and when OP2 was applied there were significant differences for PM 10 between SP and SU/WI (p < 0.001 for between SP and SU, p = 0.004 for between SP and WI). For OP3, there were also significant differences for PM 10 between SP and SU (p < 0.001) and between SU and WI (p = 0.004). Meanwhile, there was only a significant difference for PM 2.5 between SP and WI with OP1 operation of the HVAC system (p = 0.01). As a function of the operation of the HVAC system, there were significant differences for the concentrations of PM 10 between operations in the HVAC system during spring, summer and winter (except in the case of p = 0.174 between NV and OP1 for SP, and p = 0.878 between OP2 and OP3 for WI).
In Korea, natural ventilation is common. The results in Figure 3 were obtained by measuring the PM concentration during natural ventilation, when the HVAC system was not operating. The concentration of PM 10 measured in the two underground parking garages (UPG1, UPG2) for the four seasons exceeded Korea's regulation (indoors, for multi-use facilities: 100-200 µg/m 3 ).
Sustainability 2020, 12, x FOR PEER REVIEW 7 of 20 concentration was found in the spring season, and for PM2.5 the value measured during the winter season showed the highest value. Table 3 shows pairwise comparisons of PM10 concentrations by different seasons. There was a significant difference for PM10 between SP and WI (p < 0.001) for OP1 operation of the HVAC system, and when OP2 was applied there were significant differences for PM10 between SP and SU/WI (p < 0.001 for between SP and SU, p = 0.004 for between SP and WI). For OP3, there were also significant differences for PM10 between SP and SU (p < 0.001) and between SU and WI (p = 0.004). Meanwhile, there was only a significant difference for PM2.5 between SP and WI with OP1 operation of the HVAC system (p = 0.01).
As a function of the operation of the HVAC system, there were significant differences for the concentrations of PM10 between operations in the HVAC system during spring, summer and winter (except in the case of p = 0.174 between NV and OP1 for SP, and p = 0.878 between OP2 and OP3 for WI). In Korea, natural ventilation is common. The results in Figure 3 were obtained by measuring the PM concentration during natural ventilation, when the HVAC system was not operating. The concentration of PM10 measured in the two underground parking garages (UPG1, UPG2) for the four seasons exceeded Korea's regulation (indoors, for multi-use facilities: 100-200 μg/m 3 ).   Table 4.

Air Quality Monitoring and Management
As shown in Figure 4, the integrated sensor platform for real-time indoor air quality monitoring is a wireless sensor module that can independently detect the concentration of each pollutant and exhibits robustness against signal model issues and fluctuations of the signal patterns received from the analog digital converter (ADC) in the micro-controller unit. The wireless network also includes video channel sensor nodes, which are triggered upon detection of a fall, in order to analyze the fieldof-view.  Table 4.

Air Quality Monitoring and Management
As shown in Figure 4, the integrated sensor platform for real-time indoor air quality monitoring is a wireless sensor module that can independently detect the concentration of each pollutant and exhibits robustness against signal model issues and fluctuations of the signal patterns received from the analog digital converter (ADC) in the micro-controller unit. The wireless network also includes video channel sensor nodes, which are triggered upon detection of a fall, in order to analyze the field-of-view.  Based on the real-time integrated sensor platform (Figure 4), real-time indoor air quality monitoring was conducted with natural ventilation for 400 min in two underground parking garages (UPG1 and UPG2) and two indoor environments (IE1 and IE2) ( Figure 5). For each indoor air quality factor (CO2, PM10, TVOC, CO), the concentrations measured in UPG1, UPG2, IE1, and IE2 were monitored in the computer of an integrated-sensor platform, and measurements using each measurement reference device were monitored in one of the indoor environment locations ( Figure  S2): IE1 and IE2 for UPG1 and UPG2, respectively. As the results of indoor air quality monitoring with natural ventilation using an integrated sensor platform show, the CO2 concentrations over 400 minutes ranged from 556 to 887 ppm (at IE1) and from 778 to 1456 ppm (at UPG1), from 667 to 987 ppm (at IE2), and from 807 to 1322 ppm (at UPG2). The PM10 concentrations ranged from 53.4 to 87.4 μg/m 3 (at IE1), from 155.6 to 245.4 μg/m 3 (at UPG1), from 51.2 to 65.4 μg/m 3 (at IE2), and from 109.3 to 256.3 μg/m 3 (at UPG2).
For the management of air quality, the indoor air quality data were acquired with the integrated sensor unit and transmitted through the main server using the data transmission (S2) system (Supplementary material Figure S3). The main server can consolidate the data related to indoor air quality and transmit all the data being monitored. In addition, the collected data transmitted to the monitoring control room are used to determine whether the factors influencing the air quality, measured through air quality prediction (S3) (Supplementary material Figure S3), fall within the normal regulated and/or controlled range. The integrated module system uses an automatic algorithm to judge a measurement whether the HVAC system is available based on the outdoor air quality. If the values are out of the normal/regulated range, the readout is considered abnormal, and it is judged in real time whether the indoor air quality is normal by judging individual air quality influencing factors. Further, air quality prediction (S3) is performed to determine the inflow path for the factors influencing the air quality that are determined to be abnormal (S4) (Supplementary material Figure S3). Penetration estimation (S4) can be used to divide the air quality factors into particles and gases in the monitoring control room and examine the sources of the factors influencing the air quality, such as particles and gas. Further, penetration inhibition (S5) is a step for suppressing penetration of the factors influencing the air quality through the inflow path identified by penetration estimation (S4) (Supplementary material Figure S3). Each data point collected for the sensor module was linked to the Zigbee and should be linked to the (Machine to Machine) Gateway. The linked data can be collected by a gathering server based on HTTP and JSON standards. However, many sensors were installed at various measurement points for the indoor air quality measuring device (including the sensor), as the indoor and outdoor pollutant inflows are mainly caused by infiltration from the Based on the real-time integrated sensor platform (Figure 4), real-time indoor air quality monitoring was conducted with natural ventilation for 400 min in two underground parking garages (UPG1 and UPG2) and two indoor environments (IE1 and IE2) ( Figure 5). For each indoor air quality factor (CO 2 , PM 10 , TVOC, CO), the concentrations measured in UPG1, UPG2, IE1, and IE2 were monitored in the computer of an integrated-sensor platform, and measurements using each measurement reference device were monitored in one of the indoor environment locations ( Figure S2): IE1 and IE2 for UPG1 and UPG2, respectively. As the results of indoor air quality monitoring with natural ventilation using an integrated sensor platform show, the CO 2 concentrations over 400 minutes ranged from 556 to 887 ppm (at IE1) and from 778 to 1456 ppm (at UPG1), from 667 to 987 ppm (at IE2), and from 807 to 1322 ppm (at UPG2). The PM 10  For the management of air quality, the indoor air quality data were acquired with the integrated sensor unit and transmitted through the main server using the data transmission (S2) system (Supplementary material Figure S3). The main server can consolidate the data related to indoor air quality and transmit all the data being monitored. In addition, the collected data transmitted to the monitoring control room are used to determine whether the factors influencing the air quality, measured through air quality prediction (S3) (Supplementary material Figure S3), fall within the normal regulated and/or controlled range. The integrated module system uses an automatic algorithm to judge a measurement whether the HVAC system is available based on the outdoor air quality. If the values are out of the normal/regulated range, the readout is considered abnormal, and it is judged in real time whether the indoor air quality is normal by judging individual air quality influencing factors. Further, air quality prediction (S3) is performed to determine the inflow path for the factors influencing the air quality that are determined to be abnormal (S4) (Supplementary material Figure S3). Penetration estimation (S4) can be used to divide the air quality factors into particles and gases in the monitoring control room and examine the sources of the factors influencing the air quality, such as particles and gas. Further, penetration inhibition (S5) is a step for suppressing penetration of the factors influencing the air quality through the inflow path identified by penetration estimation (S4) (Supplementary material Figure S3). Each data point collected for the sensor module was linked to the Zigbee and should be linked to the (Machine to Machine) Gateway. The linked data can be collected by a gathering server based on HTTP and JSON standards. However, many sensors were installed at various measurement points for the indoor air quality measuring device (including the sensor), as the indoor and outdoor pollutant inflows are mainly caused by infiltration from the air handling unit (AHU), building interior, and elevator (escalator) in most closed buildings. Thus, sometimes, the data were obtained from a router and monitoring server in the middle of each floor, a suddenly broken corridor, and spaces at which disturbance is suspected.
Sustainability 2020, 12, x FOR PEER REVIEW 10 of 20 air handling unit (AHU), building interior, and elevator (escalator) in most closed buildings. Thus, sometimes, the data were obtained from a router and monitoring server in the middle of each floor, a suddenly broken corridor, and spaces at which disturbance is suspected.  Both TVOC and CO concentrations were monitored in the two underground parking garages (UPG1 and UPG2), and the concentrations of TVOC ranged from 778.3 to 1232.3 ppm and from 556.7 to 897.3 ppm in UPG1 and UPG2, respectively. The CO concentrations ranged from 5.6 to 13.2 ppm and from 7.2 to 10.9 ppm in UPG1 and UPG2, respectively. The concentrations for CO 2 , PM 10 , TVOC, and CO obtained from the integrated sensor platform fit well in the comparison of each measurement reference device ( Figure 5). Table 5 shows the average concentrations for CO 2 , PM 10 , TVOC and CO investigated in UPG1, IPG2 and IE1 and IE2.

Modeled Personal Exposures in an Indoor Office and Underground Parking Garage
As shown in Figure 6, the indoor PM 10 concentrations measured using the integrated sensor platform were compared to the modeled values using Equation (1). The personal PM 10 exposure was estimated by using indoor and outdoor PM 10 Figure 5). Concentrations higher than the modeled concentrations showed a high variability over the day due to the penetrations of PM 10 concentrations from UPG1. Due to the large contributions of the particle sources released from UPG1, exposure to the higher PM 10 concentration of UPG1 was in most cases in a small office where an employee works all day long. The PM 10 concentrations determined in IE1 were well approximated, except for the peak of PM 10 due to an increase in the number of cars in UPG1.
These results show that an indoor environment is influenced by the PM 10 concentrations released in UPG1. Actually, an increase in the number of cars in UPG1 meant an increase of PM 10 concentrations and resulted in higher PM 10 concentrations by pollutant penetration from UPG1.
With the PM 10 concentrations, the concentrations of CO 2 , TVOC, and CO in an indoor environment (IE1) were investigated by the real-time indoor air quality monitoring system, and for HVAC ventilation, the OP2 operation was preferred based on the results of the reduction efficiency of PM 10 (Figure 2).

PM Dosage in Respiratory System
PM doses were calculated in indoors (IE) and UPGs (UPG) for PM10 and PM2.5 (Figure 7). For PM10, DF was calculated as the fraction of inhaled particles that deposit in the entire respiratory system: the head airways (HA), tracheobronchial region (TB), and the alveolar region (AL) and PM2.5. DF was calculated only as the PM deposited in the alveolar region (AL) [50].
For all investigated locations, 85%-92% of PM deposition occurred in the entire respiratory regions while 7.0%-8.5% of PM deposited in the alveolar region.
To evaluate the health effects associated with PM (inhalable and respirable PM), we have applied a model recommended by ICRP 66, which uses the dosage for PM10 and PM2.5 by considering the auto vehicle volume and season. Dosage is a common metric used for exposure assessment for predicting the health effects [51,52]. In this study, the inhalable dosage (based on PM10) and respirable dosage (based on PM2.5) were calculated using the number concentrations of PM obtained by Grimm 1.109 OPC. As shown in Figure. 7., during spring, summer and winter seasons, the calculated dosage ranged from ~ × 10 4 -~ × 10 5 μg (except for PM2.5 for winter, ~ × 10 3 ). There are significant differences for PM10 and PM2.5 between IE and UPG (p < 0.001). However, there is no significant difference for season and auto vehicle volume.

PM Dosage in Respiratory System
PM doses were calculated in indoors (IE) and UPGs (UPG) for PM 10 and PM 2.5 (Figure 7). For PM 10 , DF was calculated as the fraction of inhaled particles that deposit in the entire respiratory system: the head airways (HA), tracheobronchial region (TB), and the alveolar region (AL) and PM 2.5 . DF was calculated only as the PM deposited in the alveolar region (AL) [50].
For all investigated locations, 85-92% of PM deposition occurred in the entire respiratory regions while 7.0-8.5% of PM deposited in the alveolar region.
To evaluate the health effects associated with PM (inhalable and respirable PM), we have applied a model recommended by ICRP 66, which uses the dosage for PM 10 and PM 2.5 by considering the auto vehicle volume and season. Dosage is a common metric used for exposure assessment for predicting the health effects [51,52]. In this study, the inhalable dosage (based on PM 10 ) and respirable dosage (based on PM 2.5 ) were calculated using the number concentrations of PM obtained by Grimm 1.109 OPC.
As shown in Figure 7, during spring, summer and winter seasons, the calculated dosage ranged from~×10 4 -~×10 5 µg (except for PM 2.5 for winter,~×10 3 ). There are significant differences for PM 10 and PM 2.5 between IE and UPG (p < 0.001). However, there is no significant difference for season and auto vehicle volume.

Prediction of Indoor PM Concentration Using an ANN
Based on the input variables and algorithms mentioned above, indoor air quality prediction modeling of indoor environments and underground parking garages was performed as a function of the season (spring, summer, winter) ( Figure S4), and the correlation between measured and predicted values is shown in Figure 8: 0.83 (IE2, the second floor indoors) and 0.87 (the underground parking garage) for spring; 0.79 (IE2, the second floor indoors) and 0.81 (the underground parking garage) for summer; 0.69 (the second floor indoors) and 0.71 (the underground parking garage) for winter. The difference between the predicted value and the measured value depended on the season, outdoor pollution level, weather, and the vehicle traffic/frequency, with different characteristics.
During the winter, it was analyzed that the correlation with the predicted PM was lower (R 2 = 0.69 and 0.71) compared to the other seasons, due to the decrease of automobile volume/frequency and the change of outdoor temperature (not shown in this study), but it showed a very high correlation (R 2 = 0.83 and 0.87) during the yellow dust period in spring or when the outdoor air quality deteriorated, and there were many cars entering the underground parking garage. This can be applied to a prediction program for operating the underground parking ventilation system, and if more diverse input variables and improved algorithms are applied, the accuracy of the actual fine dust prediction results will be improved.
Feedforward backpropagation (FFBP) was evaluated using the minimum root mean (RMSE), maximum prediction accuracy (PA), normalized absolute error (NAE), and index of agreement (IA). The calculated parameters for PM ranged from 0.15 to 0.19 for RMSE, 0.83 to 0.92 for PA, 0.17 to 0.19 for NAE, and 0.88 to 0.92 for IA.

Prediction of Indoor PM Concentration Using an ANN
Based on the input variables and algorithms mentioned above, indoor air quality prediction modeling of indoor environments and underground parking garages was performed as a function of the season (spring, summer, winter) ( Figure S4), and the correlation between measured and predicted values is shown in Figure 8: 0.83 (IE2, the second floor indoors) and 0.87 (the underground parking garage) for spring; 0.79 (IE2, the second floor indoors) and 0.81 (the underground parking garage) for summer; 0.69 (the second floor indoors) and 0.71 (the underground parking garage) for winter. The difference between the predicted value and the measured value depended on the season, outdoor pollution level, weather, and the vehicle traffic/frequency, with different characteristics.
During the winter, it was analyzed that the correlation with the predicted PM was lower (R 2 = 0.69 and 0.71) compared to the other seasons, due to the decrease of automobile volume/frequency and the change of outdoor temperature (not shown in this study), but it showed a very high correlation (R 2 = 0.83 and 0.87) during the yellow dust period in spring or when the outdoor air quality deteriorated, and there were many cars entering the underground parking garage. This can be applied to a prediction program for operating the underground parking ventilation system, and if more diverse input variables and improved algorithms are applied, the accuracy of the actual fine dust prediction results will be improved.
Feedforward backpropagation (FFBP) was evaluated using the minimum root mean (RMSE), maximum prediction accuracy (PA), normalized absolute error (NAE), and index of agreement (IA). The calculated parameters for PM ranged from 0.15 to 0.19 for RMSE, 0.83 to 0.92 for PA, 0.17 to 0.19 for NAE, and 0.88 to 0.92 for IA.

Discussion
This study conducted indoor air quality monitoring (Figures 3 and 5) based on an integrated sensor and conducted an exposure assessment (Figure 6) based on a personal model of human exposure for evaluating human health effects in an indoor environment, rather than a simple monitoring and measurement study. For indoor air quality management while considering the operation of an HVAC system and outdoor air quality, it was possible to conduct indoor air quality monitoring and predictive research using an ANN model. Due to energy issues, most underground parking garages use natural ventilation systems or separate ventilation units [3]. However, the high cost of separate ventilation systems results in relatively sparse monitoring, which provides accurate data but only in a few locations. Moreover,

Discussion
This study conducted indoor air quality monitoring (Figures 3 and 5) based on an integrated sensor and conducted an exposure assessment (Figure 6) based on a personal model of human exposure for evaluating human health effects in an indoor environment, rather than a simple monitoring and measurement study. For indoor air quality management while considering the operation of an HVAC system and outdoor air quality, it was possible to conduct indoor air quality monitoring and predictive research using an ANN model. Due to energy issues, most underground parking garages use natural ventilation systems or separate ventilation units [3]. However, the high cost of separate ventilation systems results in relatively sparse monitoring, which provides accurate data but only in a few locations. Moreover, most air quality monitoring is conducted using static monitoring stations equipped with certificates; thus, it is difficult to control the pollutants released in UPGs and estimate the potential health burden.
By considering factors such as the ventilation rates in HVAC systems, this study established real-time monitoring based on an integrated sensor platform for evaluating personal exposure to pollutants, and could estimate indoor PM concentrations using an ANN by considering the outdoor air quality, weather, automobile frequency, and other factors.
As shown in Figure 3, the concentration of PM in the underground parking garage showed a concentration (max: 336 µg/m 3 ) significantly exceeding the Korean standard value (100-200 µg/m 3 ). It can be seen from Figure 6 that the high concentration of PM 10 affects the indoor PM 10 concentration of the indoor first floor office adjacent to the underground parking garage, so pollutants in the underground parking garage enter/penetrate indoor environments through various pathways, such as the entrance doors and elevators in the building, resulting in adverse of health effects. The comparison of the PM concentrations of the underground parking garage and the indoor office shows that the air quality in both the indoor spaces and underground parking garage could be improved by using the correct HVAC operation (Figure 2). However, if the outdoor air quality deteriorates, it cannot be guaranteed that the operation of the air conditioning system, due to the inflow of external air as well as natural ventilation, could improve the indoor air quality. The outdoor air quality may still deteriorate, and in this case, the operation of the HVAC system must be customized according to the outdoor air quality and, in particular, the air quality management of the underground parking garage.
Prior studies have evaluated the concentrations of VOC in parking garages, demonstrating a strong dependence on automobile sources and the activities of building occupants, even though the VOC concentration does not tend to be significantly correlated with the garage atmosphere [8,13]. Compared to the measured indoor monitoring data, the modeled data of the profiles in Figure 5 show mostly higher personal exposure to PM 10 , which is similar to the conclusions of the penetrations of PM 10 from vehicle-induced sources. Considering the traffic volume in parking garages, the air quality must be controlled due to penetration into indoor environments with occupants. High volumes of automobiles lead to an increase in PM 10 concentrations, and more importantly, the PM 10 concentrations can be changed and/or decreased due to use of an HVAC ventilation system. This real-time air quality monitoring system could lead to the evaluation of the indoor air quality in UPGs associated with personal-exposure vehicle-induced pollutants, based on an integrated sensor platform which uses an M2M wireless sensor network, and we also showed the potential for indoor air quality prediction using an ANN (based on the calculated 0.69-0.87 for R 2 between the measured data and predicted data), based on the variable data such as outdoor PM concentrations, weather, the frequency of automobiles in parking garage, and the developed learning algorithm.
This study is from five years ago and building design (including parking garage) and pollutant emissions from vehicles today could be different. However, an algorithm applied to this study could input different pollutant concentrations in various indoor environments and be further expanded to allow altering the algorithms that differ in weather conditions, outdoor pollutants, and input parameters such as vehicle type, vehicle fuel, and vehicle driving time based on the traffic volume. This study was an attempt to expand the monitoring system, predict pollutants, and evaluate health effects by considering the application of various factors in these algorithms applying an ANN. Although this study involved measurement and monitoring five years ago, the monitoring techniques and health impact assessment based on personal exposure to pollutants presented in this study are useful for future building design and indoor air quality monitoring. As in past studies, the ANN model provided better results than other models, especially ANN models written with non-uniform variance. Thus, the ANN model with the proposed monitoring technique and personal exposure evaluation applied the current ANN. We consider our study as more important for air quality monitoring and will be effectively applicable to all kinds of monitoring and air quality prediction, including of various other components.
Due to the increase in the exhaust emission concentrations in closed areas with insufficient ventilation, commercial underground parking garages in the city are becoming an important issue. The 2013 parking regulation in Korea has determined as 100 m 2 is required per car and underground parking spaces in commercial buildings are recommended to have more spaces than the legally required number of parking spaces. Since the number of automobiles has continuously increase, various car-induced pollutants including PM accumulate in underground parking lots. Especially as underground parking lots rely on natural ventilation in a trapped structure, their pollution degree is more serious due to the low driving speed and idling the cars in these garages. In the underground parking lot regulations, the operation of the air conditioning system and the setting of the ventilation rate are calculated based on indoor and outdoor CO concentration and vehicle idling times. However, for PM, which is one of the pollutants sourced from automobile emissions, the simple concentration rule is applied only to PM 10 concentration, even though it has been the subject of continuous shortand long-term pollutant exposure management.
In this study, we evaluated respirable PM and TVOC concentrations and health effects in underground parking lots, and the PM 2.5 concentration was more than five times the standard value due to continuous vehicle inflow during the night without the ventilation system running. Currently, there is no regulation for PM 2.5 concentration in underground parking garages, so the seriousness of automobile emissions is neglected. Vehicle-induced pollutants are accumulated indoors for a long time, and the resuspension of pollutants due to driving the vehicle results in long-term exposure to the pollutants. Although efforts have been made to improve indoor air quality, such as the operation of an air conditioning system indoors, automobile-emitted carcinogens in underground parking garages could be moved to the indoor office environment along with other pollutants through the inflow path in the building. This has the potential to be detrimental to human health, posing a challenge for designing a sustainable future. More importantly, the lack of data about the quantitative exposure to pollutants based on monitoring and health effects is one of the reasons why the development of regulations for pollutants has been delayed.
The real-time monitored PM used in our study was modeled based on the average daily PM concentration. However, in our future research, we will receive weather data in real time, connect our modeling algorithm program to predict the change, and prepare hourly concentration and prediction programs based on this. In this case, a long-term prediction model will be possible. We will attempt to adapt the algorithm to predict short-term changes such as extreme weather conditions.
To estimate PM concentration associated with health effects, we considered daily input data from 2014 to 2015 for: traffic volume, indoor and outdoor PM concentration, and season. Firstly, we aimed to predict PM concentrations due to temperature and weather events, but we found no relationship of exposure to PM in indoor environments with outdoor temperature. Therefore, we determined that the exposure to PM related to health effects is dependent on season. More monitoring and evaluation may be required for the prediction of PM due to the variation of temperature, weather conditions, and short-term pollutant events. In the future study, we will provide a modified algorithm utilizing conditions used in this study as well as weather conditions in the ANN model. The future study along with current study results will bridge the gap between exposure to pollutants and evaluation of health effects, providing the technology required toward an environmentally sustainable community.
What we need to keep in mind is that simple indoor air quality monitoring cannot contribute to human health risk assessment and indoor air quality improvement, and in order to improve indoor air quality based on monitoring, pollutant exposure evaluations based on health impact assessments must be conducted. Considering these points, this study applied exposure modeling, considering indoor pollutants and outdoor exposure factors, to enable health impact assessment. If applied together with the ANN model, it is possible to predict future indoor air quality, as well as to evaluate health effects. During the development and comparison of the personal exposure model and ANN model, this study focused only on PM 10 concentration. For PM concentration, it is necessary to monitor the concentration of PM 2.5 , because PM 2.5 further affects health. Since the regulation standard for underground parking garages in Korea is defined by the PM 10 concentration, this study monitored the PM 10 concentration and compared the modeling results. In the future, the monitoring of PM 2.5 concentrations, along with other pollutants, will require similar predictions.

Conclusions
A personal exposure model and ANN model were proposed herein that are together capable of predicting indoor PM concentrations in advance. Combined with the personal exposure model for the evaluation of health effects, the core monitoring and modelling system is considered to be an effective tool to improve the control of indoor air quality.
The novel aspect of this approach is the integrated monitoring study based on indoor air pollutant factors, with estimation using the ANN model with the personal model, which together are capable of evaluating the health effects when taking into account indoor sources.
Prediction using an ANN can provide data applications for monitoring systems, to control the operation rate of HVAC systems by considering the concentration of indoor and outdoor pollutants and balancing this with energy saving. It may also provide information on the selection of operation conditions for specific air-conditioning systems in the building.
As an air quality management strategy, buildings should be equipped with a Web-based real-time monitoring system for monitoring the quality of indoor air, including information on pollutants, in order to improve the indoor environment to assure human health and quality of life by providing a safe and environmentally friendly living atmosphere.
Supplementary Materials: The following are available online at http://www.mdpi.com/2071-1050/12/9/3794/s1, Figure S1: Individual sensor module (a) and lab test for indoor pollutants (b); Figure S2: Field test using the integrated sensor platform indoors (a-c) and underground parking garage (d); Figure S3. Building air quality monitoring (S1-S5) using the integrated sensor platform. Left: block diagram, right: diagram for the indoor air quality management system with an M2M wireless sensor network.; Figure S4: Developed prediction program using an Artificial Neural Network (ANN). Initial screen (a), reverse propagation (b), prediction screen (c), and final result (d).