![Flow chart of this work [CFD: Computational fluid dynamics; LLVM: Low-dimensional linear ventilation model; ANN: Artificial neural network; RBF: Radial basis function; RMSE: Root mean square error] 2.1 CFD simulation and experiment validation A series of CFD simulations were carried out with different ACHs (Air change rates per hour) and source positions to obtain concentration databases. The ReNormalization Group (RNG) k-ε model was employed for turbulence modelling, due to its acceptable performance in terms of accuracy, computing time, and robustness for indoor ventilation (Cao et al., 2017). Fig 3 (a) corresponded to the geometry and source position and Fig 3 (b) demonstrated the mesh used in simulation. The geometry dimension is 3.5 m (X)×3.4 m (Y)×2.5 m (Z), with the total volume of 29.75 m 3 . The ventilation mode applied is up-supply and down-return with all vents in the single side of the wall. Specifically, there were three inlets near the ceiling and three outlets near](https://www.researchgate.net/profile/Shi-Jie-Cao/publication/331562887/figure/fig1/AS:733705274413057@1551940377811/Flow-chart-of-this-work-CFD-Computational-fluid-dynamics-LLVM-Low-dimensional-linear_Q320.jpg)
![Comparison of pollutant concentration distribution obtained from CFD simulation and LVM with the pollutant source position of A (ref. Fig.3) and ACH of 8. [(a): concentration of CFD; (b): the averaging concentration of each divided zone from LVM]](https://www.researchgate.net/profile/Shi-Jie-Cao/publication/331562887/figure/fig4/AS:733705274417156@1551940377950/Comparison-of-pollutant-concentration-distribution-obtained-from-CFD-simulation-and-LVM_Q320.jpg)
![Prediction comparison (input with different Zone numbers and ZLs: blue □, better performance; red ▽,worse performance; solid shape, input locations) and validation (directly from LLVM calculation, *) using LLVM-based ANN under the conditions different pollutant sources [the 1 st column: single source; the 2 nd column: two sources]](https://www.researchgate.net/profile/Shi-Jie-Cao/publication/331562887/figure/fig5/AS:733705274404872@1551940377977/Prediction-comparison-input-with-different-Zone-numbers-and-ZLs-blue-better_Q320.jpg)
Content uploaded by Shi-Jie Cao
Author content
All content in this area was uploaded by Shi-Jie Cao on Mar 07, 2019
Content may be subject to copyright.
Accepted version of SCS journal
Incorporating Online Monitoring Data into Fast Prediction
Models towards the Development of Artificial Intelligent Ventilation
Systems
Jie Ren, Shi-Jie Cao*
Academy of Building Energy Efficiency, School of Civil Engineering, Guangzhou University,
Guangzhou, China, 510006
Corresponding author: shijie.cao@gzhu.edu.cn; shijie.cao09@gmail.com
Abstract
It has proved that LLVM (Low-dimensional Linear Ventilation Models)-based ANN
(artificial neural network) method is able to realize ventilation online control (modes or
airflow rates) based on indoor pollutant response. However, it is challenging to fast
predict indoor pollutant concentration due to the difficulty of identifying pollutant
sources (location and strength). Therefore, we will incorporate monitoring techniques
to modelling that is able to efficiently predict pollutant concentration, aiming for
ventilation online control. A large database was firstly constructed using experiment-
validated CFD (Computational Fluid Dynamics) simulations considering different
ACHs (air change rates per hour) and individual pollutant sources. Next, LLVM method
was utilized to process CFD data, further yielding low-dimensional database for ANN
predictions. We then carried out a series of ANN predictions input with monitored
concentration from different sensor layouts (i.e., positions and numbers). It is found
that well-deployed sensors would provide satisfying inputs for ANN predictions.
Suggestions were also given for the sensors placement that should be located in the
well-mix zone (e.g. outlet region), but avoiding along the same or parallel with the main
flow stream region and near the inlet zone. These findings will further provide strategies
of sensor deployment and move crucial steps forward for ventilation intelligent control.
Keywords:
Ventilation; Intelligent control; Sensor deployment; CFD; Artificial Neural Network
(ANN); Low-dimensional linear ventilation models (LLVM);
Accepted version of SCS journal
1. Introduction
The essential function of ventilation is to remove indoor pollutants and improve
indoor air quality (IAQ) ( Persily, & Emmerich, 2012). Mechanical ventilation (power-
assisted) is popularly employed in the commercial buildings, which consumes a large
amount of energy (Li, et al., 2017). To improve IAQ and building energy efficiency, the
optimal design (Ai, & Mak, 2016; Seppanen, 2008) and control (Mossolly et al., 2009;
Deng et al., 2018) of ventilation systems have been investigated by many researchers.
Since indoor pollutants vary in both time and space (Zorpas, & Skouroupatis, 2016),
the conventional methods are only able to capture the mean energy demand or occupant
needs (Kumar et al., 2016). With a rapid development of sensing technology, great
interests have grown in deploying more sensors into buildings for providing high–
resolution tempo & spatial indoor environment data (e.g., measuring IAQ) to
systematically analyze (Kumar et al., 2016), further used for ventilation regulation and
control to reduce parameters (e.g. CO2 concentration) under an acceptable level (Rackes
et al., 2018; Chenari et al., 2016; Fontanini et al., 2016).
To realize online-autonomously control of ventilation system, fast and accurate
prediction of ventilation performance is necessary, i.e., ventilation assessment based on
indoor pollutant concentration distribution and its associated relationship with
ventilation modes (ventilation rates and modes) (Cao, & Ren, 2018). There is an
important challenge that a ‘faster-than-real-time’ indoor environment prediction is
expected (i.e., indoor pollutant concentration) (Cao, 2019). Considering the cost and
monitor technique of sensors (layout and accuracy), sensing technology is not sufficient
at this point. But it is possible to incorporate monitoring techniques to modelling
prediction methods given the advances of computation power and wireless
communication technology (Shan et al., 2019) etc.. To address this challenge, we will
start with modelling techniques for ventilation application by using CFD (Computation
Fluid Dynamics) simulations.
CFD modelling has been widely applied to predict indoor airflow and pollutant
Accepted version of SCS journal
distribution (e.g., to guide the ventilation design) (Zhang et al., 2013; Zhu et al. 2015;
Deng et al., 2018；Zhou et al., 2016). However, to rapidly respond to the indoor
environment variations for online control, the high precision and huge amount of data
from CFD simulations are not practical (Desta et al., 2004). Thus, Cao and Meyers
proposed a low-dimensional linear ventilation model (LLVM) to process the high-
resolution CFD data into ‘low-dimensional’ levels through volume averaging (Cao, &
Meyers, 2012). Meanwhile, the ‘linear’ property of LLVM allows for the reconstruction
of concentration fields resulting from any type of indoor pollutant source distribution.
Both computation and data storage cost are reduced. A new control strategy was
proposed by using LLVM-based ANN (artificial neuron network), which will largely
speed up the prediction but with rather lower computation cost (Cao, & Ren, 2018).
This method is able to rapidly predict indoor pollutant concentration based on the multi-
inputs, i.e., ventilation modes, ACH (Air Change per Hours) and pollutant sources etc.
Nevertheless, it is difficult to identify real-time source information considering its
location and strength, which is an important input for LLVM-based ANN prediction.
To address this issue, we will integrate techniques of monitoring and modeling to
fast and efficiently predict indoor concentration. Specifically, limited sensors will be
deployed inside the room. Hereby, limited sensors represent the minimal numbers of
sensor deployment whilst keeping the same prediction performance. Literatures has
pointed that in a typical building, placing more sensors in the space, even with optimal
rules based on perfect knowledge of indoor air processes in the office, had almost no
effect at the median level (Rackes et al., 2018). Thus, the layout strategy and monitor
technique of sensors (e.g. placement, numbers) will be one of the main focusses in this
work. Based on limited monitoring data, ANN is employed to connect monitoring and
the aforementioned LLVM methodology to predict indoor pollutant concentration in
terms of ‘low-dimension’ distribution. Data will then be processed and analyzed using
the assessing algorithm to realize intelligent and autonomous ventilation control.
Accepted version of SCS journal
2. Methodology
CFD simulation was employed to construct Low-dimensional Linear Ventilation
Model (LLVM), which has been validated with full-scale experiments in the previous
study (Cao & Ren, 2019), including velocity and CO2 concentration. Artificial neural
network (ANN) were then adopted to predict indoor pollutant concentration providing
certain monitoring concentrations using hypothetical sensors (i.e., volume-averaging
concentration of certain zones or points). Fig 1 shows the flow chart of this work.
Fig. 1 Flow chart of this work [CFD: Computational fluid dynamics; LLVM: Low-dimensional linear
ventilation model; ANN: Artificial neural network; RBF: Radial basis function; RMSE: Root mean
square error]
2.1 CFD simulation and experiment validation
A series of CFD simulations were carried out with different ACHs (Air change
rates per hour) and source positions to obtain concentration databases. The Re-
Normalization Group (RNG) k-ε model was employed for turbulence modelling, due
to its acceptable performance in terms of accuracy, computing time, and robustness for
indoor ventilation (Cao et al., 2017). Fig 3 (a) corresponded to the geometry and source
position and Fig 3 (b) demonstrated the mesh used in simulation. The geometry
dimension is 3.5 m (X)×3.4 m (Y)×2.5 m (Z), with the total volume of 29.75 m3. The
ventilation mode applied is up-supply and down-return with all vents in the single side
of the wall. Specifically, there were three inlets near the ceiling and three outlets near
LLVM calculation
(Low-dimensional Linear Ventilation Model)
CFD concentration simulation
(&experiment validation)
Monitored concentrations of LVM :
(volume-averaged concentration of zones)
Hypothetic sensors
(volume-averaged concentration of points)
database
Step 1
ANN prediction model
(RBF neural network&RMSE evaluation)
Step 2
Step 3
Step 4
Step 5
data processing
construct ANN
optimize layout
LLVM-based ANN
Prediction Model
Optimal Layout
Monitoring
Input
Accepted version of SCS journal
the floor with the same size of 18cm×18cm.
CO2 was considered as the indoor pollutant indicator. There were four different
CO2 source positions (named A, B, C, D) along the plane of Z=1.1m (located in the
breathing region of occupied zone). The ambient background CO2 concentration is 500
ppm and the release intensity of each source was set as 5×10-6 kg/s (Deng et al., 2018).
ACHs (Air change rate per hour) are varying from 4 to 12 with an interval of 2. In
general, a total number of 20 (4 sources×5 ACHs) simulation cases were executed.
Steady state conditions were considered for all the simulations. Temperature
influence was not considered in this work, corresponding to thermal-insulated
ventilation chamber in the experiment. We used uniform velocity profile at the inlet
boundary, which can be calculated from ACH, room volume and inlet size. The
turbulent viscosity ratio and the turbulent intensity were used to define initial turbulent
boundary conditions, respectively set as 10 and 5%. SIMPLE algorithm was adopted to
solve the coupling of pressure-velocity. The second-order upwind discretization scheme
was utilized for momentum, turbulent dissipation rate and turbulent kinetic energy. The
y+ value was converged to be within 5, which was sufficient to resolve the near wall
region applying enhanced wall function in RNG k-ε model. The sensitivity analysis of
grid has been performed with doubled mesh numbers, with the relative errors below 5%
(Cao, & Ren, 2018), and results won’t be presented here.
The full-scale ventilation chamber was used for CFD validations shown in Fig 4.
In the experiment, four positions of CO2 sources (A, B, C and D) under the ACH of 12
was conducted. Dry ice was used to generated CO2, which was kept in a bucket with
thick insulation layers. The x-direction air velocity and CO2 concentration were
monitored by velocity meter and CO2 sensors respectively. The detailed information
and results of the experiment have been elaborated in our previous work (Deng et al.,
2018). In general, results have shown that the CFD prediction of both airflow and CO2
concentration were well accepted when compared with experiment with the maximum
error no more than 30%. And validation won’t repeat here.
Accepted version of SCS journal
Fig 2 (a): Sketch of geometry and CO2 source, specific coordinates (m): A (0.875,2.55,1.1);
B (2.625,2.55,1.1); C (0.875,0.85,1.1); D (2.625,0.85,1.1). (b): mesh used in simulation (c) Layout
of full-scale experimental chamber
2.2 Low-dimensional linear ventilation model (LLVM)
Regarding the low-dimensional linear ventilation model (LLVM), two crucial
characteristics have to be pointed, which are ‘linear’ and ‘low-dimensional’. Low-
dimensional ventilation model (LVM) is a discrete approach to obtain the representation
of grid information of the whole volume Ω, which could largely reduce the grid data in
quantity (Cao & Meyers, 2012). The volume Ω of simulated grid was divided into a
number of cubes with the volume Ωi (i = 1~N, N << cell number), shown in Fig. 5. The
volume-averaged value based on grid data was calculated in these cubes, i.e., in
dimensional distribution. The low-dimensional distribution of CO2 concentration after
discretization were further used to represent the high-resolution ones (obtained directly
from CFD simulation). The concentration data we will use later will exclude the
background concentration.
'
background
vv
C C C
(1)
Where,
'v
C
represents the relative volume-averaged concentration,
v
C
represents the total volume-averaged concentration,
background
C
is the background
concentration which equals to 500ppm as mentioned earlier.
(b)(a)
inlet outlet
CO2
sensor
source A
X
Y
Z
source B
source C source D
velocity meter
dry ice
（c）
Accepted version of SCS journal
Fig. 3. Principle of low-dimensional discrete process (Cao & Ren, 2018).
The ‘linear’ property is simply regarded as the linear superposition of LLVM,
which means the concentration fields resulting from multiple pollutant sources are
equal to the linear superposition of ones separately induced by these individual
pollutant source (Cao & Meyers, 2012). It can be very efficient to estimate indoor
contaminant concentration distributions compared to direct CFD simulation approaches,
and largely reduce computation cost. In the current study, the geometry was divided
into 3 cubes in each axis direction and yielded 27 zones in total, which is sufficiently
fine keeping the discretization error with 10-2. Due to the ‘linear’ property of LLVM
method, we can obtain concentration field of multiple pollutant sources efficiently from
results of individual pollutant source cases, which would save much computational cost,
used to enlarge the database for the training of neural network.
2.3 Radial Basis Function of ANN for prediction
Radial basis function (RBF) network is a particular class of multilayer feed-
forward ANNs, which will be used to predict the whole concentration field due to
several key advantages including: finding the input to output map using local
approximates, rapid learning while requiring fewer examples, good approximation and
learning ability, and easier to train (Liao et al., 2007). RBF neural networks typically
have three layers: an input layer, a hidden layer using non-linear RBF function and a
linear output layer. In our RBF neural network, both input layer and output layer are
concentrations, which represent the monitored concentration and the predicted ones,
respectively corresponding to
'v monitored
C
and
'v predicted
C
.
Accepted version of SCS journal
5 Inputs
Output layer
x1
xn
······
ƒ2
ƒh-1
·····
·
ƒ1
ƒh
y2
yj-1
RBF (Radial Basis Function) ANN
······
Input layer Hidden layer
y1
yj
22 Outputs
Figure 4 model of RBF neural network for concentration field prediction
Before ANN training, a set of databases were calculated through methods of
LLVM. Considering that pollutant sources can be at 4 different locations, there will be
15 different combinations or types. Along with 5 different inlet ACHs, a total number
of 75 sets of samples would be obtained, (15 5, pollutant sources ACHs ). Each
sample contains concentrations of the entire 27 zones, corresponding to a specific
pollutant source type and ACH value. The neural network was then trained by using
samples except for the case with ACH equal to 8. In other words, 60 samples (ACH=4,
6, 10, 12) were selected for training and 15 samples (ACH=8) was used to evaluate the
performance of ANN by using the magnitude of RMSE.
2
11
1( ' '
nz
v ij v predicted ij
ij
RMSE C C
nz
）
(2)
n
—the number of samples used in performance evaluation (equals 15, the number
of source types).
z
—the number of outputs in ANN prediction (for 3, 4 or 5 inputs
prediction models,
z
equals 24, 23 or 22 respectively).
'v ij
C
represents the
volume-averaged concentration of the output zone obtained by LLVM,
'v predicted ij
C
represents the predicted concentration obtained by RBF neural network, the subscript
i
and
j
represent the specific location of the data (sample number and outputs
number respectively).
Accepted version of SCS journal
Next, the test cases for prediction were firstly conducted with certain numbers of
volume-averaged concentrations at input zones (Zone Label, ZL), shown in Fig.5(a).
Cases were considered under the ventilated conditions with single source and two
sources. For the sake of practical application, we aimed to use the sensors as few as
possible, which would represent the most information of the whole volume. The applied
hypothetic sensing numbers of the test cases were corresponding to concentrations at 3
zones, 4 zones and 5 zones. For all cases, two sensing locations were fixed at ZL of 24
near the inlet, and ZL of 6 near the outlet which is the usual-selected sensing location.
We started predictions using concentrations at three different zones, e.g., one near the
inlet (ZL 24), one close to outlet (ZL 6) and one randomly selected. Thus, five column
cases were considered for each input ZL (in Table 1), so in total there were 15 cases.
Figure 5 Schematic description of (a) Zone Labels for monitoring identification; (b) test cases
3 4 5
6, 3, 24 6, 4, 14, 24 6, 2, 15, 26, 24
6, 11, 24 6, 8, 11, 24 6, 3, 18, 20, 24
6, 13, 24 6, 9, 25, 24 6, 5, 17, 18, 24
6, 16, 24 6, 17, 19, 24 6, 7, 17, 22, 24
6, 23, 24 6, 20, 26, 24 6, 9, 14, 17, 24
Zone Labels
for Hypothetic
Sensing Locations
Input Zone Numbers
(b) Test Cases:
(a) Zone Label layout
outlet
inlet
25
19 22
26
20 23
27
21 24
16
10 13
17
11 14
18
12 15
7
148
259
36
X
Y
Z
Zone Label
Accepted version of SCS journal
3. Results
In this section, taking limited monitoring concentrations as inputs, a LLVM-based
neural network model was employed to predict the low-dimensional concentration field.
Section 3.1 described the comparison of high-resolution concentration from CFD and
low-dimensional data using LLVM method. In section 3.2, we demonstrated the ANN
prediction by inputting volume-averaged concentrations from certain low-dimensional
zones as neural network input variables. In section 3.3, the validity of ANN prediction
was testified on the base of the monitored concentrations (from hypothetic sensors at
certain points of the room volume) to better suit the real application.
3.1 LLVM validation using direct CFD simulations
LLVM methodology has been clearly elaborated in the previous work (Cao & Ren,
2018), including discretization accuracy and linear characteristics. It was appreciated
that the discretization accuracy with LLVM should be sufficiently fine for the
engineering applications with acceptable limited error of 10% when the size ratio L/Li
= 3 (
3/1
)/(/ ii
LL
) (Cao & Meyers, 2012). In this work, the whole volume of
simulated geometry is divided into 27 cubes (3 × 3 ×3 cubes), which is sufficiently fine
for such chamber preserving with limited error. Fig. 6 shows the comparison of
pollutant concentration distribution obtained from CFD simulation and LVM (Low-
dimensional Ventilation Models) with the pollutant source position of A (ref. Fig. 3)
and ACH of 8. It can be seen that the LVM result is very analogous to that of CFD
especially in the pespective of engineering application. CO2 concentration of the zone
near the pollutant source was higher compared to other regions. Pollutant concentration
is very depending on the indoor airflow characteristics, e.g., inlet ACH magnitudes
(shown in Fig. 7). It can been noticed that the volume-averaged concentrations of 27
zones show an identical tendency under cases with different inlet ACHs. With the
increasing of the ACH value, the concentration of all zones will decrease but eventually
shows an asymptotic behavior, around 12 in this case (Deng et al., 2018).
Accepted version of SCS journal
Figure 6 Comparison of pollutant concentration distribution obtained from CFD simulation and
LVM with the pollutant source position of A (ref. Fig.3) and ACH of 8. [(a): concentration of CFD;
(b): the averaging concentration of each divided zone from LVM]
Figure 7 Volume-averaged concentration of 27 zones with different inlet ACHs [with the
pollutant source location of A(Fig.3) , ACH=12:□, ACH=10:▽, ACH=8:◇, ACH=6:☆,
ACH=4:O. ]
3.2 LLVM-based ANN prediction using limited monitoring inputs (volume-
averaged concentrations of certain zones/Zone Labels)
In this section, we would evaluate the performance of LLVM-based ANN
prediction based on the limited monitoring inputs from the volume-averaged
concentrations of certain labeled zones (directly calculated from LLVM). For the
current study, three zone-input numbers (3, 4, 5) were selected. From Fig. 5(b), each
zone-input series corresponded to 5 cases and the total number of test cases for single
pollutant source was 15. Considering two sets of pollutant sources (located at positions
of A and A&C respectively), 30 test cases (15 cases×2 source distributions) were
Accepted version of SCS journal
conducted by inputting volume-averaging concentrations of 3 zones, 4 zones and 5
zones (at different locations, two sensing locations were fixed at ZL number of 6 near
the outlet and 24 near the inlet, shown in Fig. 5). The entire low-dimensional
concentration distribution of the chamber will be the output.
After training the data (from LLVM) using RBF neural network, stable prediction
of low-dimensional concentrations was obtained. The performance index (RMSE) of
15 cases was listed in Table 1, with the pollutant source at location A. It can be noticed
that the more monitoring inputs, the better the prediction results, i.e., 5 sensing inputs
series were in general better than smaller inputs with RMSE no more than 10. However,
there may be some larger discrepancies when input numbers decreased to 4 and 3, e.g.
cases with RMSE of 43, 39, 31.
Table 1 Performance index (RMSE) of ANN prediction experiment (Source A)
Input Zone numbers
3
4
5
Performance
RMSE
(Inputs ZL)
ref. Fig. 5
10 (6, 3, 24)
31 (6, 4, 14, 24)
7 (6, 2, 15, 26, 24)
12 (6, 11, 24)
10 (6, 8, 11, 24)
8 (6, 3, 18, 20, 24)
43 (6, 13, 24)
)
9 (6, 9, 25, 24)
6 (6, 5, 17, 18, 24)
13 (6, 16, 24)
)
7 (6, 17, 19, 24)
9 (6, 7, 17, 22, 24)
39 (6, 23, 24)
10 (6, 20, 26, 24)
10 (6, 9, 14, 17, 24)
To further look into this phenomena, we compared the results of the worst
performance and the best performance cases of each zone-input series with exact
concentration predicted directly from LLVM in Fig.8, cases with a single source A and
double sources A&C, corresponding locations of input zones shown at the right side.
The cases with the same zone-input labels showed the similar performance for both
single and double source cases. It is observed that all the cases with 5 monitoring inputs
25
19 22
26
20 23
27
21 24
16
10 13
17
11 14
18
12 15
7
148
259
36
outlet
inlet
25
19 22
26
20 23
27
21 24
16
10 13
17
11 14
18
12 15
7
148
259
36
X
Y
Z
Zone Label
Accepted version of SCS journal
(along the third row of Fig.8) performed well, error within 5%, which is consisted with
smaller RSME in Table 1. When looking at the first two rows (with 3 or 4 sensing
inputs), it can be seen that the layout of the input zones showing poorer performance
have identical characteristics: two or more monitoring zones (input) along or parallel
with the streamwise direction (i.e., the iso-surface of X coordinate). We will discuss
this point later. To sum up, it is possible to use 3 sensors for well prediction.
Figure 8 Prediction comparison (input with different Zone numbers and ZLs: blue □,
better performance; red ▽,worse performance; solid shape, input locations) and validation
(directly from LLVM calculation, *) using LLVM-based ANN under the conditions different
pollutant sources [the 1st column: single source; the 2nd column: two sources]
3.3 LLVM-based ANN prediction using limited monitoring inputs (volume-
averaged concentrations of certain points as Hypothetic Sensors)
From the practical point of view, we now moved to use hypothetic sensors
(concentrations at certain points directly calculated from CFD) for online monitoring.
From the above section, three monitoring inputs can be highly possible to well predict
3 input zones
Input zones of
Input zones of
Input zones of
Input zones of
Input zones of
Input zones of
3 input zones
4 input zones 4 input zones
5 input zones 5 input zones
Single source A Double sources A&C
Accepted version of SCS journal
indoor pollutant concentration as long as we avoided deploying those sensors along the
same or parallel with streamwise plane (iso-surface of X coordinate). Specifically, we
placed these three hypothetic sensors at the coordinates (X,Y) of S1(0.04,2.83),
S2(1.75,3.86), S3 (1.75,3.86), ref., Fig.9. Monitor 3 and 2 were respectively set close
to the middle of the inlet and outlet, monitor 1 near the wall. The hypothetic monitoring
concentration was calculated as the volume-averaged concentration of the cells with
distance from the monitor point smaller than 1.5cm. The ANN database was constructed
by replacing the concentration of the low-dimensional concentration to that of the
corresponding monitoring point. The prediction procedure is similar to the above
section. Figure 10 shows the comparision of concentration prediction using ANN and
LLVM calculation. The corresponding performance index RMSE equals to 16, which
is nearly 70% larger than the corresponding case 10 (Table 1). It is aso noticed that the
concentration of ZL 24 (within inlet flow region) is almost zero.
Figure 9 Layout with three hypothetic sensors (at three points M1, M2, M3)
Figure 10 Low-dimensional concentration prediction using three hypothetic sensors (□) and
validation with LLVM calculation (*)
Accepted version of SCS journal
4 Discussion
The discussion section will provide further in-depth analysis on indoor pollutant
prediction by integrating monitoring techniques and neuron network modelling.
Suggestions are highlighted for future deployment of monitoring sensors provided that
the prediction accuracy is well accepted.
From the results section, it indicated that the ANN prediction showed satisfying
results as we keep the monitoring sensor numbers equal to or larger than 5, regardless
of placement locations. This coincides with our expectations: the more sensors we set,
the more information they would represent of the room. However, prediction error
appeared occasionally when using 3 or 4 sensors. Some of RMSE values were
extraordinarily larger compared to others. Some prediction with 3 monitoring sensors
(as ANN inputs) were even superior to those with 4 sensors. When looking at Fig.8 and
10, monitoring inputs with relatively worse prediction have a common characteristic:
at least two or more monitoring zones/points located along or parallel with the main
stream plane (the same iso-surface of X coordinate). It further reminds us that
monitoring points along the same stream directions contain similar monitoring
representation of indoor pollutant concentration, resulting in repeated inputs for ANN
prediction, i.e., one sensor without functioning
Some cases with 3 monitoring inputs were showing even better prediction
performance compared to cases with 4 monitoring input. We now turn to look at the
characteristics of better case with 3 inputs (ZL: 6, 3, 24) and worse case with 4 inputs
(ZL: 6, 4, 14, 24). We noticed that both cases contained monitoring zones with an inlet
and an outlet region, so the property of the third monitoring zone would determine the
final prediction performance. From Table 1, the case (with 3 input zones) with relatively
better prediction corresponded to input ZL of (6, 3, 24). The third input ZL is 3, which
is within the outlet region. Generally, the airflow near the outlet region was close to
well mixed status, which would contain full information of both airflow and
concentration. Hence, it should be a good choice to deploy sensors within the outlet
Accepted version of SCS journal
region, further facilitating pollutant prediction.
To validate our assumption and use for the practical application, a series of cases
with hypothetical sensors (monitoring at three points instead of afforementioned zones,
named point concentration) were carried out. These three points followed the above
presumption: two points within the inlet and outlet region, the third point not along the
same main streamwise plane with other two points. Compared to the previous better
prediction cases (3 concentration inputs with ZL of 6, 3, 24), the prediction was 70%
deviated. This is due to the fact that the magnitude of point concentration is almost zero
within the inlet flow region (strong airflow distribution), which is different from the
volume-averaged concentration of the inlet zone (with ZL of 24). Thus, in the practical
application, sensors should not be deployed close to or in the inlet flow region although
easy to install.
To further prove our conclusion, two refined cases were conducted: 1) moving the
sensor away from the inlet mainstream region (18cm away in the –Y direction) for the
first refined case, ref., Fig.11 (a); 2) moving the sensor to the outlet region for the
second refined case, ref., Fig.11 (b). Fig. 12 described the comparison of the predicted
concentration with direct LLVM calculation for two refined cases. Prediction accuracy
have been improved for two cases. When looking at RMSE for three cases in Fig. 13,
it seems that the second refined case performed better with the least averaging RMSE
of 11.05. Compared to the first refined case placing two monitoring points in the outlet
region, the second case placing three monitoring points showed better prediction
performance. Thus, to better represent indoor pollutant concentration distribution, it
would be better to place more sensors close to the outlet region, which is consisted with
the litertature study of (Rackes et al., 2018).
Accepted version of SCS journal
Figure 11 Layout of two revised cases: (a) two monitoring points located within the outlet
region and one near the inlet; (b) three monitoring points located within the outlet region
(a) (b)
Figure 12 Prediction validation (with LLVM calculation, *) of two revised cases: (a) two
monitoring points located within the outlet region and one near the inlet (▽); (b) three monitoring
points located within the outlet region (□)
Figure 13 RMSE distribution of three cases. (a): the original case: the averaging RMSE of 16.24;
(b): the first refined case: the averaging RMSE of 14.75 (c): the second refined case: the averaging
RMSE of 11.05
·
(a)
(b)
(a) (b) (c)
RMSE (ppm)
RMSE
4
8
12
16
20
Accepted version of SCS journal
5 Conclusion
In this work, we aimed to rapidly predict indoor concentration by incorporating
monitoring using a limited number of sensors to LLVM-ANN modelling, further
facilitating intelligent ventilation control. The main conclusions are summarized below.
(1) High-resolution concentration calculated from CFD can be processed using LLVM
method, further enlarged database for ANN training.
(2) A properly trained RBF of artificial neural network is able to fast and efficiently
predict low-dimensional indoor concentration field on the basis a limited input
monitoring data from several sensors (i.e., at least 3 in the current study).
(3) Sensor deployment strategies were provided: sensors should be avoided to be along
the same or parallel with the main stream plane; sensors should be better placed in
or near the ‘well-mix zone (close to outlet region)’ but avoiding close to inlet.
To sum up, the main focus of the current work is to develop a faster-than-real-time
model to rapidly estimate indoor pollutant concentration in the entire indoor field by
integrating online monitoring techniques and fast prediction models, where low-
dimensional linear ventilation models and ANN methods have been applied. This model
can be further extended to more complicated indoor environments, such as human
activities induced indoor environments. Indoor activities or human behavior can be
indirect signals indicating pollutant sources, e.g. Particle Matters, PM. We can label
these activities for ANN training. Before training, we need to understand and study the
associated relationships between these activities and PMs concentration or emission
rates. Otherwise, we may also directly link the ventilation modes to these activities.
Regarding the application for industry ventilation environment, it will be more
complicated since there are more pollutants or unstable airflow distributions. Moreover,
the space is rather larger. We may think of even faster way to predict ventilation systems
by using adaptive mesh methods. For multiple pollutants, we may find the correlation
of these pollutants and finally focused on several typical pollutants, considering the
sensor cost and the representation characteristics. The deployment of sensor locations
Accepted version of SCS journal
for larger indoor environment is also worthy being investigated. These will be of great
interest in our future work.
Acknowledgements
The authors would like to acknowledge the financial support from National Natural
Science Foundation of China (Grant No.51778385).
References
Persily, A. K., & Emmerich, S. J. (2012). Indoor air quality in sustainable,
energy efficient buildings. Hvac&R Research, 18, 4-20.
Li, X., Shen, C., & F., C. W. (2017). Building energy efficiency: Passive
technology or active technology? Indoor and Built Environment, 26(6), 729–
732.
Ai ZT, Mak CM. (2016) Short-term mechanical ventilation of air-
conditioned residential buildings: A general design framework and guidelines.
Building And Environment, 108,12-22.
Seppanen, O. (2008). Ventilation Strategies for Good Indoor Air Quality
and Energy Efficiency. International Journal Of Ventilation, 6, 297-306.
Mossolly, M., Ghali, K., & Ghaddar, N. (2009). Optimal control strategy
for a multi-zone air conditioning system using a genetic algorithm. Energy, 34,
58-66.
Deng, Y. L., Feng, Z. B., Fang, J., & Cao, S. J. (2018). Impact of ventilation
rates on indoor thermal comfort and energy efficiency of ground-source heat
pump system. Sustainable Cities And Society, 37, 154-163.
Zorpas, A. A., & Skouroupatis, A. (2016). Indoor air quality evaluation of
two museums in a subtropical climate conditions. Sustainable Cities And
Society, 20, 52-60.
Kumar, P., Martani, C., Morawska, L., Norford, L., Choudhary, R., Bell,
M., & Leach, M. (2016). Indoor air quality and energy management through
real-time sensing in commercial buildings. Energy And Buildings, 111, 145-153.
Rackes, A., Ben-David, T., & Waring, M. S. (2018). Sensor networks for
routine indoor air quality monitoring in buildings: Impacts of placement,
accuracy, and number of sensors. Science And Technology for the Built
Environment, 24, 188-197.
Chenari, B., Carrilho, J. D., & Da Silva, M. G. (2016). Towards sustainable,
energy-efficient and healthy ventilation strategies in buildings: A review.
Renewable & Sustainable Energy Reviews, 59, 1426-1447.
Fontanini, A. D., Vaidya, U., & Ganapathysubramanian, B. (2016). A
Accepted version of SCS journal
methodology for optimal placement of sensors in enclosed environments: A
dynamical systems approach. Building And Environment, 100, 145-161.
Cao, S. J., & Ren, C. (2018). Ventilation control strategy using low-
dimensional linear ventilation models and artificial neural network. Building
And Environment, 144, 316-333.
Cao, S. J. (2019). Challenges of using CFD simulation for the design and
online control of ventilation systems. Indoor And Built Environment, 28, 3-6.
Shan, X. F., Xu, W., Lee, Y. K., & Lu, W. Z. (2019). Evaluation of thermal
environment by coupling CFD analysis and wireless-sensor measurements of a
full-scale room with cooling system. Sustainable Cities And Society, 45, 395-
405.
Zhang TF, Yin S, Wang SG. (2013) An inverse method based on CFD to
quantify the temporal release rate of a continuously released pollutant source.
Atmos Environ,77, 62-77.
Zhu, P. G., Tong, X. N., Chen, L., Wang, C. W., Song, H., & Li, X. Y.
(2015). Influence of opening area ratio on natural ventilation in city tunnel under
block transportation. Sustainable Cities and Society, 19, 144-150.
Deng, H. Y., Feng, Z. B., & Cao, S. J. (2018). Influence of air change rates
on indoor CO2 stratification in terms of Richardson number and vorticity.
Building And Environment, 129, 74-84.
Desta, T. Z., Janssens, K., Van Brecht, A., Meyers, J., Baelmans, M., &
Berckmans, D. (2004). CFD for model-based controller development. Building
And Environment, 39, 621-633.
Cao, S. J., & Meyers, J. (2012). On the construction and use of linear low-
dimensional ventilation models. Indoor Air, 22, 427-441.
Zhou, Y. , Deng, Y. , Wu, P. , & Cao, S. J. . (2017). The effects of ventilation
and floor heating systems on the dispersion and deposition of fine particles in
an enclosed environment. Building and Environment, 125，192-205.
Cao, S. J., Cen, D., Zhang, W., & Feng, Z. (2017). Study on the impacts of
human walking on indoor particles dispersion using momentum theory method.
Building and Environment, 126, 195-206.
Liao, S. H. and Wen, C. H. (2007). Artificial neural networks classification
and clustering of methodologies and applications-Literature analysis from 1995
to 2005. Expert Systems with Applications, vol.32, pp.1-11, 2007.
