applied

sciences
Article
Noise Prediction Using Machine Learning with
Measurements Analysis
Po-Jiun Wen 1,2 and Chihpin Huang 1, *
1 Institute of Environmental Engineering, National Chiao Tung University, Hsinchu 30010, Taiwan;
[email protected]
2 Radiation & Operation Safety Division, National Synchrotron Radiation Research Center,
101 Hsin-Ann Road, Hsinchu Science Park, Hsinchu 30076, Taiwan
* Correspondence: [email protected]; Tel.: +886-3-5712121 (ext. 55507)




Received: 28 July 2020; Accepted: 17 September 2020; Published: 22 September 2020


Abstract: The noise prediction using machine learning is a special study that has recently received
increased attention. This is particularly true in workplaces with noise pollution, which increases
noise exposure for general laborers. This study attempts to analyze the noise equivalent level (Leq)
at the National Synchrotron Radiation Research Center (NSRRC) facility and establish a machine
learning model for noise prediction. This study utilized the gradient boosting model (GBM) as the
learning model in which past noise measurement records and many other features are integrated as
the proposed model makes a prediction. This study analyzed the time duration and frequency of
the collected Leq and also investigated the impact of training data selection. The results presented
in this paper indicate that the proposed prediction model works well in almost noise sensors and
frequencies. Moreover, the model performed especially well in sensor 8 (125 Hz), which was
determined to be a serious noise zone in the past noise measurements. The results also show that
the root-mean-square-error (RMSE) of the predicted harmful noise was less than 1 dBA and the
coefficient of determination (R2 ) value was greater than 0.7. That is, the working field showed a
favorable noise prediction performance using the proposed method. This positive result shows the
ability of the proposed approach in noise prediction, thus providing a notification to the laborer
to prevent long-term exposure. In addition, the proposed model accurately predicts noise future
pollution, which is essential for laborers in high-noise environments. This would keep employees
healthy in avoiding noise harmful positions to prevent people from working in that environment.

Keywords: noise prediction; machine learning; noise equivalent level (Leq); gradient boosting model
(GBM); harmful noise

1. Introduction
Noise pollution is often overlooked in many working environments, which are very often
noise-filled [1,2]. According to the Environmental Protection Agency (EPA), the volume of human
speech is approximately 60 dBA. Moreover, people will feel irritable, nervous, unable to concentrate,
and will be affected by prolonged exposure to environmental noise at 70 dBA [3]. Long-term exposure
to noise at more than 85 dBA will cause chronic hearing damage and can indirectly cause occupational
disasters [3]. Likewise, laboratories contain many equipment that generate noise, which distracts
researchers and impairs their ability to concentrate. Thus, locating noise sources, predicting future
noise levels, and altering environmental factors are important research topics that could be improved
to protect against noise, which is important for safe and productive work environments.
Many existing prediction models for acoustical properties and traffic noise still have problems
with accuracy limitations. For example, the grey model (GM) with Fourier correction gray model (FGM)

Appl. Sci. 2020, 10, 6619; doi:10.3390/app10186619 www.mdpi.com/journal/applsci

Appl. Sci. 2020, 10, 6619 2 of 21

was proposed to predict the normal incidence sound absorption coefficient and tire/road noise [4].
Based on the analysis of the Federal Highway Administration (FHWA) traffic noise model, a new
simplified prediction method was proposed that showed the connection between the traffic noise
increments with increases in traffic volume [5]. Another study predicted the total industrial production
output value, which will help the design plan of the development city [6]. According to nonlinear
time series such as noise data, the gradient lifting technology model is recognized in the prediction of
nonlinear time series with a high accuracy rate [6]. Successful examples include predicting real estate
sales prices, which includes noise similar to that of time series data [7]. A recent study applied this
method for forecasting air quality in Taiwan and also extracted meaningful time and historical features
as input to the gradient lifting technology model [8].
Although some literature has shown that gradient lifting technology models work well for
predicting certain targets that included a decision tree getting initial values for the fitting function
with multiple regression, which treated the many input variables considered in this research.
However, the observed data information and output values are calculated error, which uses a loss
function. The frequently used loss functions include square-error, absolute-error, and negative binomial
log-likelihood functions [8]. Then, gradient lifting technology was applied to find the fitting function
where the expected value of loss function is minimized. This procedure was repeated to acquire the
optimized fitting function. Unfortunately, their application in noise prediction is very limited.
To the best of our knowledge, predicting future noise using the gradient boosting model (GBM) [9]
has not been addressed in the existing literature. In fact, the existing noise information belongs to time
series data, and their closeness to time is similar to other predicted targets. This motivates our use of the
GBM prediction model in a noisy environment, which enables efficient identification of suitable training
features in response to different environments and noise conditions, thereby achieving robust and
reliable prediction results. Following this, the method proposed in this paper can effectively select the
appropriate features as the model input for different characteristics of the noise fields. Moreover, it has
good portability, which will be useful for the conversion of many noise sources in the future.
The paper of this purpose was to analyze the noise equivalent level (Leq (dBA)) [10–14] in a
work environment that contained the most seriously affected zones. It was evident from long-term
monitoring that the highest dBA levels occurred on a certain day every week, that the dBA of certain
frequency bands was always higher, and that the dBA levels differed between morning and night.
Thus, the noise frequencies most harmful to humans were identified and machine learning was used
to target these frequency bands for prediction. Through this method, we confirmed the noise map
of the examined field, attempted to add meaningful time and historical features from the previous
analysis, and predicted the likelihood of harmful noise [15–18] at future time points in the operating
environment. The results in this paper can be used to prevent noise pollution in advance to create
better working conditions.
The main process in this paper is divided into three parts, as shown in Figure 1:

1. According to the data provided by the National Synchrotron Radiation Research Center (NSRRC),
we performed daily and monthly statistical analyses on the noise data of 12 sensors at different
frequencies. Once collected, the data were cleaned to derive useful information and analyze the
data distribution.
2. We derived and extracted the features from the data analysis. We identified the frequency, time,
and eight sensors from related history features, and then input a harmful frequency and the
noisiest dBA sensor as extracted features.
3. We extracted the Leq historical features and time-related features from 80% of the data inputted
to the machine learning model for training; the data for the remaining 20% was used for testing.
Appl. Sci. 2020, 10, 6619 3 of 21

Figure
Figure 1. The
1. The main
main process
process of the
of the method
method in three
in three parts.
parts.

2. 2.
Materials
Materials

2.1.2.1.
Information Introduction
Information andand
Introduction Data Analysis
Data Analysis
TheThenoise datadata
noise werewere
provided by theby
provided NSRRC and contained
the NSRRC more thanmore
and contained 13,000,000
than samples
13,000,000 covering
samples
thecovering
time period from 08:00 on 1 February 2019 to 23:59 on 31 August 2019. The
the time period from 08:00 on February 1, 2019 to 23:59 on August 31, 2019. The NSRRC NSRRC installed 12
noise detection sensors around the work environment. As shown in Figure 2a,
installed 12 noise detection sensors around the work environment. As shown in Figure 2a, it showed it showed the circle
building hasbuilding
the circle a 120 m has
circumference, and 24 straight
a 120 m circumference, line
and 24experimental stations that are
straight line experimental differential
stations that are
function research experiments; the locations of the sensors were divided
differential function research experiments; the locations of the sensors were divided into theinto the inner circle (1–6)
inner
and the outer circle (7–12), shown as green dots. In addition, there are many noise
circle (1–6) and the outer circle (7–12), shown as green dots. In addition, there are many noise sources sources from
vacuum pumps, cooling
from vacuum pumps,pumps,
coolingliquid
pumps, nitrogen
liquid pressure
nitrogenrelief, computer
pressure relief, servers,
computer etc.servers,
in the working
etc. in the
environment, as shown in Figure 2b. In this study, the noise sensor module
working environment, as shown in Figure 2b. In this study, the noise sensor module had more had more details including
details
theincluding
measuring therange, frequency
measuring range,range, weighing,
frequency range,andweighing,
resolution, as resolution,
and shown in Figure as shown2c. This sensor2c.
in Figure
device
This can instantly
sensor deviceconvert sound into
can instantly Leq sound
convert information in the
into Leq cloud as shown
information in the in Figure
cloud 2d. DueintoFigure
as shown the
limitations of the hardware and network, the Leq was recorded once per second
2d. Due to the limitations of the hardware and network, the Leq was recorded once per second and and the average Leq
wastheuploaded
averageonceLeq every 10 s. Using
was uploaded thisevery
once device,10the different
s. Using frequencies
this device, the fordifferent
Leq values were collected.
frequencies for Leq
The data were divided into eight different frequencies: 125 Hz, 250 Hz, 500 Hz,
values were collected. The data were divided into eight different frequencies: 125 Hz, 250 Hz, 500 1 kHz, 2 kHz, 4 kHz,Hz,
8 kHz, and 16 kHz. Furthermore, the noise monitoring system that provided
1 kHz, 2 kHz, 4 kHz, 8 kHz, and 16 kHz. Furthermore, the noise monitoring system that providedsensors, the location, time
and different
sensors, thefrequency noiseand
location, time datadifferent
are shown in Figure
frequency 2e. An
noise dataalternative
are shown function
in Figure in 2e.
ourAn monitoring
alternative
system
function in our monitoring system showed 12 groups for real-time linear charts that can be data,
showed 12 groups for real-time linear charts that can be displayed within two hours’ as
displayed
shown in Figure 2f. The segmentation for a maximum date interval of 12 group’s
within two hours’ data, as shown in Figure 2f. The segmentation for a maximum date interval of 12 linear charts was
setgroup’s
as a month,
linearascharts
shown in Figure
was set as a2g. Basedason
month, this system,
shown in Figure we2g.
kept the data
Based collection
on this system,for wefurther
kept the
analysis in future work.
data collection for further analysis in future work.

3
Appl. Sci. 2020, 10, 6619 4 of 21

(a) Noise sensors map (b) Noise working environment of the NSRRC

(c) Noise sensor module

(d) Noise detector and location

(e) Illustration of the noise monitoring system

(f) Two hours’ real-time detection chart

(g) One-month detection chart

Figure 2. (a) Noise sensor map. (b) Noise working environment of the National Synchrotron Radiation
Research Center (NSRRC). (c) Noise sensor module. (d) Noise detector and location. (e) Noise
monitoring system illustrating. (f) Two hours’ real-time detection chart. (g) One-month detection chart.
 Appl. Sci. 2020, 10, 6619 5 of 21

As the inner and outer ring sensors had different installation times, we divided the statistics of
the Leq for the 12 sensors as shown (g)inOne-month detection
Figure 3. The chart from sensors 1–6 were recorded
data collected
from February to August; the remaining sensors collected data from April to August. The detailed
Figure 2. (a) Noise sensor map. (b) Noise working environment of the National Synchrotron Radiation
time distribution and the number of data points are shown in Figure 3. In addition, all sensors had
Research Center (NSRRC). (c) Noise sensor module. (d) Noise detector and location. (e) Noise
relatively complete data in July and August; thus, we used the most recent August data for training in
monitoring system illustrating. (f) Two hours’ real-time detection chart. (g) One-month detection
the experiment.
chart. Figure 4 shows the distribution of the average Leq levels of each sensor at 125 Hz and 1000 Hz
over the different months (for the other monthly statistics from different frequencies, see Appendix A).
As the
It is innerthat
evident andtheouter ring
average Leqsensors
level forhad different
sensor installation
8 was higher than thattimes, we divided
of the other thecloser
sensors and statistics of
to 70 dBA from 125 to 1000 Hz for each month. As shown in Figure 5, there was
the Leq for the 12 sensors as shown in Figure 3. The data collected from sensors 1–6 were recordedno sensor with a
particularly
from February toprominent value
August; the from 2000~16
remaining kHz, indicating
sensors collected that
datathe low April
from frequencies were theThe
to August. maindetailed
noise sources in the environment. Therefore, we hypothesized that when the equipment was operating,
time distribution and the number of data points are shown in Figure 3. In addition, all sensors had
it caused louder low-frequency noise near sensor 8. In fact, there were more noise sources near sensor
relatively complete data in July and August; thus, we used the most recent August data for training
8 than in the other areas, which caused dBA values higher than those of the other sensors.
in the experiment.

Appl. Sci. 2020, 10, x FOR PEER REVIEW 6 of 21

Figure 4 shows the distribution of the average Leq levels of each sensor at 125 Hz and 1000 Hz
over the different months (for the other monthly statistics from different frequencies, see Appendix
A). It is evident that the average Leq level for sensor 8 was higher than that of the other sensors and
closer to 70 dBA from 125 to 1000 Hz for each month. As shown in Figure 5, there was no sensor with
a particularly prominent value from 2000~16 kHz, indicating that the low frequencies were the main
noise sources in the environment. Therefore, we hypothesized that when the equipment was
operating, it caused louder low-frequency noise near sensor 8. In fact, there were more noise sources
near sensor 8 than in the other areas, which caused dBA values higher than those of the other sensors.
Figure 3. Twelve
Figure noise
3. Twelve noisesensors collecteddata
sensors collected datain in different
different months.
months.

Figure 4. Average at 125 Hz and 1000 Hz in different months for different sensors.
Figure 4. Average at 125 Hz and 1000 Hz in different months for different sensors.
 Appl. Sci. 2020,Figure
10, 66194. Average at 125 Hz and 1000 Hz in different months for different sensors. 6 of 21

Figure 5. Different sensor averages at 2000 Hz and 16,000 Hz for different months.
Figure 5. Different sensor averages at 2000 Hz and 16,000 Hz for different months.
Next, we measured each sensor’s average daily data on a single frequency band over one week.
Next,
The Leq value we measured
for the 125each
Hz sensor’s
band wasaverage dailythat
higher than dataofonthea other
singlefrequencies,
frequency band as shownover in
one week. 6
Figures
The Leq value for the 125 Hz band was higher than that of the other frequencies,
and 7 (for the other daily statistics from different frequencies, see Appendix B). It has been established as shown in Figures
6 that
andthe 7 (for
values theforother
sensordaily
8 at statistics
125 Hz–1000 from Hzdifferent frequencies,
were obviously greater see
thanAppendix
those of the B). other
It hassensors.
been
established
In particular, it was evident that on Sunday, the sensor 8 Leq value was greater than the other daysthe
that the values for sensor 8 at 125 Hz–1000 Hz were obviously greater than those of for
other sensors. In We
all frequencies. particular, it was evident
hypothesized that thisthat
wasondue Sunday, the that
to the fact sensorthe8equipment
Leq value was neargreater
sensor than
8 was
the other days
relatively old.forIn all frequencies.
addition, We hypothesized
there were no people working that thisonwas due to
Sunday and thethe
fact that the equipment
temperature increased
near sensor 8 was relatively old. In addition, there were no people
due to the air-conditioning being shut down to save energy costs. As a result, older equipmentworking on Sunday andcloser
the
temperature increased due to the air-conditioning being shut down to save
to sensor 8 were prone to make loud noises on Sunday. In addition, each sub-image in Figures 6 and 7 energy costs. As a result,
older equipment
had two closer to sensor
lines, representing 8 were prone
the average of theto make
Leq loud
levels noises
of the on Sunday.
12 sensors Inmorning
in the addition,(red)eachand
sub-at
image in Figures 6 and 7 had two lines, representing the average of the Leq levels
night (blue) within a week in the above frequency band. It is evident that as the frequency increased, of the 12 sensors in
the
themorning
red line (red) and at than
was higher nightthe (blue)
bluewithin
line, anda week in the above
the differences frequency
between the band.
average It isLeq
evident
levelsthat as
during
the frequency increased, the
the morning and night also increased.red line was higher than the blue line, and the differences between the
average TheLeq levelsofduring
results the analyses
statistical morningfor and allnight
sensoralso increased.
averages and for sensor 8 from morning and night
differ from the Leq value at each frequency band over the week, as shown in Figure 8. The horizontal
axis is the frequency and the vertical axis is the Leq error value. We can clearly observe that the average
chart position for all sensors grows with increasing frequency in the range of 125 Hz to 16 kHz (apart
from 500 Hz). When the frequency was higher, the difference was greater. Moreover, the average
Appl. Sci. 2020, 10, x FOR PEER REVIEW 7 of 21
values for sensor 8 values were higher than those of all other sensors.

.
Figure 6. Daily average at 125 Hz and 1000 Hz in one week for different sensors.
Figure 6. Daily average at 125 Hz and 1000 Hz in one week for different sensors.
 .

Appl. Sci. 2020, 10, 6619 6. Daily average at 125 Hz and 1000 Hz in one week for different sensors.
Figure 7 of 21

Figure 7. Daily average at 2000 Hz and 16,000 Hz in one week for different sensors.
Figure 7. Daily average at 2000 Hz and 16,000 Hz in one week for different sensors.
The results for all sensor values and from sensor 8 are shown in Figure 9, which shows a decrease
Thelevels
in Leq resultswith
of statistical
increasing analyses for all
frequency. Onsensor averages
the other hand,and for was
there sensoran 8instance
from morning
of low and
dBAnight
levels
Appl. from
differ Sci. 2020,
the10,
Leqx FOR
valuePEERat REVIEW
each frequency band over the week, as shown in Figure 8. The 8 of 23
horizontal
from a high frequency. The results also show that all average sensors were lower than sensor 8
axis is the frequency
regardless of whether and
the the
levelsvertical
were axis is theinLeq
measured the error value. We can Thus,clearlywe observe that the
regardless
average chart ofposition
whetherfor theall
levels weregrows
sensors measured with in themorning
morning
increasing
or
orat
frequencyat night.
night.
in Thus,
the rangeweofcan
can add
add
125 Hz
features
features
to 16
in
in the
the experiment.
experiment.
kHz (apart
The from 500
histograms Hz). When
added belowthe frequency
show the was higher, the difference was greater. Moreover, the
average The histograms
values added
forinsensor belowwere
8 values the average
showhigher average
than
Leq
Leq changes
those changes in
in the
of all other the morning
morning and
sensors. and atat night
night for
for each
each
frequency
frequency band
band in a
a week
week where
where the
the upper
upper and
and lower
lower bounds
bounds of
of the
the vertical
vertical axis
axis differ
differ by
by 2.5
2.5 dBA.
dBA.
We The
can results
clearly for
see all sensor
that the values and
average from
noise in sensor
the 8 are shown
morning was in Figure
slightly 9, which
higher than shows
that a decreaseat
measured
We
innight can
Leq levels clearly
with see that
increasingthe average
frequency. noise
On in
thethe morning
other was
hand, there slightly
was an higher
instance than that
of low measured
dBA levels at
night on
on any
any frequency
frequencyThe band.
band. This gap
This gap was
was more
more obvious in Figures 10 and 11 (see Appendix C for
from
other a noise
high frequency.
frequency changes results
duringalsotheshow
morningthatobvious in Figures
all average
and night). sensors 10 were
and 11 (see Appendix
lower than sensor C for
8
other noise frequency changes during the morning and night).
regardless of whether the levels were measured in the morning or at night. Thus, we can add features
in the experiment.
The histograms added below show the average Leq changes in the morning and at night for each
frequency band in a week where the upper and lower bounds of the vertical axis differ by 2.5 dBA.
We can clearly see that the average noise in the morning was slightly higher than that measured at
night on any frequency band. This gap was more obvious in Figures 10 and 11 (see Appendix C for
other noise frequency changes during the morning and night).

Figure8.8. Morning
Figure Morning and
and night
night noise
noise error
errordata
datavs.
vs. frequency
frequency (all
(all sensor
sensor avg.
avg. vs.
vs. sensor
sensor 8).
8).
Appl. Sci. 2020, 10, 6619 8 of 21
Figure 8. Morning and night noise error data vs. frequency (all sensor avg. vs. sensor 8).

Figure 8. Morning and night noise error data vs. frequency (all sensor avg. vs. sensor 8).

Figure 9. Morning and night average noise data vs. frequency (all sensor avg. vs. sensor 8).
Figure 9. Morning
Morning and
and night
night average
average noise
noise data vs. frequency
frequency (all
(all sensor
sensor avg.
avg. vs.
vs.sensor
sensor8).
8).

Appl. Sci. 2020, 10, x FOR PEER REVIEW 9 of 21

Figure 10. Average morning vs. night at 125 Hz and 1000 Hz in a week for all sensors.
Figure 10. Average morning vs. night at 125 Hz and 1000 Hz in a week for all sensors.

Figure11.
Figure Averagemorning
11.Average morningvs.
vs.night
nightatat2000
2000Hz,
Hz,16000
16000Hz
Hzin
inaaweek
weekfor
forall
allsensors.
sensors.

2.2. Methods

2.2.1. Feature Extraction

We derived and extracted features from the Leq time data for model training. The continuity of
the noise generation process, which is affected by working days and working time, resulted in the
input vector Xt containing temporal features; the output variable Yt+1 is the Leq value for the next
minute. In this experiment, we chose to include days in a week, hours in a day, whether the day was
a holiday, whether it was a Saturday or a Sunday, and the previous one minute or two minutes of
Appl. Sci. 2020, 10, 6619 9 of 21

2.2. Methods

2.2.1. Feature Extraction

We derived and extracted features from the Leq time data for model training. The continuity of
the noise generation process, which is affected by working days and working time, resulted in the
input vector Xt containing temporal features; the output variable Yt+1 is the Leq value for the next
minute. In this experiment, we chose to include days in a week, hours in a day, whether the day was
a holiday, whether it was a Saturday or a Sunday, and the previous one minute or two minutes of
historical noise frequency data for sensor 8. Thus, a total of 21-dimensional features were input for
training (as shown in Table 1).

Table 1. Input features obtained from noise sensor monitoring.

Input Feature (21-Dimensional)

History feature previous 1 min of sensor * 8 previous 2 min of sensor * 8 16
Time feature Which day in a week, Which hour in a day, Holiday or not, Saturday or not, Sunday or not. 5
* Eight noise frequencies: 125 Hz, 250 Hz, 500 Hz, 1 kHz, 2 kHz, 4 kHz, 8 kHz, and 16 kHz.

2.2.2. Machine Learning Model

Artificial intelligence has recently attracted considerable attention, and various machine learning
approaches have been extensively implemented to model data in numerous applications [19,20].
For example, a travel time prediction model based on gradient boosting decision tree (GBDT) has been
proposed to improve the prediction accuracy of traffic flow [21]. A new extreme gradient boosting
(XGBoost) model with weather similarity analysis and feature engineering was proposed for short-term
wind power forecasting [22]/ Air quality prediction in smart cities was undertaken using machine
learning technologies based on sensor data [23]. This paper presented an innovative gradient boosting
decision tree (GBDT) model to explore the joint effects of comprehensive factors on the traffic accident
indicators [24]. A method was presented for predicting the broadband noise spectra of horizontal axis
wind turbine generators [25] as well as a study on noise sensitivity by machine learning algorithms [26].
We used the gradient boosting model (GBM) to predict future Leq levels. This model combines
fitting functions, loss functions, a decision tree, and gradient descent analysis [9]. The decision tree,
error function L(F(Xt ), Yt+1 ) [27,28], fitting function F(Xt ) [29], and gradient descent analysis were
applied to train the model. Specifically, the decision tree algorithm was used to generate a series of
fitting functions F(Xt , β0 ). The error function L(F(Xt , β0 ), Yt+1 ) was used to calculate the fitting value
F(Xt , β0 ), which is the error from the actual value Yt+1 , where Xt is the input vector at time t and Yt+1
is the output variable at time t+1. Next, we used the gradient descent method to find and select the
fitting function F(Xt , β) with the smallest error. The above steps are repeated until the optimal fitting
function is found. The procedure is described in the equations in detail. The testing samples are put
into the prediction model F(xt ) to calculate the prediction results [8].
The objective of machine learning is to find a mapping function F(x) between the independent
variable xi and target variable yi by using the training data. In order to find the optimal function, a loss
function L(y, F(x)) is usually set for the model [21,25]. First, initialize the learning machine by the
following equation:
XN
F0 (Xt ) = argminβ L(yi , β) (1)
i=1

where β is the estimated constant value that minimizes the loss function and N is the number of
training samples.
Then, the target is to predict the next-24 h of noise Leq, where the output is the variable shown as
yt+24 . After N pairs of the input vector xt and the output variable yt+24 are given, a fitting function
F(xt) is selected from unknown functions F(xt , β0 ) generated by the decision tree. Moreover, β0 is a
Appl. Sci. 2020, 10, 6619 10 of 21

gradient decent step size and (xit , yit+24 ) is the i-th training sample pair. When the value of the loss
function L(yt +24, F(xt , β0 )) is minimized as [8,21,25]
XN
β = argminβ0 L(yit+24 , F(xit , β0 )) (2)
i=1

the target function F(xt ) is chosen as F(xt , β)

F0 (Xt ) = F0 (xt , β0 ) (3)

In this procedure, the value of N was 39,515. This is obtained from sensor 8 during August and the
training model needed 80% of them. In addition, the gradient descent analysis is applied for optimized
fitting function F(xt ). The procedure is described as below. In the first step, the initial guess function
F0 (xt , β0 ) is produced and initial gradient descent step size β0 as [8,21,25]
XN
β0 = argminβ0 L(yit+24 , F0 (xit , β0 )) (4)
i=1

Thus, we take the gradient of loss function as a first-step base learner function f1 (xt ) as

f1 (xt ) = − 5F0 L(yt+24 , F0 (xt )), (5)

XN
β1 = argminβ0 L(yit+24 , [F0 (xit ) + β0 f1 (xit )]) (6)
i=1
In this study, M iterations was set as 500, where fm (xt ) and βm are expressed as follows:
 
fm (xt ) = − 5Fm−1 L yt+24 , Fm−1 (xt ) (7)
XN
βm = argminβ0 L(yit+24 , [Fm−1 (xit ) + β0 fm (xit )]) (8)
i=1
and the target function F(xt ) is expressed as

M
X
F(xt ) = F0 (xt ) + βm fm (xt ) (9)
m=1

Through the above formulas, description flow was calculated by Algorithm 1, that is the entire
GBM procedure. There is no doubt that F(xt ) is the target prediction model, thus, the testing samples
were put into the model to calculate the prediction results.
The algorithm flow is as follows:
Algorithm 1. GBM
Input:
1: F0 (xt , β00 )
N
L(yit+24 , F0 (xit , β0 ))
P
2: β0 = argminβ00
i=1
3: M: Iteration times
4: N: Number of data sets
Output: F(xt ) = FM (xt )
5: For m = 1 to M  
6: fm (xt ) = − 5F L yt+24 , Fm−1 (xt )
7: βm = argminβ0m N i i 0 i
P
i=1 L(yt+24 , [Fm−1 (xt ) + βm .fm (xt )])
8: Fm (xt ) = Fm−1 (xt ) + βm .fm (xt )
9: end
Appl. Sci. 2020, 10, 6619 11 of 21

3. Results and Discussion

To use the GBM model for future Leq value prediction on the 125 Hz frequency band, time
characteristics and historical Leq were used as the input data. There were 12 sensor data, which were
recorded every 10 s (sampling time = 10 s) with an average sample of 875,000. As the training time was
too long, the sampling time was lengthened in order to reduce the amount of training data. Sensor 8
was used as a sampling example and used the previous two minutes of data to make predictions at
different sampling times; the results are shown in Figure 12.
The X-axis in Figure 12 is the R2 [30], which represents the degree of curve fit between the
predicted value and the actual value (sensor 8 at 125 Hz in August). The R2 value is distributed in the
range of 0–1, and values closer to 1 indicate better prediction performance; otherwise, the prediction
performance worsens. The Y axis is the RMSE [31]. Here, the higher the value indicates a worse
prediction result; otherwise, the convergence is smaller. We found that a sampling time of 1 min
and 30 s had a higher R2 and the best prediction performance; in addition, the RMSE was below
1 dBA. However, the R2 value of the 30 s sampling time was very close to those of 1 min and better,
but the calculation requires double the time to complete. Thus, in the subsequent experimental design,
the sampling time was adjusted to 1 min. Next, according to the frequency of harmful Leq levels
at 125 Hz, a 21-dimensional feature prediction task was performed for the Leq for 12 sensors at this
frequency in August. The features included the previous one minute and the previous two minutes of
each frequency Leq value; the prediction results are shown in Figure 13.
In Figure 13, the X-axis represents the R2 . The Y-axis represents the RMSE and the prediction
results of 12 sensors at 125 Hz; R2 > 0.7 represents sensors 10, 8, and 3, among which sensor 3 had
the largest R2 value and the smallest RMSE value. This indicates that input characteristics and Leq
values are highly influence. In addition, the RMSE values of the 12 sensors were all below 1 dBA,
indicating that the difference between the predicted Leq values and the actual Leq values was minute.
Therefore, the R2 value was mainly used as an indicator to judge the quality of the prediction results.
Observing the prediction results of sensor 1, not only was the R2 only 0.0643, but the RMSE was within
0.75 Sci.
Appl. dBA and
2020, 10,the input
x FOR features
PEER REVIEWwere almost unrelated to the Leq values. 12 of 23

Figure 12.
Figure Different sampling
12. Different sampling times
times vs.
vs. predicted
predicted effect.
effect.

The X-axis in Figure 12 is the R2 [30], which represents the degree of curve fit between the
predicted value and the actual value (sensor 8 at 125 Hz in August). The R2 value is distributed in the
range of 0–1, and values closer to 1 indicate better prediction performance; otherwise, the prediction
performance worsens. The Y axis is the RMSE [31]. Here, the higher the value indicates a worse
prediction result; otherwise, the convergence is smaller. We found that a sampling time of 1 min and
Appl.
Appl. Sci. 2020, 10, 6619
x FOR PEER REVIEW 12 of 21
Appl. Sci. 2020, 10, x FOR PEER REVIEW 13 of 23

Figure 13. Predicted effect at 125 Hz for different sensors.

In Figure 13, the X-axis represents the R2. The Y-axis represents the RMSE and the prediction
results of 12 sensors at 125 Hz; R2 > 0.7 represents sensors 10, 8, and 3, among which sensor 3 had the
largest R2 value and the smallest RMSE value. This indicates that input characteristics and Leq values
are highly influence. In addition, the RMSE values of the 12 sensors were all below 1 dBA, indicating
that the difference between the predicted Leq values and the actual Leq values was minute. Therefore,
the R2 value was mainly used as an indicator to judge the quality of the prediction results. Observing
the prediction results of Figure sensor
Figure 13. 1, not
13. only was
Predicted
Predicted effect
effect the
at R
125only
at 125
2
Hz 0.0643,
Hz for
for butsensors.
different
different the RMSE was within 0.75 dBA
and the input features were almost unrelated to the Leq values.
However,
However,
In Figure 13, wethe
we added
added multiple
multiple
X-axis linearthe
linear
represents regression
regression
R2. The (MLR) (MLR) experiments
Y-axis experiments
represents thetoRMSE
to compare
compare noise
andnoise prediction
the prediction
results with
results with
of 12the the
sensors proposed
proposed
at 125 approach,
approach,
Hz; R > as
2 as
0.7shown shown
in Figures
represents in
sensors Figures
14a10,
and 14a
8,15a. and
and This 15a. This
showswhich
3, among shows
that GBM that GBM
outperforms
sensor 3 had the
outperforms
MLR inRterms MLR
ofand inthe
both terms
index, 2
ofR both
, andindex,
RMSE Rwhile
, and RMSE while using thecharacteristics
full 21atdimensions at 125 Hz.
largest 2 value smallest RMSE 2
value. Thisusing the
indicates full
that21input
dimensions 125 Hz.
andAs shown
Leq in
values
As shown
Figure 14b, in
the Figure
results14b, the
clearly results
show clearly
that the show
GBM that
algorithmthe GBM algorithm
achieved higher R 2 and showed
achieved higher aR 2 and
good
are highly influence. In addition, the RMSE values of the 12 sensors were all below 1 dBA, indicating
showed
grasp
that the athe
good
ofdifference
trend grasp
and of thethe
trend
reference
between and of
value
predicted reference
noise value
fluctuations
Leq values and ofthe
noise fluctuations
atactual
each sensor,
Leq values at each
thus sensor,
it was
was morethus
minute. it was
effectively
Therefore,
more
and
the 2 effectively
Raccurate
value was than and
MLRaccurate
mainly than
in thisas
used an MLR
task. Thisinresult
indicator thisto task.
alsoThis
judge the result
shows thatalso
quality shows
although
of the thatworking
the
prediction although the working
environment
results. Observing
environment
was
the relativelywas
prediction relatively
stable,
results aofvery
sensorstable,
simple
1, notaprediction
very
only simple theprediction
was model notmodel
may 0.0643,
R2 only work may
but thenot
well. This
RMSEwork well.
explains
was This0.75
why
within explains
we used
dBA
whyGBM
the
and weinput
the used the GBM
prediction
featuresmodelprediction
were for this model
almost problem forinthis
unrelated tothe problem
theNSRRC. in the NSRRC.
Leq values.
However, we added multiple linear regression (MLR) experiments to compare noise prediction
results with the proposed approach, as shown in Figures 14a and 15a. This shows that GBM
outperforms MLR in terms of both index, R2, and RMSE while using the full 21 dimensions at 125 Hz.
As shown in Figure 14b, the results clearly show that the GBM algorithm achieved higher R2 and
showed a good grasp of the trend and reference value of noise fluctuations at each sensor, thus it was
more effectively and accurate than MLR in this task. This result also shows that although the working
environment was relatively stable, a very simple prediction model may not work well. This explains
why we used the GBM prediction model for this problem in the NSRRC.

(a) (b)

Figure 14.
Figure 14. (a)
(a) Performance
Performanceofofmultiple
multiplelinear
linearregression
regression (MLR)
(MLR) method.
method. (b)(b) Combine
Combine MLRMLR
andand
gradient
gradient model
boosting boosting model (GBM).
(GBM).
Appl. Sci. 2020, 10, 6619 13 of 21

Furthermore, we compared three different input features set with 21, 16, and five dimensions,
respectively, to investigate the impact of input factors. The conditions are shown in Table 1, and the
prediction
Appl. Sci. 2020,results
10, x FOR inPEER
Figure 15a–c. Then, we clearly found more dimensions were better in terms
REVIEW of
13 of 21
R2 . Then, it was easily found that the input 21-dimensional data had better performance, as shown
in Figure 15d. Therefore,
Furthermore, we comparedthe results
three showed
differentthat
inputthefeatures
full dimensional
set with 21, feature performed
16, and the best,
five dimensions,
indicating that
respectively, tothe historicalthe
investigate features
impact may have more
of input information
factors. than the
The conditions aretime features.
shown Thus,
in Table 1, the
andmore
the
feature dimensions are input in the GBM enables efficient identification of
prediction results in Figure 15a–c. Then, we clearly found more dimensions were better in terms of suitable training features in
Rresponse
2 . Then, ittowasreliable
easilyprediction
found thatresults. Moreover,
the input this study
21-dimensional dataindicates thatperformance,
had better the loudest location
as shown near
in
the working
Figure environment
15d. Therefore, thewas sensor
results 8, and that
showed 125 Hz
thewasfullthe most serious
dimensional harmful
feature frequency.
performed theFor the
best,
practical issues,
indicating that thewehistorical
could pre-improve
features maythe low
havefrequency pump surrounding
more information sensor
than the time 8 by using
features. Thus,sound
the
insulators
more featureor remind
dimensions workers to prevent
are input in thelong-term
GBM enablesexposure in that
efficient area.
identification of suitable training
featuresWe in counted
response the to
prediction results of all
reliable prediction sensors
results. with a R2this
Moreover, greater
studythan 0.7 at all
indicates frequencies,
that the loudest as
shown in Table 2. Taking sensor 2 as an example, the R 2 of noise with a frequency of 500 Hz and with a
location near the working environment was sensor 8, and 125 Hz was the most serious harmful
frequency of
frequency. For1 kHz
the were bothissues,
practical greater we
thancould
0.7 (marked with anthe
pre-improve asterisk). Other sensors
low frequency pumpcan be deduced
surrounding
by analogy from the prediction results at different frequencies, as shown
sensor 8 by using sound insulators or remind workers to prevent long-term exposure in that area. in Figure 16.

(a) (b)

(c) (d)

Figure
Figure15.
15.(a)
(a)Performance
Performanceofofinput
inputfeatures
featuresset
setwith
with21 21dimensions.
dimensions.(b)(b)Performance
Performanceofofinput
inputfeatures
features
set with 16 dimensions. (c) Performance of input features set with five dimensions. (d) Performance
set with 16 dimensions. (c) Performance of input features set with five dimensions. (d) Performance of
of input
input features
features setset with
with three
three combined
combined kinds
kinds of of different
different dimensions.
dimensions.

We counted the prediction results of all sensors with a R2 greater than 0.7 at all frequencies, as
shown in Table 2. Taking sensor 2 as an example, the R2 of noise with a frequency of 500 Hz and with
a frequency of 1 kHz were both greater than 0.7 (marked with an asterisk). Other sensors can be
deduced by analogy from the prediction results at different frequencies, as shown in Figure 16.
Appl. Sci. 2020, 10, 6619 14 of 21

Here, we focused on the noise value of 125 Hz as its dBA values reached levels that are harmful to
the human body [15–18]. Moreover, we found that sensors 3, 8, and 10 achieved favorable prediction
performance, as shown in Table 2. Among them, the prediction performance of sensor 3 from 125 Hz
to 4 kHz showed a R2 greater than 0.7, and these noise frequencies were coherent with one another.
Likewise, the sensor 7 noise frequencies between 2 kHz and 16 kHz were coherent with one another.
As a result, we found that while the sensor prediction index R2 of this frequency was above 0.7,
the values near this frequency could also produce excellent prediction results (for example, 500 Hz and
1000 Hz of sensor 2, 125 Hz to 4 kHz of sensor 3, 500 Hz and 1000 Hz of sensor 4, and 500 Hz and
1000 Hz of sensor 6). Thus, we found mutual influence between similar frequencies and hypothesized
that the noise sources of similar frequencies were likely to have very similar occurrence conditions.
The coefficient of determination (R2 ) of the sensor was higher than 0.7, and the root-mean-square-error
(RMSE) was less than 1 dBA. This indicates that the proposed model could accurately predict the
trends of future Leq levels with an average error margin within 1 dBA. Therefore, we successfully
completed predictions for all sensors at other noise frequencies, and derived an effective reference
value for improving future prediction accuracy.

Table 2. All sensors with R2 greater than 0.7 at all frequencies.

Frequency
125 250 500 1k 2k 4k 8k 16k
(Hz)
Sensor1
(R2 > 0.7)
Sensor2
F F
(R2 > 0.7)
Sensor3
F F F F F F
(R2 > 0.7)
Sensor4
F F
(R2 > 0.7)
Sensor5
(R2 > 0.7)
Sensor6
F F
(R2 > 0.7)
Sensor7
F F F F
(R2 > 0.7)
Sensor 8
F F
(R2 > 0.7)
Sensor9
(R2 > 0.7)
Sensor10
F F F F
(R2 > 0.7)
Sensor11
F
(R2 > 0.7)
Sensor12
(R2 > 0.7)
F The R2 of noise with a frequency greater than 0.7.
Appl. Sci. 2020, 10, 6619 15 of 21
Appl. Sci. 2020, 10, x FOR PEER REVIEW 14 of 21

Figure16.
Figure Differentsensor
16.Different sensorprediction
predictioneffects
effects at
at different
different frequencies.

4. Conclusions
In this study, we found that as frequencies increased, the average Leq error values between
morning and night were greater, with noise in the morning returning higher and greater values than
those at night. This may be due to the fact that more people work in the morning and more noise is
generated. Moreover, the human voice has a high noise frequency, whereas machine pumps have a
noise lower frequency, indicating significant differences in noise sources.
This study focused on the prediction results for the noise frequency for one of twelve sensors
(sensor 8) at 125 Hz. This sensor was chosen because its static Leq value (>70 dBA) reached the
threshold of damaging human hearing, which affects physical and mental health. Based on this finding,
Appl. Sci. 2020, 10, 6619 16 of 21

we used the GBM model to predict future noise data. The Leq prediction results for sensor 8 at 125 Hz
showed an error rate of less than 1 dBA and a R2 value greater than 0.7, which is a favorable prediction
performance result. The poorer prediction results of the other sensors were between 2 to 1 dBA with a
R2 value that was generally below 0.7.
The results indicate that the prediction model worked well in most regions and frequencies
and particularly for sensor 8 (125 Hz), which is a serious noise zone. The results also indicate that
this working environment produced good noise prediction performance using the proposed method.
This enables the notification of laborers to prevent long-term exposure while predicting future noise
pollution. In fact, we are now collecting more characteristic data for several months for this purpose.
We believe that it would be better to have a longer observation duration to predict detailed noise
location. This would keep employees healthy for avoiding a harmful noise position to prevent people
from working in that environment. In the future, we will analyze the data structures of the noise
frequencies of more sensors, discuss noise types, and analyze the possibilities of noise-related physical
harm. We will also attempt to add new features to improve noise prediction performance.

Author Contributions: P.-J.W. wrote and conducted the experiments, analyzed the results, and organized the
layout of the paper, figures, and tables. C.H. conceived and corrected the article and was responsible for making
clear and understandable content. All authors reviewed the manuscript. All authors have read and agreed to the
published version of the manuscript.
Funding: This research was funded by National Synchrotron Radiation Research Center (NSRRC), grant
number 10811SAO01.
Acknowledgments: The authors acknowledge the financial support and equipment provided by the National
Synchrotron Radiation Research Center (NSRRC), Taiwan. We gratefully acknowledge the software support of
PTCOM Technology Co. Ltd.
Conflicts of Interest: The authors declare no conflict of interest.
Appl. Sci. 2020, 10, 6619 17 of 21
Appl. Sci. 2020, 10, x FOR PEER REVIEW 17 of 21

Appendix
Appendix A A
Average
Average at different
at different frequencies
frequencies in different
in different months
months forfor differentsensors.
different sensors.
Appl. Sci. 2020, 10, 6619 18 of 21
Appl. Sci. 2020, 10, x FOR PEER REVIEW 18 of 21

Appendix B
Appendix B
Daily average at different frequencies per week for different sensors.
Daily average at different frequencies per week for different sensors.
Appl. Sci. 2020, 10, 6619 19 of 21
Appl. Sci. 2020, 10, x FOR PEER REVIEW 19 of 21

Appendix
Appendix CC
Average
Average morning
morning vs.vs. night
night in in a week
a week forfor
allall sensors.
sensors.
Appl. Sci. 2020, 10, 6619 20 of 21

References
1. Goines, L.; Hagler, L. Noise pollution: A modern plague. South. Med. J. 2007, 100, 287–294. [CrossRef]
2. Pirrera, S.; Valck, E.D.; Cluydts, R. Nocturnal road traffic noise: A review on its assessment and consequences
on sleep and health. Environ. Int. 2010, 36, 492–498. [CrossRef] [PubMed]
3. Effects of Noise on Health. Available online: https://ncs.epa.gov.tw/noise/B-04-01.html (accessed on
25 May 2020).
4. Shen, D.H.; Wu, C.M.; Du, J.C. Application of grey model to predict acoustical properties and tire/road
noise on asphalt pavement. In Proceedings of the 2006 IEEE Intelligent Transportation Systems Conference,
Toronto, ON, Canada, 17–20 September 2006; pp. 175–180.
5. Cheng, Y.; Ying, C. Simplifying prediction method for traffic noise based on FHWA traffic noise model.
In Proceedings of the 2011 International Symposium on Water Resource and Environmental Protection,
Xi’an, China, 20–22 May 2011; pp. 2665–2667.
6. Zhang, R.; Wang, H.I. Nonlinear prediction of gross industrial output time series by Gradient Boosting.
In Proceedings of the 2011 IEEE 18th International Conference on Industrial Engineering and Engineering
Management, Changchun, China, 3–5 September 2011; pp. 153–156.
7. Sangani, D.; Erickson, K.; Hasan, M.A. Predicting Zillow estimation error using linear regression and gradient
boosting. In Proceedings of the 2017 IEEE 14th International Conference on Mobile Ad Hoc and Sensor
Systems (MASS), Orlando, FL, USA, 22–25 October 2017; pp. 530–534.
8. Lee, M.; Lin, L.; Chen, C.Y.; Tsao, Y.; Yao, T.H.; Fei, M.H.; Fang, S.H. Forecasting Air Quality in Taiwan by
Using Machine Learning. Sci. Rep. 2020, 10, 4153. [CrossRef] [PubMed]
9. Friedman, J.H. Greedy function approximation: A gradient boosting machine. Ann. Stat. 2001, 29, 1189–1232.
[CrossRef]
10. Islam, S.; Kalita, K. Assessment of traffic noise in Guwahati city, India. Int. Res. J. Eng. Technol. 2017, 4,
3335–3339.
11. Benocci, R.; Bellucci, P.; Peruzzi, L.; Bisceglie, A.; Angelini, F.; Confalonieri, C.; Zambon, G. Dynamic noise
mapping in the Suburban area of Rome (Italy). Environments 2019, 6, 79. [CrossRef]
12. Garcia, J.S.; Solano, J.J.P.; Serrano, M.C.; Camba, E.A.N.; Castell, S.F.; Asensi, A.S.; Suay, F.M. Spatial statistical
analysis of urban noise data from a WASN gathered by an IoT system: Application to a small city. Appl. Sci.
2016, 6, 380. [CrossRef]
13. Chang, T.Y.; Beelen, R.; Li, S.F.; Chen, T.I.; Lin, Y.J.; Bao, B.Y.; Liu, C.S. Road traffic noise frequency and
prevalent hypertension in Taichung, Taiwan: A cross-sectional study. Environ. Health 2014, 13, 37. [CrossRef]
14. Subramaniam, M.; Hassan, M.Z.; Sadali, M.F.; Ibrahim, I.; Daud, M.Y.; Aziz, S.A.; Samsudin, N.; Sarip, S.
Evaluation and analysis of noise pollution in the manufacturing industry. J. Phys. Conf. Ser. 2019, 1150, 012019.
[CrossRef]
15. Baliatsas, C.; Kamp, I.V.; Poll, R.V.; Yzermans, J. Health effects from low-frequency noise and infrasound in
the general population: Is it time to listen? A systematic review of observational studies. Sci. Total Environ.
2016, 557–558, 163–169. [CrossRef]
16. Lee, H.P.; Wang, Z.; Lim, K.M. Assessment of noise from equipment and processes at construction sites.
Build. Acoust. 2017, 24, 21–34. [CrossRef]
17. Reybrouck, M.; Podlipniak, P.; Welch, D. Music and Noise: Same or Different? What Our Body Tells Us.
Front. Psychol. 2019, 10, 1153. [CrossRef] [PubMed]
18. Liu, C.; Ding, D.; Zhu, Y.; Wang, H.; Cheng, X.; Zhao, Z.; Cao, J.; Zhai, S.; Yu, N. Auditory characteristics of
noise-exposed memberscrossing age-related groups. J. Otol. 2018, 13, 75–79.
19. Fang, S.H.; Chang, W.H.; Tsao, Y.; Shih, H.C.; Wang, C. Channel State Reconstruction Using Multilevel
Discrete Wavelet Transform for Improved Fingerprinting-Based Indoor Localization. IEEE Sens. J. 2016, 16,
7784–7791. [CrossRef]
20. Fang, S.H.; Yang, Y.H.S. The Impact of Weather Condition on Radio-based Distance Estimation: A Case Study
in GSM Networks with Mobile Measurements. IEEE Trans. Veh. Technol. 2016, 65, 6444–6453. [CrossRef]
21. Cheng, J.; Li, G.; Chen, X. Research on Travel Time Prediction Model of Freeway Based on Gradient Boosting
Decision Tree. IEEE Access 2019, 7, 7466–7480. [CrossRef]
22. Zheng, H.; Wu, Y. A XGBoost Model with Weather Similarity Analysis and Feature Engineering for Short-Term
Wind Power Forecasting. Appl. Sci. 2019, 9, 3019. [CrossRef]
Appl. Sci. 2020, 10, 6619 21 of 21

23. Iskandaryan, D.; Ramos, F.; Trilles, S. Air Quality Prediction in Smart Cities Using Machine Learning
Technologies Based on Sensor Data: A Review. Appl. Sci. 2020, 10, 2401. [CrossRef]
24. Wu, W.; Jiang, S.; Liu, R.; Jin, W.; Ma, C. Economic development, demographic characteristics, road network
and traffic accidents in Zhongshan, China: Gradient boosting decision tree model. Transp. A Transp. Sci.
2020, 16, 359–387. [CrossRef]
25. Grosveld, F.W. Prediction of Broadband Noise from Horizontal Axis Wind Turbines. J. Propuls. 1984, 1,
292–299. [CrossRef]
26. Kalapanidas, E.; Avouris, N.; Craciun, M.; Neagu, D. Machine Learning algorithms: A study on noise sensitivity.
In Proceedings of the 1st Balkan Conference in Informatics, Thessaloniki, Greece, 21–23 November 2003;
pp. 356–365.
27. White, G.C.; Bennetts, R.E. Analysis of frequency count data using the negative binomial distribution. Ecology
1996, 77, 2549–2557. [CrossRef]
28. Matheson, I.B.C. A critical comparison of least absolute deviation fitting (robust) and least squares fitting:
The importance of error distributions. Comput. Chem. 1990, 14, 49–57. [CrossRef]
29. Buckley, J.; James, L. Linear regression with censored data. Biometrika 1979, 66, 429–436. [CrossRef]
30. Nagelkerke, N.J.D. A note on a general definition of the coefficient of determination. Biometrika 1991, 78,
691–692. [CrossRef]
31. Willmott, C.J.; Matsuura, K. Advantages of the mean absolute error (MAE) over the root mean square error
(RMSE) in assessing average model performance. Clim. Res. 2005, 30, 79–82. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).