IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 2, April 2025, pp. 1150~1163

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i2.pp1150-1163  1150

Performance assessment of time series forecasting models for

simple network management protocol-based hypervisor data

Yuggo Afrianto1,2, Rendy Munadi1, Setyorini1, Toto Widyanto3

1
School of Computing, Faculty of Informatic, Telkom University, Bandung, Indonesia
2
Department of Informatics Engineering, Ibn Khaldun University, Bogor, Indonesia
3
Informatics Engineering Study Program, Institut Sains and Technology Al Kamal, Jakarta, Indonesia

Article Info ABSTRACT

Article history: Time series forecasting is vital for predicting trends based on historical data,
enabling businesses to optimize decisions and operations. This paper
Received Feb 10, 2024 evaluates forecasting models for predicting trends in simple network
Revised Nov 15, 2024 management protocol (SNMP)-based hypervisor data, essential for resource
Accepted Nov 24, 2024 allocation in cloud data centers. Addressing non-stationary data and dynamic
workloads, we use PyCaret to compare classical models like autoregressive
integrated moving average (ARIMA) with advanced methods such as auto
Keywords: ARIMA. We assess 30 models on metrics including CPU utilization,
memory usage, and disk reads, using synthetic and real-time datasets.
Hypervisor Results show the naive forecaster model excels in CPU and disk read
Machine learning predictions, achieving low root mean squared errors (RMSE) of 0.71 and
Pycaret 869,403.35 for monthly and daily datasets. For memory usage predictions,
Simple network management gradient boosting with conditional deseasonalisation and detrending
protocol outperforms others, recording the lowest RMSE of 679,917.6 and mean
Time series forecasting absolute scaled error (MASE) of 4.46 on weekly datasets. Gradient boosting
consistently improves accuracy across metrics and datasets, especially for
complex patterns with seasonality and trends. These findings suggest
integrating gradient boosting and naive forecaster models into cloud system
architectures can enhance service quality and operational efficiency through
improved predictive accuracy and resource management.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Rendy Munadi
School of Computing, Faculty of Informatic, Telkom University
40257 Buah Batu, Bandung, Indonesia
Email: [email protected]

1. INTRODUCTION
The field of cloud computing has experienced rapid growth, with numerous virtual machines now
operating within physical server environments [1]. As demand for virtualization technology increases,
effective time-series forecasting methods are essential, particularly for managing simple network management
protocol (SNMP)-based hypervisor data [2]. Such data supports tracking and managing performance across
virtual instances deployed in the cloud. However, time-series forecasting has certain weaknesses when dealing
with 'non-stationary' data, as it often struggles to model dynamic and volatile workloads. This challenge limits
the adaptability of forecasting methods when faced with unpredictable, fluctuating workloads, especially in
SNMP-based hypervisor contexts, where these limitations pose significant obstacles to system administrators
seeking to optimise resource management in cloud data centers [3].
Accurate workload prediction in cloud computing is crucial for optimising resource allocation and
operational efficiency. Forecasting methods are generally divided into two main categories: those based on

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  1151

fundamental time-series dynamics and those using machine learning and artificial neural networks [4].
A range of methods has been developed to address workload prediction challenges, including classical
statistical approaches such as autoregressive (AR), moving average (MA), and the autoregressive integrated
moving average (ARIMA) model [5], as well as more sophisticated models such as seasonal autoregressive
integrated moving average (SARIMA) [6], [7]. Additionally, contemporary techniques like multiple linear
regression (MLR), ridge regression (RR) [8], and adaptive neuro-fuzzy inference systems (ANFIS) [9] have
also been applied. Recent research by Kumar and Singh [10] have indicated that auto ARIMA is a promising
approach, offering high prediction accuracy for web server workloads and demonstrating its potential to
proactively optimise resource allocation. However, traditional models such as auto ARIMA often struggle
with the complexity and dynamics of cloud data centre environments, characterised by rapidly evolving data
variability. This limitation is particularly evident over longer forecasting horizons and during periods of
significant data fluctuation, underscoring the need for models that are more adaptive to changing conditions.
In recent years, automated machine learning (AutoML) frameworks, such as PyCaret, have emerged
as powerful tools to simplify model selection, evaluation, and optimisation. PyCaret is a software tool that
integrates a range of machine learning algorithms and time-series forecasting models within a Python
wrapper, which includes assemblies of several machine learning frameworks, such as scikit-learn, XGBoost,
LightGBM, CatBoost, spaCy, Optuna, Hyperopt, and Ray. Studies indicate that PyCaret’s capabilities
significantly enhance implementation, versatility, and customisation [11]. This tool has proven influential in
time-series analysis and forecasting across multiple contexts [12]. Its applications range from forecasting
trends in the COVID-19 pandemic [13] to data collection and formatting processes for various forecasting
projects [14]. However, the potential for hypervisor management in the cloud computing domain with actual
and synthetic SNMP-based hypervisor data still needs to be explored.
This research aims to evaluate the efficacy of 30 distinct time-series forecasting models using
PyCaret on variety of authentic and synthetic datasets derived from SNMP-based hypervisor systems. The
objective is to leverage the PyCaret toolkit to identify optimal forecasting methodologies for predicting
hypervisor components accurately, specifically CPU utilization, memory utilization, and the number of disk
reads. The findings of this study will support organisations in making informed decisions regarding resource
allocation and overall operational efficiency in cloud environments. The primary contributions of this paper
are as follows: first, the application of time-series forecasting models to SNMP hypervisor data through
PyCaret, followed by analysis of the time series windows and forecasting models that yielded the best results;
second, an evaluation of resource management approaches in light of the integrated forecasting measures.

2. METHOD
This study employs a machine learning approach to evaluate the efficacy of a time series forecasting
model utilizing SNMP-based hypervisor data. The research was conducted using the stages illustrated in
Figure 1. This section will discuss the dataset, experimental design, and model evaluation employed in this
study. Furthermore, Figure 1 illustrates the essential steps for utilizing Pycaret to evaluate the model. The
framework outlines procedures for acquiring load data from synthetic and real-time databases, providing an
overview of the Pycaret interface for model evaluation. This systematic approach facilitates a comprehensive
understanding of the workflow and tasks designed to ensure practical evaluation and comparison of multiple
forecasting models in performance.

2.1. Dataset
This research employed a time series dataset, as detailed in Table 1 of the research paper. The
dataset is divided into two principal categories: synthesis data, which is simulated data, and real-time data,
which is obtained directly from the source. This information is crucial for accurately interpreting the data
analyzed in the study and verifying the forecasting models [15].
‒ Synthesis: the scenario-based system generated a dataset containing CPU load, memory usage, and small
computer system interface (SCSI) disk bytes as variables collected at different times.
‒ Real-time: a few of the datasets available include CPU utilization, memory usage, and SCSI disk bytes
obtained for one month, one week, and one day, respectively.
Synthetic data is associated with a robust correlation to elevated prediction accuracy, as the
generation of synthetic data adheres to a desired pattern. In contrast, data from real-time sources typically
exhibit a weaker correlation with forecasting accuracy, as numerous external factors beyond an individual's
control can influence their nature. The following characteristics are present in the dataset: The initial step is
to evaluate CPU utilization for the monitored device by examining the CPU idle value. A low CPU idle value
may indicate that the device operates under a high load. Secondly, to assess the memory load of the
monitored device, a low memory-free value has the same significance, whereby a high load may be placed on

Performance assessment of time series forecasting models for simple network … (Yuggo Afrianto)
1152  ISSN: 2252-8938

the device. Thirdly, through SCSI disk bytes, to assess the number of bytes written to the SCSI disk on the
monitored device, a high byte count may suggest that the device is busy.

Figure 1. Framework assesment

Table 1. Time series dataset

Dataset Metrics Workload description
Synthetic and realtime CPU utilizations Percentage of CPUs Idle
(month, week, day) Memory usage Percentage of memory free
Combined SCSI disk bytes Bytes read

2.2. Experimental setup

2.2.1. Hypervisor workload forecasting
The dataset employed in this research is centred on data extracted from SNMP sources, obtained
from network management servers and illustrative of the diverse workloads observed in cloud data centres.
The dataset is employed to evaluate the forecasting models across a range of prediction windows, thereby
facilitating the assessment of their accuracy. Figure 2 illustrates the general cloud system architecture
incorporating a forecasting module. In this context, the resource manager receives input workload data from
the hypervisor data centre and analyses historical data from real-time or synthetic workloads to forecast cloud
resource management effectively.

Figure 2. Load forecasting architecture hypervisor [10]

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1150-1163

Int J Artif Intell ISSN: 2252-8938  1153

Equation for time-series analysis of data in SNMP is described by the variable Ý, which is a
collection of workload values over time t, where 𝑦𝑡 is the workload at time t. Function ƒ represents the
previous workload analysis used to estimate outcome of future events. This equation can be expressed as (1):

Ý𝑡+1 = ƒ(𝑦𝑡 , 𝑦𝑡−1 , . . . , 𝑦1 ) (1)

Where Ý𝑡+1 is represents the predicted workload at the next time step (t+1), ƒ is a function that uses the
previous workload values to estimate the future workload1, and 𝑦𝑡, 𝑦𝑡−1,…,𝑦1 are the workload values at
the current and previous time steps.
This approach is pertinent in the context of SNMP, where data collection is continuous. It is
noteworthy for its capacity to illustrate the evolution of the network system and the monitored devices with
greater clarity. However, it is important to recognise a potential limitation, namely that the predicted values
(Ý𝑡+1) must be validated against the actual values. This can be achieved by measuring the forecast error (𝑒𝑡+1 ).

2.2.2. Scenario of creating synthetic dataset

Synthetic datasets have utility in research as they are able to imitate scenarios that do not always
occur in real practice. In the evaluation of time-series forecasting models for SNMP-based hypervisor data
context, the synthetic dataset is employed as follows: i) model testing: the primary objective of synthetic
datasets is to facilitate model testing under controlled and repeatable scenarios. As discussed in [15], [16], the
use of synthetic data is a key aspect of optimising data-driven prognostic models and modelling time series
forecasting. In this manner, researchers are able to replicate simulated workloads that are otherwise
unfeasible in the real world for a variety of reasons; ii) performance analysis: the models are exercised on
synthetic workloads for the purpose of evaluating the forecast model's performance against some known
workloads. Different models are tested and best practices are recorded to ascertain which model would work
best under a particular situation; and iii) model validation: research is also conducted to test the effectiveness
of the synthetic techniques in forecasting models on workload that may exist in a real-world context. This
increases the confidence level of the applicability of the models across various parameters.
A single physical machine and four virtual ones (vm1, vm2, vm3, and vm4) were employed in the
execution of a series of test scenarios, the objective of which was the generation of a synthetic dataset. Each
test scenario was composed of two phases: a stress test phase and a pause phase. During the stress test, the
load code was used on the virtual machine to imitate extensive usage. During the pause phase, the virtual
machine was idle to recuperate. The scenario for creating a synthetic dataset is illustrated in Table 2.

Table 2. Scenario of creating synthetic dataset

Interval testing (minutes)
Load VM
0 10 20 30 40 50 60 70 80
vm1 40 10 40
vm2 10 30 20 30
vm3 20 20 30 20
vm4 30 10 40 10
Descriptions:
Stress test time
Pause time
Lag time between stress tests

2.2.3. Scenario of colecting real-time dataset

The real-time dataset collection process is illustrated in Figure 3 for this research. Figure 3 depicts
the system architecture and components involved in gathering real-time data for analysis. Real-time datasets
play an important role in research to understand, analyze, and optimize system performance based on actual
data obtained from the environment being monitored. Real-time datasets were collected using energy
monitoring and information systems. Workload energy represents the collection of energy data through
sensors. Workload CIT deals with monitoring components such as CPU, memory, and I/O units, with data
sent via TCP/IP (UDP), SNMP, the MySQL database, and Python. Cacti is a monitoring system designed for
network and system monitoring. It allows users to visualize and analyze performance metrics, track trends,
and generate reports. Database monitoring workload focuses on database performance metrics, analyzing
workloads, and accessing datasets for further evaluation. Both monitoring systems play crucial roles in
managing IT infrastructure components, providing valuable insights to maintain optimal performance and
troubleshoot issues.

Performance assessment of time series forecasting models for simple network … (Yuggo Afrianto)
1154  ISSN: 2252-8938

This research utilises the same dataset employed by Kumar and Singh [10] to facilitate a comparison
of the findings with the results of other literature. This comprises a real-world dataset of web server traces,
namely the number of hyper text transfer protocol (HTTP) requests for the NASA server, Calgary server, and
Saskatchewan server, which is used to predict web server workload. The experiments are performed with a 5,
10, 20, 30, and 60-minute duration period window (PWS).

CPU, Memory, I/O Unit,

Througput,System
Delay, Number
/ Content DB Of
Sensor Collector Switch Power Server
Connections

TCP/IP (UDP)

Figure 3. Collecting real-time dataset

2.3. Evaluation
2.3.1. Time series forecasting models
PyCaret is a machine learning framework that does not require the use of supplements and is
straightforward to use and functional, thereby making it accessible to those new to the discipline [17].
The framework assists users at each stage of the machine learning process, from data preparation to model
analysis and execution. The primary objective of analysing time-series data is to identify trends, which are
captured by pertinent statistics and characteristics of the data being handled. There are various methods of
forecasting time series, and this research utilised 30 models from the Pycaret library, as presented in Table 3.
Various studies have utilized PyCaret, a machine learning library, in various applications. These
include diabetes classification and prediction [18], intrusion detection system performance analysis on the
UNSW-NB15 dataset [19], hyperparameter tuning for image classification [20], and AutoML
implementation on the PowerBI application [21]. PyCaret was also used in the evaluation of the predictive
power of multiple regression models for groundwater contamination [22], intrusion detection using
supervised and unsupervised learning methods on the Cicids 2017 dataset [23], machine learning-based
network-based intrusion detection system (NIDS) analysis and modelling for IoT networks [24], and URL
detection for phishing websites [25]. This research demonstrates the versatility and capabilities of PyCaret in
various fields and contexts.

2.3.2. Performance measure indicators

It is challenging to comprehend the underlying causes of individual forecasting errors due to the
complex and multifaceted nature of the process, which is not merely a consequence of a single numerical
value. A number of methods have been proposed in the literature for the evaluation of prediction models and
their relationships. Each of these methods has its own merits and demerits. For example, the root mean
squared error (RMSE) is frequently declared to be influenced heavily by outlying values. In this instance,
three measures were used for the evaluation of the model's predictive accuracy: the RMSE, the mean absolute
error (MAE), and the mean absolute scaled error (MASE) [10]. The time required for processing each
forecasting model and making predictions is TT.

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1150-1163

Int J Artif Intell ISSN: 2252-8938  1155

Table 3. Time series forecasting models

Name Reference
Naive Forecaster sktime.forecasting.naive.NaiveForecaster
Grand Means Forecaster sktime.forecasting.naive.NaiveForecaster
Seasonal Naive Forecaster sktime.forecasting.naive.NaiveForecaster
Polynomial Trend Forecaster sktime.forecasting.trend.PolynomialTrendForeca...
ARIMA sktime.forecasting.arima.ARIMA
Auto ARIMA sktime.forecasting.arima.AutoARIMA
Exponential Smoothing sktime.forecasting.exp_smoothing. ExponentialSm...
ETS sktime.forecasting.ets.Auto ETS
Theta Forecaster sktime.forecasting.theta.ThetaForecaster
STLF sktime.forecasting.trend.STLForecaster
Croston sktime.forecasting.croston.Croston
BATS sktime.forecasting.bats.BATS
TBATS sktime.forecasting.tbats.TBATS
Linear w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Elastic Net w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Ridge w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Lasso w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Lasso Least Angular Regressor w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Bayesian Ridge w/ Cond. Deseasonalize and Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Huber w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Orthogonal Matching Pursuit w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
K Neighbors w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Decision Tree w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Random Forest w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Extra Trees w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Gradient Boosting w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
AdaBoost w/ Cond. Deseasonalize and Detrending pycaret.containers.models.time_series.BaseCdsD...
Extreme Gradient Boosting w/ Cond. Deseasonali... pycaret.containers.models.time_series.BaseCdsD...
Light Gradient Boosting w/ Cond. Deseasonalize... pycaret.containers.models.time_series.BaseCdsD...
CatBoost Regressor w/ Cond. Deseasonalize and De... pycaret.containers.models.time_series.BaseCdsD...

3. RESULTS AND DISCUSSION

The PyCaret time-series forecasting module is equipped with preprocessing features and offers a
variety of algorithms, including over 30 statistical, time-series, and machine learning models. Beyond model
training, it provides capabilities for automated hyperparameter tuning, ensembling, model analysis, and
deployment. The typical workflow in PyCaret involves five steps: setup, computation, analysis, prediction,
and model saving.

3.1. Analysis workload dataset

The results of all the workload dataset analysis were reproportioned with one of the weekly memory
free flow dataset workloads and the results showed a changing trend. The data were visualized in a plot that
displayed the flow size values on the y-axis and time on the x-axis, representing a type of time-series data.
The analyzed dataset covered the period from October 25 to November 22, 2023. The term "memory free" is
used to describe the amount of unutilised memory that is available for use within a system. In Figure 4 of the
research article, the authors present a memory free datasets trend analysis, which illustrates trends in the
amount of free memory over a specified time frame. The memory free flow trend analysis plot provides
insights into the fluctuations in memory free flow over time and any changes in the regular deposit of
memory. Memory trend analysis is a crucial aspect of system performance, resource and capacity planning in
the field of information technology. It enables the formulation of well-informed decisions regarding memory
management forms, performance enhancement and the deployment of resources in alignment with historical
trends and patterns.
This case study on hypervisor data analyzes multiple time series of diverse work metrics. A
statistical analysis is performed, as shown in Table 4, using the following attributes: i) test: name of the
statistical test applied; ii) test name: brief description of the statistical test; iii) data: the type of data tested in
the research; iv) property: properties of the data being tested; v) settings: settings used during the testing
process; and vi) value: the value of the statistical test result.
The time series characteristics are used to determine the most appropriate estimation method for each
time series. The prediction task focuses on several hypervisor performance metrics characteristic of a time
series. This summary outlines the approach to evaluating and selecting the most appropriate forecasting models
for time series analysis in various application contexts, using the following summary statistics tests [26]:

Performance assessment of time series forecasting models for simple network … (Yuggo Afrianto)
1156  ISSN: 2252-8938

Figure 4. Trend analysis of memory free datasets

Table 4. Statistical analysis of memory free datasets

Test Test Name Data Property Setting Value
Summary Statistics Transformed Length None 97
Summary Statistics Transformed # Missing values None 0
Summary Statistics Transformed Mean None 1.8765E+11
Summary Statistics Transformed Median None 2.10314E+11
Summary Statistics Transformed Standard deviation None 37482887512
Summary Statistics Transformed Variance None 1.40497E+21
Summary Statistics Transformed Kurtosis None -1.33902
Summary Statistics Transformed Skewness None -0.774785
Summary Statistics Transformed # Distinct values None 97
White Noise Ljung-Box Transformed Test statictic {'alpha': 0.05, 'K': 24} 982.632081
White Noise Ljung-Box Transformed Test statictic {'alpha': 0.05, 'K': 48} 1062.022633
White Noise Ljung-Box Transformed p-value {'alpha': 0.05, 'K': 24} 0
White Noise Ljung-Box Transformed p-value {'alpha': 0.05, 'K': 48} 0
White Noise Ljung-Box Transformed White noise {'alpha': 0.05, 'K': 24} FALSE
White Noise Ljung-Box Transformed White noise {'alpha': 0.05, 'K': 48} FALSE
Stationarity ADF Transformed Stationarity {'alpha': 0.05} FALSE
Stationarity ADF Transformed p-value {'alpha': 0.05} 0.873311
Stationarity ADF Transformed Test Statistic {'alpha': 0.05} -0.589619
Stationarity ADF Transformed Critical value 1% {'alpha': 0.05} -3.502705
Stationarity ADF Transformed Critical value 5% {'alpha': 0.05} -2.893158
Stationarity ADF Transformed Critical value 10% {'alpha': 0.05} -2.583637
Stationarity KPSS Transformed Trend stationarity {'alpha': 0.05} FALSE
Stationarity KPSS Transformed p-value {'alpha': 0.05} 0.01
Stationarity KPSS Transformed Test statistic {'alpha': 0.05} 0.304834
Stationarity KPSS Transformed Critical value 10% {'alpha': 0.05} 0.119
Stationarity KPSS Transformed Critical value 5% {'alpha': 0.05} 0.146
Stationarity KPSS Transformed Critical value 2.5% {'alpha': 0.05} 0.176
Stationarity KPSS Transformed Critical value 1% {'alpha': 0.05} 0.216
Normality Shapiro Transformed Normality {'alpha': 0.05} FALSE
Normality Shapiro Transformed p-value {'alpha': 0.05} 0

3.1.1. White noise test

To evaluate the randomness of the time series data and to confirm the absence of significant
autocorrelations, we performed a white noise test utilizing the Ljung-Box statistic. This test is essential for
determining whether the data is suitable for time series modeling or if it represents purely random noise without
any predictable patterns [27]. The following components were analyzed in this test:
‒ Ljung-box transformed test statistics: transformed Ljung-box test statistics.
‒ p-value: the p-value for the Ljung-box test.
‒ White noise: whether the data follows a white noise pattern.

3.1.2. Stationarity test

Determining the stationarity of time series data is essential before applying forecasting models, as
many time series techniques assume stationarity to produce reliable results. Non-stationary data can lead to
misleading inferences and poor predictive performance. To assess the stationarity of our dataset, we
conducted the augmented dickey-fuller (ADF) test, which evaluates the presence of a unit root in the data.
The following components were examined in this test:

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1150-1163

Int J Artif Intell ISSN: 2252-8938  1157

‒ Azure data factory transformed: data transformed with Azure data factory transformation.
‒ Stationarity: conclusion whether the data is stationary.
‒ p-value: the p-value for the ADF test.
‒ ADF transformed test statistics: transformed ADF test statistics.
‒ Critical value: critical value for the ADF test at a certain significance level.
‒ Trend: whether there is a trend in the data.

3.1.3. Normality test

Evaluating the normality of the data distribution is essential in time series analysis, as many
statistical methods assume that the underlying data is normally distributed. To determine if our dataset meets
this assumption, we conducted the Shapiro-Wilk test, which is effective for detecting departures from
normality in small to medium-sized samples. The following components were assessed in this test:
‒ Shapiro transformed: data transformed with Shapiro transformation.
‒ Normality: conclusion whether the data follows a normal distribution.
‒ p-value: the p-value for the Shapiro test.
The results of the statistical analysis are presented in Table 5. Reveals a consistent trend, indicating
that the disk model frequently exhibits white noise and stationary characteristics across all periods. In
contrast, normal distributions are infrequent, with only one instance observed for the memory model under
synthetic dataset conditions. This suggests that temporal changes may significantly impact hypervisor
performance metrics. The absence of normal distribution in real-time data suggests possible irregularities in
data homogeneity and symmetry over extended durations. To address these irregularities, we performed
further analysis and forecasting models to assess the performance of different forecasting models on the time
series data [28].

Table 5. Results of statistical analysis

Dataset Model White noise Stasioner Normal distributions
Monthly CPU False False False
Memory False False False
Disk True True False
Weekly CPU False False False
Memory False False False
Disk True True False
Daily CPU False True False
Memory False False False
Disk True True False
Syntetic CPU False False False
Memory False False True
Disk False True False

3.2. Analysis of forecasting performance

The performance of 30 forecasting models, which are time series in nature and are based on
hypervisor data, was evaluated using real and synthetic datasets. The models were monitored via SNMP, and
participants and/or caregivers, researchers, and models were assessed on key performance indices, including
RMSE, MAE, and MASE, across a daily, weekly, and monthly period. Furthermore, the time required to
process each model was evaluated in order to assess the efficiency of their computation.

3.2.1. Performance of idle CPU forecasting accuracy

A summary of the accuracy of the various models in predicting the CPU idle rate is provided
in Table 6. Additionally, Table 6 presents other evaluation metrics, including RMSE, MASE, and MAE,
which were employed to assess the performance of each model on the monthly, weekly, daily, and
synthetic datasets. The forecasting accuracy of the monthly dataset the 'naive forecaster' model exhibited the
lowest RMSE of 0.71, which was lower than that recorded by models 'auto ARIMA' and box-cox, ARMA
error, trend, and seasonality ('BATS'), where RMSE=1.19 and 1.17, respectively. However, it had lower
ranks of 11 and 12 out of the overall dataset. The weekly dataset 'exponential smoothing' and error, trend and
sea-sonal ('ETS') models have identical performance with an RMSE of 0.65 and rank 7, while
'theta forecaster' has a slightly higher RMSE of 0.66 and rank 9. The daily dataset model 'grand means
forecaster' shows the best performance with an RMSE of 0.68 and rank 4, followed by 'auto ARIMA' and
'ARIMA' with an RMSE of 0.73 and rank 5. The synthetic dataset was used to benchmark each real-time
dataset model 'AdaBoost w/ cond. Deseasonalize and detrending' ranked first with the best performance,

Performance assessment of time series forecasting models for simple network … (Yuggo Afrianto)
1158  ISSN: 2252-8938

followed by 'linear w/ cond. Deseasonalize and detrending' and 'ridge w/ cond. Deseasonalize and detrending'
ranked 2nd and 3rd respectively.
From the results listed in Table 6, it can be seen that the grand means forecaster model on the daily
dataset has the lowest RMSE and the 4th rank, indicating good prediction performance. On the monthly
dataset, the naive forecaster model has the lowest RMSE but is ranked lower. On the synthetic dataset, the
AdaBoost w/ cond. Deseasonalize and detrending ranked first with the best performance. This information is
essential for selecting the best model for forecasting the availability of unused CPUs to help plan and manage
system resources efficiently. Overall, the daily dataset and the grand means forecaster model with lower
RMSE, MASE, and MAE and close to the synthetic ranking value show better CPU forecasting performance.

Table 6. Performance accuracy forecasing of CPU idle

Dataset Model RMSE MASE MAE Ranking
Monthly Naive forecaster 0.71 2.56 0.56 10
Auto ARIMA 1.19 5.36 1.06 11
BATS 1.17 5.41 1.07 12
Weekly Exponential smoothing 0.65 2.03 0.54 7
ETS 0.65 2.03 0.54 7
Theta forecaster 0.66 2.06 0.55 9
Daily Grand means forecaster 0.68 1.36 0.58 4
Auto ARIMA 0.73 1.47 0.63 5
ARIMA 0.73 1.47 0.63 5
Syntetic AdaBoost w/ Cond. deseasonalize and detrending 0.59 0.93 0.51 1
Linear w/ Cond. deseasonalize and detrending 0.61 0.93 0.51 2
Ridge w/ Cond. deseasonalize and detrending 0.61 0.93 0.52 3

3.2.2. Performance of free memory forecasting accuracy

The memory free forecasting accuracy performance on the monthly dataset model 'Huber w/ cond.
The deseasonalize and detrending overall, daily dataset, and grand means forecaster model with lower
RMSE, MASE, and MAE and close to the synthetic ranking value show better CPU forecasting performance.
model had a very high RMSE, indicating a less-than-optimal performance. The extra trees with cond.
deseasonalize and detrending had the highest MASE, and the model rank ranged from 10 to 12 out of the
entire dataset. The weekly theta forecaster model has a significantly lower error value than the BATS and
TBATS models, which have identical error values but different rankings of eight for BATS and seven
TBATS. Daily dataset gradient boosting w/ cond. Deseasonalize and detrending shows the best performance
in all error metrics among the daily dataset models, while 'Bayesian ridge w/ cond. Deseasonalize and
detrending' performed below it but still ranked 6th overall. Synthetic dataset the 'exponential smoothing'
model performed the best with the lowest error in all metrics among the synthetic dataset models, while the
'ETS' model performed below it but still ranked 2 nd overall. The results of the accuracy of the different
models in predicting memory free as shown in Table 7.
From the results listed in Table 7, it can be seen that the Huber w/ cond. model. Deseasonalize and
detrending on the monthly dataset has the highest RMSE and ranks 10 th, indicating less than optimal
prediction performance. On the weekly dataset, the theta forecaster model has a low MASE and ranks 7 th,
indicating better performance than the BATS model. The analysis from Table 7 provides insight into which
models are most effective in predicting memory-free rates at various data frequencies. Overall, the daily
dataset and gradient boosting are done using cond. Deseasonalizatione and detrending with lower RMSE,
MASE, and MAE and close to the synthetic rank value show better forecasting memory performance.

Tabel 7. Performance accuracy forecasing of memory free

Dataset Model RMSE MASE MAE Ranking
Monthly Huber w/ Cond. Deseasonalize and Detrending 3271875524.78 6.61 2731561697.07 10
Auto ARIMA 3767137446.90 7.56 3135176015.51 11
Extra trees w/ Cond. Deseasonalize and Detrending 4421809533.71 8.88 3697388235.58 12
Weekly Theta forecaster 1347869846.82 0.94 1164727393.11 7
BATS 1710867790.86 1.14 1434401664.64 8
TBATS 1710867790.86 1.14 1434401664.64 8
Daily Gradient boosting w/ Cond. Deseasonalize and Detrending 679917621.37 4.46 562941495.97 4
CatBoost regressor w/ Cond. Deseasonalize and Detrending 817036128.95 5.20 656224572.32 5
Bayesian ridge w/ Cond. Deseasonalize and Detrending 845857410.17 5.37 676449659.13 6
Syntetic ETS 531650395.06 0.69 477258423.00 2
Exponential Smoothing 527319448.26 0.69 480880807.52 1
Bayesian Ridge w/ Cond. Deseasonalize and Detrending 565535362.33 0.70 488370640.92 3

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1150-1163

Int J Artif Intell ISSN: 2252-8938  1159

3.2.3. Performance accuracy forecasing of disk read

The forecasting accuracy performance of the disk read dataset on the monthly dataset; naive
forecaster model has better forecasting performance with the lowest rank compared to the decision tree w/
cond model. Desesasonalize and detrending and BATS models. However, the monthly dataset was not
superior to the weekly and daily datasets. The Weekly dataset in the naive forecaster model has better
forecasting performance than the extra and decision trees with cond. Desesasonalize and detrending with.
However, the monthly dataset wais not superior to the weekly and daily datasets. The daily dataset and the
naive forecaster model performed best by excelling in all monthly and weekly models and datasets. The
results of the accuracy of the different models in predicting disk read as shown in Table 8.

Table 8. Performance accuracy forecasing of disk read

Dataset disk read Model RMSE MASE MAE Ranking
Monthly Decision Tree w/ Cond. Deseasonalize and Detrending 882218.20 6.23 384421.61 9
Naive Forecaster 869403.35 6.27 386646.91 7
BATS 873858.07 6.23 385887.01 8
Weekly Naive Forecaster 36034.60 0.12 23248.43 4
Extra Trees w/ Cond. Deseasonalize and Detrending 36923.95 0.16 29859.68 5
Decision Tree w/ Cond. Deseasonalize and Detrending 36923.95 0.16 29859.68 5
Daily Naive Forecaster 15855.61 0.52 13257.11 1
Croston 16753.33 0.59 15243.52 2
Grand Means Forecaster 20927.18 0.73 19268.03 3
Syntetic Grand Means Forecaster 5216888.90 0.73 4039774.44 12
TBATS 5001328.06 0.76 4189157.24 10
BATS 5026264.73 0.77 4204561.59 11

Overall, the daily dataset and naive forecaster model were the best performing datasets and models.
This model achieved the lowest rank among all tested models, indicating a smaller prediction error rate and
more accurate performance. The analysis of this table provides a deeper understanding of which models are
most effective in predicting disk read rates at various data frequencies. However, the synthetic dataset had the
highest rank because it had a large prediction residual value compared with the other datasets, as shown in
Figure 5.

Figure 5. Residual forecast disk read syntetic data

3.2.4. Performance of forecasting process time efficiency

The time efficiency of various forecasting models is shown in Table 9, presenting the processing
times for predicting CPU, memory, and disk read performance. Table 9 includes the most accurate
forecasting models and datasets from the performance tests. The results indicate that the best dataset for
forecasting hypervisor resource load is the daily dataset across multiple forecasting models. The naive
forecaster model is the most efficient for forecasting CPU performance, with a runtime of 1.71 seconds,
aligning with Yoo and Sim [7] findings that simple models often perform well in highly volatile scenarios.
The naive forecaster makes predictions using straightforward strategies [29]. For memory datasets, gradient
boosting w/ conditional deseasonalization and detrending demonstrates the best performance, with a
processing time of 0.13 seconds.

Performance assessment of time series forecasting models for simple network … (Yuggo Afrianto)
1160  ISSN: 2252-8938

Our results align with existing research on the effectiveness of framework for comparing accuracy
of time-series forecasting methods. Similar to our findings, a study by Hyndman and Athanasopoulos [30]
also identified gradient boosting had the highest accuracy in gradient boosting, which is often used in
machine learning competitions, although it was lower than naive, a simple prediction method. This reinforces
the notion that gradient boosting ability to handle complex relationships within data can be advantageous for
resource load prediction.
The gradient boosting technique was proposed as a one-of-a-kind applied gradient boosting
machine, particularly for regression and classification trees. The “boosting” concept is the root of gradient
boosting, which merges the forecasting of weak learners with additive training methods to develop a strong
learner [31]. Gradient boosting with conditional deseasonalization and detrending is a specific application of
gradient boosting that incorporates the removal of seasonal and trend components from time-series data
before model training [32].
The findings of this study further indicate that PyCaret, as an AutoML framework, simplifies the
process of selecting optimal models and tuning hyperparameters, essential for practical use in cloud data
centres. This contrasts with the manual approach in previous studies, which required model selection and
testing to be conducted individually. With AutoML capabilities, processing time and computational resources
can be optimised, enabling practitioners to efficiently select the best model without repeated adjustments to
model parameters, as discussed in the research by Westergaard et al. [12]. This information is essential for
selecting a model that is accurate in prediction and efficient in using computing resources.

Table 9. Performance of forecasting process time efficiency

Dataset Model forecasting TT (Sec)
CPU (Daily) Naive Forecaster 1.71
Memory (Daily) Gradient Boosting w/ Cond. Deseasonalize and Detrending 0.13
Disk read (Daily) Naive Forecaster 1.90

3.2.5. Comparison of workload prediction models on web server trace datasets

The results of testing the HTTP request datasets from NASA, Calgary, and Saskatchewan servers
are presented here to evaluate the performance of the proposed gradient boosting model compared with Auto
ARIMA. Performance evaluation uses error metrics, including RMSE, MAE, and MASE. These tests provide
a comprehensive overview of each model's accuracy and suitability in the context of the research conducted
by Kumar and Singh [10].
The NASA server dataset was tested with Auto ARIMA and gradient boosting models using
conditional deseasonalisation and detrending across four prediction windows, presented in Table 10 for
periods of 5, 10, 30, and 60 minutes. As the prediction window lengthens, Auto ARIMA demonstrates
relatively consistent performance, though errors (RMSE and MAE) increase. The lowest MASE value, 4.42,
is achieved at a 10-minutes window, highlighting its medium-term accuracy. Gradient boosting with
deseasonalisation and detrending generally outperforms Auto ARIMA across nearly all prediction windows,
exhibiting notably lower RMSE and MAE, especially in short- to medium-term predictions (PWS 10 and 30).

Table 10. NASA server dataset

Model PWS RMSE MAE MASE
Auto Arima 5 249.99 218.28 5.16
10 380.20 334.67 4.42
30 1818.54 1664.29 8.36
60 4478.16 4173.28 9.79
Gradient Boosting w/ Cond. Deseasonalize and Detrending 5 215.27 190.71 4.51
10 251.42 220.73 2.91
30 855.47 758.42 3.81
60 2854.49 2544.58 5.97

Testing on the Calgary server dataset involved the same models and settings, as shown in Table 11.
Auto ARIMA, RMSE, and MAE values increase with longer prediction windows, though it achieves a low
MASE value of 1.51 at a 10 minutes window, indicating optimal medium-term fit. The gradient boosting
model with deseasonalisation and detrending achieves lower RMSE and MAE at the shortest prediction
window of 5, underscoring its high accuracy. However, its RMSE and MAE values slightly increase over
longer windows, while remaining competitive.

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1150-1163

Int J Artif Intell ISSN: 2252-8938  1161

Table 11. Calgary server dataset

Model PWS RMSE MAE MASE
Auto Arima 5 23.53 14.67 2.03
10 27.19 18.83 1.51
30 70.76 53.08 1.91
60 109.16 81.24 1.74
Gradient Boosting w/ Cond. Deseasonalize and Detrending 5 11.75 9.22 1.27
10 30.36 27.70 2.22
30 76.60 69.97 2.52
60 94.45 80.88 1.73

A comparative analysis using the Saskatchewan server dataset similarly assessed the efficacy of Auto
ARIMA and gradient boosting across four prediction windows, as shown in Table 12. Auto ARIMA, RMSE,
and MAE values rise with window length, though MASE remains stable around 1.9, indicating consistent
performance. Gradient boosting with deseasonalisation and detrending outperforms Auto ARIMA across
almost all prediction windows, achieving significant accuracy improvements in the 5 and 60 minutes windows.

Table 12. Saskatchewan server dataset

Model PWS RMSE MAE MASE
Auto Arima 5 44.57 32.25 1.76
10 84.84 61.73 2.02
30 180.71 132.64 1.91
60 347.07 259.83 1.92
Gradient Boosting w/ Cond. Deseasonalize and Detrending 5 38.33 28.97 1.58
10 71.30 57.61 1.88
30 159.69 133.85 1.92
60 296.16 248.67 1.84

In terms of overall accuracy and efficiency, gradient boosting consistently outperforms Auto
ARIMA across all datasets and prediction windows, demonstrating a superior ability to capture complex
patterns and manage variability in web server workloads. Although Auto ARIMA performs adequately in
medium-term predictions, its effectiveness diminishes in longer timeframes, as indicated by rising RMSE and
MAE values, revealing limitations in adapting to dynamic changes. Gradient boosting, on the other hand,
demonstrated greater scalability and adaptability at varying time scales, maintaining lower error metrics for
both short- and long-term predictions. Its ability to integrate deseasonalisation and detrending enhances
predictive accuracy, making it well-suited for datasets with inherent seasonality and trends. These findings
suggest that future research could focus on refining gradient boosting models further, exploring hybrid
methods, and optimising model complexity to balance accuracy with computational efficiency.

4. CONCLUSION
This study presents a performance assessment of time-series forecasting models for SNMP-based
hypervisor data using Pycaret. We evaluated the performance of our method on two types of datasets:
real-time data collected from a physical server environment running multiple virtual machines and synthetic
data created based on a specific workload scenario. We compared 30 different time-series forecasting models
using three evaluation metrics, namely, RMSE, MASE, and MAE, and the computation time required.
The evaluation results obtained the best dataset using daily data with several forecasting models, according to
the resource load on the hypervisor. The best naive forecaster model for forecasting CPU performance with a
runtime of 1.71 seconds and disk read runtime of 1.90 seconds. Meanwhile, gradient boosting w/ cond.
deseasonalize and detrending the best for memory dataset with process times of 0.13. Based on the test
results with NASA, Calgary, and Saskatchewan server datasets, from the comparison between Auto ARIMA
and gradient boosting with deseasonalize & detrending models, while both models have their merits, gradient
boosting with deseasonalize and detrending provides a more robust solution for forecasting in cloud data
centre environments, adapting well to the complexities and dynamics of real-world server workloads.
In the future, researchers can concentrate on enhancing and confirming the effectiveness of our approach by
using more extensive and varied datasets. They can also integrate it with other resource management in
decision support systems for use in real-time live migration settings in hypervisor clusters. Our research has
the potential to improve the quality and efficiency of resource management maintenance in data centres,
which can ultimately enhance the quality of service (quality of service) and service level agreements (SLA).

Performance assessment of time series forecasting models for simple network … (Yuggo Afrianto)
1162  ISSN: 2252-8938

ACKNOWLEDGEMENTS
The authors gratefully acknowledge the financial support from DRTPM Kemdikbudristek and
Telkom University for this research under the Doctoral Dissertation Research Program scheme 2024, under
Research Contract Number 106/E5/PG.02.00.PL/2024; 043/SP2H/RT-MONO/LL4/2024; 034/LIT07/PPM-
LIT/2024.

REFERENCES
[1] M. C. Martachinnicieneait, “SNMP for cloud environment energy eficiency,” Research Square, pp. 1-27, 2021, doi:
10.21203/rs.3.rs-720622/v1.
[2] S. B. Shaw, C. Kumar, and A. K. Singh, “Use of time-series based forecasting technique for balancing load and reducing
consumption of energy in a cloud data center,” in 2017 International Conference on Intelligent Computing and Control (I2C2),
IEEE, 2017, pp. 1–6, doi: 10.1109/I2C2.2017.8321782.
[3] S. Steiner, “Harnessing data to make better-informed decisions,” Scientia, 2022, doi: 10.33548/SCIENTIA768.
[4] S. Jadon, A. Patankar, and J. K. Milczek, “Challenges and approaches to time-series forecasting for traffic prediction at data
centers,” in 2021 International Conference on Smart Applications, Communications and Networking, SmartNets 2021, IEEE,
2021, pp. 1–8, doi: 10.1109/SmartNets50376.2021.9555422.
[5] H. Jdi and N. Falih, “Comparison of time series temperature prediction with auto-regressive integrated moving average and
recurrent neural network,” International Journal of Electrical and Computer Engineering, vol. 14, no. 2, pp. 1770–1778, Apr.
2024, doi: 10.11591/ijece.v14i2.pp1770-1778.
[6] V. Savchenko et al., “Network traffic forecasting based on the canonical expansion of a random process,” Eastern-European
Journal of Enterprise Technologies, vol. 3, no. 2–93, pp. 60–69, 2018, doi: 10.15587/1729-4061.2018.131471.
[7] W. Yoo and A. Sim, “Time-series forecast modeling on high-bandwidth network measurements,” Journal of Grid Computing,
vol. 14, no. 3, pp. 463–476, 2016, doi: 10.1007/s10723-016-9368-9.
[8] S. N. Wahyuni, E. Sediono, I. Sembiring, and N. N. Khanom, “Comparative analysis of time series prediction model for
forecasting covid-19 trend,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 28, no. 1, pp. 600–610,
Oct. 2022, doi: 10.11591/ijeecs.v28.i1.pp600-610.
[9] P. Shastri, B. R. Dawadi, and S. R. Joshi, “Intelligent approach to switch replacement planning for internet service provider
networks,” Sustainable Futures, vol. 2, 2020, doi: 10.1016/j.sftr.2020.100036.
[10] J. Kumar and A. K. Singh, “Performance assessment of time series forecasting models for cloud datacenter networks’ workload
prediction,” Wireless Personal Communications, vol. 116, no. 3, pp. 1949–1969, 2021, doi: 10.1007/s11277-020-07773-6.
[11] M. Ali, “PyCaret: an open source, low-code machine learning library in python,” ACS Omega, vol. 6, p. 6791−6797, 2021.
[12] G. Westergaard, U. Erden, O. A. Mateo, S. M. Lampo, T. C. Akinci, and O. Topsakal, “Time series forecasting utilizing
automated machine learning (autoML): a comparative analysis study on diverse datasets,” Information, vol. 15, no. 1, 2024, doi:
10.3390/info15010039.
[13] S. Maurya and S. Singh, “Time series analysis of the covid-19 datasets,” in 2020 IEEE International Conference for Innovation in
Technology, INOCON 2020, IEEE, 2020, pp. 1–6, doi: 10.1109/INOCON50539.2020.9298390.
[14] J. Siebert, J. Groß, and C. Schroth, “A systematic review of packages for time series analysis †,” Engineering Proceedings, vol. 5,
no. 1, 2021, doi: 10.3390/engproc2021005022.
[15] V. Cerqueira, L. Torgo, and I. Mozetič, “Evaluating time series forecasting models: an empirical study on performance estimation
methods,” Machine Learning, vol. 109, pp. 1997–2028, 2020, doi: 10.1007/s10994-020-05910-7.
[16] T. Lindgren and O. Steinert, “Low dimensional synthetic data generation for improving data driven prognostic models,” in 2022
IEEE International Conference on Prognostics and Health Management, ICPHM 2022, IEEE, 2022, pp. 173–182, doi:
10.1109/ICPHM53196.2022.9815660.
[17] Y. K. Phua, T. Fujigaya, and K. Kato, “Predicting the anion conductivities and alkaline stabilities of anion conducting membrane
polymeric materials: development of explainable machine learning models,” Science and Technology of Advanced Materials, vol.
24, no. 1, 2023, doi: 10.1080/14686996.2023.2261833.
[18] P. Whig, K. Gupta, N. Jiwani, H. Jupalle, S. Kouser, and N. Alam, “A novel method for diabetes classification and prediction
with pycaret,” Microsystem Technologies, vol. 29, no. 10, pp. 1479–1487, 2023, doi: 10.1007/s00542-023-05473-2.
[19] Abdullah, F. B. Iqbal, S. Biswas, and R. Urba, “Performance analysis of intrusion detection systems using the pycaret machine
learning library on the unsw-nb15 dataset,” B.Sc. Thesis, Department of Computer Science and Engineering, Brac University,
Dhaka, Bangladesh, 2021.
[20] K. Arai, J. Shimazoe, and M. Oda, “Method for hyperparameter tuning of image classification with pycaret,” International Journal
of Advanced Computer Science and Applications, vol. 14, no. 9, pp. 276–282, 2023, doi: 10.14569/IJACSA.2023.0140930.
[21] D. J. C. Sihombing, J. U. Dexius, J. Manurung, M. Aritonang, and H. S. Adinata, “Design and analysis of automated machine
learning (autoML) in powerbi application using pycaret,” in 2022 International Conference of Science and Information
Technology in Smart Administration, ICSINTESA 2022, 2022, pp. 89–94, doi: 10.1109/ICSINTESA56431.2022.10041543.
[22] T. Huynh, H. Mazumdar, H. Gohel, H. Emerson, and D. Kaplan, “Evaluating the predictive power of multiple regression models
for groundwater contamination using pycaret,” in WM2023 Conference, Arizona, USA: WM Symposia, Inc., 2023, pp. 1–15.
[23] S. Krsteski, M. Tashkovska, B. Sazdov, L. Radojichikj, A. Cholakoska, and D. Efnusheva, “Intrusion detection with supervised
and unsupervised learning using pycaret over cicids 2017 dataset,” in Artificial Intelligence Application in Networks and Systems,
2023, pp. 125–132, doi: 10.1007/978-3-031-35314-7_12.
[24] M. Karanfilovska, T. Kochovska, Z. Todorov, A. Cholakoska, G. Jakimovski, and D. Efnusheva, “Analysis and modelling of a
ml-based nids for iot networks,” in Procedia Computer Science, 2022, pp. 187–195, doi: 10.1016/j.procs.2022.08.023.
[25] P. Rani, “PyCaret based url detection of phishing websites,” Turkish Journal of Computer and Mathematics Education
(TURCOMAT), vol. 11, no. 1, pp. 908–915, 2020, doi: 10.17762/turcomat.v11i1.13589.
[26] J. E. Monogan, “Time series analysis,” in Political Analysis Using R, Springer, Cham, 2015, pp. 157–186, doi: 10.1007/978-3-
319-23446-5_9.
[27] M. Mahan, C. Chorn, and A. P. Georgopoulos, “White noise test: detecting autocorrelation and nonstationarities in long time series
after arima modeling,” in Proceedings of the 14th Python in Science Conference, 2015, doi: 10.25080/Majora-7b98e3ed-00f.
[28] S. N. Rao, G. Shobha, S. Prabhu, and N. Deepamala, “Time series forecasting methods suitable for prediction of cpu usage,” in
CSITSS 2019 - 2019 4th International Conference on Computational Systems and Information Technology for Sustainable

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1150-1163

Int J Artif Intell ISSN: 2252-8938  1163

Solution, Proceedings, IEEE, Dec. 2019, pp. 1–5, doi: 10.1109/CSITSS47250.2019.9031015.

[29] R. J. Hyndman and G. Athanasopoulos, Forecasting: principles and practice. OTexts: Melbourne, 2021.
[30] J. Sekitani and H. Murakami, “Framework for comparing accuracy of time-series forecasting methods,” in 2022 12th
International Congress on Advanced Applied Informatics (IIAI-AAI), IEEE, 2022, pp. 669–672, doi:
10.1109/IIAIAAI55812.2022.00136.
[31] O. Alshboul, A. Shehadeh, G. Almasabha, and A. S. Almuflih, “Extreme gradient boosting-based machine learning approach for
green building cost prediction,” Sustainability, vol. 14, no. 11, May 2022, doi: 10.3390/su14116651.
[32] M. Przybyla, “New time series with pycaret,” Towards Data Science. 2021. Accessed: Dec. 25, 2023. [Online]. Available:
https://towardsdatascience.com/new-time-series-with-pycaret-4e8ce347556a

BIOGRAPHIES OF AUTHORS

Yuggo Afrianto is currently a permanent lecturer at Ibn Khaldun University,

Bogor-West Java, Informatics Engineering bachelor’s Study Program. He holds a bachelor’s
degree in informatic engineering from the Ibn Khaldun University in Bogor, a master's degree
in computer science from IPB University in Bogor. In 2017. Research in the fields of net
centric computing he is a member of cyber physical system (CPS) Telkom University and a
member of the Association for Informatics and Computer Higher Education. He can be
contacted at email: [email protected].

Prof. Rendy Munadi has been a lecturer at the Faculty of Electrical Engineering
at Telkom University (formerly, STT Telkom) Bandung, Indonesia since 1993. The teaching
materials taught include wireless sensor networks, data and protocol networks, broadband
networks, internet of things in the master field of study while in the field of undergraduate
studies include traffic engineering, future new network, and seminar proposals. The field of
research that is being carried out is manufacture of a hemoglobin measuring device by
utilizing the internet of things (IoT) based machine learning method, partner: Hasan Sadikin
Hospital. He has serving experience in universities since 1998 is as Head of Academic
Administration (formerly STT Telkom) and Head of Telecommunication Engineering Study
Program in 2005-2006, served as Vice Chancellor for Academic Affairs in 2006-2010. Since
2018, he has served as an assessor for FTE lecturer certification. Currently, he is a senior
lecturer in the network cyber management (NCM) expertise group. He can be contacted at
email: [email protected].

Dr. Setyorini is currently a permanent lecturer at Telkom University, Bandung

West Java, in an informatics master’s study program. She received a B.S. degree in
informatics from Telkom University, and received M.S. and doctoral degrees in computer
engineering from Bandung Institute of Technology. Her research interest in parallel computing
has also focused on distributed systems. She can be contacted at email:
[email protected].

Dr. Toto Widyanto is a lecturer at Al Kamal Institute of Science and

Techcnology (ISTA). The teaching materials taught include machine learning, mobile
computing, artificial intelligence, telecommunications, and innovation. Besides education
field, he is a consultant for more than two decades in green data centers and green buildings
design and development. He is also a chairman of Indonesia Green Data Center Professional
Association (IPUSTAH-ID). He holds professional certifications for greenship professional
(GP) and certified data center specialist (CDCS). He can be contacted at email:
[email protected].

Performance assessment of time series forecasting models for simple network … (Yuggo Afrianto)