Machine Learning based Material Demand Prediction of Construction Equipment for Maintenance

Abstract

Construction managers faced Construction Equipment (CE) challenges related to running repair and replacement of spare part materials as well as shortage of materials, sudden damage of spare parts and unavailability of necessary materials at job sites frequently. Regular follow up and track of materials availability and their usage at each stage of requirement phase becomes essential. This study presents Machine Learning (ML) based material demand prediction. Training of ML models utilizes historical maintenance, and procurement periodic data related to materials of the CE. This study highlights the use of Multiple Linear Regression (MLR), Support Vector Regression (SVR), Decision Tree (DT) Regressor and ensemble boosting models as Random Forest (RF) Regressor and Gradient Bosting Regressor (GBR). According to the performance measurement of each model, RF performs better and is used for prediction. Material demand prediction helps in maintenance and operational planning of CE. Subsequently, approach assists in addressing issues early by involving operators and site owners, enabling preventive actions to be taken before the scheduled procurement process. This study addresses the corrective measurement of the model using periodic data. The model performance results indicate that early prediction of maintenance costs based on the quantity of essential materials withdrawn from demand is helpful for budgeting expenditures.

Full text

International Journal of Computing and Digital Systems
2025, VOL. 17, NO. 1, 1–12
http://dx.doi.org/10.12785/ijcds/1571018142
Machine Learning based Material Demand Prediction of
Construction Equipment for Maintenance
Poonam Katyare1, Shubhalaxmi S. Joshi1and Mrudula Kulkarni2
1Department of Computer Science and Application, Dr.Vishwanath Karad MIT World Peace University, Pune, India
2Department of Civil Engineering, Dr.Vishwanath Karad MIT World Peace University, Pune, India
Received 11 Apr. 2024, Revised 18 Sept. 2024, Accepted 19 Sept. 2024
Abstract: Construction managers faced Construction Equipment (CE) challenges related to running repair and replacement of spare part
materials as well as shortage of materials, sudden damage of spare parts and unavailability of necessary materials at job sites frequently.
Regular follow up and track of materials availability and their usage at each stage of requirement phase becomes essential. This
study presents Machine Learning (ML) based material demand prediction. Training of ML models utilizes historical maintenance, and
procurement periodic data related to materials of the CE. This study highlights the use of Multiple Linear Regression (MLR), Support
Vector Regression (SVR), Decision Tree (DT) Regressor and ensemble boosting models as Random Forest (RF) Regressor and Gradient
Bosting Regressor (GBR). According to the performance measurement of each model, RF performs better and is used for prediction.
Material demand prediction helps in maintenance and operational planning of CE. Subsequently, approach assists in addressing issues
early by involving operators and site owners, enabling preventive actions to be taken before the scheduled procurement process. This
study addresses the corrective measurement of the model using periodic data. The model performance results indicate that early
prediction of maintenance costs based on the quantity of essential materials withdrawn from demand is helpful for budgeting expenditures.
Keywords: Construction Equipment, Machine Learning, Material Demand, Maintenance
1. INTRODUCTION
Construction Equipment is a key driver for executing
successful construction projects. The management of CE
concerns efficiently overseeing equipment resources to meet
the equipment requirements and to gain maximum returns
on equipment for the construction project, which is tar-
geted to be executed in a scheduled and economically
viable fashion. Major contractors often can acquire, operate,
and manage a substantial collection of heavy CE units.
Therefore, making decisions regarding routine equipment
management responsibilities is essential for overall project
management. The daily routine involves the procurement
process, maintenance process, equipment allocations, equip-
ment operational activity, and replacement and repair activ-
ities of equipment spare parts.
The day-to-day execution of these activities has financial
implications for fleet owners because cost is involved in
every activity. Proper and effective budgeting of any con-
struction project focuses on CE cost bifurcations on various
aspects, which involve the initial acquisition cost of the
equipment, operating cost, maintenance and repairs cost,
operator and labor wages, depreciation cost, financing costs,
interest payments, transportation cost, regulatory compli-
ance costs, technology integration costs, and disposal cost of
CE. A critical piece of maintenance costs is fundamentally
credited to essential materials in form of spare parts for CE
[6]. Site owners should maintain an inventory of spare parts
associated with all equipment present and currently working
on the job site. It is a big challenge to handle equipment
failures and face downtime while performing tasks on-site
using equipment. They need to keep records of all materials
in the system with their availability quantity, order details,
required quantity, withdrawn quantity, and special operating
run hours of the equipment. The cost sheet for each quantity
is recorded with the date and time. Large numbers of lists of
materials are available that are distributed in similar groups
of equipment for simplicity of cost computations.
This investigation aims to avoid manual work for com-
puting the demand quantity of materials. The proposed
study emphasizes ML-based essential materials demand
prediction of CE in advance from a maintenance perspec-
tive. This study focuses on the rational study of various ML
algorithms.
This paper is arranged as follows. The existing study
E-mail address: poonam.katyare@gmail.com

2Poonam Katyare, et al.
with limitations is elaborated in part 2. The proposed
methodology for predicting the demand quantity of ma-
terials with data preprocessing and model fitting is given
in part 3. Results and discussion with a comparison of ML
algorithm performance are indicated in part 4. The inference
with the concluded work is illustrated in part 5.
2. EXISTING STUDY
A significant study was identified related to CE cost
prediction, residual value prediction, and sensor-based data
analysis. Maintenance of CE study observed using different
methods, such as by reviewing the existing techniques used
for reliability and fault analysis of CE. ML techniques,
graphical methods, fault tree analysis, and probability dis-
tribution models have been used; however, ML models
have the best accuracy for failure prediction and reliability
estimations of CE [3], [38]. The researcher presented a
related study on the implications of the Internet of Things
with sensor-based technologies attached to the equipment
for capturing real-time information of CE with location
tracking, movement tracking, working condition of engines,
fuel data updates, distance travelled, and battery updates
from equipment working on construction job sites. This
would help managers analyse the data collected from re-
mote sensors and make proper decisions regarding the
equipment’s performance. Remote sensing devices identify
information related to construction material tracking to
handle the supply chain management process along with
cloud computing, radio frequency identification, augmented
reality, and big data technologies [1],[5],[7].
The existing study presented the prediction of residual
values of CE by an Autoregressive Tree algorithm of data
mining using equipment age, make, model, region, horse-
power, auction year, condition rating, annual construction
investment, and Gross Domestic Product features to predict
equipment price. This study compared the performance
of data mining algorithms with those of neural networks,
linear regression algorithms, and deep learning algorithms
[11]. The authors demonstrated fuel consumption prediction
using ML and live data parameters from smart sensing
devices, which indirectly impacted the maintenance cost
of CE. Real-time data ensemble methods provide better
accuracy than other regression models [2],[4],[35], [36],
[37].
Another study demonstrated a prototype model that
effectively reduces labor costs and mitigates challenges
associated with equipment maintenance decision-making
by presenting a data-driven methodology that integrates
three key skills reliability maintenance focused on relia-
bility, modeling of building data, and live tracking system.
This includes critical components for ensuring the optimal
functionality of buildings [8],[9]. Data-driven approaches
have been observed in many studies that attempt to manage
equipment information data using data analysis from huge
amounts of data and manage the data for decision making
[10].
Manufacturers in the construction machinery parts in-
dustry must manage inventories promptly, optimize produc-
tion processes for efficient and swift product manufacturing,
and promptly deliver finished products to customers. To
solve this problem, an existing study elaborated demand
estimating for spare parts in the construction machinery
industry using regression and artificial intelligence models
[12]. Similarly, the heavy equipment of specific group
demand forecasting is also performed by the researcher
using the Support Vector Machine Regressor, which is
very useful for the equipment owner [13]. Multivariant
time series analysis performs better for the constriction
raw materials of steel products prediction [14]. A related
study investigated the prediction of heavy equipment prices
with precision by employing ML algorithms on sales data
obtained from a website [15].
Another study uses an artificial neural network-based
methodology to measure uncertainties and generate fore-
casting intervals for predicting prices of construction ma-
terials, with a specific focus on asphalt and steel. This
study provides supplementary information to enhance the
effective management of project cost-related risks through
estimate intervals to project managers. The proposed opti-
mal Lower Upper Bound Estimation (LUBE) cost function
yields highly precise estimate intervals [16]. The Analytical
Hierarchy Process within the thematic domain involves
the development of a modified decision model in CE
procurement. This model is designed to order parameters
that influence equipment procurement. The approach is
particularly tailored to address the unstructured aspects of
the selection method [17].
Research showcased the prediction of maintenance costs
related to breakdown and planned maintenance activity
events for essential plant resources, and the developed
model exhibited strong predictive accuracy. The methodol-
ogy integrates a stochastic mathematical modeling approach
that considers both unplanned breakdowns and scheduled
maintenance. This technique generates a pseudo-random
number to simulate the magnitude of an impending main-
tenance cost event [18]. Time series maintenance and fuel
consumption data were used to anticipate the CE cost of
maintenance using a neural network time series model [19],
[21]. The existing study employs time series and multiple
regression models to predict construction material prices.
The combination of these statistical methods allows for
capturing both time-series trends and relationships between
different economic factors, providing a robust prediction
framework. The integration of multiple models enhances
predictive accuracy and robustness, catering to complex
market dynamics [29]. This review discusses various artifi-
cial intelligence methods for demand forecasting in supply
chain management, including machine learning algorithms
like neural networks, support vector machines, and en-
semble methods. AI methods can handle large datasets
and complex patterns, improving forecast accuracy over
traditional statistical methods [30].

International Journal of Computing and Digital Systems 3
Figure 1. Prediction Model Flow Diagram

4Poonam Katyare, et al.
This systematic review examines the adoption of ma-
chine learning technology for failure prediction in industrial
maintenance, emphasizing the use of algorithms such as
decision trees, random forests, and neural networks. Ma-
chine learning models can analyze historical data to predict
equipment failures, thus optimizing maintenance schedules
and reducing downtime [33]. This study investigates the
impact of increasing raw material prices on construction
costs, providing insights into economic factors affecting
the construction sector in specific regions. Understanding
the economic impact helps in better budgeting and cost
management strategies in construction projects [32].
Another study provides a comprehensive review of arti-
ficial intelligence and ML techniques used for performance
monitoring and failure prediction in industrial equipment.
The study highlights the increasing importance of predictive
maintenance to improve operational efficiency and reduce
downtime in industrial settings. The study discusses various
AI and ML algorithms such as neural networks, support
vector machines, decision trees, and ensemble methods.
It examines their applications in identifying patterns and
anomalies in equipment performance data to predict poten-
tial failures [31].
A. Limitations of Existing Study
Existing studies have demonstrated the prediction of
residual values of equipment, prediction related to the cost
of equipment, prediction of failure or breakdown of equip-
ment, fuel consumption, and maintenance cost estimation of
equipment. This study presents qualitative and quantitative
data analysis using ML models, time series analysis, and
factor analysis. Contractors or site owners need to maintain
records of equipment spare parts or materials available,
required or demand quantity, and how much quantity is
utilized manually. They face the issues of failure of mate-
rials, replacement of materials, and damage of materials at
the site. They should add all these records manually to log
sheets, order the materials as per the requirement and wait
for procurement of those materials. Meanwhile, there could
be downtime at the site for that equipment because of the
unavailability of the materials required for that equipment
aligned with the working conditions. The frequency of
such situations or challenges frequently occurred at the
job site. There is a need to estimate such a material
demand quantity. Many predictive models require high-
quality, granular data for accurate forecasting. Inconsistent
or incomplete data can lead to less reliable predictions.
The existing system has limitations of Data Quality and
Availability. The proposed study predicts the quantity of
equipment required for material demand from the operating
hours and history data, which would help to maintain
the required materials stock in advance at the job site.
Machine learning (ML) models are increasingly relevant
and effective in addressing a variety of challenges in the
construction equipment industry. These challenges range
from predictive maintenance and equipment optimization to
safety and operational efficiency. Predictive Maintenance is
major challenge of equipment downtime due to unexpected
failures that can be costly and disruptive. Predictive mainte-
nance uses ML algorithms to analyze data from sensors and
historical maintenance records to predict when equipment is
likely to fail. This allows for timely maintenance, reducing
downtime and repair costs. Techniques such as anomaly
detection and time-series analysis are particularly useful
here. Managing the supply chain and inventory effectively to
avoid delays and excess costs is another challenge. ML can
forecast demand for materials and equipment, optimizing
inventory levels and supply chain operations. Forecasting
models and optimization algorithms are particularly useful
in this area. The integration of ML models in the construc-
tion equipment industry addresses numerous challenges by
improving predictive maintenance, optimizing equipment
utilization, enhancing safety, reducing fuel consumption
and emissions, managing fleets efficiently, ensuring quality,
and optimizing supply chain and inventory management.
As data availability and computational power continue to
grow, the relevance and effectiveness of ML in this sector
are expected to expand, driving further innovation and
efficiency.
3. PROPOSED METHODOLOGY
Analyzing time series data for construction equipment
material prediction involves a systematic approach to ensure
accurate and reliable forecasting. Analyzing time series
data for construction equipment material prediction involves
collecting and preprocessing data, performing exploratory
analysis, selecting and training appropriate models, eval-
uating their performance, and deploying and monitoring
the models in a production environment. This structured
approach ensures accurate material forecasts, leading to
optimized resource management and reduced operational
costs in the construction industry. This emphasizes analysis
of data related to CE materials, detailing the quantity
available for each machine, used run hours, the specifics
of each machinery order, and the anticipated future demand
for each. The goal is to predict the demand for construction
machinery. Figure 1. illustrates the Prediction model flow
diagram with the detailed steps for estimating the quantity
of materials. This flow diagram interprets the Data Acqui-
sition, Data Preprocessing and Data Modelling steps.
A. Acquisition of Records
Daily logs of the repair and replacement of materials are
maintained at the construction site. New orders are placed
to purchase replacement materials. Industries keep these
records in their Enterprise Resource Planning system. The
proposed study acquires order and material quantity data
from the organizational system from 2017 to 2023 from
various sources. Interviews were conducted with experts and
contractors working on job sites, and literature review data
were used to finalize the features required for the proposed
study. The acquired dataset is in a daily basis format con-
sidering the days when orders are placed. The collected data
has features related to order, material, and material quantity
details. Order details include order number, creation date,

International Journal of Computing and Digital Systems 5
TABLE I. Material Groups clusters
No. Material Groups
1 Spares
2 Structural Steel
3 Welding Materials
4 Pipe and Pipe Fittings
5 Hardware, Painting and Chemicals
6 Electrical Items
7 Rubber Goods
8 Lubricant and Oil
9 Tools
10 Miscellaneous
11 Consumables (Anchor/Pilling/Drilling)
completion date, and requirement date. Material details
features represent Site number, Material number, Equipment
number, Material group, Operating run hours, and equip-
ment manufacturer. The major material groups are classified
into various categories. Material quantity-related features
denote the available quantity of material in the inventory, the
required quantity at the time of replacement and repair, and
the withdrawn quantity of materials representing the total
quantity of materials used. Frequently consumed materials
were observed during the study. Major material categorial
groups of materials are highlighted in the dataset. Material
group codes are present in the dataset, and mapping of all
materials under groups will be used in future studies. Table
I presents the material groups used in the dataset.
B. Preprocessing of Records
The demand quantity of material estimation is related to
the quantity of material withdrawn from historical records.
The available and required quantities are major contributors
to predicting the withdrawn quantity demanded. Table II
Statistical Description of parameters represents statistical
values for major parameters. The attributes relation is
identified from the correlation matrix denoted in Table III
Correlation matrix of parameters where P1 is withdrawn
quantity, P2 is site, P3 is material number, P4 is equipment
number, P5 is material group, P6 is run hours, P7 is quantity
available and P8 is required quantity of material. Correlation
Matrix tests can be used to check whether the information
focuses are independent and indistinguishably distributed.
It observes the relationship of the independent parameters
with the target parameter [4],[13]. In this study, the material
was highly correlated with the quantity available, quantity
required, and operating run hours. The withdrawn quantity
is also related to the available and required material quan-
tity. Finally, the major features selected for the predictive
modeling of records are material number, Equipment num-
ber, Material group, Operating run hours, quantity available,
and required quantity to predict the withdrawn quantity
demand of material. Outliers are identified and removed
from the data using the quantile method of outlier removal
[18].
1) Outlier Removal Method
A statistical approach, Interquartile Range (IQR) is used
to remove the outliers [26]. This approach identifies the
distribution of the mid-fifty percentage of the records.
Equation 1 represents the formula for computing IQR as
the subtraction of the 75th record percentile as QT3 with
the 25th record percentile as QT1.
IQR =QT 3−QT 1 (1)
Where,
QT1=Upper bound with value less than 25% of records lie.
QT3=Upper bound with value less than 75% of records lie.
This approach handles the skewed record distribution with
outliers and provides a list of outliers.
2) Feature Scaling
It is an approach of transforming values of features from
records into similar scales, which helps to define the equal
contribution of all features. Scaled features have a greater
impact on performing ML models accurately.
Standardization is an approach that denotes that the
values of features are central to the mean with a unit of
standard deviation [24]. This supports the retention of the
relationship between record points from the data mentioned
in equation 2. It is computed as
(DT –mean(allDT s))/S D (2)
Where,
DT =Data Point
DTs =All Data Points
SD =Standard deviation of all DTs
C. Preliminary Analysis of Records
Preliminary analysis of the pre-processed data helps to
observe year-wise material usage. This dataset is real-time
data of maintenance which involves materials details that
were repaired and replaced at the time of maintenance.
The issues are related to running repairs, breakdown orders,
breakdown repairs, calibration changes, maintenance after
specific run hours, order of the machinery, defective mate-
rial indication, regular servicing, and handling of damaged
materials. Every record of the issue along with order
details of materials were kept as logs in the ERP system
of the organization. This dataset contains Site number,
Equipment number, Material number, Material Group, Run
hours, Available quantity of materials in stock, Requirement
quantity and withdrawal quantity is the quantity used as the
major features along with order number, order creation date,
and order completion date as the minor features. Figure
2 Year-wise material usage from 2018 to 2023. As per
the increment in project scheduling, the increase in the
order of material usage is listed. The years 2022 and 2023
highlighted more use of materials than prior years. Key

6Poonam Katyare, et al.
TABLE II. Statistical description of parameters
Withdrawn Qty Material Group Run Hrs Quantity Available Requirement Qty
Mean 1.88 209 6552.90 2.33 2.13
Std 2.14 45.55 4246.44 3.91 3.36
Min 0 200 2 0.004 0
25% 1 200 3350 1 1
50% 1 200 5639 1 1
75% 2 200 8983 2 2
Max 39 500 21022 286 91
TABLE III. Correlation matrix of parameters
P1 P2 P3 P4 P5 P6 P7 P8
P1 1 0.03 0.32 -0.06 0.32 0 0.31 0.96
P2 0.03 1 0.11 0.19 0.11 0.3 0.05 0.03
P3 0.32 0.11 1 -0.06 1 -0.04 0.39 0.32
P4 -0.06 0.19 -0.06 1 -0.06 -0.21 -0.04 -0.05
P5 0.32 0.11 1 -0.06 1 -0.04 0.39 0.32
P6 0 0.3 -0.04 -0.21 -0.04 1 -0.01 0
P7 0.31 0.05 0.39 -0.04 0.39 -0.01 1 0.32
P8 0.96 0.03 0.32 -0.05 0.32 0 0.32 1
insights from Figure 2 are related to Trends and Patterns
with Significant Increase, Fluctuations in Earlier Years, and
High Usage. Significant Increase in 2022, with material
usage more than doubling compared to 2021.This sharp rise
indicates a significant surge in construction activity, likely
due to an increase in project scheduling or the initiation
of several large-scale projects. Fluctuations in Earlier Years
happened Between 2018 and 2021, material usage shows
notable fluctuations in a significant increase from 2018
to 2019 and a decline in 2020 and a further drop in
2021. High Usage in 2022 and 2023 with despite a slight
decrease in 2023, material usage remains high compared to
previous years, indicating a sustained period of increased
construction activity.
D. Modelling of Records
ML is a strategy for changing information into note-
worthy information. Different directed ML methods are
accessible for expectation, which is related to the verifi-
able information for anticipating new occasions of data of
interest with the connection of target factors alongside inde-
pendent information values. ML model follows information
assortment, Information preprocessing, and modeling with
different algorithms. Finally, the model with the better
measurement is chosen for predicting new occasions of
information. We used different ML regression
models, for example, MLR, SVR, DT along with ensemble
regressors as RF and GBR models for determining the
demand quantity of materials.
1) Multiple Linear Regressor (MLR)
MLR is a basic and generally involved method for
displaying the association of a dependent variable with at
least one independent variable. The model anticipates a
linear relationship between the dependent and independent
factors, suggesting that they can be represented as a straight
line. MLR is a measurable investigation technique used to
determine the dependent quantitative connection between
at least two factors. Target factors in the regression exam-
ination are perceived or assessed [24]. Independent factors
are the factors that are remembered to significantly affect
the target variable attempted for assessment. Forecasts can
be made by estimating the connections within factors using
examination. Considering input parameters as X the basic
statistical model of MLR is stated by Equation 3.
Y(x,c)=c0+c1x1+− − +cnxk=c0+
K
X
i=1
cixi(3)

International Journal of Computing and Digital Systems 7
Figure 2. Year wise Materials Utilization
where c can be projected using the least squares method as
in Equation 4
ˆc=argmin{
N
X
i=1
(yj−c0−
K
X
i=1
cixji)2}(4)
where x1,x2,. . xkare the observed values of independent
parameters, c1,c2,. . ckare the regression coefficients, c0
is the intercept term, N is the sample count size with K
representing input parameters, and y are stated value of the
dependent parameter.
2) Support Vector Regressor (SVR))
A function provided by the SVR signifies the relation-
ship existing within the dependent and independent
parameters with a reducing error factor. The fundamental
aim of SVR is to discover a hyperplane with the largest
number of points within the decision boundary line or
support vectors that should be present within that boundary
line. Decision boundaries are used with hyperplanes to
anticipate continuous values. This assortment of numerical
activities known as kernels is used to change input infor-
mation into important configurations. SVR attempts to fit
between the boundary lines and the hyperplane [4], [24].
The formula for a SVR can be expressed as follows:
ˆy=
nsv
X
i=1
aiK(xi,x)+b(5)
Where,
ˆ
y - the predicted dependent value.
nsv- the count of support vectors.
xi- the ith support vector.
b– bias term
K(xi,x)– function kernel, which calculates similarity be-
tween the i-th support vector and the input sample x
allowing for nonlinear relationships between features. aiare
coefficients associated with the support vectors. A hyper-
plane is calculated to fit the training data while minimizing
margins. This aims to find coefficients aiand the
bias term bthat minimize the empirical risk as the variation
in the anticipated and real values subject to a margin of tol-
erance ϵ. This optimization problem is typically described
as a quadratic programming problem and is solved using
optimization techniques. Common kernel functions include
sigmoid, linear, and polynomial
kernels. The choice of the kernel function varies de-
pending on the complexity of the relationships between
features and the nature of the data. SVR is intensely efficient
for datasets with dense relationships and high-dimensional
feature spaces. SVR ensures robust predictions and reduced
sensitivity to outliers by expanding the margin among the
hyperplane and the data points.
3) Decision Tree Regressor (DT))
It is an extensively applied supervised learning algo-
rithm. It supports regression and classification analysis. A
DT is a progressive model used in portraying decisions and
their expected results, consolidating chance occasions, asset
costs, and utility. This algorithmic model uses contingent
control proclamations in the form of statements. It is a
nonparametric supervised learning method helpful for both
regression and classification analysis. The tree structure
contains a root node and subtrees with branches followed
by interior nodes, and leaf nodes frame a hierarchical, tree-
like construction [25]. The DT regression model can be
represented by the following formula:
ˆy=
N
X
i=1
wi·I(x∈Ri) (6)
Where,
ˆ
y - forecasted target value.
N=total count of leaf nodes in the DT.
Ri=region as leaf node of the feature space stated as the
ith leaf node.
wi is the anticipated value correlated with the leaf node.
Input value when lies in the region, Indicator function
proceeds to success. The anticipated value of the leaf node
is treated as the final prediction. Each region with an

8Poonam Katyare, et al.
associated leaf node with an anticipated value represents
the average of the dependent estimates of the training
falling within that region [24]. The DT regressor formula
essentially represents a piecewise constant function, where
feature space is partitioned into non-overlapping regions and
each section is linked with a constant predicted value. The
last estimate for a given input sample is the sum of the
predicted values of the leaf nodes into which the sample
falls [25].
4) Random Forest Regressor (RF)
Ensemble learning models impact the finding of solu-
tions to very complex regression problems. Ensemble learn-
ing can be characterized as the method involved in creating
different models, such as classifiers, and then accumulating
their outcomes to acquire better prescient execution. Two
notable outfit-learning techniques are boosting and bagging.
In supporting, progressive models add additional load to
preparing cases that were erroneously predicted by past
models. While making the forecast, a weighted vote is con-
sidered. Although progressive models are not reliant upon
prior models in bagging, each model is freely developed by
a bootstrap test of information. Forecasting is created by
considering a basic larger part vote. The ensemble predictive
model, RF, is built on a set of decision/regression trees.
Rather than basing the forecast on a single tree, a group of
trees is used to make the determination. RF adds an extra
element of randomization to bagging, which sets it apart
from other approaches. Like other bagging models, RF uses
a bootstrap of sample data to build each decision/regression
tree. The process for creating trees is different [8]. Because
of this technique, the RF can withstand overfitting and
excel in various problems. In addition, working with the as-
sessment of variable significance and exception recognition
are different advantages of this calculation [24]. Moreover,
RF is sensibly quick to obtain and can be effortlessly
parallelized. By backward eliminating predictors according
to the specified variable relevance, RF can be improved.
The formula for a RF can be stated as:
ˆy=1
N
N
X
i=1
Ti(x) (7)
Where,
ˆ
y - anticipated target value.
N is the entire trees in the RF.
Ti(x) is the prediction of the i-th decision tree for the
sample x. This model aggregates estimates of all multiple
DTs to make a final prediction. Every DT is trained using
a bootstrap sample extracted from the training data and
allows the splitting of features randomly. At last estimate is
computed by averages of all individuals’ predictions. Over-
fitting reduces using averaging and enhances performance.
The final prediction depends on the contribution of every
tree, regardless of its individual performance. This ensemble
approach makes RFs robust and capable of handling noisy
data while providing reliable predictions.
5) Gradient Boosting Regressor (GBR))
It has a place in the class of ensemble learning tech-
niques that explicitly boost calculations. It is known for its
high prescient accuracy. It functions admirably with both
linear and nonlinear connections between the dependent
and target factors. It can deal with complex information
with high dimensionality and countless factors. It can de-
tect complex communications among factors and precisely
model non-direct connections. It can deal with missing
qualities in the dataset without requiring attribution. It
divides information by considering accessible elements and
continues to prepare the model. It is hearty handles to
anomalies in the information. This method uses a collection
of weak learners and limits the effect of exceptions to
the iterative process. The regressor includes significance
scores, allowing comprehension of the elements that are
most compelling in forecasting. This can be useful for
highlighting determination and identifying hidden examples
in the information. The regressor is less inclined to overfit
in contrast with other complex models like profound brain
organizations. This is because it assembles trees succes-
sively, improving the blunders made by the past trees. The
regressor considers tweaking hyperparameters like the num-
ber of trees, tree profundity, learning rate, and misfortune
capability, giving adaptability in model streamlining [39].
It tends to be used for an extensive variety of regression
undertakings, including the expectation of nonstop factors.
GBR is a flexible and strong model reasonable regression
undertaking, particularly when high prescient precision and
interpretability are required [22][23]. The formula for a
GBR model stated as:
ˆy=
M
X
i=1
γihi(x) (8)
where:
ˆyis the predicted target value.
Mis the total count of trees.
hi(x) is the estimate of the i-th base learner for the input
sample x.
γiis the learning rate associated with the i-th base learner.
In GBR, the model successively develops an ensemble
of weak learners, typically DTs, and joins them with strong
learners. Each subsequent base learner focuses on residuals
as the variation involving the actual with predicted estimates
of the preceding predictions. By iteratively fitting new base
learners, GBR gradually improves the model’s ability to
trap complicated data associations. The key idea behind
gradient-boosting regression is to minimize a loss function
as squared error loss. Each base learner training helps to
reduce the loss concerning residuals of previous predictions.
GBR is a compelling method for building predictive models
for handling complex datasets. However, it is important
to adjust trees to boost iterations and the learning rate to
prevent overfitting and achieve optimal performance.

International Journal of Computing and Digital Systems 9
6) Cross Validation (CV) Technique
K-fold CV is a strategy utilized to assess performance
by dividing the first dataset into k-equivalent estimated
subsamples, called folds. The cycle includes iteratively
preparing the model k times. This permits us to obtain
k arrangements of assessment measurements, ordinarily
finding the middle value to obtain a more reliable predic-
tion. This cross-validation technique of data splitting at the
training-validation split can mitigate the overfitting issues
and retain a consistent estimate of the model execution. A
more robust estimate of the model performance is provided
by this technique. It uses multiple training validation splits
and averages the performance. This is specifically used to
select the most suitable model performance and perform a
comparative evaluation of the model measurement [20].
4. Results And Discussion
The study depicts the expectation task completed to
examine a bunch of elective models for predicting the
material demand quantity of the chose dataset. We assessed
the suitability of MLR, SVR, GBR, DT, and RF models for
predicting the material demand quantity of CE. Using the
real dataset, regression models were trained and evaluated.
Mean Squared Error (MSE), Mean Absolute Error (MAE),
Root Mean Squared Error (RMSE), and Coefficient of
Determination (R2 score) are used to assess performance
measurement of the regression models presented in equa-
tions (3), (4), (5), and (6), respectively. The mean squared
variance of actual with projected values assigned as MSE,
the mean-variance within the original and estimated values
denotes MAE, and the square root of the MSE associated
with error rate along with the coefficient of estimated values
about the original values is indicated by the R2 score. The
percentages represent values between 0 and 1.
MS E =1
N
N
X
i=1
(PREDi−ACT i)2(9)
MAE =1
N
N
X
i=1
|PREDi−ACT i|(10)
RMS E =v
u
t1
N
N
X
i=1
(PREDi−ACT i)2(11)
R2=1−PN
i=1(PREDi−ACT i)2
PN
i=1(ACT i−¯
ACT )2(12)
where PREDiand ACT idenote the i-th predicted and
actual material demand quantity values. The comparative
examination of the models’ performances is shown in Table
IV. Performance measurement of models. The MLR, SVR,
GBR, DT regression and RF regression models with k-fold
cross-validation are measured with MAE, MSE, RMSE, and
TABLE IV. Performance measurement of models
Model MAE MSE RMSE R2
MLR 0.82 3.12 1.74 0.32
SVR 0.28 1.35 1.13 0.52
GBR 0.21 0.37 0.59 0.62
DT 0.06 0.24 0.46 0.65
RF 0.08 0.22 0.42 0.66
R2score values and compared to predict the withdrawn
material quantity demand of the CE.
MLR can be utilized for material quantity assessment
when there is a reasonable direct connection within the
input factors as equipment number, material available quan-
tity along with run hours and the target variable as the
withdrawn quantity of material. In MLR, the weighted
amount of the variables’ coefficients is used to predict
material quantity. MLR might give a decent beginning
stage, yet its capacity to discover complex connections
between different highlights might be restricted when pre-
dicting the withdrawn quantity of material demand. More
complex models may be expected to represent nonlinear
impacts. Figure 3 Performance Measurement of models
represents a visualization of models with R2 score. SVR
with RBF kernel meets a useful ability for anticipating
material quantities by really discovering complex relation-
ships between input parameters and the target parameter.
Appropriate information preprocessing, model preparation,
hyperparameter tuning, and assessment are critical stages
in utilizing this methodology for precise expectations in
material quantity prediction assignments. Decision trees can
deal with different categories of data. The CE materials
dataset represents numerical and categorical parameters im-
plications. DT is a very simple method of decision-making
at each stage of splitting nodes. DT is inclined to overfitting,
particularly when the tree develops intensely in the training
data. This can prompt unfortunate speculation execution on
inconspicuous information. Ensemble methods such as RF
and GBR help for resolving data overfitting and provide
better results for forecasting material withdrawn quantities
and demand. RF enhances decision trees by joining various
trees’ forecasts. It acquires complex associations between
parameters and material quantities. It can handle multi
feature data. Key insights from Comparative Analysis from
Table IV
A. Predictive Accuracy representing R² (Coefficient of De-
termination):
Random Forest (RF) and Decision Tree (DT) models
exhibit the highest R² values (0.66 and 0.65, respectively),
indicating they explain most of the variance in the data and
provide the most accurate predictions. Gradient Boosting
Regressor (GBR) performs well with an R² of 0.62.Support
Vector Regressor (SVR) shows moderate accuracy with an

10 Poonam Katyare, et al.
Figure 3. Performance Measurement
R² of 0.52. Multiple Linear Regression (MLR) has the
lowest R² (0.32), suggesting it is less effective at capturing
the underlying patterns in the data.
B. Robustness with Consistency and Outlier Sensitivity:
RF and DT models tend to be more robust to outliers and
variations in the data due to their ensemble and hierarchical
nature. GBR, as an ensemble method, also demonstrates
robustness. SVR can be sensitive to the choice of hyperpa-
rameters and may not perform as robustly across varying
datasets.MLR is the least robust, often influenced by outliers
and assumptions about linearity.
C. Computational Efficiency with Training and Prediction
Time:
MLR generally the fastest to train and predict due to
its simplicity and linear nature. SVR computationally more
intensive, especially with larger datasets, due to the kernel
trick. GBR has moderate computational efficiency, balanc-
ing between accuracy and training time. DT is efficient in
training and prediction but can suffer from overfitting if
not pruned. RF is Computationally intensive due to training
multiple trees, but parallel processing can mitigate this to
some extent.
For predictive tasks in material usage forecasting, Ran-
dom Forest (RF) and Decision Tree (DT) models are the
most effective in terms of accuracy and robustness. Gradient
Boosting Regressor (GBR) serves as an excellent alternative
with a balance of high accuracy and moderate computa-
tional demands. Support Vector Regressor (SVR) can be
considered for moderate performance needs, while Multiple
Linear Regression (MLR) is less suitable for capturing the
complexity in this context.
Using machine learning (ML) algorithms for material
demand prediction in construction settings can significantly
improve efficiency, cost-effectiveness, and project manage-
ment. Improved Accuracy in Demand Forecasting, Opti-
mized Inventory Management, Enhanced Project Schedul-
ing, and Cost Savings are the practical implications. A
large construction company used ML models to predict the
demand for materials. By analyzing historical data, weather
patterns, and project timelines, the ML model reduced
material shortages and overages by 20%, leading to cost
savings and smoother project execution. A construction
firm with multiple ongoing projects may use the prediction
of the exact quantities of materials required at different
stages of each project. This enables just-in-time delivery,
reducing storage costs and minimizing the risk of material
degradation or theft. The accurate predictions allowed for
better scheduling of deliveries, avoiding delays caused by
material shortages. Furthermore, these accurate forecasts
allow companies to optimize inventory management by
predicting the exact quantities of materials needed at various
stages of a project. This enables a just-in-time delivery ap-
proach, reducing storage costs and minimizing risks such as
material degradation, theft, or obsolescence. Additionally,
by knowing when materials will be required, construction
firms can better schedule deliveries, ensuring all resources
are available when needed, thereby minimizing delays and
enhancing coordination between teams. Ultimately, the in-
tegration of ML in material demand prediction not only
streamlines operations but also contributes to substantial
cost savings and smoother project execution.
5. Conclusions and Future Work
This study focuses on the machine-learning-based
material-demand prediction of CE. Maintenance data
records were analyzed in this study. The limitations of
the existing study were acknowledged and summarized
using ML technologies, and a model to predict the material
demand quantity was proposed. This study helps to estimate
the material demand in advance for maintenance and to
maintain the maintenance cost associated with the estimated
materials in planning. This study provides various ML-
based regression models, such as MLR, SVR, GBR, DT,
and RF regression model performance. The results reveal
the viability of utilizing ML techniques to overcome the
difficulties in predicting material quantities. The RF model
predicts material quantities accurately and performs better
than other regression models. It is critical to handle real-
time data for preprocessing, which involves outlier removal,
handling missing values, and scaling the features to acquire
accurate data for modeling. ML models are very sensitive
to the quality of the dataset. This study demonstrates ML
applications for material quantity prediction of CE in the
construction industry for maintenance. The estimation of
maintenance and operating costs of materials for CE leads
to the financial budgeting of the overall construction project
at the job site. This study presented the work for limited
construction materials data. The Study can be expanded
using large materials with similar behavior. There is a
challenge to handle the real time data with large volume.
Future research would help in providing material prediction
for various categories of construction equipment with large
volume of data.
References
[1] A. Kumar and O. Shoghli, “A review of iot applications in supply
chain optimization of construction materials,” in ISARC 2018 - 35th

International Journal of Computing and Digital Systems 11
International Symposium on Automation and Robotics in Construc-
tion International AEC/FM Hackathon Future Building Things, July
2018.
[2] P. Katyare, S. Joshi, and M. Kulkarni, “Utilizing machine learning
approach to forecast fuel consumption of backhoe loader equip-
ment,” International Journal of Advanced Computer Science and
Applications, vol. 15, no. 5, pp. 1194–1201, 2024.
[3] P. Odeyar, D. B. Apel, R. Hall, B. Zon, and K. Skrzypkowski, “A
review of reliability and fault analysis methods for heavy equipment
and their components used in mining,” Energies, vol. 15, no. 17, pp.
1–27, 2022.
[4] P. Katyare, S. S. Joshi, and S. Rajapurkar, “Real time data modeling
for forecasting fuel consumption of construction equipment using
integral approach of iot and ml techniques,” Journal of Information
and Optimization Sciences, vol. 44, no. 3, pp. 427–437, 2023.
[5] P. Katyare and S. S. Joshi, “Construction industry digitization
using internet of things technology,” in Proceeding of International
Conference on Computational Science and Applications. Algorithms
for Intelligent Systems. Springer, Singapore, 2022, pp. 243–249.
[6] H. Fan, H. Kim, and O. R. Za¨
ıane, “Data warehousing for construc-
tion equipment management,” Canadian Journal of Civil Engineer-
ing, vol. 33, no. 12, pp. 1480–1489, 2006.
[7] P. Katyare and S. Joshi, “Construction productivity analysis in
construction industry: An indian perspective,” in Proceeding of In-
ternational Conference on Computational Science and Applications.
Algorithms for Intelligent Systems. Springer, Singapore, 2022.
[8] Z. Ma, Y. Ren, X. Xiang, and Z. Turk, “Data-driven decision-making
for equipment maintenance,” Automation in Construction, vol. 112,
p. 103103, 2020.
[9] J. C. P. Cheng, W. Chen, K. Chen, and Q. Wang, “Data-driven pre-
dictive maintenance planning framework for mep components based
on bim and iot using machine learning algorithms,” Automation in
Construction, vol. 112, p. 103087, 2020.
[10] H. Fan, H. Kim, S. AbouRizk, and S. H. Han, “Decision support in
construction equipment management using a nonparametric outlier
mining algorithm,” Expert Systems with Applications, vol. 34, no. 3,
pp. 1974–1982, 2008.
[11] O. Alshboul, A. Shehadeh, M. Al-Kasasbeh, R. E. Al Mamlook,
N. Halalsheh, and M. Alkasasbeh, “Deep and machine learning
approaches for forecasting the residual value of heavy construction
equipment: a management decision support model,” Engineering,
Construction and Architectural Management, vol. 29, no. 10, pp.
4153–4176, 2022.
[12] A. Aktepe, E. Yanık, and S. Ers¨
oz, “Demand forecasting application
with regression and artificial intelligence methods in a construction
machinery company,” Journal of Intelligent Manufacturing, vol. 32,
no. 6, pp. 1587–1604, 2021.
[13] A. Kargul, A. Glaese, S. Kessler, and W. A. G¨
unthner, “Heavy
equipment demand prediction with support vector machine regres-
sion towards a strategic equipment management,” International
Journal of Structural and Civil Engineering Research, pp. 137–143,
2017.
[14] C. Lee, J. Won, and E.-B. Lee, “Method for predicting raw material
prices for product production over long periods,” Journal of Con-
struction Engineering and Management, vol. 145, no. 1, pp. 1–8,
2019.
[15] N. Boyko and O. Lukash, “Methodology for estimating the cost of
construction equipment based on the analysis of important charac-
teristics using machine learning methods,” Journal of Engineering
(United Kingdom), 2023.
[16] M. Mir, H. M. D. Kabir, F. Nasirzadeh, and A. Khosravi, “Neural
network-based interval forecasting of construction material prices,”
Journal of Building Engineering, vol. 39, p. 102288, 2021.
[17] K. Petroutsatou, I. Ladopoulos, and D. Nalmpantis, “Hierarchizing
the criteria of construction equipment procurement decision using
the ahp method,” IEEE Transactions on Engineering Management,
pp. 1–12, 2021.
[18] D. J. Edwards and G. D. Holt, “Predicting construction plant
maintenance expenditure,” Building Research Information, vol. 29,
no. 6, pp. 417–427, 2001.
[19] H. L. Yip, H. Fan, and Y. H. Chiang, “Predicting the mainte-
nance cost of construction equipment: Comparison between general
regression neural network and box-jenkins time series models,”
Automation in Construction, vol. 38, pp. 30–38, 2014.
[20] D. Berrar, “Cross-validation,” Encyclopedia of Bioinformatics and
Computational Biology: ABC of Bioinformatics, vol. 1-3, no. April,
pp. 542–545, 2018.
[21] N. Makhathini, I. Musonda, and A. Onososen, “Utilisation of remote
monitoring systems in construction project management,” Lecture
Notes in Civil Engineering, vol. 245, pp. 93–100, 2023.
[22] G. Guo, W. Zhu, Z. Sun, S. Fu, W. Shen, and J. Cao, “An
aero-structure-acoustics evaluation framework of wind turbine blade
cross-section based on gradient boosting regression tree,” Composite
Structures, vol. 337, no. June 2023, p. 118055, 2024.
[23] A. Shehadeh, O. Alshboul, R. E. Al Mamlook, and O. Hamedat,
“Machine learning models for predicting the residual value of heavy
construction equipment: An evaluation of modified decision tree,
lightgbm, and xgboost regression,” Automation in Construction, vol.
129, p. 103827, 2021.
[24] Y. Alzubi, “Comparison of various machine learning models for
estimating construction projects sales valuation using economic vari-
ables and indices,” Journal of Soft Computing in Civil Engineering,
vol. 8, no. 1, pp. 1–32, 2024.
[25] O. Ers ¨
oz, A. F. ˙
Inal, A. Aktepe, A. K. T¨
urker, and S. Ers¨
oz,
“A systematic literature review of the predictive maintenance from
transportation systems aspect,” Sustainability, vol. 14, no. 21, 2022.
[26] P. Parmar. (2021) Outlier detection and re-
moval using the iqr method. Accessed: 2024-09-
15. [Online]. Available: https://medium.com/@pp1222001/
outlier-detection-and-removal-using-the- iqr-method- 6fab2954315d
[27] M. Guerrero Cano, A. Luque Sendra, J. R. Lama Ruiz, and
A. C´
ordoba Rold´
an, “Predictive maintenance using machine learn-
ing techniques,” Proceedings from International Congress on
Project Management and Engineering, 2019.
[28] S. Hosny, E. Elsaid, and H. Hosny, “Prediction of construction
material prices using arima and multiple regression models,” Asian
Journal of Civil Engineering, vol. 24, no. 6, pp. 1697–1710, 2023.

12 Poonam Katyare, et al.
[29] M. A. Mediavilla, F. Dietrich, and D. Palm, “Review and analysis
of artificial intelligence methods for demand forecasting in supply
chain management,” Procedia CIRP, vol. 107, pp. 1126–1131, 2022.
[30] M. K. Das and K. Rangarajan, “Performance monitoring and fail-
ure prediction of industrial equipments using artificial intelligence
and machine learning methods: A survey,” Proceedings of the
4th International Conference on Computational Methodologies and
Communication (ICCMC), vol. 2020, pp. 595–602, 2020.
[31] N. Paviˇ
ci´
c, Z. Reˇ
setar, and F. Luki´
c, “The impact of the increase
in raw material prices on costs in the construction sector in the
city of osijek,” 1st International Scientific Conference on Economy,
Management and Information Technologies – ICEMIT 2023, 2023.
[32] J. Leukel, J. Gonz´
alez, and M. Riekert, “Adoption of machine
learning technology for failure prediction in industrial maintenance:
A systematic review,” Journal of Manufacturing Systems, vol. 61,
no. October, pp. 87–96, 2021.
[33] M. A. Musarat, W. S. Alaloul, A. M. Khan, S. Ayub, and
N. Jousseaume, “A survey-based approach of framework devel-
opment for improving the application of internet of things in the
construction industry of malaysia,” Results in Engineering, vol. 21,
no. January, p. 101823, 2024.
[34] O. T. Sanchez et al., “An iiot-based approach to the integrated
management of machinery in the construction industry,” IEEE
Access, vol. 11, no. January, pp. 6331–6350, 2023.
[35] R. Hidayawanti and Y. Latief, “Raw material optimization with
neural network method in concrete production on precast industry,”
International Journal of GEOMATE, vol. 24, no. 102, pp. 10–17,
2023.
[36] L. Zhang, J. Guo, X. Fu, R. L. K. Tiong, and P. Zhang, “Digital
twin enabled real-time advanced control of tbm operation using
deep learning methods,” Automation in Construction, vol. 158, no.
December 2023, p. 105240, 2024.
[37] J. Brozovsky, N. Labonnote, and O. Vigren, “Digital technologies
in architecture, engineering, and construction,” Automation in Con-
struction, vol. 158, no. November 2023, p. 105212, 2024.
[38] O. Alshboul, A. Shehadeh, M. Al-Kasasbeh, R. E. Al Mamlook,
N. Halalsheh, and M. Alkasasbeh, “Deep and machine learning
approaches for forecasting the residual value of heavy construction
equipment: a management decision support model,” Engineering,
Construction and Architectural Management, vol. 29, no. 10, pp.
4153–4176, 2022.
[39] H. Yang et al., “Optimization of tight gas reservoir fracturing
parameters via gradient boosting regression modeling,” Heliyon,
vol. 10, no. 5, p. e27015, 2024.