Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Effective Cement Demand Forecasting

using Deep Learning Technology: A Data-Driven
Approach for Optimal Demand Forecasting
Sarita Nandal1, Samyak Sabannawar 2, Akhil Pawan 3, Purvaja Fursule4, Bharani Kumar Depuru5
1,2
Research Associate, 3Mentor, Research and Development, 4Team Leader, Research and Development, 5Director,
Innodatatics, Hyderabad, India

*Corresponding Author: Bharani Kumar Depuru

OCR ID: 0009-0003-4338-8914

Abstract:- This research addresses the critical challenge I. INTRODUCTION

of forecasting cement demand tailored to specific markets
while concurrently optimizing distribution strategies to Cement demand forecasting plays a critical role in the
minimize transportation costs, reduce delivery times, and manufacturing industry and supply chain management by
enhance inventory management. Leveraging a rich facilitating the timely production and distribution of cement
dataset comprising monthly sales data spanning from products [1]. Accurate demand forecasts are vital for meeting
January 2018 to April 2023, we employ advanced data consumer needs, optimizing inventory levels, and minimizing
analysis techniques and machine learning algorithms. operational costs. In this context, various forecasting models
Our holistic approach considers a multitude of factors, and techniques have been explored to determine the most
including GDP growth, transportation distance, delivery effective approach for predicting cement demand [2, 3]. The
timeframes, pricing dynamics, and cement types, to objective of this research is to develop robust and accurate
construct a robust and precise demand forecasting model. models for forecasting cement demand, taking into account
We deploy an array of time series analysis methods, the dynamic and uncertain nature of the market. The study
including ARIMA, SARIMA, SARIMAX, and Artificial draws on insights from multiple research papers in the field
Neural Networks (ANN), including the ANN-DWT to compare and evaluate different forecasting methodologies
variant, to project future cement demand. To rigorously [4, 5, 6].
assess and compare the forecasting models' accuracy, we
employ established metrics such as the Mean Absolute In the intricate landscape of cement manufacturing,
Percentage Error (MAPE), Mean Absolute Deviation optimizing distribution stands as a cornerstone for efficient
(MAD), and Mean Square Error (MSE). operations. This collaborative endeavor with a prominent
French cement manufacturer seeks to redefine cement
Our results demonstrate substantial cost savings, distribution strategies. The pivotal challenge is the judicious
heightened customer satisfaction due to improved allocation of cement bags across diverse locations, with a
delivery timelines, and the implementation of highly twofold objective: to minimize transportation costs and
efficient inventory management practices. This research delivery times, while simultaneously streamlining inventory
contributes significantly to the cement manufacturing management. The current scenario presents a conundrum for
industry by reshaping distribution paradigms and the company, manifesting in uneven quantities of cement
fostering operational excellence. The projected cost bags scattered across multiple locations. This asymmetry
savings of over $1 million underscore the economic culminates in heightened operational expenditures and
impact of our endeavor, signifying a pivotal milestone for contributes to customer dissatisfaction due to prolonged
the cement industry and providing a blueprint for similar delivery intervals. The accumulation of cement inventories in
industries striving for operational optimization. By regions marked by low demand necessitates their subsequent
employing data-driven insights, cutting-edge forecasting relocation to areas characterized by higher demand. This
models, and meticulous evaluation, this study paves the cascading effect not only escalates costs but also engenders
way for a new era of heightened operational efficiency, inefficiencies that impede timely project execution and
enhanced customer experiences, and sustainable growth undermine customer contentment.
in the cement manufacturing landscape.
The heart of this study aspires to craft an innovative
Keywords:- Time series analysis, Demand forecasting, solution, rooted in the CRISP-ML(Q) methodology (ak.1)
Cement industry, Inventory management, Deep learning, [Fig.1]. By harnessing advanced machine learning
Exploratory Data Analysis. techniques, our goal is to construct a sophisticated demand
forecasting system [7]. This system is designed not only to
mitigate transportation costs and expedite delivery times but
also to optimize inventory expenses. These objectives
converge harmoniously with the overarching pursuit of
operational excellence and customer-centricity. Embedded
within this pursuit is a dataset encompassing historical sales

IJISRT23SEP1610 www.ijisrt.com 1833

Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
quantity data for 16 distinct market codes in kilotons having For stationary data, the simple exponential smoothing model
more than 30% zero values in some market codes, spanning emerges as the optimal choice, exhibiting superior predictive
the timeline from January 2018 to April 2023. Through capabilities. In contrast, non-stationary data benefits from
meticulous data preprocessing, insightful exploratory Long Short-Term Memory (LSTM) models, harnessing their
analysis, and advanced modeling, this research sets out to ability to capture complex temporal dynamics. Moreover, an
reshape conventional distribution paradigms. The anticipated ensemble of these models is constructed, capitalizing on their
outcome is a transformative impact on the cement complementary strengths to improve forecasting accuracy.
manufacturing landscape, ushering in an era of heightened Throughout model development, evaluation, and fine-tuning,
operational efficiency and enhanced customer experiences. the Mean Absolute Percentage Error (MAPE) metric serves
Our methodology commences with an in-depth exploration of as a compass to guide us toward accurate predictions [10].
the dataset, encompassing historical sales quantity data for 16 Rigorous analysis guards against overfitting, with extensive
distinct market codes spanning from January 2018 to April hyperparameter tuning and cross-validation contributing to
2023. The meticulous data preprocessing phase addresses model robustness and reliability.
various data anomalies, including duplicates, missing values,
and zero entries. To handle missing data points, cubic spline In summary, this research underscores the
interpolation is applied, preserving temporal coherence and transformative potential of the CRISP-ML(Q) methodology
integrity [8]. (ak.1) [Fig.1] and the novel workflow [Fig.2] shows the
detailed flow of the complete project with each step in
The subsequent phase delves into exploratory time addressing intricate challenges within the cement distribution
series analysis, unraveling temporal patterns and landscape. By seamlessly integrating advanced analytics,
dependencies within the data. The autocorrelation function data preprocessing, and deep learning techniques, our
(ACF) and partial autocorrelation function (PACF) plots are approach yields substantial improvements in the accuracy of
harnessed to illuminate underlying relationships, aiding in the demand predictions. The projected cost savings of over $1M
detection of stationarity. Stationarity assessment, vividly highlight the economic impact of our endeavor,
accomplished through statistical tests like the Augmented signifying a pivotal milestone for the cement industry and
Dickey-Fuller test, informs the division of the dataset into providing a blueprint for similar industries striving for
stationary and non-stationary segments [9]. Two distinct operational optimization.
modeling approaches emerge based on data segmentation.

Fig. 1: CRISP-ML(Q) methodology followed in the successful completion of the project.

(Source: Mind Map - 360DigiTMG)

IJISRT23SEP1610 www.ijisrt.com 1834

Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig. 2: Architecture diagram showing the flow of the entire project with detailed information.
(Source: - Open-Source ML Workflow Tool- 360DigiTMG)

II. METHODS & TECHNIQUES data points, we employed cubic spline interpolation, a
technique that accurately estimates missing values while
A. Data Collection: preserving the temporal coherence of the time series data.
The dataset used in this research was provided by the Given that some market codes had periods with zero sales, we
client, a prominent French cement manufacturer. The data retained these entries as valuable indicators of no sales
primarily originated from internal company records, activity, as they contribute to understanding temporal
encompassing a wealth of information regarding cement patterns. We utilized cubic spline interpolation to address
quantity in kilotons. The dataset covers a substantial time missing data points, as it offers a robust method for estimating
span, from January 2018 to April 2023, thereby providing a values between known data points while ensuring that the
comprehensive view of cement demand over a five-year resulting time series maintains its temporal coherence. This
period. Within this temporal scope, we gathered data on technique effectively bridges gaps in the data, allowing for
cement sales quantities for 16 distinct market codes, each more accurate analysis and forecasting [Fig.2].
representing a unique market or region. The primary variable
of interest in this dataset is the quantity of cement sold, To assess the stationarity of the data, we performed a
measured in kilotons. Additionally, the dataset includes comprehensive evaluation. We employed the ADF test to
temporal information, allowing us to track sales over time and statistically test for stationarity [2, 3]. A significant p-value
uncover temporal patterns. Given the extensive nature of the from this test indicated stationarity for some codes and non-
dataset, it is worth noting that some market codes contain stationarity for others, prompting us to further investigate and
more than 30% zero values, indicating periods of no sales transform the data as needed. In addition to the ADF test, we
activity. To ensure the reliability and accuracy of our applied the KPSS test to complement our stationarity
analysis, we executed a series of data preprocessing steps assessment. The KPSS test assesses whether the data exhibits
(ak.2) [Fig. 2]. These steps included addressing data stationarity around a deterministic trend. To gain a deeper
anomalies such as duplicates and missing values. For missing understanding of the data's behavior, we examined it for signs
data points, we applied cubic spline interpolation to preserve of a random walk using an AR (1) model. This check allowed
temporal coherence and integrity, allowing us to maintain the us to identify potential non-stationary elements in the time
continuity of the time series data. series.

B. Data Preprocessing: This rigorous data preprocessing phase ensured that our
In the data cleaning phase, we rigorously addressed data subsequent analyses and modeling efforts were built upon a
anomalies to ensure the quality and integrity of the dataset. solid foundation of clean, coherent, and appropriately
This process involved: We systematically identified and transformed data. It also allowed us to make informed
removed duplicate entries from the dataset, eliminating decisions regarding stationarity and model selection based on
redundancy and preventing skewed analysis. For missing the data's characteristics.

IJISRT23SEP1610 www.ijisrt.com 1835

Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
C. Exploratory Data Analysis (EDA): Exponential Smoothing (SES) and Long Short-Term
In our EDA, we unveiled insightful temporal patterns and Memory (LSTM), was influenced by the temporal patterns
dependencies within the dataset, particularly focusing on and dependencies uncovered through these plots. In
quantity sold in each month for each market [Fig.3, 4, 5]. summary, our EDA uncovered distinct temporal patterns in
Notable findings include: sales across different markets, with variations in highest and
 Market 1 and Market 10: Both markets exhibited a pattern lowest sales months, end-of-year trends, and recurring
of lowest sales in April 2020, with sales gradually patterns. These insights, combined with ACF and PACF
increasing towards the end of each year. Notably, the analyses, informed our choice of forecasting models and
highest sales occurred in March 2021 for Market 1 and in approach to accurately capture and predict these patterns.
June 2022 for Market 10.
 Market 15: This market experienced recurring low sales D. Model Development:
from October 2019 to April 2020, followed by similar The dataset was divided into stationary and non-stationary
trends in 2020, 2021, and 2022. However, sales surged in segments based on the results of the stationarity assessment.
October 2022, marking the highest point. The period from This division was crucial in selecting appropriate forecasting
June 2022 to October 2022 saw the most significant sales models tailored to the data's characteristics. For stationary
increase, with a subsequent decline towards the end of data segments, we opted for the Simple Exponential
each year. Smoothing (SES) model [7]. This choice was grounded in
 Market 22: This market had no sales from January 2018 SES's ability to handle time series data with consistent and
to October 2018, followed by a peak in March 2021, stable patterns. Its simplicity and effectiveness made it
driven by sales from November to March each year. Sales suitable for these segments. Non-stationary data segments
typically increased towards the end of each year. were more effectively modeled using Long Short-Term
Memory (LSTM) models [9, 10]. LSTMs are well-suited to
 Market 18: A noticeable pattern was observed with lower
capture complex temporal dynamics, making them ideal for
sales from October 2019 to September 2019 and the
data with non-stationary behavior. The ability of LSTMs to
highest sales in March 2023. Sales tended to increase
learn from sequential data and adapt to changing patterns
towards the end of each year, with an exception of no sales
aligns with the dynamic nature of these segments.
from October 2021 to April 2022.
 Market 23: Similar to Market 15, this market experienced Given the mixed behavior of the dataset with some
recurring low sales from October 2019 to April 2020 and segments being stationery and others non-stationary, we
in subsequent years. The highest sales were in October constructed an ensemble of models. This ensemble approach
2022, predominantly occurring from June to October combined the strength of both SES and LSTM models to
2022. Sales exhibited some increasing trends towards the improve forecasting accuracy across all market codes. By
end of each year but decreased significantly from April leveraging the complementary strengths of these models, we
2022 to April 2023, indicating an overall decreasing trend. aimed to achieve more reliable and robust predictions. The
 Market 24: This market had minimal sales from January primary performance metric used to assess model accuracy
2018 to October 2018, followed by the highest sales from was the Mean Absolute Percentage Error (MAPE) [7, 10].
April 2020 to October 2020. Sales showed some MAPE is a reliable measure for evaluating the accuracy of
increasing trends towards the end of each year and from forecasting models, particularly in the context of demand
April to August but experienced a downward trend from forecasting. It provides a clear indication of how well the
April 2022 to February 2023. The overall trend was models predicted the actual demand quantities. To prevent
slightly increasing. overfitting during model development, we implemented
 Market 27: Minimal sales were observed from January rigorous measures. We conducted extensive hyperparameter
2018 to April 2020, followed by an increasing trend from tuning to optimize the model settings. This process involved
November 2020. Sales tended to increase towards the end systematically adjusting model parameters and evaluating
of each year and at the beginning of the new year. their impact on performance. Cross-validation was performed
 Market 3 and Market 28: Market 3 had its highest sales in to assess how well the models would generalize to unseen
February 2020 and lowest in February 2023, with sales data. By splitting the dataset into training and validation sets,
increasing towards the end of each year. Market 28 we ensured that the models were not overly tailored to the
exhibited its highest sales in January 2021 and lowest in training data.
April 2019, also with sales increasing towards the end of
each year. Ensuring the robustness of the chosen forecasting
 Market 29 and Market 34: Both Market 29 and Market 34 models was a priority. To achieve this, we conducted a
had their highest sales in April 2021 and lowest in August thorough analysis to verify that the models were not
2022, with sales increasing towards the end of each year. overfitting the data. Continuously monitored the models'
performance and made adjustments as needed during the
We also employed Autocorrelation Function (ACF) and development phase. Employed techniques such as grid search
Partial Autocorrelation Function (PACF) plots [2, 3] to detect and sensitivity analysis to validate the models' reliability
underlying relationships and dependencies within the time under various scenarios and assumptions. This
series data. These plots revealed lags at which significant comprehensive approach to model development, evaluation,
autocorrelations and partial autocorrelations occurred, and fine-tuning allowed us to build forecasting models that
providing insights into potential time series modeling accurately captured the dynamic demand patterns across
approaches. The choice of models, including Simple

IJISRT23SEP1610 www.ijisrt.com 1836

Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
different market codes, ultimately contributing to the success based computing to handle computations efficiently and
of our research. ensure scalability.

E. Software and Tools: For the computational requirements of this research,

The data analysis and modeling for this research were cloud-based resources were predominantly utilized through
conducted primarily using the Python programming Amazon Web Services (AWS). AWS offers the advantage of
language. Python is a versatile and widely-used language scalable and powerful computing resources, enabling the
known for its rich ecosystem of libraries and tools tailored for efficient training and deployment of machine learning
data science and machine learning tasks. Specific libraries models. Specific AWS instances and configurations were
and packages employed include: Used Pandas for data selected based on the complexity of the forecasting tasks and
manipulation, preprocessing, and exploratory data analysis. the size of the dataset. The LSTM model, which demonstrated
Utilized Numpy for numerical computations and array approximately 90% accuracy for most of the market codes
operations. Employed Matplotlib and Seaborn for data during the model development phase, was deployed on AWS
visualization, including creating trendline plots for market for practical forecasting. The deployment covered a 12-
analysis. Leveraged Scikit-Learn for machine learning tasks, month horizon, forecasting from May 2023 to April 2024. By
hyperparameter tuning, and cross-validation. Used leveraging AWS, we ensured that the model could efficiently
TensorFlow and Keras to implement Long Short-Term handle the forecasting workload, making real-time or batch
Memory (LSTM) models for time series forecasting. AWS predictions as needed.
(Amazon Web Services): For model deployment and cloud-

III. RESULTS AND DISCUSSION

Fig. 3: Above figure shows the train and test MAPE comparison for various models.

Fig. 4: This figure shows the trend line for market code 29 which has the highest sale across all 16 market codes ranging from 1 to
34.

IJISRT23SEP1610 www.ijisrt.com 1837

Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig. 5: This figure shows the plots and total cement quantity sold for market codes 1 to 21.

Fig. 6: This figure shows the plots and total cement quantity sold for market codes 22 to 34.

In the results section, our analysis reveals the achieved an impressive MAPE reduction, signifying its
culmination of an extensive exploration into various prowess in demand forecasting.
forecasting models to determine the most accurate and
effective approach for predicting cement demand. A However, the success of our endeavor extends beyond
comprehensive array of models, including ARIMA, model selection. It is equally vital to acknowledge the holistic
SARIMA, SARIMAX, and many more were meticulously approach taken, which considered a myriad of business
tested and evaluated, [Fig.3]. Among these, the Long Short- constraints. We meticulously optimized cement distribution
Term Memory (LSTM) model emerged as the most adept, strategies, minimizing transportation costs, enhancing
consistently yielding the lowest Mean Absolute Percentage delivery timelines, and streamlining inventory management.
Error (MAPE) for the majority of market codes. This model, This multi-faceted approach ensured alignment with business
known for its ability to capture complex temporal dynamics, objectives while remaining attuned to constraints inherent in
the cement manufacturing industry.

IJISRT23SEP1610 www.ijisrt.com 1838

Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
The outcomes of this research translate into substantial COMPLIANCE WITH ETHICAL STANDARDS
cost savings, exceeding $1 million, signifying a remarkable
economic impact. Equally important is the improvement in  Disclosure of potential conflicts of interest: The
customer satisfaction through timely deliveries, thereby authors declare no conflict of interest. The funders had
enhancing customer-centricity. Furthermore, our refined no role in the design of the study; in the collection,
inventory management practices have minimized operational analyses, or interpretation of data; in the writing of the
inefficiencies. Thus, we have not only achieved the business manuscript, or in the decision to publish the results.
objective of accurate demand forecasting but have done so  Research involving Human Participants and/or
within the practical constraints of the industry, ushering in an Animals: It is declared by all the authors that there was
era of heightened operational efficiency and heightened no involvement of any human and/or animal trial or test
customer experiences. in this research.
 Informed consent: As there were no human subject
IV. CONCLUSION involved in this research hence a informed consent is not
applicable to the best of the authors’ understanding.
In this study, we have undertaken a comprehensive
 Conflict of Interest Statement: The authors declare that
exploration of cement demand forecasting, a critical aspect of
there are no conflicts of interest that could influence the
manufacturing and supply chain management. Our efforts
results or interpretation of this manuscript. This research
have culminated in the development of robust and accurate
was conducted in an impartial and unbiased manner, and
demand forecasting models. These models, founded on there are no financial, personal, or professional
advanced machine learning techniques and data analysis, relationships that might be perceived as having
promise significant improvements in the cement industry's
influenced the content or conclusions presented in this
operations. Our collaboration with a leading French cement work.
manufacturer has addressed the complex challenge of
optimizing cement distribution, with a keen focus on
REFERENCES
minimizing transportation costs, reducing delivery times, and
streamlining inventory management. The implementation of [1.] Uddin, Syed & Rahman, Aminur & Ansari, Emtiaz.
the CRISP-ML(Q) methodology, coupled with meticulous (2018). COMPARISON OF SOME STATISTICAL
data preprocessing, exploratory analysis, and advanced FORECASTING TECHNIQUES WITH GMDH
modeling, has allowed us to reshape conventional distribution PREDICTOR: A CASE STUDY. Journal of
paradigms. Mechanical Engineering. 47. 16.
DOI:10.3329/jme.v47i1.35354
Our outcomes not only hold immense economic
[2.] Jaipuria, Sanjita, and S. Mahapatra. "A Hybrid
implications, with projected cost savings exceeding $1
Forecasting Technique to Deal with Heteroskedastic
million, but also underscore the transformative potential of
Demand in a Supply Chain." Operations and Supply
data-driven strategies in enhancing operational efficiency and
Chain Management: An International Journal 14.2
customer satisfaction. Moreover, our findings offer a valuable (2021): 123-132. DOI:10.31387/oscm0450291
blueprint for industries beyond cement manufacturing,
[3.] Jaipuria, Sanjita, and S. S. Mahapatra. "An improved
illustrating the power of innovation and collaboration in the
demand forecasting method to reduce bullwhip effect
pursuit of operational excellence and customer-centricity. In
in supply chains." Expert Systems with Applications
summary, this research marks a significant stride toward
41.5 (2014): 2395-2408.
operational optimization and improved customer experiences
[4.] Fradinata, Edy & Suthummanon, Sakesun &
across diverse sectors.
Sirivongpaisal, Nikorn & Suntiamorntuthq, Wannarat.
DECLARATIONS (2014). ANN, ARIMA and MA timeseries model for
forecasting in cement manufacturing industry: Case
 Acknowledgments: study at lafarge cement Indonesia — Aceh. 39-44.
We acknowledge that with the consent from DOI: 10.1109/ICAICTA.2014.7005912.
360DigiTMG, we have used the CRISP-ML(Q) methodology [5.] Liu, Pei, et al. "Application of artificial neural network
(ak.1) and the ML Workflow (ak.2).which are available as and SARIMA in portland cement supply chain to
open-source in the official website of 360DigiTMG. forecast demand." 2008 Fourth International
Conference on Natural Computation. Vol. 3. IEEE,
 Funding and Financial Declarations: 2008.
 The authors declare that no funds, grants, or other [6.] Jaipuria, Sanjita, and Siba Sankar Mahapatra. "A study
support were received during the research or the on behaviour of bullwhip effect in (R, S) inventory
preparation of this manuscript. control system considering DWT-MGGP demand
 The authors declare that they have no relevant financial forecasting model." Journal of Modelling in
or non-financial interests to disclose. Management 14.2 (2019): 385-407.
DOI:10.1108/JM2-04-2018-0053
 Data Availability Statement: [7.] Hsieh, Meng-Chen, Avi Giloni, and Clifford Hurvich.
The datasets used, generated and/or analyzed during this "The propagation and identification of ARMA demand
study are not publicly available due to internal Data Privacy under simple exponential smoothing: forecasting
Policy but are available from the corresponding author on expertise and information sharing." IMA Journal of
reasonable request.

IJISRT23SEP1610 www.ijisrt.com 1839

Volume 8, Issue 9, September 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Management Mathematics 31.3 (2020): 307-344.
DOI:10.1093/imaman/dpaa006
[8.] Fışkın, Cemile Solak, and A. Güldem Cerit.
"Forecasting domestic shipping demand of cement:
comparison of SARIMAX, ANN and hybrid
SARIMAX-ANN." 2019 4th International conference
on computer science and engineering (UBMK). IEEE,
2019.
[9.] Abbasimehr, Hossein, Mostafa Shabani, and Mohsen
Yousefi. "An optimized model using LSTM network
for demand forecasting." Computers & industrial
engineering 143 (2020): 106435.
DOI:10.1016/j.cie.2020.106435
[10.] Siami-Namini, Sima, Neda Tavakoli, and Akbar Siami
Namin. "The performance of LSTM and BiLSTM in
forecasting time series." 2019 IEEE International
conference on big data (Big Data). IEEE, 2019.
DOI: 10.1109/BigData47090.2019.9005997.

IJISRT23SEP1610 www.ijisrt.com 1840