Available online at www.sciencedirect.
com
ScienceDirect
Available online at www.sciencedirect.com
Procedia
Procedia Manufacturing
Manufacturing 00
00 (2019)
(2019) 000–000
000–000
ScienceDirect www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
Procedia Manufacturing 38 (2019) 256–263
29th International Conference on Flexible Automation and Intelligent
Manufacturing (FAIM2019), June 24-28, 2019, Limerick, Ireland.
Inventory management and cost reduction of supply chain processes
using AI based time-series forecasting and ANN modeling
Umamaheswaran Praveen, Ganjeizadeh Farnaz, Ghasib Hatim
California
California State
State University
University ,, East
East Bay,
Bay, United
United States
States of
of America
America
Abstract
Reducing waste, and therefore cost, within a supply chain can be a very challenging process due to the large
number of variables involved. One of the biggest wastes that are usually present in a supply chain are unnecessarily
high inventory costs and shortage costs which are caused by errors in demand forecasting. A high variance between
the forecasted demand and the actual demand results in costs that could have been avoided. To eliminate that
waste, we developed a model that utilizes artificial neural network to accurately forecast the demand. The model
performs forecasting analysis based on multilayer feed-forward neural network with backpropagation. The use of
machine learning can assist with the rapid changes in customer demand. Our holistic solution will minimize the
supply/demand mismatch and its associated costs and consequently increase profit margins
© 2019 The Authors. Published by Elsevier B.V.
© This
2019isThe
an open access
Authors, article under
Published the CCB.V.
by Elsevier BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review
Peer underthe
review under responsibility of the
responsibility scientific
of the committee
scientific of theofFlexible
committee Automation
the Flexible and Intelligent
Automation Manufacturing
and Intelligent 2019 2019
Manufacturing
(FAIM 2019)
Keywords: ANN;Supply
ANN;Supply chain
chain management;
management; Time-series
Time-series Forecasting;
Forecasting; Multilayer
Multilayer Feed-Forward
Feed-Forward Neural
Neural Network;
Network; Back
Back
propagation;
propagation; Optimization;
Optimization; process
process enhancement;
enhancement; R-Studio;
R-Studio; Predictive
Predictive Neural
Neural Network
Network
2351-9789
2351-9789 ©© 2019
2019 The
The Authors,
Authors, Published
Published byby Elsevier
Elsevier B.V.
B.V.
Peer
Peer review
review under
under the
the responsibility
responsibility of
of the
the scientific
scientific committee
committee of
of the
the Flexible
Flexible Automation
Automation and
and Intelligent
Intelligent Manufacturing
Manufacturing 2019
2019
2351-9789 © 2019 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the Flexible Automation and Intelligent Manufacturing 2019 (FAIM 2019)
10.1016/j.promfg.2020.01.034
Umamaheswaran Praveen et al. / Procedia Manufacturing 38 (2019) 256–263 257
2 Author name / Procedia Manufacturing 00 (2019) 000–000
1. Introduction
Manufacturing and distribution companies with their various supply chain networks often operate at below
maximum potentials and lose out on millions of dollars annually. Thus, utilizing the ideal software intelligence in
today's time can help increase the performance of these supply chain networks to their max potential. One way to
achieve this is by implementing machine learning into Enterprise Resource Planning (ERP) systems. ERP software
integrates various functions such as goods, employees and customers into one complete system to streamline
processes and information across the entire organization. Machine learning has many tools and the one being used
in our system is Artificial Neural Networks, considered as “nonlinear statistical data modeling tools where the
complex relationships between inputs and outputs are modeled or patterns are found.” (Techopedia, 2018). This
marriage of systems can lead to optimizations in current demand, demand in the next 3 months, the current
warehouse values purchase pricing and transport cost.
2. Literature Review
Literature search is conducted to compare various methodologies and to provide a background to this study. A
summary of the literature review, topic, method, strengths and weaknesses of these sources used for this study is
given in Table 1, categorized by the method used and strength and weakness of each paper.
Table 1: literature review summary
Reference Topic Method Strength Weakness
Jad Farhata, ERP Neural Network Embedding ANN Increased forecasting accuracy Systematic Input
Michel Inventory Control algorithm with and religious
Owayjan ERP tracking
required
Sven F. Crone Artificial Neural Seasonal cost and Seasonal trends taken into account, Considers only
Networks for Time Series inventory Increased forecasting accuracy under estimation
Prediction - a novel forecasting using and not over
Approach to Inventory ANN forecasting estimation of
Management using techniques and inventory
Asymmetric Cost ARIMA models
Functions
Bo Xing, Wen- Artificial Intelligence In Reverse Model able to assist practitioners Only a certain
Jing Gao, Reverse Supply Chain engineering using decision AI algorithms
Kimberly Management: The State AI algorithms tested.
Battle, Tshilidzi Of Art
Marwala,
Fulufhelo V.
Nelwamondo
Tereza Šustrová A Suitable Artificial Using AI Takes into account all aspects One single over
Intelligence Model for forecasting to required to make an order the top bad or
Inventory Level determine including demand, forecast, price good season
Optimization Quantity to order and seasonality could topple the
model
Erik Fosser, Ole ERP systems and Link between ERP ERP system can be used to feed Changing from
Henrik Leister, competitive advantage: systems and data warehouse and gain SAP to ERP
Carl Erik Moe, some initial results competitive competitive advantage would be
Mike Newman advantage difficult due the
high associated
costs
Andrew Harvey, Structural time series Using statistical Models are used for deriving Kalman filtering
Ralph D. models in inventory models in formulae for estimating the mean can be used for
Snyder control inventory and variance of the lead time only linear state
management demand distribution under both transitions
constant and stochastic time
assumptions.
Aburto, L., Improved supply chain Hybrid Detailed instructions shown to Learning
Weber, R. management based on Forecasting improve supply chain management process can lead
hybrid demand forecasts. Models systems using ARIMA type to an overfitted
Applied Soft Computing. process. Backpropagation is used model, thus
to adjust network’s weights. losing its ability
to generalize.
G. Dorffner Neural Networks for Time Forecasting Ability to approximate arbitrary Model only has
Series Processing technique using nonlinear functions. Classical short term use.
ARIMA model pattern recognition, detecting Nonlinear
systematic patterns in an array of functions cannot
measurements which do not be used for
change in time. larger problems.
Mohamed T. Back-Propagation Neural network Cost estimation without defining Model is limited
Kotb, Moutaz Artificial Neural Network model the function points. This model can in data
Haddara, and for ERP Adoption Cost be extended and can be applied in processing.
Yehia T. Kotb Estimation various environments
258 Umamaheswaran Praveen et al. / Procedia Manufacturing 38 (2019) 256–263
Author name / Procedia Manufacturing 00 (2019) 000–000 3
This summary indicates that although ANN model has some weaknesses but comparatively is more
effective modelling method to be implemented in time series forecasting .
Umamaheswaran, Praveen ; Ganjeizadeh, Farnaz ; Ghasib, Hatim / Procedia Manufacturing 00 (2018) 000–000 3
3. Methodology
The prediction technique used is Artificial Neural Network (ANN) modeling as it is capable of handling data
with high volatility more accurately. ANN thus overcomes the shortcomings of traditional forecasting models.
The algorithm and implementation of this approach is presented in the following sections.
3.1. Algorithm
To use Artificial Neural Network (ANN) to accurately forecast the future demand, we needed to develop an
algorithm. The following steps are used to develop our proposed model.
Step 1: Determine the neural network architecture to be used to design the system. Determine the number of
input nodes, hidden nodes, hidden layers and output nodes.
Step 2: Data cleaning
• Remove NAs, which causes error in the model
• Eliminate unnecessary columns and add columns based on the time series forecasting model- linear trend,
quadratic trend
• Normalize the data after all information has been filtered
• Divide dataset in two halves for training (2010-2011, 2011-2012) and testing (2012-2013)
Step 3: Create a training sample with feed-forward network and resilient backpropagation (RPROP) with weight
backtracking
• Use Logistics function as the activation function
• Set the learning rate
• Calculate the actual
output Step 4: Testing
the dataset
• Validate the result by comparing the prediction with actual data
(2013) Step 5: Minimize the error
• Feed the difference between true and predicted value into the system (backward pass)
• Update weights and bias values
Upon completing the application via above steps a mode was generated that minimizes the error in the demand
forecast. The details of the modules is presented in the following sections.
3.2. Multilayer Feed-Forward Neural Network with Back-propagation
The type of neural network used in the proposed system is feedforward network wherein information moves
Umamaheswaran Praveen et al. / Procedia Manufacturing 38 (2019) 256–263 259
4 Author name / Procedia Manufacturing 00 (2019) 000–000
only in one-direction, which is from input layer to output layer as shown in Figure 1. The input units or
computational units pass through multiple hidden layers which are stacked between the input and output layer. All
three layers are interconnected. Initially, random weight and bias values are assigned to the data points. The
predicted output is compared with the actual data to calculate the error.
Supervised learning technique is used to design this system. In this technique network is trained with a dataset
with defined output for every input neuron. The weights are then adjusted to get the observed value closest to the
true value which is achieved by the method of backpropagation. Backpropagation refers to propagating the error
back to the network.
4 Umamaheswaran, Praveen ; Ganjeizadeh, Farnaz ; Ghasib, Hatim / Procedia Manufacturing 00 (2019) 000–000
Figure 1: Sample feed forward and back propagation neural network with input nodes, hidden nodes, hidden layers and output nodes
The weights are then adjusted to get the observed value closest to the true value which is achieved by the method
of backpropagation. Backpropagation refers to propagating the error back to the network. Weight and bias values
can be updated using this feedback. End goal is to keep changing the weight and bias until minimum error value is
achieved.
3.3. Research Design
The primary goal of this investigation is to develop a new forecasting method that performs more robust thus
make the supply chain more efficient. The ANN algorithm executes at the most evident level producting promising
results as outlined in the proceeding sections.
3.4. Data Collection
The dataset used here is obtained from Kaggle (give reference number). This dataset represents information
about Walmart. The dataset includes past sales of 45 Walmart stores arranged in various areas, wherein each store
contains different departments. The information is gathered week by week for various kind and size of Walmart
stores.
The week after week deals incorporate the limited time markdowns which go before the occasions consistently.
Four biggest occasions incorporate Super Bowl, Labor Day, Thanksgiving and Christmas. Weeks including such
occasions have five times more weightage than the standard weeks.
3.5.1. Training
260 Umamaheswaran Praveen et al. / Procedia Manufacturing 38 (2019) 256–263
Author name / Procedia Manufacturing 00 (2019) 000–000 5
The training dataset is historical data for 45 Walmart stores located in different regions. Each store has 8
departments. The train data covers the time from January 7, 2011 till November 26,2012 (01/07/2011 to
10/26/2012). It comprises of various fields as:
• Store - The store number
• Dept - The department number
• Date - The precise date of the recorded Sales
• Actual Data - Sales for the given department in the given store
• Linear Trend - Linear time series demand forecast
• Quadratic Trend - Quadratic time series demand forecast
• Actual % of trend - Trend evaluation in percentage
The training sets were used to construct the model and using similar data sets to test the model.
3.5.2. Test
The Data set used for testing is similar to the training sets, but it only excludes the Actual Sales Feature. The
dataset ranges from November 2, 2011 till June 26, 2013 (11/02/12 to 7/26/13). The analysis was conducted by
constructing the model via training first. The accuracy of the model was evaluated via test sets.
4. Analysis
4.1. Application of Methods
In this section we will discuss the various methods used in our approach for a desired result. As discussed
above, we have mainly used two platforms for the deployment of our process. Excel to process and assemble our
dataset as an input; and R Studio for running the code and testing the forecasting results obtained.
4.1.1. Excel
Initially, the raw dataset obtained was incomplete comprising of only store number, department, date of sale,
and actual sale as shown in Table 2. To compute results, TREND function in EXCEL was used. Two Types of
trends were identified as Linear Trend and Quadratic Seasonal Trends as shown in Table 2 as:
• Linear Trend: Linear time series demand forecast
• Quadratic Trend: Quadratic time series demand forecast
Table 2: Screenshot of dataset in MS Excel
We added trends in the datasets to increase number of inputs in the hidden layer of the neural network, to elevate
Umamaheswaran Praveen et al. / Procedia Manufacturing 38 (2019) 256–263 261
6 Author name / Procedia Manufacturing 00 (2019) 000–000
the results and obtain a more accurate and realistic response. Also, utilizing “lookups “ and “pivot” tables in Excel
to clean the dataset by removing null values and organizing the large entry list. The filtered dataset was used in R
studio as input.
4.1.2. R Studio (Packages)
The following packages are used to analyze data in R studio. The Screenshot of the output from R-Studio is
shown in figure 3.
Tidyverse: A collection of R packages designed for data science.
Neuralnet: Neural net package is used to train neural networks using backpropagation, resilient backpropagation
(RPROP) with or without weight backtracking. This function allows flexible settings through the choice of error
and activation function and calculation of generalized weights.
Caret: It is a set of functions that attempt to streamline the process for creating predictive models.
CaTools: Contains several basic utility functions including: moving (rolling, running) window statistic
functions, base64 encoder/decoder, round-off error free sum etc.
Amelia: Uses a Bootstrap+EM algorithm to insert missing values from a dataset and produces multiple output
datasets for analysis. The above listed packages were used to filter, edit, and graph the data as shown in Figure 2.
4.2. Result Analysis
After analyzing the results obtained from the R model, a neural network schematic comprising of three hidden
layers and a total of six hidden nodes was generated as shown in Figure 2. Every layer holding an equal weightage
of 1 in this network.
Figure 2: Screenshot of the output from R-Studio
All the predicted sale figures show linear characteristics compared to the erratically distributed original sale data
as shown in figure 3. The black lines show the connection between each layer and the weights on each connection,
while the blue lines show the biased terms added in each step.
By visually inspecting the plot generated by R Studio in Figure 3, we can conclude that the predictions made by
the neural networks are more concentrated around the fitted line with certain errors away from the line. A perfect
alignment around the line would indicate zero error and thus an ideal prediction.
262 Author name /Praveen
Umamaheswaran ProcediaetManufacturing 00 (2019) 000–000
al. / Procedia Manufacturing 38 (2019) 256–263 7
Figure 3: Screenshot of the graph from R-Studio
The forecasting accuracy and performance achieved is around 70-80% higher than the average performance of
most supply chain networks worldwide.
7
4.3. Strengths and Weaknesses
The pros and cons of the proposed approach is presented in Table 3. ANN Models can be applied to non-
linear complex problems and can make predictions by inferencing hidden relationships. The models are also a good
predictor when high variability and non-constant-variance is present in the data sets. It should be noted that ANN-
based models generate better results when input /output relationship is simple.
Table 3. Pros and cons of the proposed approach
Strengths Weaknesses
ANN model handles non-linear If there are too many hidden nodes, the
and complex problems model might produce inferior results
They are capable of inferring the hidden ANN’s are best applied to problems where
relationships and make predictions input and output data are fairly simple. With
large dataset, the network topology can
become too complex and difficult to
interpret
Such models handle heteroskedasticity; data It can be a time-consuming process
with non- constant variance and high
volatility in a better way
5. Conclusion
The goal of this study was to improve the overall performance and efficiency of the supply chain management,
and this has been achieved by using the Artificial Neural Network algorithm. The average performance for the
supply chain network is about 72%, the accuracy achieved in this study by implementing an AI system which uses
the ANN algorithm is about 75%-80% which is two to three percent higher than the average accuracy, hence
improves the overall performance and efficiency of the supply chain network. This model improves the forecasting
Umamaheswaran Praveen et al. / Procedia Manufacturing 38 (2019) 256–263 263
8 Author name / Procedia Manufacturing 00 (2019) 000–000
analysis and gives a better prediction of future sales.
The overall improved efficiency of the supply chain networks also results in various other benefits. It reduces the
overall operational costs which includes storage and transportation costs. This eventually improves the supply
chain surplus.
References
[1] Aburto, L., Weber, R. (2007). Improved supply chain management based on hybrid demand forecasts. Retrieved from
Applied Soft Computing,7(1), 136-144. doi:10.1016/j.asoc.2005.06.001.
[2] Crone, S. F. (n.d.). Artificial Neural Networks for Time Series Prediction - a novel Approach to Inventory Management using
Asymmetric Cost Functions. Retrieved from https://pdfs.semanticscholar.org/2c33/0dbff659d50eaa22b163a12b63ae4594b549.pdf.
Hamburg.
[3] Dorffner, G. (1996). Neural Networks for Time Series Processing. Retrieved from Neural Network World, 4, 1996, 6, 447-468.
[4] Xing, B., Gao, W., Battle, K., Marwala, T., & Nelwamondo, F. V. (n.d.). Artificial Intelligence In Reverse Supply Chain
Management: The State Of Art. Retrieved from https://pdfs.semanticscholar.org/60fc/df407a55224a7d7e56ab443afc4b409990cd.pdf.
[5] Šustrová, T. (n.d.). A Suitable Artificial Intelligence Model for Inventory Level Optimization. Retrieved
from https://dspace.vutbr.cz/handle/11012/63168.
[6] Farhat, J., Owayjan, M. (2017, November 17). ERP Neural Network Inventory Control. Retrieved from www.sciencedirect.com.
[7] Faggella, Daniel. “What Is Machine Learning? - An Informed Definition.” TechEmergence, 3 Sept. 2017,
www.techemergence.com/what-is- machine-learning/. ThippeSwamy, K., Amardeep, R. (2017, January). Training Feed forward Neural
Network With Backpropagation Algorithm. Retreived from https://www.ijecs.in/index.php/ijecs/article/download/1905/1760/
[8] Techopedia.com. (2018). What is an Artificial Neural Network (ANN)?. Retrieved
from https://www.techopedia.com/definition/5967/artificial-neural-network-ann
[9] Gao, Wen-Jing, et al. “Computational Intelligence in Used Products Retrieval and Reproduction.” Research Methods, pp. 1188–
1230., doi:10.4018/978-1-4666-7456-1.ch052.
[10] Stergiou, C., Siganos, D. (n.d.). Neural Networks. Retrieved from https://www.doc.ic.ac.uk/~nd/surprise_96/journal/vol4/cs11/report.html
[11] Stakem, Fred & Alregib, Ghassan. (2008). Neural Networks for Human Arm Movement Prediction in CVEs. Retrieved
from https://ieeexplore.ieee.org/document/4674314/
[12] Lantz, B. (2015). “Machine Learning with R”. Birmingham, UK: Packt Publishing Ltd. Hastie, T., Tibshirani, R., Friedman, J. (2001).
“The Elements of Statistical Learning Data Mining, Inference, and Prediction”.
