Spare Parts Demand Forecasting in Maintenance, Repair & Overhaul

Abstract

Despite a high degree of uncertainty about the scope of future orders and the corresponding capacity and material demands, Maintenance, Repair & Overhaul (MRO) service providers face high expectations regarding due date reliability by their customers. To meet these requirements while at the same time keeping delivery times short, the availability of the required spare parts or pool parts is an essential success factor. As these cannot be kept in stock in large quantities due to their high monetary value, reliable spare parts demand forecasts are of vital importance for the profitability of MRO service providers. As a result of a high degree of information uncertainty and the mostly lumpy demand patterns, conventional time-based and statistical methods do not show sufficient forecasting quality for application in the MRO industry. Data-based approaches incorporating machine learning methods offer promising capabilities to achieve improved predictive accuracy but still need to be adequately linked to production planning and control to realize their full potential. This paper first analyses potential approaches to spare parts demand forecasting in the MRO industry, focusing on forecast accuracy and potential for integration into material and production planning. Based on this, a classification of demand forecasting approaches is presented and an approach for order-based material demand forecasting with two-step feature selection is proposed. Finally, the presented approach is applied on a real dataset provided by an MRO service provider.

Full text

CONFERENCE ON PRODUCTION SYSTEMS AND LOGISTICS
CPSL 2022
__________________________________________________________________________________
DOI:
3rd Conference on Production Systems and Logistics
Spare Parts Demand Forecasting in Maintenance, Repair & Overhaul
Torben Lucht1, Volodymyr Alieksieiev1, Tim Kämpfer1, Peter Nyhuis1
1 Leibniz University Hannover, Institute of Production Systems and Logistics (IFA), Germany
Abstract
Despite a high degree of uncertainty about the scope of future orders and the corresponding capacity and
material demands, Maintenance, Repair & Overhaul (MRO) service providers face high expectations
regarding due date reliability by their customers. To meet these requirements while at the same time keeping
delivery times short, the availability of the required spare parts or pool parts is an essential success factor.
As these cannot be kept in stock in large quantities due to their high monetary value, reliable spare parts
demand forecasts are of vital importance for the profitability of MRO service providers. As a result of a high
degree of information uncertainty and the mostly lumpy demand patterns, conventional time-based and
statistical methods do not show sufficient forecasting quality for application in the MRO industry. Data-
based approaches incorporating machine learning methods offer promising capabilities to achieve improved
predictive accuracy but still need to be adequately linked to production planning and control to realize their
full potential. This paper first analyses potential approaches to spare parts demand forecasting in the MRO
industry, focusing on forecast accuracy and potential for integration into material and production planning.
Based on this, a classification of demand forecasting approaches is presented and an approach for order-
based material demand forecasting with two-step feature selection is proposed. Finally, the presented
approach is applied on a real dataset provided by an MRO service provider.
Keywords
MRO; spare parts demand; forecasting; Machine Learning; Artificial Neural Networks.
1. Introduction
Maintenance, Repair and Overhaul (MRO) of complex capital goods, such as aero engines or wind turbines,
is also known as “regeneration” [1]. This process comprises the disassembly, inspection, repair, reassembly,
and test (quality control) of mostly high value products [2]. In addition to this, there are up to two pooling
stages in the regeneration supply chain (see Figure 1) to provide repairable or serviceable spare parts to their
downstream processes and by this improve robustness against disturbances or material shortage along the
regeneration process. [3]. These pools are filled either from the respective upstream processes or via the
procurement of new or used parts. The availability of the pool parts and the precision of the corresponding
demand forecast thus have a significant influence on the punctuality of the material supply for the reassembly
and the achievable adherence to delivery dates of the MRO service provider to its customers [4]. In turn, the
on-time delivery by MRO service providers is complicated with the high degree of uncertainty about the
future work scope at the beginning of the regeneration process. Due to the complexity of goods to be repaired,
it is not possible until the end of the inspection to recognize all existing damages and thus to plan repair
operations and forecast the material demand. Furthermore, it is uncertain, whether a component can be
repaired or has to be replaced (e.g. due to heavy damage) [1].

Figure 1 – Universal supply chain structure for the regeneration of aero engines [5]
As the spare parts cannot be kept in stock in large quantities due to their high monetary value, reliable
forecasting is a crucial factor to ensure the profitability of the MRO service provider. Because of the lumpy
patterns of spare parts demand, which will be described in the next section, traditional time-series and
statistical forecasting methods do not provide sufficient forecasting quality for application in the MRO
industry [6]. However, today more and more condition data, e.g. oil pressure or temperatures are measured
during operation, which can be indicators regarding the wear of components [7]. Besides these quantitative
parameters, also qualitative parameters, such as region, climate, maintenance politics of aircraft operator or
owner have to be considered while forecasting material demand. This is possible e.g. using Machine
Learning (ML)-based methods, which thus are the focus of this paper. Based on a brief introduction to spare
parts demand classification a brief analysis of characteristics of spare parts demand in the MRO industry and
potential methods for spare parts demand forecasting is performed in section 3. Based on this, section 4
presents a hybrid approach to spare parts demand forecasting and outlines the first prediction results
obtained. Finally, conclusions are given in section 5.
2. Spare parts demand in the MRO industry
Spare parts demand can be categorized, using periodicity (inter-demand intervals) and quantity variation.
Typical demand structures are smooth, erratic, intermittent and lumpy demand (see Table 1) [8,9].
Table 1 – Demand categorization according to [8], [9]
Demand Type
Inter-demand
intervals
Quantity
variation
Smooth
Low
Low
Erratic
High
Intermittent
High
Low
Lumpy
High
Smooth and erratic demand patterns can be distinguished according to quantity variation, which is relatively
low in the case of smooth demand patterns and relatively high in case of erratic demand. Periods between
demand occurrence are small in both cases. Intermittent and lumpy demand is characterized by the mostly
random appearance of demand and many periods of zero demand. Furthermore, lumpy demand, in
comparison to intermittent demand, shows high variance in spare parts quantity [11,10]. Cut-off values
regarding the separation of these demand patterns are proposed in [10]. Considering complex capital goods
like aircraft about 80% of the demand for repair, and corresponding material demand comes up unplanned
[12]. Due to this and corresponding uncertainties regarding damage pattern, work scope and spare parts
demand of unplanned MRO-activities can mostly be categorized as intermittent or lumpy (cf. [6] for sources
of intermittency and lumpiness for aircraft spare parts). Hence, different forecasting methods and potential
fields of application in forecasting of intermittent or lumpy demand are analyzed in the next section.
Reassembly
Disassembly
Capital good in service
Inspection QA
Repair
Pool
SA
Pool
RA
Repairable parts
RA: Serviceable parts
SA: Quality assurance
QA: Procurement
Diagnosis
Procurement Procurement

3. Literature review: Forecasting of material demand
Methods for demand forecasting methods overall can be grouped in deterministic, stochastic demand
assessment and subjective estimation methods [13]. [14] categorizes forecasting approaches depending on
the influencing variables in causal, lifecycle, time series and consumption analysis. A differentiation between
qualitative and quantitative approaches is used in [15], whereby the quantitative methods are subdivided in
uni- and multivariate methods. [16] uses a similar structure but subdivides quantitative methods in time-
series and causal forecasts. An alternative classification is presented in [17] that distinguishes between past-
based and future-based methods, each divided into qualitative and quantitative methods. These are further
differentiated in methods for forecasting of time and quantity of material demand by [18]. These approaches
to classification of material demand forecasting form the basis for the classification scheme (see Figure 2)
that is presented in the following sections.
3.1 Deterministic approaches
Deterministic demand forecasting methods are methods by which material demand is determined solely
based on an existing independent primary demand [13]. These methods comprise analytical and synthetic
approaches [13]. Analytical methods rely on the bills of material of the finished product. Based on them, the
demand on finished product (primary demand) is disassembled in demand for subassemblies and
components [13]. Synthetical methods to forecasting make use of parts usage lists as a forecast basis and are
suitable especially for long-term planning [13]. Another deterministic approach is e.g. consumption analysis.
This method is based on maintenance measures planning [14]. Due to their inability to consider uncertainties
and thus unplanned material demand in the regeneration process, deterministic approaches are only suitable
for spare parts provision during planned regeneration events (e.g. mandatory replacements of components).
For intermittent and lumpy demands, which are in the focus of this paper, stochastic methods are commonly
used [19].
3.2 Stochastic approaches
Stochastic demand forecasting can be defined as "mathematical-statistical methods, in which past
consumption values are used to infer future demand” [13]. These methods can again be grouped into
quantitative and qualitative. Quantitative stochastic methods include univariate and multivariate approaches
[15] that are presented separately in the following sections.
3.2.1 Quantitative univariate approaches
Univariate approaches are those based on consideration of only one independent variable and include e.g.
time-series and life-cycle analysis. Time-series methods are methods by which the forecasting for a future
time horizon is made based on a demand history from the past. Among others, the approaches based on well-
known statistical methods, such as exponential smoothing or moving average, are to be emphasized.
Statistical methods for forecasting intermittent and lumpy demands were first studied by CROSTON [20]. In
his work, he found that exponential smoothing does not provide sufficient forecast quality to forecast
intermittent demand and proposed his method, based on exponential smoothing, in which demand rate and
time intervals between its occurrence are analyzed and forecasted separately [20], [21]. [10] and [22]
identified a bias in CROSTON’s method and introduced an additional correction factor to avoid this bias.
Further statistical methods for predicting intermittent and lumpy demand are also presented and discussed
in [23], [24], [25] and [26].
Life-cycle analytical methods for demand forecasting are based on an "estimation of the time until failure of
the corresponding component" [14]. These methods are based on failure rates or, in other words, the
probability of a failure as a function of its lifetime [15]. Practical studies on these methods are presented e.g.
in [27] and [28].

Time-series and life-cycle analytical methods are easy to use and require a relatively small amount of input
data. Nevertheless, the increasing number of influencing factors that MRO service providers are provided
with, e.g., from condition monitoring systems, cannot be taken into account completely with the help of
these approaches, which leaves potential for improvements of the forecast quality unused.
Figure 2 – Classification of demand forecasting methods (based on [13], [14], [15], [16], [17], [18])
Analytical methods
Synthetic methods
e.g. bills of material
e.g. parts usage lists
Stochastic
methods
Qualitative
Methods
Univariate
methods
(time-based
forecasting)
Multivariate
methods
(order-based
or time-
based
forecasting)
Coefficient-based
methods
Causal analysis
methods
Data basis Examples
Consume analytical
approach
Consumption
coefficient, Number of
planned repairs
Summed failures within
a period of time, time
periods between failures
Deterministic
methods
Lifecycle-analysis
methods
Methodic basis
Failure rate
(Probability
distribution)
Time-series methods Past spare parts demand
Croston-
based
methods
[20], [22],
[23]
Exponential
Smoothing
Moving Average
Influencing factors
(quantitative und
qualitative) ANN
MLP
Machine
Learning
ANN
RNN
H y b r i d a p p r o a c h e s
H y b r i d a p p r o a c h e s
Q u a n t i t a t i v e m e t h o d s
Influencing factors BN
H y b r i d a p p r o a c h e s
Expert knowledge
Interview
Brainstorming
Delphi-method
Scenario-
technique
Available information Relevance tree
analysis
Intuitive approaches
(future-based)
Graphical
approaches
(past-based)
Reg-
ression
Consumption
coefficient *
Number of
planned repairs
[27], [28]
[25], [26]
[24]
[35]
[8]
[30]
[3]
[36]
[34]
[34]

3.2.2 Quantitative multivariate approaches
As forecasting material demand is usually dependent on more than one variable, multivariate forecasting
methods are gaining more and more importance over the recent years. These methods include
coefficient-based and causal analysis methods [15]. These approaches typically apply data from the use
phase, for example using condition monitoring systems, or the maintenance phase of the goods (cf. [7]).
Coefficient-based methods consider several influencing factors (quantitative and qualitative) to determine a
wear coefficient (cf. [15] for definition). These methods include, for example, ML-based methods, such as
Artificial Neural Networks (ANN) that represent simplified representations of the biological neural network
(cf. [29] for the definition of ANN). They consist of several information processing units (“neurons”) that
contain mathematical functions and are interconnected. The signals entering a neuron are weighted and
converted into the output signals using an activation function. To do so, the ANN is trained based on a
training data set, e.g. to achieve desired prediction results. Lumpy demand forecasting using a multilayer
perceptron (MLP) type of ANN is explored and analyzed in [8], [30], and [31]. The analysis of 60
contributions related to ANN-based intermittent demand forecasting in [8] reveals, that MLP-based methods
provide the best forecasting performance compared to other types of ANN. Above mentioned research also
proves that the forecasting accuracy of MLP outperforms that of time series analytical methods. Other
ANN-based methods for forecasting material demand are investigated in [32] (e.g. Recurrent Neural
Networks (RNN)), that also show good results in the forecasting of non-stationary demand in the field of
aircraft spare parts management. Through good forecasting performance, big input-data requirements as well
as poor traceability auf causal relationships can be highlighted as disadvantages of ANN-based methods.
These can be identified using causal analysis forecasting methods [15]. One of the most common causal
forecasting methods are Bayesian networks (BN). BN are a set of variables (nodes) and directed edges
between them, that form a directed acyclic graph (DAG). Edges of this graph represent potentially causal
dependencies between the nodes [3], [33]. First applications of different types of BN (expert-initiated BN,
data-based BN, and hybrid ML-based BN, which combines the first two approaches) for forecasting lumpy
spare parts demand are performed in [34]. Here, the hybrid BN outperforms the expert-initiated BN and the
data-based BN as well as logistic regression in terms of prediction accuracy [34]. First applications of BN
in regeneration logistics can be found in [3]. In this context, they are used to determine the probability with
which regeneration orders are required for a component or an assembly. For this purpose, the product
structure of the regeneration good is represented in form of a BN, in which the assemblies and components
are mapped as its nodes. The edges are derived from the product structure and existing influencing factors.
For the determination of the initializing probability distribution, existing service data from the past is
provided as the basis for the BN. In both [3] and [34] good causality determination performance of BN is
reported. In contrast to time-series based methods, quantitative multivariate forecasting can be used for
order-specific forecasting to predict the demand for a certain regeneration order, e.g. based on conditional
or operating data of a certain regeneration good. However, this only allows the total demand per order to be
predicted and does not include information about the specific time this demand occurs. This is illustrated
graphically in Figure 3 using a fictitious demand time-series.
Figure 3 – Types of material demand forecasting
1 2 3 …
…t-1 t0
t-2
t-3
t-4 T+1 T+n
…regeneration order / -
time / days
forecasting uncertainty
demand per period
material demand
/ pieces
forecasting horizon
demand history
Time-series based forecasting
material demand
/ pieces
Order-specific forecasting
uncertainty of demand forecast
demand per order

3.2.3 Qualitative approaches
Qualitative approaches are methods based on expert estimates or the analysis of existing information
(without causality determination) about the forecasting asset. These can be subdivided in past-based and
future-based qualitative forecasting methods. Methods based on past data include, for example, relevance
tree analysis. The future-based methods include, among others, questioning, brainstorming, Delphi method,
and scenario technique. [17]. Although qualitative methods are widely used for spare parts demand
estimation in the MRO sector due to their simplicity, they are still strongly dependent on individual,
subjective estimations and thus can neither be proven by data, nor can they be reproduced or even automated.
Due to high financial risks, the high variability in demand as well as the complexity of the goods, the quality
of the forecasts is often insufficient, which is why they are not the focus of this paper.
3.3 Hybrid approaches to material demand forecasting
Hybrid forecasting approaches combine different forecasting methods to improve forecast accuracy. In [35]
hybrid approach for intermittent demand forecasting in the semiconductor supply chain is proposed, which
combines RNN-based and time-series-based methods. In this study, the presented method outperforms
time-series and RNN-based forecasting methods in terms of demand prediction accuracy. In [36] a hybrid
approach for material demand forecasting dedicated to the mining industry is proposed. It combines
regression modeling and ANN-based method and which also shows better forecasting performance
compared with time-series and ANN-based methods as standalone approaches.
The overview of relevant literature has shown, that advanced ANN MLP-based approaches outperform
conventional statistics methods in forecasting accuracy. Hence, in the following section an ANN-based
order-specific approach dedicated to the MRO industry is presented. This order-specific forecast could
afterwards potentially be distributed over the demand time periods, which could be a topic of further
research.
4. Overview of ANN MLP-based approach for material demand forecasting
As mentioned in section 3, ML-based and, especially, ANN MLP-based methods provide better forecasting
performance in comparison to the time-series methods. In this section, hence, an approach for systematic
application of ANN for order-specific material demand forecasting in the MRO industry is presented. First
the approach functionality and general process is presented in section 4.1. Afterwards its software-based
implementation based on real dataset provided by MRO service provider is presented in section 4.2.
4.1 Overview of approach functionality
The performance of ANN MLP-based methods can be significantly improved by the selection of relevant
input-features (cf. [38,37]). The approach presented in this section (see Figure 4) is focused on sufficient
data preparation and feature selection for ANN MLP-based order-specific demand forecasting for the MRO
industry. It combines qualitative and causal analysis methods into a two-step process to select relevant
features and, by this, increase forecasting accuracy. The structure of the approach is based on typical
structure of data analytics project, presented e.g. in [39]. Consequently, the first step of the approach is data
preparation based on typical datasets available to MRO service providers. This usually comprises condition
parameters, contractual information, customer related data and data from previous regenerations of similar
or the same product. To apply ANN-processing this data needs to be prepared accordingly (e.g. through
normalization). Afterwards the data irrelevant to the subject area needs to be excluded – usually in
corporation with a subject area expert (e.g. internal customer numbers). This may help to decrease
computational time and costs for the next FS-step (see Figure 4). After this assessment and basic filtering of
irrelevant features, systematic techniques for feature selection have to be applied to avoid redundancy [37],

which can not be identified during expert evaluation, as well as to enhance the understandability and to
minimize the effort of further data processing [38].
Figure 4 – ANN MLP-based approach for material demand forecasting in the MRO industry
This paper focuses on Forward Feature Selection (FFS) only as one of the most popular feature selection
methods. This represents an iterative approach, that progressively adds features that improve the model’s
forecasting accuracy the most until no additional accuracy can be gained [40]. Due to the increasing number
of features available in regeneration this needs to be supported systematical. To do so, BN are chosen as a
model learner, due to their good performance in the identification of interdependencies as reported in [3] and
[34]. After relevant features have been selected, forecasting can be performed and analyzed using statistical
failure rates. This assessment allows for a preliminary evaluation of forecasting results. If the applied
forecasting method did require the normalization of data during pre-processing, data has to be denormalized
to obtain forecast values usable in practice.
4.2 Software-based application of ANN-based order-specific demand forecasting
For validation of the proposed approach functionality, it was applied to a real data set, provided by an MRO
service provider. The data provided comprises more than 600 datasets with 22 qualitative and quantitative
parameters each. The data preparation and forecasting method were implemented using the open-source
visual-programming tool KNIME Analytics Platform v4.5. Using above described two-step-FS the following
features were selected: cycles since new and last regeneration and (partially) product owner, region of
operation, regeneration project type. To analyze the forecasting accuracy the results obtained by forecasting
with one-step (only expert estimation) feature selection is compared with results obtained using the presented
two-step (expert estimation and FFS) feature selection based on typical statistical measures: Mean Absolute
Error (MAE), Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) (see Table 2). In this
comparison the normalized values are used for better understanding of the range of the forecasted value. The
training algorithm was repeated ten times to determine the achievable range of forecasting accuracy.
Table 2 – Comparison of forecasting accuracy of ANN MLP with one-step and two-step of feature selection
Error
MAE
MSE
RMSE
ANN MLP (1 St.)
0,292…0,300
0,115…0,135
0,339…0,367
ANN MLP (2 St.)
0,288…0,299
0,112…0,120
0,340…0,346
The comparison confirms that ANN MLP-based approach with two stages of feature selection outperforms
the similar order specific approach with only one step of feature selection (expert evaluation) in forecasting
accuracy. It needs to be mentioned, that in this example only required demand for serviceable components
were forecasted, as there was no information on capacity demands per regeneration order, which have to be
included for demand forecasting and demand-oriented inventory dimensioning of repairable spare parts. For
comparison with conventional time series-based approaches, it also has to be taken into account that the
prediction results obtained with the ANN-MLP approach so far only forecast order-specific demands without
their demand timing. Consequently, it requires a scheduling of the demands based on the probability of
occurrence of the regeneration events as well as the delay in demand based on the date of occurrence of the
regeneration events. A potential approach to this estimation is described in [41] that uses a hybrid approach
Selection of relevant features
Pre-
processing Expert
estimation BN-based
FFS Forecasting Score
analysis Post-
processing
Dataset
= mandatory step = optional step FFS = Forward Feature Selection BN = Bayesian Network

of data mining and logistics models to predict throughput times of regeneration orders. As mentioned in
section 4, this coupling should be focused next to allow for an application in the MRO industry.
5. Summary and outlook
Despite various research regarding the prediction of mostly intermittent or lumpy spare parts demand in the
MRO-industry service providers still lack suitable and applicable approaches to spare parts demand
forecasting using available quantitative and qualitative information. In this paper, different methods for
material demand forecasting are analyzed, compared and systematically structured. Here it needs to be
differentiated between time-based and order-based forecasting. The literature review has shown, that
ML- and especially ANN-based forecasting methods significantly outperform conventional time-series
methods in terms of forecasting non-stationary demand. Taking into account MLP as the best performing
approach among other ANN-based methods, a systemic approach for application of ANN MLP to forecast
material demand in the MRO industry was proposed afterwards. Further this approach was applied to a real
dataset provided by an MRO service provider for the prediction of required quantity of serviceable
components with two stages of feature selection (expert estimation and FFS). Its performance was compared
with the similar approach, using one-step feature selection (expert estimation) only, afterwards. This
comparison has shown, that using two-stage feature selection with FFS technique, based on a BN learner,
better forecasting accuracy can be achieved. Further research needs to be dedicated to the hybridization of
time-based and order-based forecasting approaches with the purpose of distributing precise ANN-based
demand forecasts over time periods. In this context, the material demand forecast must also be extended to
include the expected demand for repair, so that inventories of repairable components can also be
systematically taken into account for the purpose of meeting the total material demand. An additional
direction of research is the comparison of alternative feature selection methods and different selection model
learners to further improve forecast accuracy.
Acknowledgements
Funded by Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) ”SFB 871/3” -
119193472.
References
[1] Eickemeyer, S.C., 2014. Kapazitätsplanung und -abstimmung für die Regeneration komplexer Investitionsgüter.
Dissertation, Hannover.
[2] Guide, V.R., Kraus, M.E., Srivastava, R., 1997. Scheduling policies for remanufacturing. International Journal
of Production Economics 48 (2), 187–204.
[3] Berkholz, D.A., 2012. Grundmodell zur Kapazitäts- und Belastungsabstimmung eines Arbeitssystems in der
Regeneration. Dissertation, Hannover.
[4] Hermeier, B., Platzköster, C., 2006. Ergebnisse der ersten bundesweiten FOM-Marktstudie „Industrie-
Dienstleistungen“, Essen.
[5] Lucht, T., Kämpfer, T., Nyhuis, P., 2019. Characterization of supply chains in the regeneration of complex
capital goods, in: Dimitrov, D., Hagedorn-Hansen, D., Leipzig, K. von (Eds.), International Conference on
Competitive Manufacturing (COMA 19) proceedings. 30 January 2019-1 February 2019, Stellenbosch, South
Africa. Department of Industrial Engineering Stellenbosch University, Stellenbosch, South Africa, pp. 444–449.
[6] Ghobbar, A., Friend, C., 2002. Sources of intermittent demand for aircraft spare parts within airline operations.
Journal of Air Transport Management 8 (4), 221–231.
[7] Andersson, J., Jonsson, P., 2018. Big data in spare parts supply chains. IJPDLM 48 (5), 524–544.
[8] Amirkolaii, K.N., Baboli, A., Shahzad, M.K., Tonadre, R., 2017. Demand Forecasting for Irregular Demands in
Business Aircraft Spare Parts Supply Chains by using Artificial Intelligence (AI). IFAC-PapersOnLine 50 (1),
15221–15226.

[9] Eaves, A.H.C., Kingsman, B.G., 2004. Forecasting for the ordering and stock-holding of spare parts. Journal of
the Operational Research Society 55 (4), 431–437.
[10] Syntetos, A.A., Boylan, J.E., Croston, J.D., 2005. On the categorization of demand patterns. Journal of the
Operational Research Society 56 (5), 495–503.
[11] Syntetos, A.A., 2001. Forecasting of intermittent demand. Ph.D. Thesis.
[12] Mensen, H., 2013. Handbuch der Luftfahrt. Springer, Berlin, Heidelberg.
[13] Hartmann, H., 2002. Materialwirtschaft: Organisation, Planung, Durchführung, Kontrolle, 8., überarb. und erw.
Aufl. ed. Dt. Betriebswirte-Verl., Gernsbach, 740 pp.
[14] Markiewicz, M., 1988. Bedarfsermittlung für Ersatzteile, in: Besters, H., Colbe, W.B. von, Engelhardt, W.,
Jaeger, A., Laßmann, G., Maßberg, W., Schwark, E., Wartmann, R., Markiewicz, M. (Eds.),
Ersatzteildisposition im Maschinenbau, vol. 32. Gabler Verlag, Wiesbaden, pp. 35–67.
[15] Loukmidis, G., Luczak, H., 2006. Lebenszyklusorientierte Planungsstrategien für den Ersatzteilbedarf,
in: Barkawi, K., Baader, A., Montanus, S. (Eds.), Erfolgreich mit After Sales Services. Springer, Berlin,
Heidelberg, pp. 251–270.
[16] Becker, J., 2016. Dynamisches kennliniengestütztes Bestandsmanagement. Dissertation, Hannover.
[17] Schmidt, M., 2019. Modellgestützte Bewertung der Kapazitätsabstimmung im Umfeld veränderlicher
Nachfrage. Dissertation, Hannover.
[18] Sendler, M., 2020. Ganzheitliche Konfiguration der PPS in der Instandhaltung hochwertiger Investitionsgüter.
Dissertation, Braunschweig.
[19] Biedermann, H., 2008. Ersatzteilmanagement: Effiziente Ersatzteillogistik für Industrieunternehmen, 2., erw. u.
aktualisierte Aufl. 2008 ed. Springer, Berlin, Heidelberg, 156 pp.
[20] Croston, J.D., 1972. Forecasting and Stock Control for Intermittent Demands. Journal of the Operational
Research Society 23 (3), 289–303.
[21] Xu, Q., Wang, N., Shi, H., 2012. Review of Croston’s method for intermittent demand forecasting, in: 2012 9th
International Conference on Fuzzy Systems and Knowledge Discovery. 2012 9th International Conference on
Fuzzy Systems and Knowledge Discovery (FSKD), Chongqing, Sichuan, China. 29.05.2012 - 31.05.2012.
IEEE, pp. 1456–1460.
[22] Syntetos, A.A., Boylan, J.E., 2005. The accuracy of intermittent demand estimates. International Journal of
Forecasting 21 (2), 303–314.
[23] Teunter, R.H., Syntetos, A.A., Zied Babai, M., 2011. Intermittent demand: Linking forecasting to inventory
obsolescence. European Journal of Operational Research 214 (3), 606–615.
[24] Romeijnders, W., Teunter, R., van Jaarsveld, W., 2012. A two-step method for forecasting spare parts demand
using information on component repairs. European Journal of Operational Research 220 (2), 386–393.
[25] Ghobbar, A.A., Friend, C.H., 2003. Evaluation of forecasting methods for intermittent parts demand in the field
of aviation: a predictive model. Computers & Operations Research 30 (14), 2097–2114.
[26] Hemeimat, R., Al-Qatawneh, L., Arafeh, M., Masoud, S., 2016. Forecasting Spare Parts Demand Using
Statistical Analysis. AJOR 06 (02), 113–120.
[27] Voigt, J.P., 1973. Erfassung, Auswertung und Nutzung von Schadendaten in der Eisen- und Stahlindustrie.
Dissertation, Braunschweig.
[28] Wilke, F.L., 1981. Bestimmung von Ausfällen und Störungen an Baugruppen des im Metallerzbergbau
eingesetzten Fahrlader, Untersuchung ihrer Charakteristiken und Auswirkungen auf die Kosten des
Gesamtbetriebes. Forschungsbericht Nr. 4397.
[29] Jacob, H., Hansmann, K.-W., Layer, M., Preßmar, D.B., Alex, B., 1998. Künstliche neuronale Netze in
Management-Informationssystemen. Gabler Verlag, Wiesbaden.
[30] Gutierrez, R.S., Solis, A.O., Mukhopadhyay, S., 2008. Lumpy demand forecasting using neural networks.
International Journal of Production Economics 111 (2), 409–420.
[31] Kozik, P., Sęp, J., 2012. Aircraft Engine Overhaul Demand Forecasting using ANN. Management and
Production Engineering Review, 21–26.
[32] Şahin, M., Kızılaslan, R., Demirel, Ö.F., 2015. Forecasting Aviation Spare Parts Demand Using Croston based
Methods and Artificial Neural Networks. Computer Science.
[33] Jensen, F.V., Nielsen, T.D., 2007. Bayesian networks and decision graphs, Softcover reprint of the hardcover 2nd
ed. 2007 ed. Springer, Berlin, 447 pp.

[34] Boutselis, P., McNaught, K., 2019. Using Bayesian Networks to forecast spares demand from equipment
failures in a changing service logistics context. International Journal of Production Economics 209, 325–333.
[35] Fu, W., Chien, C.-F., Lin, Z.-H., 2018. A Hybrid Forecasting Framework with Neural Network and Time-Series
Method for Intermittent Demand in Semiconductor Supply Chain, in: Moon, I., Lee, G.M., Park, J., Kiritsis, D.,
Cieminski, G. von (Eds.), Advances in Production Management Systems. Smart Manufacturing for Industry 4.0,
vol. 536. Springer International Publishing, Cham, pp. 65–72.
[36] Rosienkiewicz, M., Chlebus, E., Detyna, J., 2017. A hybrid spares demand forecasting method dedicated to
mining industry. Applied Mathematical Modelling 49, 87–107.
[37] Pavya, K., Srinivasan, B., 2017. Feature Selection Techniques in Data Mining: A Study. International Journal of
Scientific Development and Research 2 (6), 594–598.
[38] Arauzo-Azofra, A., Aznarte, J.L., Benítez, J.M., 2011. Empirical study of feature selection methods based on
individual feature evaluation for classification problems. Expert Systems with Applications 38 (7), 8170–8177.
[39] Runkler, T.A., 2012. Data Analytics. Vieweg+Teubner Verlag, Wiesbaden.
[40] Guyon, I., Elisseeff, A., 2003. An introduction to variable and feature selection. The Journal of Machine
Learning Research (3), 1532–4435.
[41] Hiller, T., Lucht, T., Kämpfer, T., Vinke, L., Holtsch, P., Nyhuis, P., 2021. Hybride Lieferzeitprognose:
Verbesserte Termin- und Auftragsplanung im volatilen MRO-Umfeld. ZWF 116 (12), 882–888.
Biography
Torben Lucht, M.Sc. (*1991) studied industrial engineering with the focus on
production technology at RWTH Aachen University. Since 2018, he works as a
research associate in the field of production management at the Institute of Production
Systems and Logistics (IFA) at the Leibniz University Hannover.
Volodymyr Alieksieiev, B.Sc. (*1999) studied engineering of logistic systems at
National Technical University “Kharkiv Polytechnic Institute” (Ukraine). Since 2020
he studies production and logistics at Leibniz University Hannover. He also works as
student assistant at the IFA since 2021.
Tim Kämpfer, M.Sc. (*1992) studied production engineering and logistics at Leibniz
University Hannover and works as a research associate in the field of production
management at the IFA at the Leibniz University Hannover since 2019.
Prof. Dr.-Ing. habil. Peter Nyhuis (*1957) studied mechanical engineering at Leibniz
University Hannover and subsequently worked as a research assistant at IFA. After
completing his doctorate in engineering, he received his habilitation before working as
a manager in the field of supply chain management in the electronics and mechanical
engineering industry. He is heading the IFA since 2003. In 2008 he became managing
partner of the IPH - Institut für Integrierte Produktion Hannover gGmbH.