Refine
Year of publication
Document Type
- Doctoral Thesis (21)
- Part of Periodical (5)
- Article (1)
- Book (1)
Has Fulltext
- yes (28)
Is part of the Bibliography
- no (28)
Keywords
- Supply Chain Management (6)
- Lehrstuhlbericht (5)
- Luftverkehr (4)
- Risikomanagement (4)
- Supply chain management (3)
- Air traffic (2)
- Air transport (2)
- Business plan (2)
- Container shipping (2)
- Geschäftsmodell (2)
Institute
- Kühne Foundation Endowed Chair of Logistics Management (28) (show_all)
This dissertation investigates the application of machine learning to improve decision making in airline operations. In an introductory overview, we1 discuss the airline industry and the challenges of decision making in airline operations: e. g., complex IT infrastructure, interconnected resources, delay management, fuel price volatility and future environmental regulation. Machine learning can efficiently integrate a large volume of data from a variety of data sources and formats to generate accurate predictions. To assess the viability of using machine learning models for decision making in airline operations, we develop a model based on linear regression and gradient boosting to predict aircraft arrival time. Furthermore, we integrate cost index optimization to model the impact of aircraft speed on arrival time. While we find that machine learning can improve prediction accuracy by more than 30 %, the optimal cost index varies according to fuel cost, flight distance and delay costs. We propose an overall reduction in cost index for short-haul flights to reduce fuel cost while maintaining punctual operations. Arrival time predictions serve as input for daily operations planning and control. For network carriers accurate arrival time predictions are key for efficient hub operations. In a next step, we focus on aircraft arrival time prediction for intercontinental flights. We analyze the accuracy of en-route weather data provided by the flight plan, generate features based on en-route weather data and integrate them in our prediction model. We evaluate three machine learning models: linear regression, random forest and gradient boosting. Through our approach, we can assess the impact of en-route weather data. Overall, an increase in prediction accuracy of 25 % is achieved. By including en-route weather data, prediction accuracy is improved by 5 %. Our model outlines the essential features for intercontinental arrival time predictions and assess the value of en-route weather data. Furthermore, we outline organizational challenges in implementing predictive analytics. Future environmental regulations are a challenge for the airline industry. From 2020 onwards net growth in CO2 emissions is prohibited. Thus, airlines need to focus on initiatives to limit fuel consumption. We develop a prediction model for fuel consumption considering ten different aircraft types. Our results show that fuel consumption can be improved by more than 30 % for short and long-haul flights resulting in an annual reduction of 10.5 million e in fuel costs and a reduction on green house gas emissions of 66.000 tons. Furthermore, we assess the pilots’ willingness to integrate our prediction model in their fuel decision. Our analysis shows that high prediction accuracy, model understanding and a long testing phase are essential for acceptance of the prediction model. The main implication is that decision making in airline operations can be substantially improved through machine learning. Therefore, more prediction.
1The term “we” refers to the authors of the respective chapters as denoted at the beginning of each chapter. For the abstract, this refers to the authors of Achenbach and Spinler (2018a,b), Achenbach et al. (2017).
The trucking industry is at the beginning of a radical change due to several megatrends which will reshape the industry significantly. Based on the targets of the Paris Climate Agreement, the German government adopted its own Climate Action Plan 2050 which includes sector- specific reduction targets to reach a greenhouse gas neutral society by the middle of the 21st century. By 2030, the German Climate Action Plan specifies a reduction target of 40% from transportation compared to the reference year 1990. While emissions from other sectors such as energy or industry have decreased significantly since then, emissions from transportation remained stable. Among the various modes of transportation, passenger cars and commercial vehicles are by far the largest emitters of greenhouse gas emissions. As of January 2019, 99.7% of heavy-duty trucks registered in Germany run on diesel while the number of alternative fuel- powered passenger cars increases steadily. Apart from rising emissions, the industry faces a severe shortage of qualified truck drivers. According to the German Association of Freight Forwarders and Logistics Companies, the industry was facing a shortage of 45,000 drivers in Germany in 2017 with increasing tendency due to higher trade volumes and e-commerce.
This dissertation aims to discuss the transition of road transport in Germany toward innovative heavy-duty trucks. The main body of this dissertation consists of three research papers each of them focusing on autonomous and/or alternative fuel-powered heavy-duty trucks. The first research paper presents the results of a Delphi study with experts from industry and academia on factors affecting the purchasing decision and operation of alternative fuel-powered heavy-duty trucks in Germany. In the second study, a choice-based conjoint experiment with employees from freight companies was conducted to test how customers value the main attributes of innovative heavy-duty trucks. The Generalized Bass diffusion model was applied in the third study to investigate the future diffusion of battery electric heavy-duty trucks considering total-cost-of-ownership reduction effects.
Supply chain risk management
(2012)
This dissertation investigates different applications of data analytics in supply chain planning. In the last years, data analytics became more important, because of the increase of computational power and the larger availability of data. Data analytics is used in various domains to improve operations performance, increase customer satisfaction and revenues. However, both the research and the application of data analytics in supply chain management is still lacking behind other industries. We1 analyze the potential of data analytics in the field of supply chain planning in three exemplary fields: demand forecasting, partial defection prediction and price discrimination. In addition, we demonstrate how to deal with three common challenges in the field of data analytics: the manual effort for method selection and hyperparameter tuning, the difficult interpretability of machine learning methods and the risks associated with data collection through randomized experiments. In the first paper, we develop a method selection approach in the field of intermittent demand prediction. Our model combines high predictive performance with automation and calculation efficiency. Unlike common practice, the prediction method gets automatically chosen for each data set without any manual selection. Our results are stable across three different data sets that come from different sources but all contain intermittent demand time series. We showcase the impact of the proposed forecasting approach with a warehouse operation simulation. We thereby prove the financial benefit with empirical data. In the second paper, we deal with partial defection prediction in a business-tobusiness environment in the logistics industry. The predictions must combine predictive performance with interpretability and profit maximization. Our model uses a large variety of customer-based and time-series-based features to predict the probability of partial defection for each customer. We use a data permutation approach to make the best performing, black-box models interpretable. Furthermore, we use a profit assessment to identify the method that leads to the highest revenue through successful retention actions. In the third paper, we study price sensitivity prediction. We do not use any randomized experiments, because the risk of loosing customers through such experiments is too high. Thereby, we address the challenge of data availability
1The term “we” refers to the authors of the respective chapters as denoted at the beginning of each chapter. For the abstract, this refers to the authors of Faber and Spinler (2019a,b,c).
In 2020, approximately 151,000 warehouses were operating worldwide, with a total annual expenditure of e300 billion, representing roughly half of total supply chain costs. Optimized warehouse management may provide competitive advantages from both cost and customer service perspectives. One way to achieve these is to leverage the abundance of data collected in supply chains in combination with powerful algorithms. This dissertation investigates how novel data sources and optimization algorithms, such as machine learning, can be applied in the context of warehouse advancement. In this research, we1 analyze the warehouse environment from two perspectives. On the one hand, we examine two available resources in warehouses—equipment and employees—and explore how predicting breakdowns and productivity, respectively, may improve warehouse performance. On the other hand, we investigate whether the warehouse concept of crossdocking can be applied virtually to allow dynamic transfers between delivery vehicles.
In our first paper, we partner with one of the largest logistics service providers to examine how master, usage and sensor data on material handling equipment can be incorporated into a predictive maintenance model. Existing literature focuses on either time- or condition-based variables, whereas we show, in a comprehensive study of statistical learning methods, that both variable types can be included simultaneously. Our predictive maintenance model is able to capture the majority of breakdowns (> 85%), while maintaining a low false-positive ratio (< 7%).
In our second paper, we work with the same logistics service provider and apply Extreme Gradient Boosting to predict the productivity of new hires. We include operator, shift and product data to show that productivity can be predicted on an individual employee basis while substantially decreasing the forecasting error (50%), which translates into cost savings.
In our third paper, we look at dynamic and synchronized transshipments during direct deliveries. This concept uses transfers between vehicles, as carried out in cross-docking, but without the need for a physical warehouse. This reduces the proportion of empty return trips by increasing the proximity of vehicles to their location of origin. Our easy-to-implement multi-algorithm reduces the total distance by 15% on average compared with simple direct deliveries, and solves large problem instances within reasonable computational times.
This dissertation with its individual research contributions highlights how novel data sources and optimization algorithms can contribute to advancing warehouse management, and highlights the managerial implications of all three topics.
1The term “we” refers to the authors of the respective chapters, as denoted at the beginning of each chapter.
Supply chain fit
(2009)
This dissertation investigates capacity and technology choice decisions in maritime container shipping under demand and regulatory uncertainty. In an introductory overview, we discuss the industry and the challenges that complicate investment decisions in shipping: e. g., the multitude of decisions, market volatility, excess capacities and the trend of new environmental regulation. Real option valuation methods can account for strategic options and the uncertainties in capacity decisions in shipping. To assess the impact of chartering on maritime investment, we analyze investment and charter options individually in a continuous-time model. We combine both in a discrete-time approach taking into account key features of the industry: investment with time to build, divestment, chartering, an endogenous charter rate, layup, and demand uncertainty. While we find demand volatility to increase optimal capacities if only investment with time to build is possible, chartering reduces this effect. It adds value to the overall project, should be mainly applied to compensate unexpected capacity shortages and needs to be considered in decision-making. Uncertainty about future eco-regulation is a further challenge for the industry. In an approximate dynamic programming model extension, we account for a stochastic introduction of operating cost-increasing regulation to assess optimal capacity choice under regulatory uncertainty. Regulation can allow for grandfathering, affecting only newly acquired vessels. We find uncertainty about future regulation with grandfathering to induce heavy up-front investment to secure a low cost base even in regulated markets. Such uncertainty may increase excess capacities and industry emissions. Uncertainty without grandfathering, however, reduces overall investment and emissions. In this case, the market may contract as investors resort to chartering until uncertainty is resolved. To also assess technology choice in light of uncertain future regulation, we develop a two-phase regime-switching model. We derive analytical solutions and study the effects of regulatory uncertainty on technology choice in a numerical extension that relaxes restrictive assumptions. We find that such uncertainty can increase optimal capacities and that a single-technology strategy is preferable over a fleet of mixed technologies in most cases. We further extend the model and compare the effectiveness of two different regimes—an emissions cap and an emissions tax. Results suggest that an emissions cap is more effective at reducing overall emissions while a tax causes lower regulation cost. The regime choice also determines the optimal technological fleet composition. The main implication is that chartering and regulatory uncertainty have a strong effect on optimal investment in shipping and should be considered in project valuation. Further, regulatory uncertainty can lead to unintended investment behavior that undermines regulatory goals. Lastly, regulatory regimes are not equally effective/efficient in reducing the environmental footprint of shipping and constitute varying incentives for investing in eco-friendly technology.