Theses supervised by Prof. Dr. Metin Türkay
26 theses · Koç University
Multi-objective optimization in cement and textile industries using the triple-bottom-line accounting for sustainable production
With the increasing environmental disasters and growing social awareness sustainability has become the cornerstone/key to the operation and design of supply chain management for organizations and governments. The triple bottom line accounting is a comprehensive tool for researchers and practitioners in incorporating the three pillars of sustainability i-e Economy, Environment, and Society. So far research has been more focused on economic objectives and recently to large extent on environmental pillars. Research on all three pillars simultaneously is still lagging behind. Therefore, in order to analyze the complete picture of sustainability a more comprehensive approach is to be applied. The TBL is a methodology that takes all three dimension of sustainability at the same time for a sustainable decision-making process. In this thesis, a systematic approach is used to incorporate sustainability considerations in resource consuming industries of textile and cement production. The proposed framework is based on identifying sustainability indicators in textile and cement industries. Then validating these indicators if they are not previously validated in literature and practice, developing a MOMILP and then generating Pareto optimal solutions. We have used the 3 S methods for validation. This method use the self, scientific and society (hence the 3S) for indicator validation. Next, we formulated a generalized MILP for both Textile and Cement Industries incorporating our sustainability indicators. Lastly, we have used two different solution strategies for multi-objective optimization problems,(i) we have used AUGMECON2 for solving the MOMILP of textile industries and generated a Pareto frontier for balanced decision making. We performed a sensitivity analysis of the model to give further insights into the resultant solutions. (ii) For cement industry we have used a recently proposed multi-objective optimization method GoNDEFF. We have generated a set non-dominated solution points and efficient binary solution set.
Aggregate planning with sustainability consideration and business model assessment
Industrialization and technology have wreaked havoc on the planet during the previous century. As humans have grown conscious of the devastation they have created and the threats they have posed to future generations, sustainability has risen to the fore as a method of decreasing it. The three bottom lines (TBL) of sustainability are economic, social, and environmental. These three aspects should simultaneously involve in the sustainable decision-making process. The literature review has shown that many academic and practical studies have been done for the economic objective comprehensively, yet few studies have been done analytically considering the three bottom lines. An inclusive business model, which encourages sustainable growth, aims to derive value for low-income communities by engaging them into the value chain. Although the inclusive business model has been studied extensively by academicians and practitioners, there is a lack of systematic research on how the inclusive business model affects the companies and base of the pyramid (BOP) quantitively; therefore, many inclusive business models model implementations turned into philanthropy. The TBL accounting for the aggregate planning problem is addressed in this thesis using a methodological framework based on mathematical programming. The framework is a multi-objective mixed-integer linear programming model with the following objectives: maximizing profit, minimizing environmental damage, and maximizing social benefit. Additionally, the inclusive business model is integrated into the sustainable aggregate planning problem. The multi-objective mixed-integer model was examined in four different phases, and a realistic case study was conducted for each phase. The proposed framework demonstrates how sustainability bottom lines and business strategies can be included in the aggregate planning model. Framework proves that decision-makers will make a profit while being environmentally and socially conscious. Although companies may have to sacrifice some profits initially to enhance their social and environmental impact, their reputation and profitability will increase in the long run.
Forecasting brent oil futures prices using machine learning
Accurate forecasting is needed to define strategy and profitable future business operations, particularly of price data which leads to highly profitable trades and investments especially for liquid goods or assets. This thesis project researches successful financial and sales univariate data time-series forecasting models, and applies various methods in the forecasting of Brent Crude Oil Futures price while analyzing in a detailed way which proposes a proper data analysis process which involves three main parts: data examination, model evaluation and result analysis. That data analysis process can be applied to any data in order to have detailed information about the underlying commodity of the data, test forecasting models, make decisions, make inferences about related data, etc. The performance of the forecasting methods is then evaluated according to the following error measures: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE) and Tracking Signal. A number of forecasting models are used: Moving Average by Statsmodel library, manually coded Moving Average, Holt's Method, Holt's Winters Method, Long Term Short Memory, Auto-regressive Integrated Moving Average (ARIMA), Seasonal Auto-regressive Integrated Moving Average (SARIMAX), Simple Exponential Smoothing by Statsmodel library, manually coded Simple Exponential Smoothing, Deep Neural Network with 2 hidden layers, Support Vector Regression, Kth Nearest Neighbors regression.
Proactive return management in E-commerce supply chains: Predictive analytics approach
In e-commerce, product returns management has become a critical concern for online retailers. With the drastic growth of online shopping, the increasing volume of returns poses significant challenges to the efficiency of supply chain operations. To effectively address these challenges, proactive return management strategies are essential. This MS thesis aims to predict product returns before the product is sold, leveraging open-source customer-item rating and feedback data. The study adopts a novel approach combining Natural Language Processing (NLP) techniques, Matrix Factorization based on Bayesian personalized ranking loss, and deep learning methodologies to uncover latent factors behind shopping and rating behavior to make accurate return predictions, aiding online retailers in proactive decision-making and optimizing their supply chain and inventory management processes.
Advanced algorithms and solution techniques for U-shaped assembly line balancing problems
Assembly line balancing problems entail allocation of work content to different work stations, optimizing certain criteria like minimizing the number of stations, fair allocation of work content and cycle time minimization etc. The U type layout is known for its exceptional efficiency and flexibility for line balancing problems. It also provides workers with opportunities to enhance their skills and work experience through cooperation and continuous learning. However, most of the models in literature ignore the full potential of U-lines and focus on a few narrow aspects for line balancing. In this dissertation, we address the issue of assigning tasks to workstations in ways that offer more choices to line managers and workers. In the first part of this dissertation, we develop effective logic cuts that exploit the logical structure of the problems, to improve the performance of integer programming model in terms of computational time. Proposed logic cuts enable the model to utilize bin-packing bounds at each station. Moreover, idle time information, and the knowledge about combinations of task assignments to stations, that produce solutions dominated by alternative assignments, are also exploited. Computational experiments demonstrate that our enhanced model outperforms the previously developed integer programming models in solving U-type assembly line balancing problems In the second part, we present novel ways of assigning fair workload to stations. Our model is able to fulfill different fairness criteria like minimization of mean absolute deviation and sum of squared differences. We also develop a method for distributing certain types of workloads regularly along the assembly line. This way, work sharing and benefits of communication between the workers can be enhanced. Our approach also enables the decision makers to allocate appropriate idle time to certain groups of tasks. This third part deals with uncertain task times. We develop a new variance bounds based approach to deal with the stochastic U-line balancing problem. The variance based approach is considerably simpler to implement and solves the problems efficiently, outperforming other chance constrained models used for U-lines in literature. Overall, we present methods that can be adapted easily to more complex assembly line balancing problems and provide additional choices to managers, researchers and decision makers.
Advantage actor-critic deep reinforcement learning approach for paint shop planning and scheduling
Paint shops usually act as bottlenecks in production facilities requiring a painting procedure. To enhance efficiency and optimize the process by minimizing color batch changes that can decrease productivity, it is essential to develop optimization algorithms. Traditionally, these problems have been addressed using a mixed-integer linear programming (MILP) approach. However, mathematical optimization methods face challenges in adapting to dynamic production planning environments and the real-time nature of a production facility. This is due to its memoryless structure and search for exact and optimal solutions by solving the entire every time a schedule is required. To overcome the issues, this study proposed a deep reinforcement learning algorithm to solve and optimize a paint shop scheduling and planning problem that can adapt to dynamic environments. The actor-critic approach was the best method amongst the other policy-based state-of-the-art deep reinforcement learning algorithms. To train a DRL agent, a real-life simulation model of a paint shop in a household appliance factory was built to act as an environment. After the training and inference processes, the outcome was a paint shop production plan that minimizes the inventory cost and bottlenecks while maximizing the productivity of washing machine production and achieving energy efficiency through planned production stops. Besides some advantages of linear programming methods, DRL models performed well on the selected application of paint shop scheduling and planning problems. It is seen that DRL methods are superior in terms of efficiency and computational performance on inference step while obtaining at least sub optimal solution.
Spare parts demand forecasting and inventory management using machine learning models: A comprehensive application
This thesis was conducted at Türkiye's first and largest integrated air conditioning factory, focusing on spare parts management. The study addresses key challenges in demand forecasting and inventory management of spare parts, with the goal of minimizing the risks of overstocking and stock shortages. A comprehensive framework from Bacchetti and Saccani (2012), merging spare parts classification, demand forecasting, inventory management, and performance evaluation, was applied. Aligning with the forecasting framework proposed by Boylan and Syntetos (2010), the study encompassed pre-processing, processing, and post-processing phases. In the preprocessing phase, demand was classified into categories such as Erratic, Intermittent, Lumpy, and Smooth, with Intermittent being the most prevalent. Demand forecasting methods, including Na¨ıve Forecast, Holt-Winters' Seasonal Method, Croston Modification (SBA), ARIMA, MLR, MNLR, SVR, ANN, Random Tree, REP Tree, Random Forest Regressor, XGBoost, and LightGBM, were evaluated for their suitability. In the post-processing phase, total inventory cost calculations and the optimal service level ratio were determined using the Newsvendor model. Additionally, a data-driven approach employing Sample Average Approximation was utilized for optimization inspiring from Huber et al. (2019). XGBoost outperformed all other models by achieving the minimum cost while meeting the target service level. The main contribution of this study lies in incorporating demand classification as a feature in forecasting models, emphasizing the balance between service levels and costs, and demonstrating its practical significance through real-world application in spare parts management.
Classification of surgical and recovery durations in healthcare settings utilizing machine learning models: A case study at Koç University Hospital
Efficient operating room (OR) scheduling is essential for minimizing patient wait times, optimizing resource use, and reducing operational costs. However, accurately predicting surgery and postoperative recovery durations is challenging due to patient variability, procedural complexity, and limited data availability. This study addresses these challenges by converting continuous duration data into categorical intervals, enabling classification-based machine learning methods suitable for small datasets. Using real-world data from Koç University Hospital, several algorithms—including Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machines, K-Nearest Neighbors, Decision Trees, and ensemble combinations—were evaluated under Random Oversampling and Synthetic Minority Oversampling Technique strategies. Results showed that ensemble models consistently outperformed individual classifiers. For surgery time prediction, the best ensemble achieved 0.71 accuracy and 0.698 F1 score, while Random Forest reached 0.79 accuracy and 0.768 F1 score for recovery time. SMOTE proved effective in mitigating class imbalance, improving recall and F1 scores across models. Targeted feature grouping further enhanced interpretability and predictive reliability. This study provides a practical, data-driven framework for hospitals with limited records to improve OR scheduling. Reliable categorical predictions of surgery and recovery durations can enhance resource allocation, reduce scheduling conflicts, and improve patient outcomes. Keywords: Surgical duration prediction, recovery time prediction, machine learning, classification algorithms, SMOTE, ensemble models, healthcare resource optimization
Çok bloklu i̇malat için senaryo tabanli dayanikli tesis yerleşim tasarimi: endüstriyel uygulamali veri odakli bir karma tamsayili doğrusal programlama (MILP) çerçevesi
This dissertation addresses the Unequal-Area Facility Layout Problem (UA-FLP) under uncertainty, with a particular focus on multi-block manufacturing environments where material flows are complex and subject to disruption. The motivation for this research stems from the limitations of traditional deterministic layout models, which often fail to accommodate the operational variability and spatial constraints characteristic of modern industrial systems. In response, the study proposes a structured and data-driven modeling framework based on mixed-integer linear programming (MILP), progressively incorporating realistic constraints and robustness features to enhance layout feasibility and long-term performance. The thesis begins with the development of a deterministic MILP model that assigns departments of unequal area to two fixed-size layout blocks. The model minimizes total transportation cost while integrating key industrial features such as AEIOUX-based closeness preferences, ramp-based inter-block material transfer, and capacity-constrained flow routing across multiple ramps. Building on this foundation, the second stage introduces a Gamma-robust optimization model that protects against material flow uncertainty by incorporating worst-case deviation terms. The uncertainty parameters are derived from a three-year historical flow dataset using a weighted two-sigma rule, and a protection level is applied to ensure robust yet cost-efficient layouts. In the final stage, a scenario-based robust MILP model is developed. Ten disruption scenarios are defined to represent potential operational disturbances, including ramp closures, peak demand surges, and supply shortages. Scenario-dependent flow matrices, ramp capacities, and external transportation costs are incorporated while maintaining a single, unified layout across all scenarios. The model allows for flexible flow redistribution and evaluates spatial configuration through penalty terms, offering a balanced trade-off between cost minimization and layout resilience. Scenario probabilities are used to assess expected performance and solution stability. The proposed framework is applied to a real-world case study involving the washing machine production plant of a leading home appliance manufacturer. The layout problem involves assigning 21 departments to two connected blocks under realistic spatial and operational constraints. Historical data from 2022 to 2024 is used to quantify uncertainty and validate the robustness of the proposed solutions. The scenario-based results are analyzed through a detailed cost breakdown, ramp utilization metrics, AEIOUX separation penalties, and layout visualizations. The findings demonstrate that the proposed robust layout configurations maintain feasible and efficient performance across a wide range of disruption scenarios. This dissertation contributes a comprehensive and implementable framework for solving robust multi-block facility layout problems under uncertainty. By integrating practical layout requirements, historical data-driven uncertainty modeling, and scenario-based robustness, the study advances both the theoretical understanding and industrial applicability of robust facility layout design. The resulting methodology provides a scalable decision-support tool for long-term facility planning in dynamic manufacturing environments.
Kompleks tedarik zinciri aglarinin modelleme, cozum ve uygulamalari
In this thesis, the main focus is the supply chain networks in terms of modeling, solution and applications. Networks are one of the main representation tools for the systems in engineering, science and management. Supply chain networks are the specific networks that are used to represent the flows of materials, information and financials from one physical node to others. The supply chain context is very broad, we consider it in 3 parts: Modeling, solution and applications, during this thesis. In the modeling part the development of new modeling technique is described in detail. In conventional supply chain network modeling, the flows and the inventories are modeled separately for each flow and node. This makes modeling and re-modeling difficult to maintain. In this new modeling technique, the network topology that represents the flows, entities and inventories are stored in multi-dimensional matrix. Therefore model-data independence is succeeded. This makes models lean and easy to maintain. In the solution part, we have focused on the non-integer representation of the batch behaviors in the supply chain network. In conventional modeling, this batch behavior is modeled by using integer variables. However this makes the model combinatorial and difficult to solve. Therefore, in this thesis, we have developed new solutions techniques to model batch behavior without using integer variables. Finally, in applications part, we have developed a new version of the VRP (Vehicle Routing Problem). The capacities for this new version of the VRP show combinatorial behavior. A new algorithm is developed to solve this new VRP problem. It is known that VRP is one of the most difficult problems in the literature. With addition of new combinatorial capacity constraints, some models can become computationally intractable. With this new developed algorithm, the capacity constraints can be handled in polynomial time.
Atık pillerin tersine lojistik ağ tasarımı ve Türkiye uygulaması
Reverse logistics has received increasing attention in the last decade. This is mainly due to the environmental concerns, forcing legislations, and potential economic benefits. Ever increasing amount of waste batteries which are classified as hazardous waste poses a significant environmental problem; thus it requires an effective management. In this thesis, we develop a mixed-integer linear programming model (MILP) to design reverse logistics network for waste batteries and implemented the model to the Turkey case. The proposed model has various realistic features compared to existing models; it is a multi-period optimization model that considers various battery-types, different capacity options, capacity expansion of facilities, sale prices of recycled materials, variable operational cost, construction cost, and existing infrastructure of the facilities in the network. There are two disposal options for waste batteries: recycling and landfill. Using the developed optimization, the policy makers responsible for the design and operation of the reverse network of waste batteries can decide on the network configuration, while maximizing the profit. Uncertainty in the return of waste batteries is a major issue in the reverse logistics network design. As the network design requires high investment costs, the uncertainty has to be taken into account. In addition to the deterministic network optimization model, a stochastic programming approach is presented by adding scenarios to consider the uncertain amounts of waste battery returns explicitly. All models are programmed and implemented in GAMS (General Algebraic Modeling System) optimization package and solved using the CPLEX solver. Finally, we conduct sensitivity analyses to determine which parameters are critical for the network structure and to analyze the impact of variations in the critical parameters.
Elektrik dağıtım ağlarının verimliliğinin modellenmesi
In 2009, power loss occupied approximately fifteen percent of the total generated electricity in Turkey. These losses can be classified into two subgroups, technical and non-technical losses. Technical losses occur mainly because of the electricity system components; (i)transformers, (ii)cables which are being used in transmission and distribution lines, and (iii)measurement systems. Among these components, cables that carry electricity remain as the major factor for the technical power losses. In cables, corona loss and resistive loss are the primary reasons behind the losses during transmission of power. Non-technical losses can be listed as electricity theft, non-payment by customers and errors in record keeping and accounting. Technical losses can be minimized by selecting the right type of material and optimizing the network parameters, which include resistance, size and length of the cables. In this thesis, main motivation is to minimize power generation cost by reducing power loss during transmission while satisfying capacity and balance constraints. Whole analysis is implemented by the GAMS software. All data used for the analysis is taken from TEIAS, Turkish Electricity Transmission Company.
Iki amaçlı karışık tamsayılı programlama problemlerinin nondominated noktaları
The nondominated frontier in the objective space of biobjective mixed-integer linear/nonlinear programming problems consists of points that cannot be improved in value of one of the objectives without degrading the other objective value. This frontier is usually very involved consisting of many isolated points and open, closed, or half-open/half-closed line segments or curves. Some researchers considered specific classes of these problems to reduce the complexities in nondominated frontier. Some algorithms have been also proposed to find a subset of nondominated set. Several mathematical models for nonlinear process network problems have been developed and solved using epsilon-constraint, weighted sum, and minimum distance. This thesis outlines some possible complexities in nondominated frontier of BOMILPs and drawbacks of existing algorithms, and proposes an effective algorithm, EnpoBomip, to find the exact nondominated frontier of general BOMILPs, as well as all possible values of integer variables associated with each nondominated point. We also investigate biobjective mixed-integer nonlinear problems that are formulated using generalized disjunctive programming for nonlinear network synthesis problems and propose an effective algorithm, epsilon-OA, based on augmented epsilon-constraint and logic-based outer approximation (OA). We provide theoretical characterization of the proposed algorithm and show that the solutions generated are efficient. An experimental study is conducted to present a comparative analysis between EnpoBomip and the existing algorithms on three well-known problems, and show that our novel algorithm significantly outperforms others with respect to solution quality and computational performance. We also illustrate the effectiveness of epsilon-OA compared to the augmented epsilon-constraint with/without OA, and the traditional epsilon-constraint. Based on the results, epsilon-OA is very effective in solving the biobjective generalized disjunctive programming problems in the synthesis of nonlinear process networks.
Optimization of energy mix considering sustainability characteristics and its application to Turkey.
Artan enerji talebi ve birincil enerji kaynakları nı n yol a ctı gı zararl ı gaz salı nı mı çevresel s ürd ür ülebilirli ğin en önemli problemidir. E ğer enerji politakalar ı de ği stirilmezse, bu durum gelecekte daha b üy ük problemler te şkil edecektir. Birincil enerji kaynakları n ın çevresel etkilerini kontrol etmede "enerji mix" üzerinde durulmas ı gereken en hayati kararlardan biridir. Dikkatli bir enerji mix dizayn ı s ürd ür ülebilirlik g östergelerini ve farklı enerji kaynakları n ın maliyetini de ğerlendirme fı rsatı n ı sunar. Biz bu tezde elektrik üretim sisteminin gelece gini, CO2, NOx ve SO2 salı nı mları nı , enerji talep projeksiyonları nı ve enerji kaynakları nı n maliyet tahminlerini g öz ön une alarak analiz ettik. G ü c üretim sistemi i cinde, yedi farklı enerji kayna ğı nı , fuel-oil, do galgaz, k öm ür ve linyit, r üzgar, g üne s, jeotermal ve hidro enerji olmak üzere inceledik. Optimizasyon modeli çok ama cl ı tamsay ılı-kar ışı k do ğrusal programlama problemi olarak form ule edildi. Ama ç fonksiyonlar ise talep ve b ütçe kı sı tları alt ında maliyet ve sera gaz sal ın ım ı olarak d ü ş ün üld ü. Elde edilen optimizasyon modeli epsilon k ısı t metodu (epsilon constraint method) kullanı larak çöz üld ü. Modelin ampirik olarak test edilebilmesi i cin T ürkiye'nin g üç üretim sistemi vaka çal ı şmas ı olarak kullanı ldı . Modelin ç öz üm ü sonucunda hangi g üç teknolojisinin nerede ve ne zaman kurulması gerekti ği ve k öm ür yakan bir termik santralin yak ıtı nı de ği stirmenin ama ç fonksiyonları bazı nda tercih edilebilir olup olmad ığı belirlendi. Tayin edilen parametrelerden birinde yap ılacak bir de gi şikli gin optimal sonucu ne derece etkiledi ğini g örebilmek amacı yla bir duyarl ıl ık analizi yapı ldı . R üzgar t ürbininin g üç katsayı sı nı n, yatı rı m miktarı nı n ve do ğal gaz fiyatı nı n etkileri analiz kapsamı nda incelendi.
Elektirikli araçlar ve geleneksel araçların çevreye olan etkileri ile İstanbul'da şarj istasyonlarının yerlerinin planlanması
Traffic congestion is one of the most important problems in metropolitan regions including Istanbul. Vehicles running with conventional internal combustion engines have low energy efficiency and release harmful substances into the atmosphere. In order to decrease the effects of conventional vehicles, it is beneficial to use environmentally friendlier cars such as phev, bev, hev and ev. Our work includes the analysis of the current situation on the environmental effects and the design of a network for the charging stations in Istanbul. The design of such a system requires extensive data analysis on the origin destination matrices for urban transportation networks as well as the charging requirements of the vehicles. First part of this thesis focuses on environmental effects of conventional vehicles and how their replacement will affect the environment. In this perspective, we analyse environmental friendlier cars and their infrastructure need. Most important thing for promoting the usage of environmentally friendlier vehicles is opening charging stations in optimal places. We also analyse the effects of passenger vehicles in main roads in aspect of noise and air pollution. Current situation is shown with graphs and tables. In the second part, we analyse the data for Istanbul and develop a MILP formulation to design the network of charging stations to satisfy energy requirements. The objective is to minimize the total station opening costs and electricity needs by deciding how many stations to open in each district of Istanbul. We also develop a stochastic programming model to incorporate the effect of a random acceptance rate parameter which denotes the proportion of vehicle owners who plan to use electrical vehicles. We perform a sensitivity analysis on different parameters of the model.
Havayolu taşımacılığında çok periyodlu dinamik gelir yönetimi ve kapasite optimizasyonu
Air cargo as an industry dealing in across the border shipping of commodities has experienced much success as well as remarkable growth in the recent past, thanks to the evolutionary developments the sector has enjoyed over the past few years. Technically speaking, there would be no profit maximisation by companies offering such services if there was not a perfect technique of selecting shipments to transport. The secret behind ensuring that an air cargo company gets the most out of its services, in terms of profits earned, is a function of its ability to identify the correct shipments to accept as well as the capacity required to spare for each type of cargo amongst those shipped. This study performs a careful analysis in this sector with the aim establishing effective results. The key factors this thesis covers are the design of two different models and their corresponding practical applications.The incoming booking requests for every given time unit are characterised by their capacities, which are in the form of weight and volume and a third parameter of profit rate. These parameters have the sole objective of distributing the available requests in such a manner that maximum profit is earned at the end of the booking period. It is then upon the carriers to make a decision as to whether they should accept or reject the incoming booking requests based on the cargo's capacity, weight, volume and nature. The bid-price mechanism was employed such that whenever the income generated by a given request exceeds its opportunity cost, as indicated by the profit rate, then the request is automatically approved. This procedure is simulated using a dynamic environment. The programme monitors the entire process and summarises the results at the end of every trading season. At this point, the programme selects the most profitable booking requests in terms of weight, volume and profit rate. These selected parameters are then used to calculate the new bid-price, which is then adopted for the subsequent trading period. This way, the subsequent year is characterised by more profitable requests than the previous season. Simulations will be developed to bring an understanding of how such models can be used to achieve the intended objectives. First, the model deals with generating a bid-price mechanism for a single-leg flight such that whenever the income generated by a given request exceeds its opportunity cost as indicated by the profit rate, then the request is automatically approved. Mixed-Integer Non-Linear Programming (MINLP) is developed to calculate dynamic bid-prices that are adjustable at each time unit and solved in a dynamic environment. Bid-prices are calculated closer to departure time when a booking is assumed to be more urgent and charged higher than regularly bookings. However, when a flight is close to its departure time, certain bookings may be accepted and charged regularly to prevent the flight taking off under occupied. Results are compared with static bid-prices, which are calculated once over the decision horizon and results obtained from applying FCFS policy. Second, we suggested adapting a fare-class approach, which is a common technique in passenger revenue management. The customers' requests can only be accepted if there is space in the particular fare class to which they wish to apply. The problem could be approached using the nonlinear programming model (MINLP) hence the revenue function is first linearised before proceeding. The results are compared with the FCFS policy and a booking limit model by introducing dynamic threshold values, which is achieved by limiting the capacity available for shipments in each fare class. Challenges associated with these models will also be discussed to create a clear justification as to why certain methods are recommended and others discouraged. The thesis will end by testing the performance of each of these models. This way, the effectiveness becomes testable and hence the models can be graded.
Rüzgar santrali yatırım kararlarında kesikli sürekli yaklaşım
Harvesting energy from wind is becoming an effective and worthwhile way of meeting the electricity demand. Due to the carbon free structure of the wind energy, some countries started to invest more funds into the construction of wind farms on suitable areas. Before the construction of a wind farm, some variables must be chosen carefully to acquire the optimum results from both the investment cost and efficient design. The location of the wind farm must be convenient in terms of its yearly average wind speed. Other restrictions of wind farm structure include wind turbine design. Rotor length (or blade) is one of the major parameters of a wind turbine in the amount of energy generated. In addition, since wind farms contain multiple wind turbines, their placement also introduces another decision variable. Considering all of the aspects of wind farm, we modeled a multi-period mixed-integer non-linear optimization problem to decide on the investment decisions of a wind farm. We illustrate the efficiency and accuracy of our model on a real example of Turkey case and implemented the proposed model into GAMS (General Algebraic Modeling System) optimization package by using BARON solver. The result of 40.2 GW of energy that can be generated in the short term with immediate investments is obtained. Due to the uncertainty on wind speed values, we implement two-stage scenario based stochastic approach into the model and compared the results to see the value of perfect information. Finally, we conclude with the sensitivity analysis in order to see the affects of variations within the variables of the model.
Rafineri utilite sisteminin çok zamanlı tam sayılı doğrusal programlama ile optimizasyonu
Energy efficiency is one of the most important factors in refineries which affect the cost and the competitiveness. In order to deal with economical difficulties and compete with other refineries, a refinery must fulfill its energy demand on its own. Moreover securing the continuity of the processes is also another concern. In a refinery processes must be held online whatever the situation is. So it is unreliable to run such huge system only depending on outsources. In this concern we have provided a decision support system for determining the optimum working criteria of equipments in a Tüpraş İzmit Refinery power plant. Refineries use different sources for fulfill their energy demands. Generally power plants are the main providers. However for electricity demand refineries are usually connected to national grids in order to buy or sell electricity depending on the amount of production and the prices of the electricity in "day ahead" and "daily" electricity market. We illustrate the efficiency and accuracy of our model on a real example of İzmit Refinery Power Plant and implemented the proposed model into GAMS (General Algebraic Modeling System) optimization package by using "Cplex" solver. The result of 0.45 % of cost reduction can be obtained without any investment. Finally, we conclude with the sensitivity analysis in order to see the affects of variations in electricity prices in daily electricity market.
Destek vektör regresyon yöntemi ile bir atmosferik damıtma kolonu ürünlerinin fiziksel özelliklerini tahmin etme
Atmospheric distillation column is one of the most important units in an oil refinery where crude oil that is little use as it is, is fractioned into its more valuable constituents. Different fractions of the crude distillation unit are then further processed in downstream conversion and purification units to produce final products such as gasoline, diesel and jet fuel. The physical properties and the quantity of the final products vary depending on the physicochemical properties of the crude oil being processed and the operation parameters of the atmospheric distillation column. The process operators should keep the physical properties of hydrocarbon products in specified limits and operate the crude distillation unit according to instructions from the production planning department for maximizing the profit from operations. The physical properties of different crude distillation unit fractions must be measured by taking a sample from the stream periodically and analyzing these samples in a laboratory with appropriate equipment or by online analyzers that are very expensive to install, operate and maintain. Almost all of the state-of-the art online equipment has a time lag to complete the analysis in real time due to complexity of the analyses. Therefore, the laboratory or on-line measurements become available to decision makers and operator with a time lag. The intermittent nature of the measurements leads to sub-optimal control of the unit or in some cases leads to off-spec products that have considerably less economic value than the products that are within specified limits. As a result, estimation of the physical properties from online plant data and implementation of a soft sensor has a potential benefit in improving the profit. Among different approaches, the support vector regression shows great promise in machine learning due to its ability in generalizing well to unseen test data. The objective of this study is to fully estimate the physical properties of the hydrocarbon products of an atmospheric distillation column by support vector regression. Linear, Polynomial and Gaussian Radial Basis Function (Gaussian RBF) kernels are tested and the SVR parameters are optimized by embedding k-fold cross validation into a variety of algorithms including genetic algorithm (GA), grid search (GS) and non-linear programming (NLP). The performance of SVR is compared against artificial neural network (ANN) method and robust quality estimator (RQE) already functioning in the ADC. The testing results suggest that SVR with any of the integrated kernels performs well in estimating the property and generalizes well to blind test data. Compared to RQE, the mean testing error of estimation is improved by 31% with SVR from 6.4˚C to 4.4˚C and the standard deviation of estimation error is improved by 23% with SVR from 7.3 ˚C to 5.6 ˚C.
Urban transportation network design problem with sustainability considerations
Traffic congestion and environmental issues associated with transportation are considered to be serious problems faced by modern cities because of their negative effects on productivity, health and living conditions. Half a million people in developing countries die each year from transport-related air emissions, with a similar death toll from traffic accidents. Therefore, designing sustainable land use transportation systems is considered one of the most pressing issues faced by modern cities. In this dissertation, I develop optimization models and solution algorithms to design sustainable land use and transportation systems for urban regions to improve sustainability. In this study, two novel stochastic bi-objective land use and transportation optimization models are formulated to design a sustainable transportation network to minimize carbon monoxide emissions, traffic congestion, and travel costs: covering environmental, social, and economic aspects of sustainability. The goal is identifying the residential areas and roads in a city that should be expanded. Travel demand is considered as a random variable which affects the probability of traffic congestion and the travel cost. I formulate a closed form expression to calculate the Cholesky factorization matrix that is implemented as a set of constraints in the optimization model, to estimate the probability of traffic congestion by multivariate normal distribution. ϵ-constraint method is implemented to solve the multi-objective optimization problems. In addition, some useful insights on the relationship among land use, transportation network and environmental impact associated with them are found. The presented models are developed by considering a multi-modal assumption. The problem is formulated as a bi-level optimization problem. In the lower level, the transportation design problem is formulated to minimize traveler transportation costs and in the upper level I consider minimizing carbon monoxide emission and minimizing the probability of traffic congestion as the objective functions. I formulate the closed form optimality expressions of the lower level problem using Karush–Kuhn–Tucker (KKT) conditions. These optimality conditions are incorporated in the upper level problem as a set of constraints, and a single level model is obtained. The formulated single stage model is a Mixed Integer Non-linear Programming (MINLP) problem with two objective functions. An exact solution algorithm based on outer approximation and ϵ-constraint method is implemented to solve the bi-objective MINLP problem. The developed models are applied for Istanbul, Turkey to minimize emissions and overloading flows. The amount of emissions is calculated as a function of the real average speed in each road based on traffic congestion. In this study, the travel production and attraction of 38 districts and detail information of 17,663 Istanbul main roads such as numbers of lanes, length and maximum speed limit in each main road are implemented as the input to the optimization model. The large data set is successfully handled and the formulated models are solved optimally. The optimal results show the selected roads for capacity expansion and installing additional shared public vehicles. The result shows specific roads have significant effect on emission production and traffic congestion in Istanbul metropolitan area. Furthermore, I analyze the reason of traffic congestion and pollution in these roads. In addition, in this study I analyze the effect of different fuel types on different emission production. This work presents the first optimization framework at micro scale level to analyze traffic emission problem in the Istanbul metropolitan area. In addition, in this dissertation I explore the problem of selecting the optimal locations of electric vehicle charging stations in urban regions. I propose an optimization model based on vehicle travel patterns to capture public charging demand and select the locations of public charging stations to maximize the amount of vehicle miles traveled (VMT) being electrified. The formulated model is applied to Beijing, China as a case study using vehicle trajectory data of 11,880 taxis over a period of three weeks. The mathematical problem is formulated in GAMS modeling environment and Cplex optimizer is used to find the optimal solutions. Formulating the mathematical model properly, inputting data transformation, and adjusting the Cplex options are considered for accommodating large-scale data. I show that, compared to the 40 existing public charging stations, the 40 optimal ones selected by the model can increase electrified fleet VMT by 59% and 88% for slow and fast charging, respectively.
Synchromodal ulaşım ağları tasarımı ve işletilmesi için ayrık-sürekli optimizasyon yaklaşımı
Modern supply chain systems have become mostly integrated and complex due to the constant pressure to improve their efficiency thresholds. In Europe, some of the important challenges including traffic congestion, carbon emission rates, intermodal transportation integration, the need for new infrastructure developments and maintenance of existing infrastructures as well as ensuring quality, reliability and efficiency of transportation and mobility services have always been a concern for the transportation systems. These challenges are emerging as an urgent issue and there are many visions and solution methodologies addressed by different stakeholders in the transportation domain. The main objective of this thesis is the elaboration, testing and validation of new models for the analysis, design, evaluation, and management of complex sustainable transportation activities. This thesis presents the multi-objective mixed-integer programming problem for integrating specific characteristics of synchromodal transportation (such as real-time planning and bundling of shipments). The problem includes different objective functions including the total transportation cost, travel time and CO2 emissions of proposed networks while optimizing the network structure. The traffic congestion, time-dependent vehicle speeds and vehicle filling ratios are considered in an integrated manner and computational results for different illustrative cases are presented with real data from the Marmara Region of Turkey. The defined non-linear model is converted into linear form and solved by using a customized implementation of the ϵ-constraint method for the multi-objective mixed-integer linear programming problem. Then, the analysis of Pareto solutions and sensitivity analysis of proposed mathematical models with different pre-processing constraints are summarized for decisions makers. It is shown that the synchromodal transportation model presented in this paper is very affective in determining transportation network structures and optimized planning and operation of these networks.
Türkiye'deki atık piller stokastik geri dönüşüm içın ters lojistik ağı tasarımı ve problem çözme için sağlam optimizasyon
In recent decades, there has been an increasing interest in the mobile electronic devices that use batteries as their source of power. The growing demand for batteries creates environmental concerns for governments and companies because of their content of heavy metals that are usually significantly harmful to the environment. Waste batteries are known as hazardous waste and should therefore be appropriately disposed or recycled. Since finding suitable landfill areas is very hard—because people prefer to not live close to the landfill areas—recycling has become even more popular. In addition to the environmental benefits, recycling waste batteries also has economic advantages. There are a number of mathematical models in the literature that address the design of recycling systems for different products. The design of a reverse logistics system for recycling waste batteries is expressed as a multi-period mixed integer linear programming (MILP) problem to address the collection, transportation, sorting, recycling, and landfill operations of the waste batteries. This model considers various types of collected batteries, several existing and potential facilities with different types and capacity options, various construction and operational costs, and revenue from selling recycled batteries at a secondary market. The objective of this MILP model is to maximize the profit—i.e., the total revenue minus the total cost. The amount of used batteries that are returned by the end users is stochastic. Two-stage stochastic optimization method is used in the literature to address uncertainty in the model. In the first stage, the model makes strategic decisions—e.g., the capacity of the sorting and recycling facilities—while in the second stage, it makes the tactical decisions—e.g., inventory levels in warehouses. We extend the deterministic MILP model by using robust optimization approach to incorporate the uncertainty in the model. We consider four different scenarios for the amount of waste-batteries that are collected from the end users. We use GAMS (General Algebraic Modeling System) optimization package and the CPLEX solver to compare the results of these two approaches. We observe that the two-stage approach has worse results than the robust optimization method with respect to the worst-case performance. We also conduct a sensitivity analysis to show the impact of the amount of collected waste-batteries on the situation of facilities in the model.
Rafineri enerji ağı optimizasyonu
Energy consumption is a critical factor in refinery operations, having a significant impact on production costs. The effective management of the energy system in the refinery can improve the economic performance significantly. Energy demand in refineries does not stay constant due to change in crude oil properties, operation conditions of process units and cost of fuels. TUPRAS Izmit Refinery operates a complex utility system to satisfy its energy demand in the form of steam and electricity. The design of the utility plants allows multiple operational configurations to secure the continuity and flexibility of refinery processes in different operating conditions. The main objective of the work is developing a decision support system to manage the complex energy network of the refinery by determining the optimum operational combinations of the equipment for achieving a global minimum in terms of energy costs. In the scope of the work, first of all, the steam and power production equipment in the refinery are analyzed, thermodynamically to determine the variables which have an impact on the equipment efficiency. The determined variables are evaluated by regression analysis and the efficiency models of the equipment are developed. In the second part of the work, the optimization problem is formulated as mixed integer linear programming model. The model contains all operational constraints, efficiency models, mass balances, operational status of the equipment and demand satisfaction constraints of the refinery. The developed approach is analyzed and tested with a real case of the refinery. MILP model scenarios are solved by using GAMS CPLEX solver. In scenario analysis, the economic impact of optimization is evaluated by comparing the optimal solutions with online refinery operations. As a result, up to 3.5 % cost reduction is achieved without making an investment.
Çok gruplu veri sınıflandırması problemi için eniyileme tabanlı çokyüzlü bölge yaklaşımı
Multi-class data classification is a supervised machine learning problem that involves assigning data to multiple groups. There are various methods for data classification problems that are based on the separation of training sets by means of hyperboxes, hyperplanes and polyhedral regions. Polyhedral approaches are either designed for binary classification problem or do not focus on global optimal solutions. Hyperbox methods are restrictive in fitting a model to the data. In addition, the training models are complex to build optimal classifiers and they are applicable on instances up to a certain size. We address the multi-class data classification problem by a mixed integer linear programming model (MILP). We split data set of each class into subsets such that the subsets of different classes are separable by a hyperplane. The hyperplanes that separate a subset form a polyhedral region and the regions of different classes are disjoint. A MILP model is used to find the optimal separation by minimizing the total number of regions and misclassified data points. A preprocessing step, which is based on graph theory, is proposed to decompose or simplify the problem considering pairwise separation of classes. We have shown that there is an undirected graph for each dataset representing the linear separability relation of class pairs. Connected components of the graph is formed and MILP model is solved for each connected component with more than one element in the vertex set. We have shown that for the collection of the optimal solutions to the connected components, there is an equivalent optimal solution to the main MILP model. In addition, we present a novel MILP-based algorithm to form the linearly separable subsets. At each iteration we form a subset of samples out of the set of unassigned samples by a MILP model that maximizes the cardinality of the new subset. The generated subset which is linearly separable from the subsets of other classes is removed from the unassigned sample set. The rest of the samples are used for generating new subsets and the algorithm terminates when all the samples are assigned. We have shown that the subsets of different classes generated by the algorithm are linearly separable and can be used to construct a feasible solution to the general MILP model. Furthermore the algorithm can be used to reduce the dataset in few iterations so that MILP model is simplified. We build classifiers based on the convex hulls of the subsets and the polyhedral regions defined by the hyperplanes for the testing phase. The hyperplanes are generated by maximizing the margin as in the support vector machines. We evaluated our approach on 15 artificial datasets and 52 benchmark problems and compared with the methods from the literature. We conclude that our optimization based approach complemented with the proposed classifiers, provides competitive results in terms of prediction accuracy.