Geri Dön

Makine öğrenmesi tabanlı akıllı şehir çözümleri: Bina türü tanıma, su tüketimi öngörüsü ve sızıntı tespiti

Machine learning-based smart city solutions: Building type identification, water consumption forecasting, and leak detection

PDF İndir

  1. Tez No: 1018567
  2. Yazar: RAUF YASİN
  3. Danışmanlar: DR. ÖĞR. ÜYESİ SEÇKİN ARI
  4. Tez Türü: Doktora
  5. Konular: Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
  6. Anahtar Kelimeler: Belirtilmemiş.
  7. Yıl: 2026
  8. Dil: Türkçe
  9. Üniversite: Sakarya Üniversitesi
  10. Enstitü: Fen Bilimleri Enstitüsü
  11. Ana Bilim Dalı: Bilgisayar Mühendisliği Ana Bilim Dalı
  12. Bilim Dalı: Bilgisayar Mühendisliği Bilim Dalı
  13. Sayfa Sayısı: Belirtilmemiş.

Özet

Küresel nüfus ve kentleşme oranlarındaki hızlı artış, modern şehirleri hava kirliliği, ulaşım, enerji ve su yönetimi gibi karmaşık zorluklarla karşı karşıya bırakmaktadır. Bu sorunları çözmek için geleneksel mühendislik yöntemlerinin ötesine geçmek ve veri odaklı akıllı şehir yönetim sistemleri geliştirmek şarttır. Bu doktora tezi kapsamında, kentsel altyapı sistemlerinin heterojen yapısını analiz etmek ve kaynak yönetimini optimize etmek için makine öğrenmesine dayalı bir“Akıllı Şehir Tasarım Modeli”geliştirilmiştir. Çalışma, Fransa'daki Lille Üniversitesi kampüsünde bulunan“SunRise Akıllı Şehir Gösterim Projesi”nden elde edilen gerçek dünya verileri kullanılarak doğrulanmıştır. Tezin temel metodolojisi üç ana sütun üzerine kurulmuştur. İlk olarak, yalnızca su tüketim verilerinden bina fonksiyonlarını (örneğin, konut, laboratuvar, restoran) belirlemek için“Bina İmzası”kavramı önerilmiştir. Bu aşamada, günlük yaşam döngüsü beş farklı zaman dilimine (Uyku, Uyanma, Çalışma, Öğle Yemeği ve Eve Dönüş) bölünerek insan davranış modelleri özellik mühendisliğine entegre edilmiştir. Sekiz makine öğrenme modellenin karşılaştırıldığı bir analizde, Karar Ağacı algoritması saatlik verilerde %94 doğruluk oranına ulaşarak en başarılı model olarak ortaya çıkmıştır. Çalışmanın ikinci aşamasında, gelecekteki kaynak tüketimini tahmin etmek için regresyon modelleri geliştirilmiştir. 15 farklı regresyon algoritması kullanılarak yapılan bu tahminler, akıllı su yönetimi için stratejik bir temel oluşturmuştur. Üçüncü ve son aşamada ise, altyapı güvenliğini artırmak amacıyla“DBSCAN-Leak”adı verilen, DBSCAN kümeleme algoritmasının geliştirilmiş bir versiyonu önerildi. Bu yöntem, hem günlük hem de saatlik gerçek zamanlı verileri kullanarak su şebekelerindeki sızıntıları ve anormal su tüketimini yüksek hassasiyetle tespit edebilmiştir. Sonuçlar, önerilen akıllı şehir tasarım modelinin yalnızca su gibi belirli şebekeler için etkili olmakla kalmayıp, ölçeklenebilir mimarisi sayesinde farklı kentsel alt sistemlere de uygulanabilir olduğunu göstermektedir. Geliştirilen modeller, sızıntı tespitinden tüketim tahminine kadar geniş bir uygulama yelpazesinde akıllı şehir yönetim platformlarına entegre edilebilir. Bu araştırma, akıllı şehirlerin yalnızca bir teknoloji yığını değil, insan davranışının ve altyapı verilerinin uyum içinde olduğu yaşayan sistemler olduğunu göstermektedir. Geliştirilen tasarım örüntüsü, geleceğin sürdürülebilir ve dirençli şehirleri için evrensel bir mühendislik yol haritası niteliğindedir.

Özet (Çeviri)

The rapid increase in global population and urbanization rates confronts modern cities with complex challenges across energy and water management. As populations concentrate in metropolitan areas, the strain on critical infrastructure has surpassed the operational limits of traditional management systems. Modern cities face a paradoxical challenge: while they generate vast amounts of data through massive Internet of Things (IoT) deployments, this data often remains trapped in 'silos,' hindering a holistic understanding of urban dynamics. To address these inefficiencies, it is essential to move beyond traditional engineering toward data-driven systems. This doctoral thesis argues that the primary obstacle to a true 'Smart City' is not a lack of data, but the absence of a unified framework capable of integrating heterogeneous streams. Consequently, this research develops a standardized Smart City Design Pattern that shifts the focus from reactive maintenance to proactive governance, centered on three critical pillars: Identity, Forecasting, and Security.  Identity: Automatically identifying the purpose of a building based on consumption data (Building Signature).  Forecasting: Predicting future resource demands to enable proactive management.  Security: Identifying infrastructure failures, such as water leaks, in real-time. The empirical foundation of this research is the SunRise Smart City Demonstrator at the University of Lille. This facility represents a unique large-scale urban laboratory, encompassing a 100 km water distribution network and 93 high-precision Automatic Meter Reading (AMR) sensors. The dataset utilized in this study comprises over 4 million hourly consumption records, providing a rigorous testing ground for academic schedules, seasonal changes, and aging physical infrastructure, which introduce real-world complexity often missing from purely theoretical models. This environment enables a systemic analysis of urban metabolism, where every liter of water recorded becomes a data point reflecting the campus's pulse. A significant challenge in modern urban planning is the lack of up-to-date metadata. In many rapidly developing cities, official building use classifications—such as residential, commercial, or industrial—often change without being updated in central records, leading to inefficiencies in resource allocation. This research introduces the concept of the“Building Signature”to solve this identity crisis. The methodology posits that water consumption is a direct, high-fidelity proxy for human behavior. To increase classification accuracy, the study moved beyond simple averages and instead divided the 24-hour cycle into five distinct“Human Behavior Windows”: Sleep, Wake-up, Work, Lunch, and Return Home. This feature engineering enables algorithms to distinguish between a residential building, where usage peaks in the morning and evening, and a laboratory or office, where usage is sustained throughout the workday. Eight supervised learning algorithms were rigorously tested and compared to classify buildings solely on the basis of these temporal consumption patterns. The algorithms included Decision Tree, Naive Bayes, Random Forests, Extremely Randomized Trees, Gradient Boosting, k-Nearest Neighbors, Adaboost and Logistic Regression. Among these, the Decision Tree algorithm emerged as the superior model, achieving a 94.2% accuracy rate at hourly intervals. The evaluation used confusion matrices to visualize the performance across different time granularities—hourly, daily, weekly, monthly, and seasonal. This high level of precision proves that buildings possess a unique“digital DNA”or signature. By implementing such a model, city managers can perform automated urban auditing, identifying changes in building usage in real-time without the need for manual data entry or intrusive on-site inspections. This capability is essential for dynamic zoning and the optimization of city-wide utility services. Forecasting resource demand constitutes the second significant technical contribution of this thesis, transitioning the focus from classification (what is the building?) to regression (how much will it consume?). Accurate prediction is the backbone of urban sustainability, enabling utilities to optimize supply, manage water network pressure to prevent bursts, and reduce operational costs. This study implemented and compared 15 different regression algorithms. These algorithms are Linear Regression, Ridge Regression, Lasso, Elastic Net, Least Angle Regression (LARS), Orthogonal Matching Pursuit (OMP), Bayesian Ridge Regression, Nearest Neighbors Regression, Decision Tree Regressor, Gradient Boosting Regressor, Random Forest Regressor, Extremely Randomized Trees, Bagging meta-estimator, AdaBoost, and Voting Regressor. To evaluate these models, the research utilized metrics such as Mean Squared Error (MSE) and Variance Score (R2). The performance was measured across multiple time horizons. The findings revealed a critical trade-off between the level of detail and the reliability of the prediction. While hourly models are prone to“noise”from the erratic behavior of a few users, the study found that seasonal and monthly models provided the highest reliability. These longer-term models successfully accounted for exogenous variables that are often ignored in standard algorithms, such as university holidays, academic breaks, and climate-driven shifts in consumption. For example, during summer breaks, the campus consumption drops significantly, a pattern that the seasonal models captured with high precision. This multi-layered forecasting approach provides a roadmap for infrastructure planning, allowing city planners to determine when a network needs expansion versus when it simply needs more intelligent management during peak hours. The most direct impact on urban sustainability is the reduction of resource loss through advanced security measures. Water leakage accounts for significant economic and environmental waste, often totaling up to 30% of treated water in older urban networks. This thesis proposes DBSCAN-Leak, an advanced adaptation of the Density-Based Spatial Clustering of Applications with Noise algorithm. The operational efficiency of the DBSCAN-Leak framework is centered on a dual-mode detection strategy designed to mitigate the inherent volatility of urban water consumption. By moving beyond static thresholds, the framework utilizes density-based spatial clustering to differentiate between stochastic“noise”and genuine infrastructure failure. The first layer of defense is the Daily Analysis mode, which focuses on macro-level anomalies by clustering total daily consumption volumes. The second and more sophisticated layer is the Hourly Real-Time Analysis mode, engineered to address“silent leaks”—small, continuous flows that do not trigger high-volume alarms. The innovation here lies in calculating a Leak Probability index based on temporal persistence. In this network's architecture, the concept of a single, static threshold is replaced by a Multi-Dimensional Thresholding Strategy. This recognizes that“normal”consumption is a fluid concept, shifting dramatically by the hour and the day of the week. The Hourly Thresholding mechanism mirrors the city's circadian rhythm. During the early morning hours, the threshold is calibrated to its most sensitive state; because human activity is at a statistical minimum during this“Sleep Window,”even a minor flow is likely to represent a leak. As the city transitions into“Work”periods, the system automatically elevates the threshold to accommodate legitimate usage. Beyond the hourly cycle, the framework applies Day-of-the-Week Thresholding. In a campus environment, consumption on Monday is fundamentally different from that on Sunday. On weekdays, high-volume usage is permitted in labs and offices, while on weekends, the“normal”activity cluster shifts toward residential zones. Validated against historical records and physical leak reports from the University of Lille campus, the DBSCAN-Leak model successfully identified several leaks that traditional systems had missed for weeks. One specific case study in the 4Canton building demonstrated that the algorithm could pinpoint a physical leak event during the September-October period by observing the shift in cluster density. By catching these incidents early, the framework demonstrates a quantifiable reduction in infrastructure waste and operational expenditure. The research concludes that a pattern-based approach is the only viable path to truly integrated smart cities. Rather than purchasing proprietary software for every new sensor, cities can adopt this open design pattern that is scalable and applicable to any urban utility, including gas and electricity. The primary contributions of this work include the development of a scalable open design pattern, demonstrating that integrating human behavior time windows is critical for machine learning accuracy, and creating a robust real-time leak detection framework. By applying principles of Transfer Learning, the insights gained from the water network at the University of Lille can be applied to other city systems, such as electricity grids, by mapping similar“signatures”to consumption data. This work provides a technical and strategic roadmap for municipalities to transition from reactive,“break-fix”maintenance to proactive, data-driven governance. By treating the city as a living, data-generating organism rather than a collection of static pipes and wires, we can ensure the sustainability and resilience of future urban environments in the face of rapid global growth. The modularity of the proposed pattern ensures that as new sensors or“smart”devices are added to the city, the underlying ML models can be retrained or adapted, maintaining the city's intelligence over decades of evolution.

Benzer Tezler

  1. Tez No

    968555

    Yapay zeka ile kentsel sistemler mühendisliğinde sürdürülebilir yapıların oluşturulması ve akıllı şehirlerin planlanması

    Creating sustainable structures and planning smart cities in urban systems engineering with artificial intelligence

    EMRE TONKUŞ

    Yüksek Lisans

    Türkçe

    Türkçe

    2025

    İnşaat Mühendisliğiİstanbul Arel Üniversitesi

    İnşaat Mühendisliği Ana Bilim Dalı

    DR. ÖĞR. ÜYESİ MEHDİ ÖZTÜRK

  2. Tez No

    973253

    Akıllı şehirler için kentsel yangın risk yönetimine yönelik model tasarımı

    Model design for urban fire risk management for smart cities

    TUĞRUL URFALI

    Doktora

    Türkçe

    Türkçe

    2025

    Jeodezi ve FotogrametriErciyes Üniversitesi

    Harita Mühendisliği Ana Bilim Dalı

    PROF. DR. ABDURRAHMAN EYMEN

  3. Tez No

    974281

    Determining the barriers and opportunities for the integration of extended reality in disaster management using MCDM methods

    Afet yönetiminde genişletilmiş gerçeklik teknolojilerinin entegrasyonu önündeki engeller ve fırsatların ÇKKV yöntemleriyle belirlenmesi

    EREN MERT KAŞ

    Yüksek Lisans

    İngilizce

    İngilizce

    2025

    Bilim ve Teknolojiİstanbul Teknik Üniversitesi

    Afet ve Acil Durum Yönetimi Ana Bilim Dalı (disiplinlerarası)

    DOÇ. DR. ÖMER EKMECİOĞLU

  4. Tez No

    102

    Diyabetus mellitus'da gözlenen mikrovasküler komplikasyonların prostaglandinler ile olan ilişkisi

    The Relation of microvaskular complications observed in diabetes mellitus with prostoglandins

    CANAN NEBİGİL

    Doktora

    Türkçe

    Türkçe

    1985

    Eczacılık ve FarmakolojiCumhuriyet Üniversitesi

    Farmakoloji Ana Bilim Dalı

    YRD. DOÇ. DR. YUSUF SARIOĞLU

  5. Tez No

    353

    Kam yapma aparatı

    Cam milling attachment

    ORHAN UZUN

    Yüksek Lisans

    Türkçe

    Türkçe

    1986

    Makine MühendisliğiGazi Üniversitesi

    Makine Mühendisliği Ana Bilim Dalı

    DOÇ. DR. FEVZİ ERCAN

Geri Dön