Geri Dön

Türkiye enerji piyasasının makine öğrenmesi uygulamasıyla düşük-karbon salımına geçişte ajan temelli modellenmesi

Turkish energy market agent-based modeling for low-carbon transition with machine learning application

PDF İndir

  1. Tez No: 995537
  2. Yazar: BURAK GÖKÇE
  3. Danışmanlar: PROF. DR. GÜLGÜN KAYAKUTLU
  4. Tez Türü: Doktora
  5. Konular: Enerji, Bilim ve Teknoloji, Energy, Science and Technology
  6. Anahtar Kelimeler: Elektrik piyasası, Enerji piyasası, Karbon emisyonu, Makine öğrenmesi, Sıfır emisyon, Yenilenebilir enerji kaynakları, Çok ajanlı sistemler, Electric market, Energy market, Carbon emission, Machine learning, Zero emission, Renewable energy resources, Multiagent systems
  7. Yıl: 2026
  8. Dil: Türkçe
  9. Üniversite: İstanbul Teknik Üniversitesi
  10. Enstitü: Lisansüstü Eğitim Enstitüsü
  11. Ana Bilim Dalı: Enerji Bilimi ve Teknolojileri Ana Bilim Dalı
  12. Bilim Dalı: Enerji Bilim ve Teknoloji Bilim Dalı
  13. Sayfa Sayısı: Belirtilmemiş.

Özet

Yeşil enerjiye geçiş süreci, giderek karmaşıklaşan enerji sistemlerini analiz edebilecek gelişmiş modelleme tekniklerini gerekli kılmaktadır. Bu teknikler arasında, çok ajanlı sistemler veya ajan temelli modelleme (ATM), enerji piyasalarındaki etkileşimleri ve belirsizlikleri ayrıntılı biçimde temsil edebilmesi nedeniyle öne çıkmaktadır. Makine öğrenmesi (MÖ) yöntemleri ise, yenilenebilir enerji santralleri, teşvik mekanizmaları, elektrikli araçlar ve batarya enerji depolama sistemleri gibi yeni piyasa ajanlarının davranışlarını daha gerçekçi şekilde simüle etmek amacıyla ATM'ye entegre edilmektedir. Başlangıçta ML yöntemleri çoğunlukla pekiştirmeli öğrenme (reinforcement learning, RL) yaklaşımlarıyla ajan öğrenmesini geliştirmek için kullanılırken, zamanla dışsal veri tahminine duyulan ihtiyaç doğrultusunda çok daha geniş ML yöntemlerinin kullanımı yaygınlaşmıştır. Bu tez çalışması, 2014–2024 dönemindeki literatürü inceleyerek MÖ–ATM entegrasyonuna ilişkin kapsamlı bir değerlendirme sunmaktadır. İnceleme sürecinde, piyasa veri türleri, simülasyon süreleri ve ML katkı tipleri gibi nitelikler kullanılarak özdüzenleyici haritalar (self organizing maps, SOM) üzerinde kümeleme yapılmıştır. Bulgular, ajan öğrenmesi ve geleneksel RL yöntemlerinin hâlen baskın olduğunu, ancak uzun dönemli analizler ile veri tahmini çalışmalarının literatürde yetersiz temsil edildiğini göstermektedir. Ayrıca, elektrik dışındaki enerji piyasalarına ilişkin gözlemlerin entegrasyonu da önemli bir araştırma ihtiyacı olarak ortaya çıkmaktadır. Türkiye'nin enerji piyasasının 2053 net-sıfır hedefi doğrultusunda düşük karbonlu bir yapıya dönüşümü, uzun vadeli belirsizlikleri ve karmaşık piyasa etkileşimlerini yakalayabilecek bütünleşik modelleme yaklaşımlarını gerektirmektedir. Bu kapsamda, çalışmada uzun dönemli bir ATM çerçevesi geliştirilmiş; talep tahmini için mevsimsel otoregresif bütünleşik hareketli ortalama (seasonal autoregressive integrated moving average, SARIMA) modeli, gün öncesi piyasa (GÖP) fiyat tahmini için ise MÖ tabanlı yöntemler kullanılmıştır. Test edilen MÖ modelleri arasında CatBoost, XGBoost ve Rastgele Orman'a kıyasla en yüksek doğruluk düzeyine ulaşmış ve yatırım analizlerinde kullanılan net bugünkü değer (net present value, NPV) hesaplamalarını desteklemiştir. Modelde üretim santralleri, yatırımcılar ve piyasa işletmecisi gibi temel piyasa aktörleri temsil edilmiş; ayrıca Türkiye'nin ilk nükleer santrali ile ulusal bir emisyon ticaret sistemi (ETS) kurulması senaryolarının etkileri de modele dahil edilmiştir. Tüm model bileşenleri geçmiş verilerle doğrulanmış ve güçlü tahmin performansı göstermiştir. 2053 yılına kadar saatlik çözünürlükte yürütülen farklı politika ve talep senaryoları, yenilenebilir enerji üretiminin özellikle karasal rüzgâr ve güneş yatırımları sayesinde istikrarlı biçimde arttığını göstermektedir. Nükleer kapasitenin devreye alınması elektrik piyasa fiyatlarını düşürürken, ETS uygulaması fiyatları önemli ölçüde artırmakta; her ikisi de karbon emisyonlarını azaltmaktadır. Talep varsayımları ise üretim karışımı, fosil yakıt bağımlılığı ve emisyonlar üzerinde belirleyici rol oynamaktadır. Elde edilen senaryo sonuçları, yenilenebilir enerji yatırımlarının ölçeklendirilmesi, ETS tasarımı, nükleer enerji yol haritası ve enerji verimliliği politikaları gibi alanlarda uygulanabilir politika önerilerine dönüştürülmüştür. Sunulan ATM çerçevesi, politika yapıcılar, yatırımcılar ve piyasa katılımcıları için ileriye dönük bir karar destek aracı niteliği taşımakta olup; gelecekte diğer enerji piyasalarının eklenmesi, depolama sistemlerinin entegrasyonu ve daha zengin senaryo tasarımlarıyla genişletilebilir.

Özet (Çeviri)

The global shift toward low-carbon energy systems has intensified the need for analytical tools capable of capturing the complexity, interdependence, and long-term uncertainty of modern electricity markets. Traditional optimization or equilibrium-based models often fall short in representing heterogeneous decision-making processes, adaptive behavior, and policy-induced structural changes. In contrast, agent-based modeling (ABM) provides a bottom-up framework that simulates the interactions of decentralized actors such as generators, investors, consumers, and policymakers, thereby offering a powerful method to analyze transition pathways under evolving technological and regulatory conditions. As energy systems become increasingly data-rich and decentralized, machine learning (ML) methods have also gained prominence for learning and forecasting. The integration of ML techniques within ABM environments allows for more realistic, data-driven simulations that reflect both the behavioral and structural dynamics of contemporary energy markets. Motivated by these developments, this thesis investigates Turkey's transition toward a low-carbon electricity system in the context of its 2053 net-zero target. Turkey's electricity market is undergoing rapid changes, driven by accelerated renewable deployment, planned nuclear expansion, evolving market mechanisms, and the upcoming introduction of a national emissions trading system (ETS). These dynamics create a unique environment that demands sophisticated modeling approaches capable of integrating long-term policy commitments, short-term market behavior, investment decision processes, and the interactions between thermal, renewable, and emerging technologies. Addressing these challenges, the thesis develops a comprehensive modeling platform that integrates econometric demand forecasting, ML-based price prediction, and an agent-based simulation framework capable of evaluating energy market evolution at hourly resolution. The research begins with a systematic literature review to assess the current state of ABM and ML applications in energy systems. 65 academic articles published between 2014 and 2024 are examined and clustered using self-organizing maps (SOM). The clustering is based on attributes such as data type (real versus artificial), simulation horizon (short-term versus long-term), and ML contribution (agent learning versus exogenous prediction). The review reveals several key insights. First, although ABM has become a popular tool for modeling electricity market behavior, most studies focus on short-term bidding strategies and rely heavily on synthetic datasets. Second, ML integration typically centers on reinforcement learning (RL) to optimize strategic bidding, rather than using ML for system-level forecasting. Third, only a small number of studies investigate long-term dynamics such as investment decisions, policy evaluation, or carbon pricing. This review therefore highlights the need for long-term ABM frameworks that incorporate ML-based predictive models, real-world market data, and policy instruments relevant to emerging energy transitions. To address these gaps, the thesis develops a multi-layered modeling architecture grounded in real and publicly accessible data from national and international sources. Hourly electricity demand and generation records, day-ahead market (DAM) prices, historical and projected fuel prices, installed capacity data, investment cost parameters, and renewable support scheme tariffs are compiled and processed using Python and SQL. The data pipeline includes missing value handling, anomaly detection, normalization, and structural validation to ensure consistency across historical data. This robust data foundation forms the backbone of the demand forecasting, price prediction, and investment modeling components. The econometric component employs a Seasonal Autoregressive Integrated Moving Average (SARIMA) model to forecast Türkiye's long-term electricity demand. The model is trained on monthly demand data from 1975 to 2021. Validation against 2022 out-of-sample data yields a mean absolute percentage error (MAPE) of 3.5%, confirming the suitability of SARIMA for long-term forecasting. The monthly forecasts are subsequently disaggregated into hourly profiles that reflect historical intraday consumption patterns, producing realistic hourly demand inputs for the ABM simulations. ML-based price prediction forms the second major methodological pillar. Three tree-based algorithms—CatBoost, XGBoost, and Random Forest—are trained on hourly market data from 2017 to 2020 and tested on 2021 data. Input features include fuel prices, renewable output, and residual demand. Among the models evaluated, CatBoost achieves the highest performance with a 6.09% MAPE and an annual mean error of 1.94%, outperforming other ML benchmarks and providing reliable hourly price forecasts. These model outputs feed into the ABM system to capture the impact of future fuel price variations, renewable expansion, and market conditions on day-ahead price formation. The core ABM platform is implemented using AnyLogic and Python, simulating Turkey's electricity market from 2022 to 2053 under multiple scenarios. The model includes generator agents representing renewable, thermal, and nuclear power plants; investor agents making capacity decisions based on net present value (NPV) calculations; and a market operator agent representing Energy Exchange Istanbul (EXIST, EPİAŞ), responsible for DAM clearing and ETS operations where applicable. Investor agents operate under an economic rationality approach, using NPV evaluations to determine capacity additions or retirements. The model simulates generator bidding behavior based on marginal cost calculations that include fuel prices, operation and maintenance (O&M) costs, and carbon costs. The scenario engine enables hourly market clearing over the entire study horizon, capturing long-term dynamics such as price trends, emissions trajectories, investment flows, and generation mix evolution. The scenario set encompasses a wide range of policy and demand trajectories. The Stated Policies Scenario includes planned nuclear expansion (Akkuyu) but no ETS implementation. The No-Nuclear case removes nuclear capacity to evaluate its counterfactual impact. The Emissions Trading Scenario introduces a national ETS from 2026 onward, with carbon prices aligned with the EU ETS trajectory. High and Low Demand Scenarios apply ±10% adjustments to the baseline demand forecast to assess system sensitivity. Additional sensitivity analysis include emission factor reductions for thermal plants, and discount rate changes between 4%, 6%, and 8% for NPV analysis. The simulation results provide deep insights into Türkiye's potential energy transition pathways. Under Stated Policies, renewable energy generation share rises to 63.1% in 2053, driven primarily by large increases in solar and wind capacity. Solar PV and onshore wind jointly reach 35.8% of generation, enabled by declining technology costs and strong investor interest. Natural gas plants remains crucial for system flexibility, accounting for 14.8% of generation in 2053, while coal-based generation declines significantly. Total CO₂ emissions peak in 2038 at 147.2 Mt and then fall to 101.4 Mt by 2053—a 25.9% reduction from 2022 level. Nuclear power plays a strategically important role: excluding Akkuyu increases average prices by 6.28 USD/MWh and raises 2053 emissions by 20.2 Mt, demonstrating the value of low-carbon baseload capacity. The Emissions Trading Scenario accelerates decarbonization. By adding carbon costs to marginal bids, ETS shifts dispatch decisions toward renewable and low-carbon resources, reducing emissions by an additional 8.4 Mt CO₂ in 2053 relative to the baseline. However, carbon pricing increases average wholesale market prices by 57.91 USD/MWh from 2026 onward. Sensitivity analysis reveals that reducing emission factors by 10% or 20% significantly mitigates these price effects, suggesting that efficiency improvements and partial abatement technologies could soften ETS-induced price shocks. Demand uncertainty is another critical factor. In the High Demand Scenario, increased consumption requires greater reliance on natural gas, elevates average prices by 6.46 USD/MWh, and lowers renewable shares. In contrast, the Low Demand Scenario reduces prices by 7.11 USD/MWh and accelerates decarbonization by increasing renewable penetration to 68% in 2053. Regardless of scenario, emissions decline over time, indicating that Turkey's long-term decarbonization trajectory is robust, though its pace and cost vary. Based on the simulation results, the thesis provides several actionable recommendations. Strengthening renewable support mechanisms by extending and enhancing existing programs is essential to maintaining investor confidence and accelerating renewable deployment. Facilitating affordable financing can further stimulate investment. Upgrading Türkiye's transmission and distribution networks, deploying energy storage, and digitalizing grid operations are critical for integrating high shares of variable renewables. Enhancing regional market integration with the EU can provide additional flexibility and cost efficiency. Implementing an ETS with transparent, predictable pricing rules aligned with the EU ETS would provide credible decarbonization signals. Demand-side measures, including stronger enforcement of energy efficiency obligations, widespread adoption of demand-response mechanisms, and deployment of smart grid technologies, are important for shaping demand patterns and mitigating the need for costly capacity expansions. Nuclear planning should continue with the timely commissioning of Akkuyu and the preparation of long-term strategies for additional nuclear capacity. Methodologically, the thesis contributes an integrated energy system modeling framework that combines ABM, ML, and econometrics into a unified long-term simulation environment. The model uses real market data, transparent assumptions, and validated components, ensuring both empirical reliability and policy relevance. The future work can incorporate distributed generation, microgrids, battery storage systems, endogenous demand elasticity, more sophisticated representations of investor risk and behavioral heterogeneity, and generative AI for enhanced prediction and scenario generation. In conclusion, this thesis demonstrates that integrating agent-based modeling with machine learning and econometric forecasting provides a powerful, data-driven, and flexible framework for assessing long-term energy market evolution under diverse policy and demand conditions. The findings show that renewable energy expansion, nuclear power deployment, carbon pricing mechanisms together form the cornerstone of Turkey's long-term decarbonization strategy. The model reveals that while Turkey can significantly reduce emissions under current policies, achieving deep decarbonization requires coordinated efforts across renewable support mechanisms, grid modernization, ETS implementation, nuclear planning, and demand management. The modeling framework developed in this thesis offers a comprehensive ex-ante decision-support tool for policymakers, investors, and market stakeholders, helping guide Türkiye's transition to a sustainable, secure, and low-carbon energy future.

Benzer Tezler

  1. Tez No

    714401

    Comparative analysis of XGBoost and LightGBM methods for day ahead spot natural gas price forecasting

    Gün öncesi spot doğalgaz fiyatı tahminlemesinde LightGBM ve XGBoost metodlarının karsılastırılmalı analizi

    DOĞUKAN ŞAHİN

    Yüksek Lisans

    İngilizce

    İngilizce

    2022

    Enerjiİstanbul Teknik Üniversitesi

    Enerji Bilimi ve Teknolojileri Ana Bilim Dalı

    DR. ÖĞR. ÜYESİ MUSTAFA BERKER YURTSEVEN

  2. Tez No

    782821

    YEKDEM katılım kararı için arıma modeli kullanılarak uzun dönemli elektrik fiyat tahminlemesi

    Long-term electricity price forecast using the arima model for the decision to participate of YEKDEM

    ULVİ FİKRİ DOĞAN

    Yüksek Lisans

    Türkçe

    Türkçe

    2023

    Endüstri ve Endüstri Mühendisliğiİstanbul Teknik Üniversitesi

    Elektrik Mühendisliği Ana Bilim Dalı

    PROF. DR. AHMET CANSIZ

  3. Tez No

    826973

    Yapay zekâ yöntemleri ile elektrik piyasa takas fiyatının tahmini ve enerji maliyet optimizasyonu

    Electricity market clearance price forecast and energy cost optimization with ai methods

    MESUT AKSOY

    Yüksek Lisans

    Türkçe

    Türkçe

    2023

    EnerjiIsparta Uygulamalı Bilimler Üniversitesi

    Enerji Sistemleri Mühendisliği Ana Bilim Dalı

    PROF. DR. AHMET KABUL

  4. Tez No

    827198

    Evolution of electricity markets from the perspective of production and organized markets

    Organize toptan satış piyasaları ve üretim perspektifinden elektrik piyasalarının dönüşümü

    MUSTAFA ÇAĞRI PEKER

    Doktora

    İngilizce

    İngilizce

    2023

    EkonomiHacettepe Üniversitesi

    İktisat Ana Bilim Dalı

    PROF. DR. AYŞEN SİVRİKAYA

  5. Tez No

    196783

    Electricity generation investment planning under uncertainity: A real options application for Türkiye

    Belirsiz şartlar altında elektrik üretimi yatırım planlaması: Türkiye için bir reel seçenekler uygulaması

    MUSTAFA LÜTFÜ DEMİREL

    Yüksek Lisans

    İngilizce

    İngilizce

    2005

    Endüstri ve Endüstri MühendisliğiBoğaziçi Üniversitesi

    Endüstri Mühendisliği Ana Bilim Dalı

    YRD. DOÇ. DR. GÜRKAN KUMBAROĞLU

    PROF. DR. KUBAN ALTINEL

Geri Dön