Thesis by Kutlay Kizil

Abstract

EN

TR

This study presents a comprehensive time series analysis of wind energy production in Türkiye, focusing on two primary objectives: quantifying the performance degradation of aging wind turbines and developing a robust short-term power forecasting model. A national-scale Wind Farms Database (WFD) was compiled, integrating publicly available data from Turkish institutions (EPIAS, EPDK, TUREB) with ERA5 meteorological reanalysis data.

The aging analysis, conducted on a filtered dataset of 74 wind farms over a four-year period (2020-2023), revealed a statistically significant mean annual capacity factor degradation of -0.90% (95% CI [-1.04%, -0.76%]). This degradation rate did not show a significant correlation with turbine manufacturer, installation year, or other physical plant characteristics within the analyzed timeframe.

For short-term forecasting, a Long Short-Term Memory (LSTM) network was developed to be used with autoregressive inference. The model demonstrated high predictive accuracy, achieving an average 1-hour-ahead normalized Mean Absolute Error (nMAE) of approximately 6% and an R² score of 0.91.

The study concludes by providing a crucial performance degradation benchmark for the Turkish wind fleet and a validated, efficient LSTM model for forecasting. It also highlights the critical performance gap between idealized academic studies using reanalysis data and real-world applications, suggesting future work on techniques like scheduled sampling to improve model robustness against the uncertainties of autoregressive inference.

Bu çalışma, Türkiye'deki rüzgar enerjisi üretimine yönelik kapsamlı bir zaman serisi analizi sunmakta ve iki temel hedefe odaklanmaktadır: yaşlanan rüzgar türbinlerinin performans düşüşünü nicelemek ve güvenilir bir kısa vadeli güç tahmin modeli geliştirmek. Türk kurumlarından (EPİAŞ, EPDK, TÜREB) kamuya açık olarak temin edilen veriler ile ERA5 meteorolojik yeniden analiz verilerini entegre eden ulusal ölçekli bir Rüzgar Enerji Santralleri Veritabanı (WFD) oluşturulmuştur.

Dört yıllık bir dönemi (2020-2023) kapsayan ve filtrelenmiş 74 rüzgar santralinden oluşan bir veri seti üzerinde yürütülen yaşlanma analizi, istatistiksel olarak anlamlı, yıllık ortalama -%0.90'lık bir kapasite faktörü düşüşü olduğunu ortaya koymuştur (%95 GA [-1.04, -0.76]). Bu düşüş oranının, analiz edilen zaman dilimi içinde türbin üreticisi, kurulum yılı veya diğer fiziksel santral özellikleri ile anlamlı bir korelasyon göstermediği tespit edilmiştir.

Kısa vadeli tahmin için bir Uzun Kısa Süreli Bellek (LSTM) ağı geliştirilmiştir. Optimize edilmiş otoregresif model, 1 saatlik ileriye dönük tahminde yaklaşık %6'lık bir normalize Ortalama Mutlak Hata (nMAE) ve 0.91'lik bir R² skoru elde ederek yüksek bir öngörü doğruluğu sergilemiştir.

Çalışma, Türkiye’de rüzgar enerjisi için kritik bir performans düşüşü referans noktası ve verimli bir LSTM modeli sunarak sonuçlanmaktadır. Aynı zamanda, yeniden analiz verileri kullanan idealize edilmiş akademik çalışmalar ile gerçek dünya uygulamaları arasındaki kritik performans farkını vurgulamakta ve modelin otoregresif modelin dezavantajlarına karşı direncini artırmak için ‘scheduled sampling’ gibi teknikler üzerine gelecekteki çalışmalar için önerilerde bulunmaktadır.

1. Introduction

The global imperative to transition towards renewable energy is more pressing than ever. Wind energy stands as a major contributor, with Türkiye's installed capacity exceeding 13.2 GW, representing a significant portion of its total renewable energy portfolio. However, the large-scale integration of this variable resource introduces challenges for grid operators and energy markets, impacting stability and operations.

This research addresses two critical aspects of wind energy operations in Türkiye:

Performance Degradation: The operational efficiency of wind turbines naturally declines over time due to aging, a crucial factor for investors in assessing the lifecycle cost and profitability of wind projects.
Intermittency & Forecasting: The fluctuating nature of wind requires accurate power forecasts to ensure grid balancing, optimize maintenance, and maximize economic value.

This study aims to provide a quantitative benchmark for turbine aging in Türkiye and to develop a robust, practical short-term forecasting tool.

2. Methodology

The foundation of this study is the comprehensive Wind Farms Database (WFD), a national-scale database created by integrating public data from Turkish institutions (EPİAŞ, EPDK, TÜREB) with the ERA5 meteorological reanalysis dataset from ECMWF.

2.1. Turbine Aging Analysis

A rigorous data preprocessing pipeline was implemented. Production data from 2020-2023 was filtered based on availability. Then, a density-based clustering algorithm, DBSCAN, was employed to identify and remove anomalous data points, ensuring the reliability of the trend analysis. The annual performance degradation trend was then determined by applying a linear regression model to the capacity factor data for each farm.

Illustration of the DBSCAN clustering algorithm separating core points, border points, and noise. — Figure 1: Illustration of the DBSCAN algorithm used for outlier removal.

2.2. Short-Term Forecasting

A machine learning model with a multi-layered Long Short-Term Memory (LSTM) network was developed. The model was trained to predict future power generation values using an autoregressive method, where the model's own recent predictions are used as inputs for subsequent forecasts. Key input features included 100-meter wind components from ERA5, calculated air density, and the lagged production value from the previous hour.

3. Results

3.1. Turbine Aging

The analysis of 74 wind farms reveals a statistically significant mean annual capacity factor degradation of -0.90%. This provides a crucial, data-driven benchmark for the Turkish wind fleet. The result is highly significant (t(74) = -12.74, p < .001) with a 95% Confidence Interval of [-1.04%, -0.76%].

Box plot showing the distribution of annual degradation trends for 74 wind farms. — Figure 2: Box plot of the wind farm degradation trends.

Further analysis showed no statistically significant correlation between the degradation rate and turbine manufacturer, installation year, or other physical factors within the 2020-2023 timeframe.

Box plot showing degradation trends grouped by turbine manufacturer. — Figure 3: Trend distribution grouped by turbine brands (single-brand farms only).

3.2. Production Forecasting

A systematic hyperparameter optimization process resulted in a final model with 3 LSTM layers (256-128-64 neurons). The model demonstrates high predictive accuracy, achieving an average 1-hour-ahead normalized Mean Absolute Error (nMAE) of approximately 6% and an R² score of 0.91.

Static chart showing average model performance vs forecast horizon — Figure 4: Average error metrics for WPPs (24-hour forecast horizon).

The autoregressive model exhibits a self-regulating behavior. It gives significant weight to its previous prediction, causing a delay in its response to sharp changes. However, the model has also learned to distrust its own anomalous predictions when they conflict with plausible weather data, preventing cumulative error from spiraling out of control.

Animated GIF showing continuous inferences for the testing period — Figure 5: Continuous inferences for the testing period.

Animated GIF showing time series view of the forecast horizons — Figure 6: Time series view of the forecast horizons.

The following figures illustrate the model's behavior in different scenarios. The first set of plots shows the model accurately tracking relatively stable, high-production periods. The second set demonstrates the model's ability to capture sharp ramps up and down in generation, albeit with a slight delay, which is characteristic of autoregressive models reacting to new information.

More examples of single forecast inferences

Figure 7: Examples of individual 24-hour autoregressive forecast inferences.

4. Discussion

A critical limitation of this study is its reliance on ERA5 reanalysis data, which serves as a near-perfect, error-free representation of past weather conditions. This creates an idealized scenario for model training and evaluation.

A critical "performance gap" exists between these academic results and real-world applications, which must use weather forecasts containing their own inherent and growing uncertainty.

This gap means that while the model architecture is proven effective, its operational accuracy will inevitably be lower when deployed with real, imperfect weather predictions. Understanding and quantifying this gap is essential for managing expectations and for the practical application of this forecasting tool in the energy sector.

5. Conclusion

This study makes two primary contributions. Firstly, it establishes a crucial performance degradation benchmark of -0.90% per year for the Turkish wind fleet. Secondly, it presents a validated, high-performing autoregressive LSTM model for short-term forecasting, achieving a 1-hour-ahead nMAE of approximately 6% and an R² score of 0.91. The research highlights the critical "performance gap" between idealized models and real-world applications, outlining a clear path for future work.

6. Future Work

To bridge the performance gap, future research should focus on:

Validating the model's performance using actual, imperfect weather prediction data to quantify the real-world performance drop.
Implementing techniques like 'scheduled sampling' to train the model to be more robust to its own errors during autoregressive inference.
Extending the aging analysis period to capture long-term effects and integrating operational data.