Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Nadir Subaşı has a total of 16 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Adaptive Hybrid GWO–SSA Optimized Deep Learning Framework for Accurate Power Forecasting of Next-Generation Perovskite Photovoltaic Systems Under Desert Climate Conditions

    by Ali Mahmood Aswad, Maftun Aliyev, Aysel Ersoy, Nadir Subaşı, Ertuğrul Adıgüzel

    Accurate photovoltaic (PV) power forecasting is essential for enhancing grid stability, optimizing energy management, and facilitating the large-scale integration of renewable energy resources. Although deep learning techniques have demonstrated promising results in PV forecasting, their predictive performance is highly dependent on effective hyperparameter optimization. Furthermore, studies dedicated to field-deployed perovskite photovoltaic systems operating under semi-arid desert climatic conditions remain limited. To address this gap, this study proposes an Adaptive Hybrid Grey Wolf Optimizer–Sparrow Search Algorithm (AH-GWOSSA) deep learning framework for single-step-ahead (5 min lead time) and multi-step power forecasting of a perovskite PV system installed in Mosul, Iraq. A real-world dataset comprising 21,456 valid daytime observations (filtered for solar irradiance > 5.0 W/m2) was collected between November 2025 and April 2026 and partitioned strictly chronologically (70% training, 10% validation, 20% held-out testing). The proposed framework was benchmarked against Persistence, an unoptimized Base LSTM, Standard GRU, CNN-LSTM, standalone GWO, SSA, Random Search, and an ablation Fixed-Weight Hybrid under an equivalent evaluation budget of 160 candidate network trainings. Over 30 independent optimization runs, AH-GWOSSA achieved the lowest mean RMSE of 15.55 W (std. 0.34 W), an MAE of 8.79 W (std. 0.25 W), and the highest mean R2 of 0.885 (std. 0.008). Non-parametric Wilcoxon signed-rank testing across 30 paired run-wise RMSE values confirmed a statistically significant improvement over the unoptimized Base LSTM (W=78.0, p=1.8×10−6 at α=0.01). Multi-horizon evaluations (5, 15, 30, 60, and 120 min) demonstrated consistent superiority over Persistence and Base LSTM, with RMSE improvements of up to 39.8% at the 30 min horizon. The findings demonstrate that adaptive hybrid metaheuristic optimization provides a competitive framework for ultra-short-term perovskite PV forecasting, with RMSE improvements of up to 39.8% over Persistence at the 30 min horizon, offering a solid foundation for future smart-grid and battery storage management applications.

  • Open Access

    Comparative Analysis of Ensemble Learning Methods for Software Reliability Prediction

    by Nadir Subaşı, Özen Özer

    Ensuring software reliability is a fundamental requirement in the software development lifecycle, as early detection of potential failures significantly reduces maintenance costs. The literature documents several mathematically based classification techniques that provide developers with a versatile toolkit. Among these are Logistic Regression, Random Forest, Extra Tree Classification, Gaussian Naive Bayes, XGB Classifier, LGBM Classifier, CatBoost Classifier, and Voter Classifier. The present research aims to identify which classification method is most suitable for predicting software reliability based on code metrics such as Cyclomatic Complexity and Halstead Volume. Therefore, the research will use relevant datasets and graphics in its analysis with the goal of identifying the most effective classification approach from the techniques cited. The study evaluates various models including Logistic Regression, Random Forest, XGBoost, and Voting Classifier. The results demonstrate that ensemble methods provide superior performance in identifying reliability levels. Its primary purpose is to ensure the resulting data pattern is immune to compromise. Many options are available when it comes to mathematical classification types. As indicated in the literature, these include, but are not limited to, the examples previously listed.

  • Open Access

    AI-driven solar energy forecasting using meteorological variables: Insights from a 22 MW PV plant in Nakhchivan

    by Ertuğrul Adıgüzel, Nadir Subaşı, Javanshir Zeynalov, Maftun Aliyev, Sevinj Rzayeva, Rüya Şamlı, Aysel Ersoy

    This study presents a systematic comparative evaluation of ten regression-based machine-learning models for day-ahead photovoltaic (PV) energy forecasting under semi-arid climatic conditions. The analysis is conducted using a limited six month dataset (May–October 2024) of real operational production data obtained from a 22 MW grid-connected PV plant in Nakhchivan, Azerbaijan, integrated with key meteorological variables including solar irradiance, air temperature, relative humidity, and wind speed. Linear Regression, Ridge, Lasso, ElasticNet, Support Vector Regression (SVR), Decision Tree, Random Forest, Gradient Boosting, XGBoost, and a Multi-Layer Perceptron (MLP) were benchmarked using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R 2 ), and five-fold cross-validated R 2 . The results indicate that regularised linear models—particularly Lasso Regression—provide the most consistent balance between predictive accuracy and generalisation stability under moderate data availability, while Random Forest demonstrates strong cross-validated robustness and the MLP exhibits overfitting behaviour, highlighting sensitivity to limited training data. These findings demonstrate that increased model complexity does not necessarily translate into improved forecasting reliability in data-constrained semi-arid environments. The novelty of this work lies in its unified benchmarking framework that simultaneously evaluates predictive accuracy, interpretability, and generalisation performance using real-world utility-scale operational data. By explicitly linking forecasting reliability with sustainability-oriented planning, the study contributes to more reliable life-cycle cost and emission assessment of large-scale PV systems in emerging renewable-energy markets. The findings offer practical guidance for energy planners and policy-makers seeking transparent, computationally efficient forecasting strategies in semi-arid climates.

  • Open Access

    Optimizing XGBoost via mSMA_plus: A Novel Meta-Heuristic Approach for High-Precision Multiclass Dry Bean Classification

    by Nadir Subaşi

    Precise classification of dry bean varieties holds critical importance for agricultural sustainability, food security, and the preservation of seed quality standards. Traditional classification methods rely on human intervention and exhibit significant error rates; this necessitates the use of high-performance machine learning models and effective optimization strategies. This study aims to propose an innovative framework that optimizes the hyperparameters of the Extreme Gradient Boosting model for classifying seven different bean varieties on the Dry Bean Dataset using meta-heuristic algorithms. Within this study, critical parameters of the XGBoost model, such as learning rate, tree depth, and subsampling rates, have been systematically tuned using Slime Mould, Modified SMA (mSMA), mSMA_plus, Particle Swarm Optimization, and Grey Wolf Optimizer algorithms. The effectiveness of the proposed methods has been comparatively evaluated against commonly used GridSearch and RandomSearch techniques in the literature. The experimental results, assessed using accuracy, F1-score, precision, and recall metrics, reveal that the proposed mSMA_plus algorithm achieves a peak classification accuracy of 99.39% and an F1-score of 0.9939. This marks a clear architectural advancement over baseline frameworks, raising the classification accuracy baseline by approximately 1.15% compared to traditional GridSearch approaches within a total execution timeline of 507.55 s.


  • Open Access

    Reliability Analysis of Al2O3/Epoxy Nanocomposites Under Surface Tracking Degradation

    by Kaan Kırkağaç, Nadir Subaşı, Ertuğrul Adıgüzel

    Surface tracking resistance and long-term reliability are critical performance parameters for polymeric insulation materials used in high-voltage electrical applications. In this study, the tracking behavior of epoxy/aluminum oxide (Al2O3) nanocomposites was experimentally, statistically, and data-driven investigated under high electrical stress conditions. Epoxy composites containing 0, 1, 3, and 5 wt.% Al2O3 nanoparticles were fabricated and tested according to the International Electrotechnical Commission(IEC) 60112 standard at applied voltage levels of 400 V, 500 V, and 600 V, with repeated experiments conducted to ensure statistical consistency. The results demonstrated that Al2O3 incorporation significantly influences tracking resistance in a concentration-and voltage-dependent manner, where the composite containing 1 wt.% Al2O3 exhibited the highest tracking resistance at moderate voltage levels (400–500 V), achieving higher drop-to-failure values (defined as the number of contaminant droplets required to cause sustained surface tracking for more than 2 seconds or material burn-through, in accordance with IEC 60112) than pure epoxy. However, increasing the filler content beyond 1 wt.% led to a reduction in tracking lifetime, particularly under high electrical stress, which was associated with increased surface erosion and localized degradation. The degradation behavior at 600 V was further analyzed using two-parameter Weibull statistics, revealing a decrease in characteristic life with increasing Al2O3 content, while the 1 wt.% Al2O3 composite exhibited the highest Weibull shape parameter, indicating a stable and predictable wear-out dominated failure mechanism. In addition, machine learning models were developed to predict tracking lifetime with high accuracy (R2 > 98%), and feature importance analysis identified erosion-related mass loss, filler content, and applied voltage as the dominant parameters governing tracking degradation. Overall, the strong agreement between experimental results, Weibull reliability analysis, and machine learning predictions provides a combined experimental and data-driven framework for evaluating and optimizing epoxy-based insulation materials for high-voltage applications. Cite this article as: K. Kırkağaç, N. Subaşı and E. Adıgüzel, “Reliability analysis of Al2O3/epoxy nanocomposites under surface tracking degradation,” Electrica, 26, 0017, doi: 10.5152/electrica.2026.26017.

  • Open Access

    Reliability Analysis of Al2O3/Epoxy Nanocomposites Under Surface Tracking Degradation

    by Kaan Kırkağaç, Nadir Subaşı, Ertuğrul Adıgüzel

    Surface tracking resistance and long-term reliability are critical performance parameters for polymeric insulation materials used in high-voltage electrical applications. In this study, the tracking behavior of epoxy/aluminum oxide (Al2O3) nanocomposites was experimentally, statistically, and data-driven investigated under high electrical stress conditions. Epoxy composites containing 0, 1, 3, and 5 wt.% Al2O3 nanoparticles were fabricated and tested according to the International Electrotechnical Commission(IEC) 60112 standard at applied voltage levels of 400 V, 500 V, and 600 V, with repeated experiments conducted to ensure statistical consistency. The results demonstrated that Al2O3 incorporation significantly influences tracking resistance in a concentration-and voltage-dependent manner, where the composite containing 1 wt.% Al2O3 exhibited the highest tracking resistance at moderate voltage levels (400–500 V), achieving higher drop-to-failure values (defined as the number of contaminant droplets required to cause sustained surface tracking for more than 2 seconds or material burn-through, in accordance with IEC 60112) than pure epoxy. However, increasing the filler content beyond 1 wt.% led to a reduction in tracking lifetime, particularly under high electrical stress, which was associated with increased surface erosion and localized degradation. The degradation behavior at 600 V was further analyzed using two-parameter Weibull statistics, revealing a decrease in characteristic life with increasing Al2O3 content, while the 1 wt.% Al2O3 composite exhibited the highest Weibull shape parameter, indicating a stable and predictable wear-out dominated failure mechanism. In addition, machine learning models were developed to predict tracking lifetime with high accuracy (R2 > 98%), and feature importance analysis identified erosion-related mass loss, filler content, and applied voltage as the dominant parameters governing tracking degradation. Overall, the strong agreement between experimental results, Weibull reliability analysis, and machine learning predictions provides a combined experimental and data-driven framework for evaluating and optimizing epoxy-based insulation materials for high-voltage applications. Cite this article as: K. Kırkağaç, N. Subaşı and E. Adıgüzel, “Reliability analysis of Al2O3/epoxy nanocomposites under surface tracking degradation,” Electrica, 26, 0017, doi: 10.5152/electr.2026.26017.

  • Reliability Analysis of Surface Tracking Method Using Classification Methods (svm, random forest, XGBoost) with Whale Optimization Algorithm

    by Ertuğrul Adıgüzel, Nadir Subaşı, Aysel Ersoy

    This study presents a comprehensive reliability analysis of a surface tracking methodology integrating machine learning classifiers—Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)—optimized via the Whale Optimization Algorithm (WOA). The framework addresses surface degradation detection by leveraging experimental data from epoxy/MgO nanocomposite materials subjected to hydro-aging and electrical tracking conditions. Features representing physical and electrical behaviors were engineered and subjected to advanced preprocessing and classification workflows. WOA was employed for hyperparameter tuning and feature importance calibration, aiming to improve model robustness and accuracy. Results indicate that XGBoost, when optimized with WOA, outperforms SVM and RF in all metrics, achieving 92.3% accuracy and a macro F1-score of 0.94, with enhanced generalizability and resilience to class imbalance. The dataset consists of 180 samples, categorized into three reliability classes (‘low’, ‘medium’, and ‘high’) based on the quartile thresholds of reliability scores. The study underlines the critical importance of nature-inspired optimization in classification pipelines and provides a robust blueprint for predictive modeling in materials reliability applications.

  • Open Access

    Optimization of Istanbul Urban Traffic Data with Grey Wolf Optimization-Assisted Machine Learning Techniques: Comparative Analysis of Long Short-Term Memory and eXtreme Gradient Boosting

    by Nadir Subaşı

    This study examines machine learning approaches enhanced by the Grey Wolf Optimization (GWO) algorithm for analyzing Istanbul’s complex urban traffic patterns. The GWO was employed due to its capability to efficiently handle complex nonlinear traffic patterns and its robustness in achieving global optimization, outperforming several conventional metaheuristics. Using hourly traffic density data from June–December 2024 obtained from Turkey’s Ministry of Environment database, researchers developed a Unified Traffic Density Index combining average speed (40%), vehicle count (40%), and speed variation (20%) to identify the busiest traffic zones. The dataset was obtained from the Ministry of Environment, Urbanization, and Climate Change and covers the period from June 1, 2024, to December 31, 2024. The dataset contains hourly data for 2462 geohash regions within the boundaries of Istanbul province. The methodology compared long short-term memory (LSTM) and eXtreme Gradient Boosting (XGBoost) algorithms using both standard and GWO-optimized hyperparameters. Time series analysis separated trend, seasonality, and randomness components while examining hourly and daily periodicity patterns in traffic data. Results demonstrated that GWO optimization significantly enhanced both algorithms’ performance. The standard LSTM model’s systematic deviations and wave-like patterns were substantially reduced through GWO optimization. The XGBoost performed consistently in both versions, with the GWO-XGBoost combination achieving superior prediction accuracy. Performance metrics revealed that GWO-XGBoost attained the lowest mean squared error (1.6209) and mean absolute error (0.9082) values while achieving the highest coefficient of determination (R2 percentage bias = ±0.8486%, outperforming other configurations. These findings indicate that the GWO-XGBoost combination shows significant potential as a highly accurate solution for traffic density prediction applications for traffic management systems within critical high-density zones of metropolitan areas like Istanbul, particularly for traffic density prediction applications. The study concludes that advanced optimization techniques are essential for addressing traffic management challenges in rapidly urbanizing cities with increasing vehicle density. Cite this article as: N. Subaşi, "Optimization of Istanbul urban traffic data with Grey wolf optimization-assisted machine learning techniques: comparative analysis of long short-term memory and eXtreme gradient boosting," Electrica, 25, 0196, 2025. doi:10.5152/electrica.2025.25196.

  • Unraveling Data Complexity in the Metaverse for Anomaly Detection With Python on NYC Taxi

    by Özen Özer Özer, Nadir Subasi

    In the dynamic environment of the Metaverse, where virtual interactions and transactions thrive, detecting data anomalies becomes imperative for maintaining integrity and security. This paper explores the application of Pythonbased anomaly detection models, including Inter Quartile Range (IQR), Median Absolute Deviation (MAD), and Local Outlier Factor (LOF), in identifying anomalies within NYC Taxi data. Through comprehensive analysis and experimentation, we investigate the effectiveness and comparative performance of these models in detecting outliers amidst the complex and diverse data landscape of the Metaverse. In the NYC Taxi Data, which contains 10320 data, it was analyzed with the mentioned algorithms and 2 anomalies (0.019%) were found with IQR. In the same data set, 1 anomaly (0.009%) was found with MAD model and 1032 anomalies (10%) were found with LOF