WoS İndeksli Yayınlar Koleksiyonu / WoS Indexed Publications Collection

Permanent URI for this collectionhttps://hdl.handle.net/11147/7150

Browse

Filters

2025 - 20266

Settings

Start date: 2025Subject: search.filters.subject.Deep Learning

Search Results

Now showing 1 - 6 of 6
  • Article
    Univariate Deep Learning Models for Short-Term Electricity Load Forecasting from Renewables
    (Ankara University, Faculty of Science, 2025) Kabran, Fatma Basoglu; Unlu, Kamil Demirberk
    Renewable energy offers a cost-effective, carbon-free solution for energy needs, while protecting the environment. Accurate forecasting of electricity generation from renewable sources is crucial for the efficiency of modern power grids. This study employs a univariate deep learning approach to predict daily renewable energy generation, evaluating Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs) as candidate models. Five performance metrics-mean absolute error, root mean squared error, mean absolute percentage error, mean absolute scaled error and the coefficient of determination-are employed to assess the forecasting power of the algorithms. The empirical results show that CNN outperforms other models, achieving an R2 of almost 94%. This research shows that the univariate model based on historical data of electricity load generated from renewables can accurately predict day-ahead electricity load, even without meteorological data.
  • Article
    Automated Detection and Quantification of Honey Adulteration Using Thermal Imaging and Convolutional Neural Networks
    (Pergamon-Elsevier Science Ltd, 2026) Unluturk, Mehmet S.; Berk, Berkay; Unluturk, Sevcan
    Honey is a valuable natural food rich in bioactive substances beneficial to health. Despite strict regulations prohibiting adulteration, honey remains one of the most frequently adulterated foods, often with low-cost commercial syrups. Conventional detection methods require expensive instruments, expert operators, and lengthy analysis times, limiting their practical use. This study introduces a rapid and automated method for detecting and quantifying honey adulteration using thermal image analysis combined with a tailored Convolutional Neural Network (CNN) architecture. Thirty-six pure honey samples (blossom and honeydew) from different regions of T & uuml;rkiye were adulterated with inverted sugar, maltose, and glucose syrups at varying levels (3 %-60 % weight/weight (w/w)). Samples were heated to 60 degrees C and thermal images were captured during cooling using a custom image-capturing unit. The CNN model employed a multi-layer structure, starting with a shallow network for binary classification (pure vs adulterated honey) achieving 100 % accuracy, followed by specialized deeper CNN regressors to quantify adulterant levels with mean squared errors of 0.0003, 0.001, and 0.0002 for glucose, maltose, and inverted sugar, respectively. This layered CNN approach leverages thermal patterns linked to adulteration, enabling sensitive, rapid, and non-destructive quality control. Furthermore, the method is integrated into a user-friendly hardware-software system called Compact Adulteration Testing Cabinet on Honey (CATCH), requiring no specialized expertise, demonstrating strong potential for automated honey authenticity verification in practical settings.
  • Article
    Comprehensive Analysis and Machine Learning-Based Solutions for Drift Behavior in Ambient Atomic Force Microscope Conditions
    (Pergamon-Elsevier Science Ltd, 2025) Deveci, D. Gemici; Barandir, T. Karakoyun; Unverdi, O.; Celebi, C.; Temur, L. O.; Atilla, D. C.
    This study outlines the effectiveness of combining numerical methods, Computer Vision (CV) and Machine Learning (ML) approaches to analyze and predict drift behavior in high-resolution Atomic Force Microscope (AFM) scanning procedures. Using Long Short-Term Memory (LSTM) models for time series analysis and the Light Gradient Boosting Machine (LightGBM) algorithm for predictive modeling, significant progress was achieved in understanding the dynamic and variable nature of drift and mitigating its impact on scanning. The models demonstrated a robust predictive capability, achieving approximately 94% accuracy in drift predictions. The study emphasizes the nonstationary characteristics of drift and demonstrates how the selection of features directly related to the target variable enhances the efficiency of the model and enables adaptive real-time correction. These findings confirm the predictive strength of the models and highlight the potential for integrating ML predictions with real-time feedback mechanisms to improve the resolution and stability of AFM imaging in both scientific and industrial applications.
  • Conference Object
    Current Sensing in Phase-OTDR Systems Using Deep Learning
    (SPIE, 2025) Yeke, M.C.; Sirin, S.; Yüksel, K.; Gumus, A.
    Fiber optic current sensors are marked by a number of advantages such as light-weight, small-size and inherently insulated nature when compared to conventional current transformers which get bulkier and costlier as the desired values of current to be measured increase. Phase-OTDR is a widely known technology especially in acoustic and thermal sensing, but it suffers from noise that limits its usage for current sensing especially for low currents. In order to interpret the noisy data retrieved from Phase-OTDR current sensor simulator, deep learning techniques can have promising performance. In this paper, 3 different types of deep learning models were proposed and applied on the data generated by Phase-OTDR current sensor simulator tool to improve the ability to distinguish low and similar current levels. The current measurements were analyzed as a classification problem where different current ranges with different current increments are selected as different classes. The proposed method provided 100% accuracy at a difference of 20 A between the current levels. In addition, other scenarios where the current levels were increased by 15 A and 10 A were also studied. In this case, the accuracies 97% and 89% were obtained, respectively. © 2025 Elsevier B.V., All rights reserved.
  • Article
    Towards Facile Deep Learning Architectures for Chemical Processes: Simultaneous Pseudo-Global Training and Economic Synthesis
    (Institution of Chemical Engineers, 2025) Sildir, Hasan; Yalcin, Damla; Tuncer, Basak; Deliismail, Ozgun; Leblebici, Mumin Enis
    Chemical process data is usually not directly valorized in pure machine learning predictive models due to limited data availability. This limitation often caused from high sensor costs, data variety, and veracity issues. In response, this study proposes a novel formulation based on mixed-integer linear programming (MILP), called Approximated Deep Learning (ADL), to overcome these limitations and enable accurate modeling under data scarcity. The ADL simultaneously performs input selection, outlier filtering, and training of deep learning architectures within a single-level optimization problem. The method approximates the nonlinear and nonconvex components of traditional deep learning models in the mixed-integer domain through sophisticated reformulations, achieving a pseudo-global solution. A key feature of ADL is the integration of sensor pricing as a regularization mechanism, which promotes cost-efficient soft sensor design without compromising predictive performance. The proposed framework is validated on a publicly available bubble column dataset and benchmarked against four conventional deep learning methods. Results show that ADL achieves superior test accuracy with more than 50% reduction in input space, drastically reducing sensor cost. Furthermore, the optimized architecture is a high-quality initial guess for transfer learning on larger datasets. Overall, the method offers a practical and economically viable solution for data-driven chemical process modeling.
  • Article
    Citation - WoS: 5
    Citation - Scopus: 5
    Automated Deep Learning Model Development Based on Weight Sensitivity and Model Selection Statistics
    (Pergamon-elsevier Science Ltd, 2025) Yalcin, Damla; Deliismail, Ozgun; Tuncer, Basak; Boy, Onur Can; Bayar, Ibrahim; Kayar, Gizem; Sildir, Hasan
    Current sustainable production and consumption processes call for technological integration with the realm of computational modeling especially in the form of sophisticated data-driven architectures. Advanced mathematical formulations are essential for deep learning approach to account for revealing patterns under nonlinear and complex interactions to enable better prediction capabilities for subsequent optimization and control tasks. Bayesian Information Criterion and Akaike Information Criterion are introduced as additional constraints to a mixed-integer training problem which employs a parameter sensitivity related objective function, unlike traditional methods which minimize the training error under fixed architecture. The resulting comprehensive optimization formulation is flexible as a simultaneous approach is introduced through algorithmic differentiation to benefit from advanced solvers to handle computational challenges and theoretical issues. Proposed formulation delivers 40% reduction, in architecture with high accuracy. The performance of the approach is compared to fully connected traditional methods on two different case studies from large scale chemical plants.