Akademik Veri Yönetim Sistemi
7th IEEE Global Power, Energy and Communication Conference, GPECOM 2025, Bochum, Almanya, 11 - 13 Haziran 2025, ss.13-18, (Tam Metin Bildiri)
Accurate estimation of State of Charge (SOC) across different temperatures is essential for ensuring the optimal performance and dependability of electrical systems, particularly in electric vehicles, Electrical Vertical Take-Off and Landing aircraft (eVTOLs), and smart devices. In this study, both hybrid and non-hybrid models are evaluated through a dynamically structured neural network to identify the most efficient approach for SOC prediction. The model architecture, which includes layer types, neuron counts, activation functions, and batch normalization, is optimized to enhance performance. A comparative analysis demonstrates that hybrid models achieve superior efficiency in SOC estimation, with lower Mean Absolute Error (MAE) values and enhanced generalization across varying temperature conditions. The findings indicate that combining various recurrent architectures enhances both adaptability and precision, making hybrid models a more effective and reliable option for SOC estimation. To further explore model performance, a dynamic model generation strategy was employed, enabling the construction of diverse architectures using real-world battery data collected under various driving cycles and temperature conditions. Among the evaluated models, the GRU-LSTM hybrid architectures achieved the lowest MAE values, demonstrating their improved generalization capability across diverse thermal and operational conditions.