Prediksi State Of Charge (SOC) Menggunakan Machine Learning Pada Fast Charging Baterai Lithium Iron Phosphate

Christian, Alifevious (2026) Prediksi State Of Charge (SOC) Menggunakan Machine Learning Pada Fast Charging Baterai Lithium Iron Phosphate. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022241011-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Estimasi State of Charge (SOC) yang akurat sangat krusial bagi keselamatan dan efisiensi baterai kendaraan listrik, meskipun sifat elektrokimia baterai yang dinamis dan non-linear menjadi tantangan besar. Penelitian ini membandingkan tiga arsitektur Machine Learning (ML) yaitu LSTM, GRU, dan ResNet18 berbasis CNN menggunakan dataset Universitas McMaster dan dataset mandiri baterai LFP pada kondisi fast charging. Selain menguji model, penelitian ini juga mengevaluasi dampak variasi fungsi aktivasi (sigmoid, hard-sigmoid, tanh, dan softsign) serta pengaruh fluktuasi suhu dan arus pengisian terhadap akurasi prediksi.
Hasil eksperimen menunjukkan bahwa rumpun RNN (LSTM dan GRU) lebih unggul daripada ResNet18, di mana model LSTM berhasil memberikan akurasi terbaik dengan nilai RMSE terkecil. Di sisi lain, GRU menjadi solusi alternatif yang seimbang karena menawarkan akurasi yang mendekati LSTM namun dengan komputasi yang lebih cepat dan hemat memori, berbanding terbalik dengan ResNet18 yang membutuhkan kapasitas memori besar akibat ukuran parameternya. Secara spesifik, kombinasi model terakurat dalam penelitian ini dicapai oleh arsitektur LSTM yang menggunakan fungsi aktivasi Tanh untuk state activation function dan Hard-sigmoid untuk gate activation function.
=====================================================================================================================================
Accurate State of Charge (SOC) estimation is crucial for the safety and efficiency of electric vehicle batteries, although the dynamic and non-linear electrochemical nature of batteries poses a significant challenge. This study compares three Machine Learning (ML) architectures namely LSTM, GRU, and CNN-based ResNet18 using the McMaster University dataset and an independent LFP battery dataset under fast charging conditions. In addition to testing the models, this research evaluates the impact of various activation functions (sigmoid, hard-sigmoid, tanh, and softsign) as well as the effects of temperature fluctuations and charging currents on prediction accuracy.
Experimental results indicate that the RNN family (LSTM and GRU) outperforms ResNet18, with the LSTM model delivering the highest accuracy and the lowest RMSE. On the other hand, GRU serves as a well balanced alternative, offering accuracy close to LSTM but with faster computation and lower memory consumption unlike ResNet18, which demands a large memory capacity due to its parameter size. Specifically, the most accurate model combination in this study is achieved by the LSTM architecture utilizing the Tanh activation function for the state and Hard-sigmoid for the gate.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Baterai LFP, Battery LFP, GRU, LSTM, CNN, State of Charge, Fast Charging
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK2941 Storage batteries
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Alifevious Christian
Date Deposited:	02 Aug 2026 06:32
Last Modified:	02 Aug 2026 06:32
URI:	http://repository.its.ac.id/id/eprint/142006

Actions (login required)

View Item