Deep Learning Methodology for Charging Management Applications in Battery Cells Based on Neural Networks
Rolando Antonio Gilbert Zequera, Viktor Rjabtšikov, Anton Rassõlkin, Toomas Vaimann, Ants Kallaste
Abstract
A Battery Energy Storage System (BESS) plays an important role in achieving energy transition and climate change mitigation, with charging management applications being a crucial topic to improve the build, design, and operation of renewable technologies. With the continuous development of Artificial Intelligence (AI), implementing accurate algorithms that monitor Key Performance Indicators (KPIs) and provide predictive maintenance is a callenging task. This article presents a solid and robust Deep Learning methodology based on Neural Networks (NNs) in the TensorFlow framework and using Python as a programming language, all to predict the Open Circuit Voltage (OCV) and improve the state estimation of battery cells. Extensive tests on Lithium-ion cells under diverse operating conditions were carried out, with data acquisition meticulously recorded using a programmable DC Electronic Load. Various architectures were designed using Keras as a high-level Application Programming Interface (API) to build Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs). In addition, advanced computer science techniques were executed to improve the performance of the algorithms, such as cross-validation, Fine-tuning, Regularization, Bayesian optimization, Machine Learning techniques, and Data-Driven approaches. The resulting network architectures were stored in Hierarchical Data Format (HDF5) files, tested against both seen and unseen data to ensure its robustness and effectiveness in new battery measurements. The Deep Learning methodology provides remarkable testing accuracy of over 95% for all types of NNs, affirming its high adaptability and reliability in the development of AI-powered technology for battery management.