The growing demand for energy-efficient and reliable Lithium-Ion Batteries (LIBs) motivated the need to have advanced tools to monitor and predict their internal states and long-term behavior. The aim of this thesis is to address how Machine Learning (ML) and Deep Learning (DL) models can be efficiently applied to estimate five key performance indicators (KPIs) of Lithium-Ion Batteries. These are State-of-Charge (SoC), State-of-Health (SoH), Energy Efficiency (EE), Capacity Fade (CF), and Cycle Life (CL). The goal of this thesis is to evaluate the trade-off between predictive accuracy and model performance. Three Machine Learning models (Random Forest, Support Vector Regressor, and XGBoost) and three Deep Learning models (Simple Recurrent Neural Network, Recurrent Neural Network with Long Short-Term Memory architecture, and Recurrent Neural Network with Gated Recurrent Unit architecture) were trained and evaluated on a large set of battery data. The evaluation was done using accuracy metrics (RMSE, MAE, MAPE, and SMAPE) and performance metrics (training time, inference time, and model size). Results showed that no single model outperformed all others across all the target variables. However, Random Forest revealed the best accuracy for SoC and SoH, while Recurrent Neural Networks (RNNs) architectures performed better for EE, CF, and CL. Remarkably, the RNN with Long Short-Term Memory architecture achieved the best overall trade-off between accuracy and performance. In conclusion, the findings demonstrate how Machine Learning and Deep Learning approaches can be applied to battery KPIs prediction and offer insightful suggestions on model selection based on accuracy and performance requirements.
Alex Orlandi (Wed,) studied this question.