Data-Driven Capacity Modeling of 18650 Lithium-Ion Cells from Experimental Electrical Measurements

The prediction of lithium-ion battery capacity degradation is crucial for enhancing the reliability, efficiency, and sustainability of energy storage systems. This study proposes a data-driven approach to model capacity degradation in 18650 lithium-ion cells, supporting the long-term performance and responsible management of battery technologies. A systematic search was conducted to identify publicly available experimental datasets reporting charge/discharge processes, leading to the selection of the MIT-BIT Battery Degradation Dataset (Fixed Current Profiles and Arbitrary Use Profiles). This dataset was chosen for its extensive degradation data, variability, and adaptability to real-world applications. Of the 77 tested cells, 73 were included after filtering data completeness; cells with missing critical information, such as temperature, were excluded. A subset of cells tested under a 1C–2C charge/discharge profile was analyzed, and cell 52 was selected for its comprehensive structure. Using this dataset, a predictive model was developed to estimate the battery capacity based on the current, voltage, and temperature, with capacity as the target variable. A neural network was implemented using TensorFlow and Keras, incorporating ReLU activation, Adam optimization, and multiple loss functions. The dataset was standardized using MinMaxScaler, StandardScaler, and RobustScaler, and the training–test split was 75–25%. The model achieved a prediction error of 3.35% during training and 3.48% during validation, demonstrating robustness and efficiency. These results highlight the potential of data-driven models in accurately predicting lithium-ion battery degradation and underscore their relevance for promoting sustainable energy systems through improved battery health forecasting, optimized second-life use, and extended operational lifetimes of storage technologies.