Predicció de degradació en sistemes d’emmagatzematge electroquímic d’energia mitjançant intel·ligència artificial

Abstract

This work develops and evaluates a machine learning-based methodology to predict the State of Health (SOH) of lithium-ion batteries using the NASA Battery Dataset. The study is based on experimental discharge cycles, from which a tabular dataset is constructed, the SOH is defined as the target variable, and a conservative data quality review is applied. The methodology follows a progressive structure: first, a model is built within a base domain composed of batteries tested under relatively homogeneous experimental conditions; then, the analysis is extended to domains with different temperatures, currents, and discharge profiles. The results show that, within the base domain, Linear Regression achieves the best overall performance, with a mean R² of 0.9629. However, when more heterogeneous domains are incorporated, Gradient Boosting provides better relative performance, especially after feature enrichment and hyperparameter optimization. Nevertheless, the final model still shows generalization limitations, particularly in the low-temperature domain and in the external test performed on a discharge profile not included during training. Overall, this work shows that SOH prediction does not depend only on the algorithm used, but also on the representativeness of the dataset, the experimental domain, and the validation strategy applied