Forecasting the health status of energy storage systems using machine learning based on the identification of useful features of time series.

Lithium-ion batteries are a key component of modern energy storage systems and electric transportation. Accurate prediction of their state of health is critically important for ensuring reliability, safety, and optimizing lifespan. This article presents an improved approach to state of health prediction based on machine learning. The implemented approach demonstrates high accuracy on synthetic data that mimics real degradation processes, with a coefficient of determination R² " 0.95. An analysis of feature importance was conducted, identifying key parameters influencing degradation. The proposed model serves as an effective tool for non-invasive diagnostics and predicting the remaining life of battery cells. This work proposes a method for predicting SOH that combines the identification of useful features from multidimensional time sequences of operating parameters with static characteristics of the battery cell. The regression model used is the random forest algorithm, known for its resistance to overfitting and ability to assess feature importance. The methodology includes extracting informative features from multidimensional time series of operational parameters (voltage, current, temperature, internal resistance, capacity) combined with static characteristics of the cell. An ensemble algorithm, random forest, is used to build the predictive model. The implemented approach demonstrates high accuracy on synthetic data that mimics real degradation processes, with a coefficient of determination R² " 0.95. An analysis of feature importance identified key parameters affecting degradation. The proposed model is an effective tool for non-invasive diagnostics and predicting the remaining life of battery cells. The comparison chart of predicted characteristics and real data shows a tight clustering of points around the bisector, visually confirming the high accuracy of the model. The histogram of error distribution resembles a normal distribution with a zero mean, indicating the unbiased nature of the predictions. The proposed method is non-invasive and could potentially be implemented in battery quality control systems for real-time assessment of battery condition.