MODELING OF BATTERY SYSTEM OPERATION BASED ON MACHINE LEARNING METHODS

Abstract

Renewable energy integration and electromobility have become fundamental drivers of modern energy storage systems. A key difficalty in this domain is the accurate estimation of internal battery parameters—particularly the state of charge and terminal voltage—due to the nonlinearity of electrochemical processes governing battery operation. This study provides a battery behavior modeling during charging and discharging cycles, which lead to crearing application based on machine learning approach, that provide high level of prediction accuracy and improve battery system control.

To specify limitations of static modeling for  capacity prediction, a dynamic battery model was developed. It based on an equivalent electrical circuit approach. Created model allows real-time estimation of voltage and SOC paramethers by incorporating instantaneous current measurements together with historical operational data. Particular attention is focus on nonlinear characteristics of charge and discharge processes, as this characteristics are unique for each battery system and influence stability and type operating mode of systems. A detailed comparative evaluation of multiple algorithmic approaches is conducted to identify the most effective parameter estimation techniques under varying load conditions.

The results reveal a distinct dependence of algorithmic performance on current magnitude. For stable operating regimes characterized by low to medium current levels (up to 11 A), the least squares method showcased the highest metrics of efficiency, benefiting from the quasi-linear behavior of the approach and providing a high level of prediction accuracy. Conversely, in high-current operating conditions—examined in the range of 19 A to 31 A—or in scenarios involving pronounced nonlinearity and limited data availability, genetic algorithms demonstrate superior performance. Their heuristic optimization capabilities enable robust parameter estimation in complex, non-convex search spaces where conventional deterministic methods are insufficient.

Model approbation was conducted through comparison of simulation outcomes with experimental measurements, confirming high predictive accuracy, achieving a coefficient of determination of R² = 0.96. These outputs approve the relevance of the proposed approach as a way to improve performance for the next-generation battery management systems. Overall, the methodology supports scalable and adaptable approaches for predicting battery operating modes, facilitating more efficient strategies for hybrid power generation systems of different scales.