State-of-Charge Estimation of Medium- and High-Voltage Batteries Using LSTM Neural Networks Optimized with Genetic...

Area of Science:

• Battery Technology
• Machine Learning
• Electric Vehicle Systems

Background:

• Accurate state-of-charge (SOC) estimation is crucial for lithium-ion battery management systems (BMS).
• Standalone methods like Coulomb Counting (CC) and LSTM neural networks have limitations in dynamic conditions.
• Optimizing estimation algorithms is key for reliable performance in electric vehicles (EVs) and second-life applications.

Purpose of the Study:

• To develop a hybrid SOC estimation method integrating LSTM and CC for improved accuracy.
• To optimize the hybrid model using genetic algorithms (GA) for enhanced real-time performance.
• To validate the proposed method under standardized driving cycles (NEDC, WLTP) for medium-voltage battery packs.

Main Methods:

• A hybrid approach combining LSTM neural networks with Coulomb Counting (CC) for SOC estimation.
• Optimization of LSTM using genetic algorithms (GA).
• Dynamic fusion of LSTM predictions with CC estimates via a parameter α for recalibration.

Main Results:

• The hybrid model achieved a low Mean Absolute Error (MAE) of 0.181%.
• Superior performance compared to conventional standalone SOC estimation strategies.
• Enhanced prediction accuracy under dynamic driving conditions (NEDC, WLTP).

Conclusions:

• The hybrid LSTM-GA-CC method offers a robust solution for accurate SOC estimation.
• This approach effectively addresses the limitations of individual estimation techniques.
• The findings support the development of more reliable BMS for EVs and second-life battery applications.