A Transformer-CNN-BiLSTM Architecture for SOH Prediction of Battery Storage Systems

Accurate prediction of the State of Health (SOH) of batteries is essential for the reliability and efficiency of energy storage systems, particularly in electric vehicles and renewable energy applications. This paper presents a novel battery SOH prediction algorithm that combines Kalman filtering with a Transformer-CNN-BiLSTM architecture. The introduction of Kalman filtering at the data preprocessing stage effectively reduces sensor noise, resulting in cleaner input data for the model. The Transformer component captures long-term dependencies through its self-attention mechanism, while the Convolutional Neural Network (CNN) extracts local features by learning shortterm variations in the battery data. The Bidirectional Long Short-Term Memory (BiLSTM) processes temporal dependencies bidirectionally, enabling more accurate forecasting of future SOH values. Our proposed hybrid model synergizes these components, achieving multi-level feature extraction and robust noise suppression, which significantly improves prediction accuracy. Extensive experiments on real-world battery datasets demonstrate that the hybrid model outperforms traditional machine learning techniques and standalone deep learning approaches, showcasing its generalization capability across different battery chemistries and operational conditions. Our approach provides a promising solution for real-time battery health monitoring and management.