Dynamic conditions-oriented model-data fused framework enabling state of charge and capacity accurate co-estimation of lithium-ion battery

Unpredictable dynamic conditions exhibit severe challenges on state of charge (SOC) and capacity accurate co-estimation for lithium-ion battery. Combining the complex battery dynamics emulated via sophisticated model and the fruitful aging latent information hidden in raw data, this study presents an innovative dynamic conditions-oriented model-data fused framework enabling SOC and capacity accurate co-estimation. Replacing Cholesky decomposition in traditional unscented transform strategy, we firstly design a new singular value decomposition-adaptive unscented particle filter for accurate SOC online estimation during dynamic conditions, which is capable to iterate effectively and continuously even with non-positive definite error covariance matrix. To break through the limitations of aging information extraction only against static conditions, we then develop an exquisite convolutional neural network-bidirectional long short-term memory data-driven framework referring to sequence to point structure, which enables flexible aging-dependent features online extraction from battery domain knowledge under sophisticated dynamic conditions. The experimental verification results demonstrate that the designed model-data fused framework enables accurate SOC and capacity co-estimation under sophisticated dynamic conditions with sufficient generalization ability and stability ability, where both mean absolute error and root mean squared error of SOC estimation are below 0.65 %, and most relative error between real and monitored capacity is roughly controlled within ±0.2 %.