Abstract:

Remaining useful life （RUL） prediction of lithium-ion power battery can estimate the future state of batteries， which can guide battery maintenance and reduce the risk of failure. The battery cycle conditions are not controlled in real-world vehicle conditions， and RUL prediction under dynamic operating conditions still suffers from difficulties in processing messy data， poor accuracy of prediction results and inability to take aging uncertainty into account. For this case， the Attention Mechanism Sequence to Sequence-Particle Filter （Aseq2seq-PF） hybrid model is proposed， where the common State of Charge （SOC） charging interval is selected to obtain the normalized battery capacity and a fusion prediction strategy of Iteration and Direct is adopted， with the Aseq2seq model as the Iteration part to achieve accurate prediction of capacity sequences， the PF model as the Direct part to achieve uncertainty prediction of capacity fluctuations， and RUL is predicted by extrapolating the trend of battery capacity degradation. Verified by the real-world vehicle power battery data， the public SOC charging interval effectively obtains a clear trend of capacity degradation. The hybrid model improves the long-term prediction accuracy of the capacity degradation with good robustness， with reduction of average absolute error by more than 56% compared with existing models， and outputs confidence intervals to meet the needs of different application to achieve aging uncertainty description.

Key words: lithium-ion power battery, remaining useful life （RUL）, real-world vehicle data, sequence to sequence, aging uncertainty