In the field of Electric Vehicle (EV), what the driver is most concerned with is that whether the value of the battery's capacity is less than the failure threshold because of the degradation. And the failure threshold means instability of the battery, which is of great danger for drives and passengers. So the capacity is an important indicator to monitor the state of health (SOH) of the battery. In laboratory environment, standard performance tests can be carried out to collect a number of related data, which are available for regression prediction in practical application, such as the on-board battery pack.
Firstly, we make use of the NASA battery data set to form the observed data sequence for regression prediction. And a practical method is proposed to determine the minimum embedding dimension and get the recurrence formula, with which a capacity model is built. Afterwards, an optimized Relevance Vector Machine (RVM) algorithm is utilized to improve the prediction performance as well as the operating efficiency and get a tradeoff between training time and computational complexity. Finally, given the conditional probability distributions of hyper-parameters, we can capture the uncertainty in our prediction of the capacity. The calculation results make it clear that the algorithm we develop have better accuracy and performance than others.
