State of power prediction considering cell inconsistency for a series-parallel battery pack based on adaptive SRUKF and double-neural network

Accurate prediction of the state of power (SOP) for lithium-ion batteries is critical in electric vehicles, where state of charge (SOC) serves as a key constraint. Increasing battery cell count and inherent inconsistencies raise predicted SOP deviation despite accurate SOC. In this study, a SOP prediction method for a series-parallel battery pack based on adaptive square root unscented Kalman filter (ASRUKF) and double neural network is developed. First, cell inconsistency is quantified using weighted cosine similarity, and the cell with the largest coefficient in each branch is selected to establish a pack mean model. Second, since noise interference such as battery measurement noises can cause instability in SOP prediction, an ASRUKF with variable forgetting factor is developed to improve the system's noise resistance. Finally, to describe the influence of cell inconsistency on the SOP deviation and capture the power characteristics in different stages, a double-neural network model containing radial basis function and gated recurrent unit is designed. Moreover, an enhanced beluga whale optimization algorithm is presented to tune the hyperparameters for the network model. The results show the developed SOP prediction method has a maximum error below 0.32 W across time scales, while simultaneously reducing the runtime cost by at least 32.2 %.