Lightweight State of Charge Estimation of Lithium-Ion Battery Based on Pruned Neural Networks

Neural networks have been widely adopted for state of charge (SOC) estimation due to their high-efficiency non-linear mapping, adaptive and self-learning abilities. However, their application in embedded devices is restricted by significant input feature fluctuations, time-consuming hyperparameter optimization, and large model sizes. In this paper, a framework based on sequentially-connected dual-Savitzky-Golay filters, Bayesian optimization and hyperband optimization and neural network prune method is proposed for SOC estimation, sequentially-connected dual-SG filters are used to smooth the measured signals and further reduce the fluctuations of the estimated SOCs, Bayesian optimization and hyperband is adopted for low-time consumption hyperparameter optimization and neural network prune method is employed to achieve high-accuracy SOC estimation with lightweight frameworks. The adoption of the sequentially-connected dual-SG filters can significantly improve the accuracy without obvious sacrifice on time consumptions. The Bayesian optimization and hyperband could realize consistent accuracy with lower time consumption (time reduction more than 22.54% in this paper) than the Bayesian hyperparameter optimization method. The pruned convolutional neural network-long short-term memory model could realize consistent accuracies with a 57.34% smaller model size by sacrificing 15.26% more time for training per epoch than the baseline framework when sparsity is 0.7.