Particle swarm optimized data-driven model for remaining useful life prediction of lithium-ion batteries by systematic sampling

Abstract

The remaining useful life (RUL) is considered an important health indicator in lithium-ion batteries for evaluating various features such as efficiency, robustness, and accuracy. The RUL investigates the battery reliability for determining the advent of failure and further mitigating battery risk. The effective RUL prediction of a lithium-ion battery can ensure safe operation, avoid internal, external failures, and unwanted catastrophic occurrences. However, the accomplishment of accurate RUL prediction is difficult due to the occurrence of capacity degradation and performance deviation with temperature and aging impacts. Hence, this paper delivers an improved hybrid data-driven model comprising recurrent neural network (RNN) and particle swarm optimization (PSO). A systematic sampling technique is employed to construct a 31-dimensional multi-channel input data framework. Various parameters from NASA battery datasets such as charging profile voltage, temperature, current, and discharge capacity are selected as suitable health indicators to generate a 31-dimensional input framework for RNN-PSO model training. Furthermore, the validation of the presented framework is carried out with other optimized data-driven models and MIT Stanford battery datasets. Compared with other optimized data-driven models, the experimentation conducted with NASA and MIT Stanford battery datasets demonstrates that RNN-PSO delivers higher prediction accuracy with mean square error at 3.7719 × 10−8 for battery B5 and 2.9759 × 10−8 for c33. The results prove the effectiveness of the proposed RNN-PSO model for the RUL prediction of lithium-ion batteries on different battery datasets.