Remaining useful life prediction of lithium ion batteries using an optimised recurrent neural network

Abstract

The remaining useful life (RUL) is one of the important assessment index for battery energy storage system (BESS) technology. Accurate prediction of RUL ensures safe operation, prevent risk failure and unwanted catastrophic occurrence of the battery storage system. However, precise prediction for RUL is challenging due to battery capacity degradation and performance variation under temperature and aging impacts. Existing research on RUL prediction suffers from high computational complexity, requires large volume of training datasets and high training time. Therefore, the main aim of this research is to develop an optimised neural network (NN) models for predicting the RUL of lithium ion batteries. To achieve this, firstly, a multi-charging input (MCI) with 61-dimensional input data features is created. Various NN models namely Back Propagation neural network (BPNN), Fitting neural network (FNN), Feedforward neural network (FFNN), cascaded forward neural network (CFNN) and Recurrent neural network (RNN) is used. These NN models are then optimised with Particle swarm optimization (PSO) technique. The outcomes of PSO optimised NN models are tested initially with NASA battery dataset and validated with another battery dataset from MIT Stanford-Toyota Research Institute. In this study, a multi-charging input (MCI) profile with a 61-dimensional inputs consisting of voltage, current, temperature and discharge capacity is developed by extracting critical samples from various charging parameters with systematic sampling technique. 20 samples of every parameter from each charging cycle is extracted to develop a 61-dimensional input features. The NASA database consists of 4 battery datasets namely B5, B6, B7 and B18 while the battery datasets from MIT Stanford-Toyota Research Institute comprises of c33, c34, c35 and c36 battery datasets. The effectiveness of the proposed PSO-based NN models is analysed under different training ratios such as 70:30 and 50:50. During the training with 70:30 ratios, the RMSE calculated for RNN-PSO model was 0.0019 for B5, 0.0236 for B6, 0.0070 for B7 and 0.0148 for B18 whereas, the RMSE for the 50:50 ratios was 0.0266 for B5, 0.0544 for B6, 0.0424 for B7 and 0.1584 for B18. Therefore, consecutive analysis uses 70:30 training ratio as it was found that training using this ratio results in better predictions. Among the NN models used, it was found that the RNN model achieves highest accuracy. Next, it was examined that RUL prediction of the proposed RNN-PSO intelligent algorithm results in better accuracy compared to other PSO optimised NN models. It is also found that B6, B18 and c36 datasets demonstrated high error due to capacity regeneration phenomena. Validatio n work using the MIT database demonstrates that, for all NN models, the results are in agreement with the initial work using the NASA database. The outcomes of the presented PSO optimised NN-based RUL prediction framework demonstrate accurate RUL prediction results with the least RUL error in both types of battery datasets. In conclusion, the RNN-PSO-based RUL prediction framework has great potential for its implementation in real-time battery monitoring system. Additionally, the developed RUL prediction framework can be deployed towards the development of suitable failure mechanism.