PNP-Lightweight Model for Predicting Remaining Useful Life of Lithium-Ion Battery to Applying Embedded Systems

This paper proposes a Positive & Negative Perceptron (PNP)-Lightweight model prediction method for the Remaining Useful Life (RUL) prediction of Lithium-Ion Battery (LIB) for use in low computation environments such as smartphones and embedded systems. The PNP-Lightweight model is based on the PNP model and has a very low number of trainable parameters. We present a PNP-Lightweight model that can predict the RUL of the LIB using the charge and discharge data of the initial LIB. To prove the superiority of the proposed PNP- Lightweight model, we perform the performance evaluation compared with Long short-term memory (LSTM) and Convolution Neural Network (CNN) models. We build Deep-learning Model using the PNP-Lightweight model (DMP), Deep-learning Model using the LSTM model (DML) and Deep-learning Model using the CNN model (DMC) that can predict the RUL of LIB using the LIB's initial charge and discharge data. To evaluate fairly, we organize the same number of layers and the number of nodes in the hidden layer of the three deep-learning models (DMP, DML, and, DMC) As a result of the performance evaluation, we know that the number of trainable parameters of the DMP is 30 times lower than that of DML and 60 times lower than the DMC. We confirm that the DMP has a lower Root Mean Square Error (RMSE) by 16.18 [cycle] and 3.03 [cycle] compared to the DML and DMC, respectively. In addition, the Symmetric Mean Absolute Percentage Error (SMAPE) of the DMP is 3.6% and 3.96% lower than the DML and DMC, respectively. In addition, we also compare the performance between proposed DMP and CNN+LSTM model. We recognize that performance of proposed DMP is superior to CNN+LSTM model.

I. Introduction

Lithium-Ion Battery(LIB) is widely used in the field of electric vehicles, renewable energy source systems, and smart grids [1]. Specially, LIB have been widely applied to mobile devices such as mobile phone, laptops, and electric vehicles [2]–[7]. Such an operation field, LIB fault can lead to poor performance, increase maintenance costs, and fatal device failures of equipment, using LIBs[8]. The battery management system (BMS) is responsible for monitoring, balancing, and controlling the LIB to suppress the failure of the LIB [9], [10]. In general, when the capacity of the LIB decreases to 80% of the initial value, it is cost-effective to replace the LIB used in the electric vehicle. For this reason, it is necessary for the BMS to accurately and effectively predict the Remaining Useful Life(RUL) of the LIB and provide a way for the user to optimize LIB usage. In the field of RUL prediction of LIB, the data-driven model has become popular in recent years as an alternative method to overcome shortcomings of the physical-based model. In particular, many researchers have focused on the research to predict the RUL of LIB using LIB's initial operational data with deep-learning. [11]–[17]. The data-driven model requires a huge amount of training data, and most of it supports cloud-based services. With the recent combination of hardware, computing power improvements, and cloud computing systems, sophisticated data-driven methods are expected to be applied in practice. However, due to limited bandwidth and computational resources, centralized clouds are inefficient in processing and analyzing vast amounts of data [18], [19]. In recent years, we have seen a new trend towards deep-learning network processing in embedded systems such as mobile and the Internet of Things (IoT) [20]. High-specification embedded systems can support data analysis. However, in a mobile environment, most embedded systems are organized as low-specification. Thus, these embedded systems cannot handle the high computation requirements for deep-learning tasks [21]. From this point of view, the BMS of LIB needs to be studied to handle deep-learning networks in embedded systems. As the number of trainable parameters are low, the computation is also low. Thus, we can use this feature to apply the deep-learning model to the embedded system [22]. Therefore, we present a Positive & Negative Perceptron (PNP)- Lightweight model in which the number of trainable parameters is much lower than other deep-learning models. To prove the superiority of the proposed PNP- Lightweight model, we perform the performance evaluation compared with Long Short-Term Memory(LSTM) and Convolution Neural Network (CNN) models. To verify the performance of the proposed Deep-learning Model using the PNP-Lightweight model (DMP), we compare the prediction performance of RUL in LIB with the Deep-learning Model using LSTM model (DML) and Deep-learning Model using CNN model (DMC) models. We also compare the number of trainable parameters for three models.