Prediction of Li-Ion Battery State of Charge Using Multilayer Perceptron and Long Short-Term Memory Models

摘要

Lithium-ion batteries are used in different applications such as electric vehicles and grid-scale energy storage. These applications rely greatly on the accurate measurement and prediction of state of charge (SOC) to ascertain the battery's available capacity. Although multiple methods exist in the literature to predict SOC and other battery parameters, they have low accuracy, make offline predictions, and do not consider enough battery parameters. The battery's nonlinear characteristics and time-variance have direct impacts on the applications connected to it, which make the prediction a complex but more necessary problem to solve. This paper bridges the gap by comparing the performance of two widely used data-driven learning models: long short-term memory (LSTM) and a multilayer perceptron (MLP), for predicting SOC using predictors such as cell current, cell voltage, elapsed time, and cell temperature. The models are run using mean squared error as the loss function, and different loss function optimizers. The models are also applied to datasets from different charging/discharging rates to demonstrate their efficacy. A recommendation is finally made on the model for SOC prediction subject to specific conditions considered in the paper, and the groups of optimizers that work better to minimize the loss function.