Artificial Neural Network based State of Charge Estimation for Lithium-ion Batteries

Abstract

Lithium-ion batteries dominate the battery industry due to their high energy density, long life cycle, zero maintenance, minimal self-discharge rate, and versatility. These attributes have significantly advanced the green revolution, particularly through their use in Electric Vehicles (EVs) and Battery Energy Storage Systems (BESS). However, lithium-ion batteries are prone to voltage fluctuation and tend to overheat during charging. Lithium-ion batteries are also highly flammable and can undergo thermal runaway as they age. To oversee the overall operation of a lithium-ion battery during its operational life cycle, a Battery Management System (BMS) is used. The cost and complexity of BMS are dependent on its scope of applications, available features and don’t have a unique set of criteria. The BMS monitors parameters like cell voltage, current, and module temperatures in-order to estimate the batteries’ State of Charge (SOC) and State of Health (SOH). These parameters are of prime importance in the modern EVs to gauge the battery health and the range of EV. Since there is no direct measurement of SOC and SOH, they are estimated using mathematical techniques like coulomb counting, open circuit voltage, impedance measurement, etc. These techniques will require precise knowledge of battery composition. Modern EVs use lookup table and current counting method to estimate the SOC and ultimately the range of the vehicle. In certain contexts, these estimations demonstrate satisfactory performance, yet they are susceptible to errors and exhibit noise. However, the global demand for lithium-ion batteries is ever increasing and thousands of data points are measured all around the world by battery manufacturers, EV companies and most importantly, by users. These data points can be collected and made into a global battery database which can then be used to train Machine Learning (ML) and Artificial intelligence (AI) models. These models then can be used to estimate the SOC and SOH and other battery parameters that will help the BMS to make smart decisions during operation. This thesis explores the application of deep learning algorithms in the design of Battery Management Systems (BMS) integrated with artificial intelligence (AI). Two ANNs, Deep Neural Network (DNN) and Long Short Term Memory (LSTM) Network, estimated SOC of an LG 18650 HG2 cell at four temperatures: -10 ◦C, 0 ◦C, 10◦C, and 25◦C. Initially, DNN trained with voltage, current, and temperature as inputs yielded unsatisfactory results. Incorporating average voltage and average current along with those inputs, derived from 500 previous time-steps, significantly enhanced DNN performance, achieving a mean absolute error of (MAE) of approximately 0.99%. Further increasing neuron count and training epochs slightly improved performance to 0.97% MAE. LSTM, a type of recurrent neural network, was then utilized which is capable of bidirectional information flow and memory retention. Despite training for less number of epochs (200 epochs) compared to DNN (1000), LSTM exhibited competitive performance with an average MAE of 1.08%.