Estimation of State of Charge of Lithium-Ion Battery Based on LSTM Neural Network

Reasons for Recommending Quality Articles

Estimation of State of Charge of Lithium-Ion Battery Based on LSTM Neural Network

MING Tongtong, WANG Kai, TIAN Dongdong, XU Song, TIANG Hao Han

(Qingdao University)

Abstract: To address the prediction problem of the state of charge (SOC) of lithium-ion batteries, a SOC prediction model is constructed using Long Short-Term Memory (LSTM) Recurrent Neural Networks. Using a DC electronic load, various discharge conditions of the 18650 lithium-ion battery are tested, with battery voltage and discharge current as model inputs. The collected data is divided into training, validation, and testing sets; the model is trained on the training set, hyperparameters are adjusted on the validation set, and model performance is tested on the testing set. The Stochastic Gradient Descent (SGD) method with momentum is used for weight updates, and the Dropout regularization method is incorporated. Under dynamic discharge conditions, the proposed method predicts the maximum absolute error of the battery SOC to be 2.0% and the average absolute error to be 1.05%, validating the feasibility of this method. The test results indicate that incorporating the Dropout regularization method during the model training process can effectively reduce overfitting and enhance the model’s generalization ability.

Funding Project: National Natural Science Foundation Project

Author Information:

MING Tongtong (1994), male, from Qingdao, Shandong, a master’s student, majoring in development of new energy electric control systems and estimation of lithium battery state of charge, E-mail:[email protected].

WANG Kai (1985), male, from Binzhou, Shandong, associate professor, master’s supervisor, PhD in engineering, majoring in distributed microgrid and energy storage, storage of new energy, energy internet, etc., E-mail: [email protected].

TIAN Dongdong (1995), male, from Zhoukou, Henan, a master’s student, majoring in development of electric control systems for new energy vehicles, monitoring and control of high-speed train operations, renewable energy integration and smart distribution network technology, lithium-ion battery SOC prediction, E-mail: [email protected].

XU Song (1997), male, from Tai’an, Shandong, a master’s student, majoring in new energy generation technology, E-mail: [email protected].

★MING Tongtong, WANG Kai, TIAN Dongdong, et al. Estimation on State of Charge of Lithium Battery Based on LSTM Neural Network [J]. Guangdong Electric Power, 2020, 30(3):26-33.

MING Tongtong, WANG Kai, TIAN Dongdong, et al. Estimation on State of Charge of Lithium Battery Based on LSTM Neural Network [J]. Guangdong Electric Power, 2020, 30(3):26-33.

Research Background

According to the World Health Organization, 6.5 million people die prematurely each year due to air pollution. In the 2015 statistics on nitrogen oxide pollution, the transportation industry accounted for 50%, emitting a total of 5.3×107 tons of nitrogen oxides. Countries in Europe, such as Norway, are considering banning gasoline and diesel vehicles by 2025. Lithium-ion batteries are widely used in the field of electric vehicles due to their high energy density, and their applications in autonomous vehicles, drones, and smart grids are also continuously increasing.

Existing Problems

The accurate estimation of the battery’s state of charge (SOC) is crucial for the stable operation of power equipment and is essential for monitoring the remaining mileage of vehicles and battery balancing systems. However, the repeated speed changes during the operation of electric vehicles pose challenges for accurately estimating the SOC. SOC refers to the remaining capacity of the battery divided by its rated capacity. Generally, SOC is a nonlinear function of temperature and charge/discharge current rates. The basic methods for measuring battery SOC include open-circuit voltage method, ampere-hour integration method, and internal resistance-based prediction methods. The ampere-hour integration method predicts SOC by measuring battery current and integrating it, requiring high accuracy of the current measurement; the open-circuit voltage method has the disadvantage of requiring the battery to be at rest for a long time before prediction, making it unsuitable for online measurements; the internal resistance-based prediction method is affected by the type, quantity, and consistency of the battery, making it rarely used in electric vehicles. The limitations of these three methods have led to their replacement by more complex algorithms, such as adaptive observers, sliding mode observers, and Kalman filters; however, these algorithms are generally complex in computation and often require additional parameters and different models to adapt to SOC estimation in different scenarios.

In the 1990s, data and hardware limited the development of machine learning. After 2012, machine learning has made tremendous progress, driven by advancements in hardware, data collection, and algorithm improvements. From 1990 to 2010, the speed of non-custom Central Processing Units (CPUs) increased by 5000 times. In 2016, Google showcased its Tensor Processing Units (TPU) project at its annual I/O conference, which was developed specifically to run deep neural networks. The rapid development of the Internet has made it feasible to collect and distribute ultra-large datasets. Besides hardware and data, algorithm improvements have achieved better gradient propagation, leading to rapid development in deep learning.

This Study’s Content

This study proposes a method for estimating the SOC of lithium-ion batteries based on Long Short-Term Memory (LSTM) Recurrent Neural Networks. Using the TR9600 programmable DC electronic load, multi-condition discharge experiments are conducted on selected lithium-ion batteries, collecting high-precision test data. The Keras deep learning framework is used, with TensorFlow as the backend engine, to construct feedforward (back propagation, BP) neural networks, Gated Recurrent Units (GRU), and LSTM neural networks for performance comparison. During the model training process, the data is divided into training, validation, and testing sets; the model is trained on the training set, evaluated on the validation set, and optimal configuration parameters are adjusted, followed by performance testing on the testing set. Stochastic Gradient Descent (SGD) with momentum is chosen as the optimizer to prevent the model from falling into local minima and to increase the model’s convergence speed. The Dropout regularization method is added to the LSTM RNN to enhance the model’s generalization ability.

Methods to Access Articles from Guangdong Electric Power for Free

1. Mobile Reading. Scan the QR code below, no app download required, to access the full text for free.

2. Domestic databases such as CNKI (priority publication), Wanfang Data, Chongqing Weipu can be searched and downloaded.

3. Author QQ group files: 681399883, 478017233

4. You can request the complete electronic version of this issue from the editorial department by sending an email to [email protected] for free.

5. You can request a printed journal from the editorial department by sending an email or leaving a message at the end of this article, informing the mailing address; we will send a printed version of the journal for free.