State of Charge Prediction of EV Li-ion Batteries with Machine Learning: A Comparative Analysis
Authors: 
Afroditi Fouka, Alexandros Bousdekis, Katerina Lepenioti, Gregoris MentzasAuthors Info & Claims
Article No.: 20, Pages 1 - 7
Abstract
Li-ion batteries are an important component of the Battery Management System (BMS) in Electric Vehicles (EVs) and their performance is crucial for the efficient function of EVs. State of Charge (SoC) prediction is a significant task since this state of the battery is indicative for its performance. Data-driven methods for SoC prediction are gaining increasing interest in both academia and industry and have been driven by recent advances in Machine Learning (ML) algorithms and methods. In this paper, we perform a comparative analysis of ML pipelines and methods for SoC prediction based on a structured methodology. To do this, we build the general pipeline for SoC prediction, including data preprocessing, model development and evaluation, while we also define and configure the critical dimensions of this pipeline. We use two different datasets, and we evaluate and compare the various experimental approaches that we have applied. In other words, we assembled several experiments with different attributes and configurable dimensions, enabling the evaluation of the experiments based on these dimensions. This way, we were able to compare the performance of each approach in terms of different aspects.
1 Introduction
Over the past decade, industry and academia have been focusing on the development of Electric Vehicles (EVs) in order to achieve a high degree of efficiency, performance and decarbonization. In EVs, a crucial component is the Lithium-Ion (Li-ion) batteries which should provide safety and dependability during their performance [1]. Moreover, as the cost of the battery stands still high, its accurate states estimation is of outmost importance in order to enhance the efficiency and the overall performance of the vehicle [2].
As one of the key components of EVs, the Li-ion Battery Management System (BMS) is crucial to the industrialization and marketization of electric vehicles [3]. The fundamental functionalities of the BMS are: data collection, different battery's states estimation, balance control, thermal management, and communication [4]. The SoC prediction is one of the key roles of BMS since it is an indicator of how long the battery can endure without being recharged [1].
SoC can be defined as the ratio of the remaining capacity in the battery to its maximum availability. Its value can range from 0 to 100, without reaching the values on the limits [1]. Reliable and accurate SoC prediction is of outmost importance since it contributes to cell equalization stability, elimination of charging and discharging issues and BMS charging and discharging control [5].
There are various methods for SoC estimation and prediction, such as: direct methods (e.g. Coulomb Counting and Open Circuit Voltage (OCV) method), empirical models, equivalent circuit models, filter-based methods (e.g. Kalman Filter). In the literature, there is a trend towards data-driven methods exploiting the large amount of the available data, which take advantage of Machine Learning (ML) in order to achieve predictions of higher accuracy and reliability [2][6], but also flexibility due to their model-free nature [7]. In this paper, we perform a comparative analysis of ML pipelines and methods for SoC prediction based on a structured methodology. To do this, we build the general pipeline for SoC prediction, including data preprocessing, model development and evaluation, while we also define and configure the critical dimensions of this pipeline. We use two different datasets, and we evaluate and compare the various experimental approaches that we have applied.
The rest of the paper is organized as follows: Section 2 presents the related works on SoC prediction based on data-driven methods and ML. Section 3 describes the proposed methodology for comparative analysis. Section 4 presents the implementation of the proposed methodology and the comparative results. Section 5 concludes the paper and presents our plans for future work.
2 Related Work
In this Section, we review related works on the use of ML algorithms and data analytics methods for SoC prediction of EV Li-ion batteries. During the last years, the related literature has been increasing, thus leading to an extensive use of ML algorithms for SoC prediction.
In [8], the authors apply a Feedforward Neural Network (FFNN) to estimate SoC using as input the variables of voltage, current and temperature. The model was trained on five datasets with different automotive behaviors and tested in another one, with the goal to reduce the estimation error. In [9], the authors propose an approach for SoC prediction, also using a FFNN network. In this case, the input data include the internal resistance along with voltage, current, and temperature in order to increase the estimation accuracy.
Another algorithm which is applied for SoC estimation is the Support Vector Machine (SVM), due to its suitability on multiple learning tasks [10]. For example, in [11], the authors propose a model based on the SVM algorithm for reducing the dependence of SoC prediction on battery model accuracy and speed. The problem which is analyzed in this paper is formulated as a regression problem, allowing Support Vector Regression (SVR) to deal with its non-linearity [11].
Several research works have used a Deep Neural Network (DNN) because it presents a strong ability to learn any function given the suitable data to the application [6]. For example, in [12], the authors utilize a DNN in order to predict the SoC for a variety of driving cycles. Based on their results, the author concludes that the increasing number of hidden layers minimizes the error rate of SoC estimation. In [13], the authors develop a DNN using drive cycle datasets at various temperatures. The proposed model is capable of being generalized to various variable dynamics by adjusting the time and temperature dependencies into the network, thus providing more accurate SoC predictions comparing to other approaches.
In [14], the authors develop a Long-Short Term Memory (LSTM) neural network which is trained on the nonlinear relationship between SoC and OCV curve. This approach was validated with two datasets with different driving characteristics, such as temperature and battery material. In [15], the authors combine a LSTM model with a slow time-varying information sliding window referring to the average voltage. In this way, they increase the ability of the model to correlate the nonlinear characteristics of the battery to the SoC estimation. In [16], the authors combine a LSTM model with a Transfer Learning approach for providing robust and suitable for real time SoC predictions. The results of this study showed that the driving conditions that were examined, affected the accuracy of the model.
In [17], the authors combined a Convolutional Neural Network (CNN) and a LSTM model for predicting SoC, using variables such as voltage, current and temperature. The data used in this study for training and testing the model were derived from an experimental procedure which considered different driving cycles making the model's accuracy more robust comparing to other approaches. In [18], the developed CNN model demonstrates high accuracy at variant temperatures and driving conditions without the need of training the model again. This indicates its ability to adapt to many different variables and be generalized to many types of battery applications, sustaining high accuracy.
3 Research Methodology
SoC estimation and prediction is critical to the safe and reliable operation of Li-ion battery packs since this quantity reflects a vehicle's remaining driving range [19] as well as the remaining energy inside a battery during operation [20]. In this Section, we present the methodology that we designed for the comparative analysis of SoC prediction on EV Li-ion batteries. For this purpose, we developed several experiments with differently configured experimental dimensions, following a common experimental pipeline. These are described in detail in the subsequent sections.
Figure 1:
The experimental pipeline that we designed and implemented for performing the comparative analysis consists of three components as depicted in Error! Reference source not found.: (i) Experimental Dimensions Configuration; (ii) Defined Experimental Process; and (iii) Comparative Analysis.
The Experimental Dimensions Configuration describes the definition of the specific steps, parameters, processes and configuration included in each experiment. As a result, through this component, we can create different experiments with a unique combination of these dimensions.
Figure 2:
The Defined Experimental Process component consists of the seven steps that are executed employing the defined experimental dimensions in each iteration. The Data Ingestion and Data Harmonization steps refer to the collection of the data from different data sources along with their organization and unification in a suitable way for the analysis. The Data Pre-processing step refers to all the required data cleaning, handling, structuring and validating tasks that prepare the data and facilitate the following analysis [3]. This step, also, contains the feature engineering process which can provide additional or more concrete information to the analysis where new features are extracted from raw or preprocessed data [3]. The next step is Model Building which forms the basis for instrumenting and fine-tuning the algorithms that are developed. According to the former dimensions’ definition, the appropriate algorithms along with their configuration parameters are retrieved and be prepared for further processing. After selecting and building the analytics model, the Model Training and Model Testing steps follow. Model Training refers to learning the parameters of the selected algorithm by employing a subset of the selected data, i.e., the training dataset, while Model Testing refers to predicting the SoC on a different subset of selected data, i.e., the testing dataset. Finally, the Model Evaluation step compares the predicted values with these that we have considered as ground truth values for SoC. In this step, we extract some evaluation metrics that contribute to the next step of the pipeline.
The Comparative Analysis considers all the different experimental dimensions for SoC predictions. Each experiment performed follows the experimental pipeline presented in Error! Reference source not found. It is also defined by the configuration of experimental dimensions which characterize it, as depicted in Error! Reference source not found. In other words, in order to compare the different experiments, we group according to their dimensions and look for their performance with respect to the parameters that differentiate them. These parameters are the dataset's category, the SoC estimation method, the training and testing set, the features, and the ML model. All of them are considered for the comparative analysis and are explained in detail in the subsequent sub-sections.
3.1 Dataset's category
The datasets that have been used for this analysis are open datasets and they have been created though cycling tests in a controlled environment, i.e. a battery test lab. In addition, the datasets that we selected to use are related to two main categories of datasets regarding the Li-ion batteries in EVs. The first one involves cycle ageing data and describes the whole lifecycle of the battery. In this case, the data represent the battery's condition from manufacturing until its end of life, which is defined as the state that the capacity has reached the 80% of its nominal capacity. The second category involves driving cycle data that illustrate only one cycle from their lifecycle and specifically, the discharge phase of a battery's cycle.
After examining various open datasets that belonged to the above categories, we selected two of them for our experiments. The first one, Dataset1, that belongs to the driving cycle category is the Madison dataset of McMaster University [21], which is composed from 10 different drive cycles in three different temperatures. These drive cycles have been created from different charging profiles and different features such as mean, RMS and peak power values for each temperature that can effectively simulate different vehicle driving behavior. These datasets describe one random cycle of the battery and do not represent its complete life. Because of that, there is no knowledge of the battery's previous condition and thus, the models can be more flexible and adjustable to different conditions, while their performance can be more efficient. Therefore, each drive cycle contains different variables recorded during one discharge phase of the cell at one random cycle. For the creation of this dataset Panasonic 18650PF cells with NCA chemistry and nominal capacity 2.9 Ah have been used.
The second one, Dataset2, belongs to cycle ageing category and was created by Toyota Research Institute (TRI) in partnership with MIT [22]. It consists of 124 datasets describing the whole lifecycle of each cell. Therefore, in this dataset, we can observe the degradation of the battery's maximum capacity through its lifetime and take it into consideration in our pipeline. These datasets have been created from different charging profiles enabling research to examine the influence of charging protocol to battery's cycle life. More specifically, in this dataset, the cycle lives range from 150 to 2300 cycles and includes 72 different fast charging protocols. The ambient temperature was controlled by the experimental chamber to be stable at 30oC. This dataset has been generated by Panasonic 18650PF cells with Lithium Iron Phosphate (LPF) chemistry and nominal capacity 1.1 Ah.
3.2 SoC estimation method
By definition, SoC refers to the usable capacity of the battery in its current state in relation to the capacity in its fully charged state [23]. In general, SoC is a non-linear feature, and its progress depends on temperature and current [19]. There are no direct ways to measure SoC but there are many ways to obtain an estimation for this information. Two traditional ways are: Coulomb counting and OCV techniques; however, these methods are known for their limitations which affect their accuracy and reliability [19]. For this reason, more robust and sophisticated methods are preferred to handle sensor errors and uncertain model knowledge [2].
We selected three different methods to calculate SoC in order to estimate the SoC label that characterizes the input timeseries. Initially, we apply the approach that is indicated by Equation 1, which describes the SoC as the percentage of the battery's current residual capacity compared to the nominal capacity according to the battery's manufacturer [24]. Secondly, we apply the approach indicated by Equation 2, in which the SoC is calculated as the percentage of the battery's current capacity compared to the maximum battery's capacity in each cycle representing the fully charged state of the battery in each cycle [23]. Finally, we apply an approach based on Coulomb Counting method, which is connected to the SoC's previous state, the measurement of the current, the integration of that current over time and the battery's capacity. This method, represented by Equation 3, is common for SoC estimation in batteries, even though it is known for its limitations [19].
\begin{equation}SoC = \ \frac{{{C}_{res}}}{{{C}_{nom}}}*\ 100\% \end{equation}
(1)
\begin{equation}SoC = \ \frac{{{C}_{curr}}}{{{C}_{full}}}*\ 100\% \end{equation}
(2)
\begin{equation}SoC = SoC\left( {t - 1} \right) + \ \frac{{I\left( t \right)}}{{{Q}_n}}*\ \Delta t \end{equation}
(3)
3.3 Training and Testing set
The dimension of training and testing set is necessary for training and testing the model. When the dataset refers to one cycle, a slice of it is assigned to the training set and the rest of it to the testing one. It is important to note that the training set, in our analysis, is always bigger than the testing set. The same approach can be applied to a dataset that describes the whole lifecycle of the battery and thus, several cycles in the beginning of the lifetime are considered as the training set, while the rest are considered as the testing set. In addition, another choice for the training set is using several randomly selected consecutive cycles of the dataset where, in this case, the testing set refers to the next cycle. It is also valuable to use some randomly selected and scattered cycles of the dataset for both the training and the testing set.
3.4 Features
For the features dimension configuration, several features’ combinations have been considered. Some features point out to raw data which have aroused from the data ingestion and data harmonization steps of the experimental pipeline and they refer to data that can be recorded as timeseries through a system of sensors connected to the batteries. Additional features have been derived by applying diverse statistics based on the data pre-processing step, which follows the data ingestion and harmonization steps of the pipeline. In a similar way, resampled data of the raw data with a specific frequency were considered.
3.5 ML model
In this dimension, we specify the ML model for each experiment, among the candidate models to be compared, i.e. LSTM, FFNN, Linear Regression (LR), SVR, while we also examine how Automated Machine Learning (AutoML) can be employed to support the model development process. Below, we briefly explain these ML algorithms.
One of the most common sequential modelling techniques is the Recurrent Neural Network (RNN). The main advantage of RNNs is that they do not have a finite, discrete set of states to represent the system, but they are able to capture complex non-linear relationships [25]. However, RNNs usually underperform when learning long-term dependencies of time series due to the vanishing and exploding gradient problems. LSTM, a deep learning extension of RNN, has been designed to prevent these problems and is used when time dependency exists in the input features. A LSTM network can represent nonlinear dynamic systems by mapping input to output sequences [19]. Furthermore, more advanced LSTM architectures can deal with complex conditions and degradation behaviors [26]. In LSTM, the same hyperparameters as CNN are designed [27].
From the neural network family, FFNNs are also examined. FFNN is one of the simplest types of artificial neural networks where the information moves in only one direction. Specifically, the information goes from the input layer, through the hidden layer, to the output layer, a process that corresponds to the model's architecture. Therefore, this model has no cycles or loops in the network. Despite its simplicity, FFNN is capable of modelling complex relations in the data and forms the basis for other more complicated neural networks [27].
LR can handle one dimension and multiple-dimension input. The model can be described by Equation 3 where y, x, m and c are the output, the input variables, the vector of fitting parameters according to the gradients and the fitting parameter for the shift, respectively [23]. Linear models also have low computational cost, since they can be trained offline and online [28].
SVM is a supervised and non-parametric ML algorithm based on kernels. Although it is mainly used in classification problems, SVM is capable of solving regression problems, also known as SVR. Moreover, it performs non-linear prediction problems which belong to a low-dimensional space by transforming them into linear ones in high-dimensional space. The purpose of SVR is to produce an error function in which the maximum deviation of target values during training would be under a threshold and simultaneously maintain the largest possible smoothness of the function [27].
Additionally, to the aforementioned models, in the comparative study we examined the contribution of AutoML for facilitating the development of the appropriate models. AutoML is the method of automating the entire pipeline of ML models and procedures that are going to be applied in problem solving. In general, AutoML uptakes many stages of data analysis and model development resulting in generating a ML model ready for deployment. One of the main benefits of AutoML is that it reduces the demand for time consuming and extensive data analyses; instead, it enables the automated building of ML applications [29]. The functionality of AutoML makes easier for ML experts to focus on the application details and particularities, to achieve high performance and accuracy of their approach as well as to reduce substantial resources and time.
4 Implementation and Results
In this Section, we present the experiments that were developed by following the pipeline that was described in Section 3, while examining different experimental dimensions configurations. In this way, we assembled several experiments with different attributes and configurable dimensions, enabling the evaluation of the experiments based on these dimensions. This way, we were able to compare the performance of each approach in terms of different aspects and thus conclude a thorough comparative analysis on the use of ML models for SoC prediction. Finally, for the evaluation of each method, we use the Root Mean Squared Error (RMSE).
4.1 Experimental approaches with LSTM model
The first set of experiments that were implemented focus on the application of several LSTM models. In all the approaches that were designed with LSTM, the data were normalized to the range [0,1] and the problem was turned into a supervised one. These procedures were adjusted to the analysis pipeline, in the Model Building step. Specifically, in the first approach with the LSTM model, we used datasets that describe only the discharge phase of one cycle, Dataset1 [21], and specifically 8 and 2 datasets correspondingly for the training and testing procedure. The input data that were selected to feed the model referred to raw data and described the battery's voltage (V), current (I) and temperature (T). An additional variable was given to the model, representing the ground truth value of SoC, for validation purposes, which was calculated based on the Equation 1.
In the second approach, we used 5 and 2 datasets that describe the whole lifecycle of the battery for the training and testing procedure respectively Dataset2 [22]. Moreover, the input data selected were both raw data and statistics and referred to the battery's voltage and temperature, as well as the statistics of mean voltage (Vmean) and mean current (Imean). In this setup, the ground truth value of SoC was calculated based on Equation 1.
In the last approach employing an LSTM model, 5 consecutive and randomly selected cycles from the entire lifecycle of the battery Dataset2 [22] were selected as the training set, while the next cycle of this set was defined as the testing one. As input features in this approach, we determined raw data and specifically, battery's voltage, current, capacity (C) and temperature. Finally, the SoC estimation method used was the one corresponding to Equation 2.
4.2 Experimental approaches with FFNN model
The second set of experiments was constructed by setting the dimension of ML model for the FFNNs. In the first approach, we used data that describe only the discharge phase of one cycle, Dataset1 [21], while the training and testing set was a percentage of the selected dataset. The input features were raw data, i.e. the temperature and resampled data on voltage and current in frequencies of 0.5 mHz and 5 mHz. Finally, the SoC estimation method was defined by Equation 3.
An additional analysis was determined based on FFNN model, using datasets that describe the whole lifecycle of the battery, Dataset2 [22]. We used 5 consecutive and randomly selected cycles as the training set and the next cycle as the testing one. The features that were assigned to this analysis were raw data and extracted statistics out of them, i.e. the battery's voltage and temperature, as well as the mean voltage and mean current. The ground truth value of SoC was calculated by Equation 2.
4.3 Experimental approaches with LR and SVR model
The next set of experiments was designed focused on the LR and SVR models while keeping the others dimensions constant, due to the similarities of these models’ functionality and structure. In the first approach, we used data that described only the discharge phase of one cycle, Dataset1 [21], while the training and testing setw were a percentage of the selected dataset. The features selected for this approach were raw data, i.e. the battery's voltage, current and temperature, while the SoC estimation method was based on Equation 1. Furthermore, as input variables, we defined raw data such as the battery's voltage, current, temperature and capacity, while the ground truth value of SoC was calculated based on Equation 2.
4.4 Experimental approaches with autoML method
The last set of experimental procedures referred to approaches that were implemented with AutoML. In the first approach, we used datasets that describe the whole lifecycle of the battery, Dataset2 [22], and specifically, we used only 9 non-consecutive and randomly selected cycles of the whole lifecycle of the battery for the training process as well as for the testing set. Moreover, the features were indicated by both raw data and statistics referring to battery's voltage, temperature, and internal resistance, as well as to statistics of mean voltage and mean current. Finally, the computation of SoC for the validation of the model's performance was based on the Equation 1.
In the second AutoML approach that was examined, we also used datasets that describe the whole lifecycle of the battery, Dataset2 [22], but we used 5 consecutive and randomly selected cycles for the training process and the next one for the evaluation procedure. The variables that fed into the AutoML model were raw data referring to the battery's voltage and temperature, as well as the statistics of mean voltage and mean current. The calculation of SoC for the evaluation process was based on the Equation 2.
4.5 Comparative analysis of experimental approaches
In the realm of computational science, selecting the right algorithm and parameters is crucial for achieving optimal outcomes. This section delves into a comparative analysis of different algorithms and experimental dimensions, aiming to highlight their strengths, weaknesses, and practical applications with regards to SoC prediction. In Table 1, we present a summary of the experiments along with their experimental dimensions’ configuration and the RMSE metric, which is representative for their performance. By comparing the different characteristics of each experiment, we can observe which combinations are more effective and suitable for the case of SoC prediction.
Table 1:
Summarizing these results, we notice that the use of deep neural networks, such as LSTM of FFNN, and AutoML in the pipelines, lead to a better performance comparing with the experiments conducted using LR or SVR. This behavior is an expected consequence since deep neural networks are able of selectively forgetting or retaining information due to their more complex architecture. Comparing the experiments based on LSTM with these employing FFNNs, we can notice that the first ones have better performance. This is also a reasonable conclusion since the LSTM model handles history data and builds long-term data dependencies, improving its performance. Furthermore, comparing all the experiments that were implemented with LSTM, we can observe better performance when the Dataset2 [22] as dataset was selected. As it has been already mentioned, this dataset describes the whole battery lifecycle and thus, the model can retain more information about the battery behavior. Moreover, we can conclude that the Equation 2 and specific raw data are more suitable for pipelines with LSTM and Dataset2 [22] as dataset, since this dimensions’ selection demonstrates a better accuracy, even though the training set consists of only 5 consecutive cycles and not of the whole lifecycle of the battery.
In the case of FFNN models, we can also note that the randomly selected training set is adequate for an accurate SoC prediction while, at the same time, it increases the dynamicity and the flexibility of the pipeline. Besides that, the SoC estimation method and the suitable features selection contribute to a better performance with FFNN model. Additionally, the accuracy of SoC prediction with the AutoML method is high since the model is adjusted to the input features and their behavior. Even though the dimensions’ selection is different for all of them, except the dataset's category, the accuracy is similar. This is an indicative conclusion for the performance of the AutoML method despite the randomness of the training set and the selected features.
Furthermore, using the LR or SVR model, the accuracy is smaller than the other models because of their simpler structure and weaker ability to retain information. The only exception was the experiment with a LR model and the Dataset2 [22] as dataset, in which we can observe high accuracy. This is a result of the model's ability to replicate the behavior of the features in the specific dataset, when adequate data exist, but also leading to the model's weakness to generalize its performance in other datasets. Finally, a major conclusion from these experiments about the features is that the capacity is an important indicator of SoC of the battery.
5 Conclusions and Future Work
Li-ion batteries are an important component of the BMS in EVs and their performance is crucial for the efficient function of EVs. SoC prediction is a significant task since this state of the battery is indicative for its performance. Data-driven methods for SoC prediction are gaining increasing interest in both academia and industry and have been driven by recent advances in ML algorithms and methods. Due to the large variety of ML algorithms which can be applied to the SoC prediction task, in this paper we presented a comparative analysis of pipelines that have been conducted for this purpose, considering different dimensions that configured the SoC prediction pipeline.
In our future work, we plan to apply the same pipeline for SoC prediction with different combinations of the pipeline dimensions. This will establish a new set of experiments that will be compared and analyzed and thus will emphasize the importance of the selected dimensions and how they affect the SoC prediction. Moreover, since the data used in this paper were taken from open datasets and were created under controlled conditions, we will implement this series of pipelines to data from real world's applications (e.g. BMS data from EVs). Furthermore, we intend to apply these pipelines in datasets that have been arisen from battery testing with different characteristics such as chemistry and nominal capacity. Finally, we aim at applying transfer learning methods with purpose to transfer the knowledge from one domain to another, such as the different battery chemistries or the different architectural layers of the battery, without retraining the models.
Acknowledgments
This work is funded by the European Union's Horizon 2020 project MARBEL (Grant agreement No 963540). The work presented here reflects only the authors’ view and the European Commission is not responsible for any use that may be made of the information it contains.
References
[1]
D. V. S. R. Sesidhar, Chandrashekhar Badachi, and Robert C. Green II, 2023. "A review on data-driven SOC estimation with Li-Ion batteries: Implementation methods & future aspirations." Journal of Energy Storage 72, (Nov 2023), 108420.
[2]
Carlos Vidal, Pawel Malysz, Phillip Kollmeyer, and Ali Emadi, 2020. "Machine learning applied to electrified vehicle battery state of charge and state of health estimation: State-of-the-art." IEEE Access 8, (Mar 2020), 52796-52814.
[3]
Afroditi Fouka, Katerina Lepenioti, Alexandros Bousdekis, and Gregoris Mentzas. 2022. "A Machine Learning Framework for Li-Ion Battery Lifetime Prognostics." In 13th International Conference on Information, Intelligence, Systems & Applications (IISA’22). IEEE, Corfu, Greece, 9 pages.
[4]
Mina Naguib, Phillip Kollmeyer, and Ali Emadi, 2021. "Lithium-ion battery pack robust state of charge estimation, cell inconsistency, and balancing." IEEE Access 9, (Mar 2021), 50570-50582.
[5]
JC Álvarez Antón, PJ García Nieto, F. J. de Cos Juez, F. Sánchez Lasheras, M. González Vega, and MN Roqueñí Gutiérrez, 2013. "Battery state-of-charge estimator using the SVM technique." Applied Mathematical Modelling 37, 9 (May 2013), 6244-6253.
[6]
Vedhanayaki Selvaraj, and Indragandhi Vairavasundaram, 2023. "A comprehensive review of state of charge estimation in lithium-ion batteries used in electric vehicles." Journal of Energy Storage 72, (Aug, 2023), 108777.
[7]
Xiaosong Hu, Jiuchun Jiang, Dongpu Cao, and Bo Egardt, 2015. "Battery health prognosis for electric vehicles using sample entropy and sparse Bayesian predictive modeling." IEEE Transactions on Industrial Electronics 63, 4 (Jul 2015), 2645-2656.
[8]
Carlos Vidal, Mike Haußmann, Daniel Barroso, Parisa M. Shamsabadi, Atriya Biswas, Ephrem Chemali, Ryan Ahmed, and Ali Emadi. 2018. "Hybrid energy storage system state-of-charge estimation using artificial neural network for micro-hybrid applications." In IEEE Transportation Electrification Conference and Expo (ITEC’18), IEEE, Long Beach, CA, USA, 7 pages.
[9]
Yahui Hu, and Zhiping Wang. 2019. "Study on SOC estimation of lithium battery based on improved BP neural network." In 8th International Symposium on Next Generation Electronics (ISNE’19). IEEE, Zhengzhou, China, 3 pages.
[10]
Abbas Fotouhi, Karsten Propp, and Daniel J. Auger. 2015. "Electric vehicle battery model identification and state of charge estimation in real world driving cycles." In 7th computer science and electronic engineering conference (CEEC’15). IEEE, Colchester, UK, 6 pages.
[11]
Ran Li, Shihui Xu, Sibo Li, Yongqin Zhou, Kai Zhou, Xianzhong Liu, and Jie Yao, 2020. "State of charge prediction algorithm of lithium-ion battery based on PSO-SVR cross validation." IEEE Access 8, (Jan, 2020), 10234-10242.
[12]
Dickshon NT. How, Mahammad A. Hannan, Molla S. Hossain Lipu, Khairul SM Sahari, Pin Jern Ker, and Kashem M. Muttaqi, 2020. "State-of-charge estimation of li-ion battery in electric vehicles: A deep neural network approach." IEEE Transactions on Industry Applications 56, 5 (Jun, 2020), 5565-5574.
[13]
Ephrem Chemali, Phillip J. Kollmeyer, Matthias Preindl, and Ali Emadi, 2018. "State-of-charge estimation of Li-ion batteries using deep neural networks: A machine learning approach." Journal of Power Sources 400, (Aug, 2018) 242-255.
[14]
Ma, L. Ma, C. Hu, and F. Cheng, 2021. "State of charge and state of energy estimation for lithium-ion batteries based on a long short-term memory neural network." Journal of Energy Storage 37, (Mar, 2021), 102440.
[15]
Junxiong Chen, Yu Zhang, Ji Wu, Weisong Cheng, and Qiao Zhu, 2023. "SOC estimation for lithium-ion battery using the LSTM-RNN with extended input and constrained output." Energy 262, (Jan 2023), 125375.
[16]
Paul Takyi-Aninakwa, Shunli Wang, Hongying Zhang, Huan Li, Wenhua Xu, and Carlos Fernandez, 2022. "An optimized relevant long short-term memory-squared gain extended Kalman filter for the state of charge estimation of lithium-ion batteries." Energy 260, (Dec 2022), 125093.
[17]
Xiangbao Song, Fangfang Yang, Dong Wang, and Kwok-Leung Tsui, 2019. "Combined CNN-LSTM network for state-of-charge estimation of lithium-ion batteries." IEEE Access 7, (Jul 2019). 88894-88902.
[18]
Mahammad A. Hannan, Dickson NT How, MS Hossain Lipu, Pin Jern Ker, Zhao Yang Dong, M. Mansur, and Frede Blaabjerg, 2020. "SOC estimation of li-ion batteries with learning rate-optimized deep fully convolutional network." IEEE Transactions on Power Electronics 36, 7 (Dec 2020), 7349-7353.
[19]
Ephrem Chemali, Phillip J. Kollmeyer, Matthias Preindl, Ryan Ahmed, and Ali Emadi, 2017. "Long short-term memory networks for accurate state-of-charge estimation of Li-ion batteries." IEEE Transactions on Industrial Electronics 65, 8 (Dec 2017), 6730-6739.
[20]
Xiaosong Hu, Fei Feng, Kailong Liu, Lei Zhang, Jiale Xie, and Bo Liu, 2019. "State estimation for advanced battery management: Key challenges and future trends." Renewable and Sustainable Energy Reviews 114, (Oct 2019), 109334.
[23]
Man-Fai Ng, Jin Zhao, Qingyu Yan, Gareth J. Conduit, and Zhi Wei She, 2020. "Predicting the state of charge and health of batteries using data-driven machine learning." Nature Machine Intelligence 2, 3 (Mar 2020), 161-170.
[24]
Carlos Vidal, Pawel Malysz, Mina Naguib, Ali Emadi, and Phillip J. Kollmeyer, 2022. "Estimating battery state of charge using recurrent and non-recurrent neural networks." Journal of Energy Storage 47, (Mar 2022), 103660.
[25]
Ahmed Elsheikh, Soumaya Yacout, and Mohamed-Salah Ouali, 2019. "Bidirectional handshaking LSTM for remaining useful life prediction." Neurocomputing 323, (Jan 2019), 148-156.
[26]
Zunya Shi and Abdallah Chehade, 2021. "A dual-LSTM framework combining change point detection and remaining useful life prediction." Reliability Engineering & System Safety 205, (Jan 2021), 107257.
[27]
Zicheng Fei, Fangfang Yang, Kwok-Leung Tsui, Lishuai Li, and Zijun Zhang, 2021. "Early prediction of battery lifetime via a machine learning based framework." Energy 225, (Jun 2021), 120205.
[28]
Kristen A. Severson, Peter M. Attia, Norman Jin, Nicholas Perkins, Benben Jiang, Zi Yang, Michael H. Chen, Muratahan Aykol, Patrick K. Herring, Dimitrios Fraggedakis, Martin Z. Bazant, Stephen J. Harris, William C. Chueh and Richard D. Braatz, 2019. "Data-driven prediction of battery cycle life before capacity degradation." Nature Energy 4, 5 (Mar 2019), 383-391.
[29]
Xin He, Kaiyong Zhao, and Xiaowen Chu, 2021. "AutoML: A survey of the state-of-the-art." Knowledge-based systems 212, (Jan 2021), 106622.
Index Terms
- State of Charge Prediction of EV Li-ion Batteries with Machine Learning: A Comparative Analysis
Recommendations
EV charging: separation of green and brown energy using IoT
UbiComp '16: Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing: AdjunctThe energy drivers use to charge their Electric Vehicles (EVs) comes from various sources. Of those, some are renewable green energy sources such as solar photovoltaic systems (SolarPV) and home storage battery whilst some are called brown sources such ...
State of Charge Estimation for Electric Vehicle Batteries Based on LS-SVM
IHMSC '13: Proceedings of the 2013 5th International Conference on Intelligent Human-Machine Systems and Cybernetics - Volume 01For the study of optimal control problems of battery power in electric vehicle, accurately estimating the state of charge (SOC) of the battery is a non-negligible part. This paper proposes a prediction model for state of charge of batteries Based on ...
INTELLIGENT WIRELESS EV FAST CHARGING WITH SMFIR. TECHNOLOGY
While introducing electric vehicles EVs into the massive market volume is important for the global environmental protection and reduced dependence on the petroleum depletion, the customers' acceptance on electric vehicles is one of critical elements ...
Comments
Please enable JavaScript to view thecomments powered by Disqus.Information & Contributors
Information
Published In
September 2024
437 pages
ISBN:9798400709821
DOI:10.1145/3688671
Copyright © 2024 Copyright held by the owner/author(s).
This work is licensed under a Creative Commons Attribution International 4.0 License.
Publisher
Association for Computing Machinery
New York, NY, United States
Publication History
Published: 27 December 2024
Check for updates
Author Tags
Qualifiers
- Research-article
Funding Sources
Conference
SETN 2024
SETN 2024: 13th Hellenic Conference on Artificial Intelligence
September 11 - 13, 2024
Piraeus, Greece
Contributors
Other Metrics
Bibliometrics & Citations
Bibliometrics
Article Metrics
- 0Total Citations
- 171Total Downloads
- Downloads (Last 12 months)171
- Downloads (Last 6 weeks)69
Reflects downloads up to 25 Feb 2025
Other Metrics
Citations
View Options
Login options
Check if you have access through your login credentials or your institution to get full access on this article.
Sign in