Explainable Artificial Intelligence for State of Charge Estimation of Lithium-Ion Batteries

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes an explainable artificial intelligence for lithium-ion battery state of charge estimation. Please further improve the paper based on the comments below:

 

1. The logic of the abstract needs to be reorganized and the existing issues in current research and the advantages of the proposed method have not been clearly articulated. In addition to qualitative analysis, quantitative indicators also need to be presented in the abstract.
2. Please provide more details about the "real-world dataset" used, including data sources, collection methods, data cleaning and preprocessing procedures.
3. Describe in detail the neural network architecture used, including the number of layers, number of neurons, and activation functions.
4. Provide hyperparameter settings for model training, such as learning rate, batch size, and number of training rounds.
5. Please further discuss the feasibility of this approach in practical applications, including how to integrate the model into the battery management system of electric vehicles and how to ensure the robustness of the model under different environmental conditions.
6. The Introduction should clarify the necessity and importance of the current research by evaluating and comparing the existing literatures, rather than just listing the relevant references.
7. In the Literature Review section, please add the main results of reviewed papers.
8. Some references in the Introduction section are too old. Old references are often not representative. The author needs to cite the most recent articles published in the field.
9. Neural networks show important application potential in the field of estimation and prediction. For Long Short-Term Memory, authors should cite recent papers published in the field, like Learning-based model predictive energy management for fuel cell hybrid electric bus with health-aware control; A novel deep reinforcement learning-based predictive energy management for fuel cell buses integrating speed and passenger prediction.
10. The contributions should be improved to fully explain the superiority of the proposed method compared with recent works. What is the main challenge and why previous works could not solve it?
11. The parameter settings involved in the paper need to be given in detail.
12. Contributions should be presented clearly and succinctly in bulleted form.
    13. The writing details of this paper need to be improved and the quality of the paper's writing needs to be improved.
13. Include details about parameter selection and sensitivity analysis for this study.
14. The overall picture quality is not high.
15. It is recommended to add a discussion of future research directions in the conclusion section

Comments on the Quality of English Language

The writing details of this paper need to be improved and the quality of the paper's writing needs to be improved.

Author Response

The logic of the abstract needs to be reorganized and the existing issues in current research and the advantages of the proposed method have not been clearly articulated. In addition to qualitative analysis, quantitative indicators also need to be presented in the abstract.

The abstract has been rewritten. The existing issues of SoC have been stated, the advantages of using data driven methods in combination with xAI have been explained, and quantitative metrics that were achieved within the manuscript have been included.

 

Please provide more details about the "real-world dataset" used, including data sources, collection methods, data cleaning and preprocessing procedures.

Details of the dataset used can be found in Section 4.1 Data and Section 4.2 Dataset.

 

Describe in detail the neural network architecture used, including the number of layers, number of neurons, and activation functions.

Details of the architectures used are described in Section 5.1 E1: SoC estimation by different NN architectures. Details of the parameters can be found in Section A Appendix: Hyperparameters.

 

Provide hyperparameter settings for model training, such as learning rate, batch size, and number of training rounds.

Details of parameters can be found in Section A Appendix: Hyperparameters. Details of learning rate, number of epochs, etc. can be found in the subsection Training Procedure of each experiment section.

 

Please further discuss the feasibility of this approach in practical applications, including how to integrate the model into the battery management system of electric vehicles and how to ensure the robustness of the model under different environmental conditions.

The research was conducted at Technology Readiness Level (TRL 2-3) for a project, the focus remained on theoretical development and early-stage validation. Therefore, practical implementation aspects, including integration with real-world battery management systems in electric vehicles and evaluation of model robustness under diverse environmental conditions, were not within the scope of the project. This is the reason why this is an article of type ‘Project Report’.

 

The Introduction should clarify the necessity and importance of the current research by evaluating and comparing the existing literatures, rather than just listing the relevant references.

Section 1 Introduction has been rewritten for clarity and more existing works have been used for comparison.

 

In the Literature Review section, please add the main results of reviewed papers.

Each paper evaluated their work using different metrics, different methods of comparison and different types of datasets. This makes the main results of each paper not easily comparable nor very informative. In many cases, extended descriptions of their work will be required for the results to be understandable, which would include information not related to the scope or aim of this work. References are present in the literature review section should a reader be interested in the results. 

 

Some references in the Introduction section are too old. Old references are often not representative. The author needs to cite the most recent articles published in the field.

Just because a reference is old doesn't mean the work they did is unimportant. It's essential to recognise the value and contributions of past research. That said, we have included more recent articles within Section 1 Introduction.

 

Neural networks show important application potential in the field of estimation and prediction. For Long Short-Term Memory, authors should cite recent papers published in the field, like Learning-based model predictive energy management for fuel cell hybrid electric bus with health-aware control; A novel deep reinforcement learning-based predictive energy management for fuel cell buses integrating speed and passenger prediction.

Learning based model predictive energy management for fuel cell hybrid electric bus with health-aware control, has been added to Section 2.3 Battery Modelling using Neural Networks as reference. Other recent papers regarding NN architectures and SoX, in particular SoC estimation have also been added as reference.

 

The contributions should be improved to fully explain the superiority of the proposed method compared with recent works. What is the main challenge and why previous works could not solve it?

The challenges of SoC estimation are further explained in Section 1 Introduction and Section 2.2 Battery Modelling. The differences between the work conducted and previous works are discussed in Section 2.3 Battery Modelling using Neural Networks. We have clarified that most previous works benchmarking NNs for battery modelling do not compare all SotA NN architectures and rely either on simulated or lab data. We propose a more complete benchmark and deploy it on a dataset obtained from field EV battery deployment.

 

The parameter settings involved in the paper need to be given in detail.

Details of parameters can be found in Section B Appendix: Hyperparameters. Details of learning rate, number of epochs, etc. can be found in the subsection Training Procedure of each experiment section.

 

Contributions should be presented clearly and succinctly in bulleted form.

The contributions presented by bullet points are available within Section 1 Introduction.

 

The writing details of this paper need to be improved and the quality of the paper's writing needs to be improved.

The manuscript has since been proofread again. 

 

Include details about parameter selection and sensitivity analysis for this study.

Details of parameters can be found in Section B Appendix: Hyperparameters. Details of learning rate, number of epochs, etc. can be found in the subsection Training Procedure of each experiment section.

Unfortunately, sensitivity analysis cannot be conducted because this work is based on a project that has since ended, therefore it is no longer possible to do so. This is the reason why this is an article of type ‘Project Report’.

    

The overall picture quality is not high.

The figures within the manuscript have been improved as much as possible.

 

It is recommended to add a discussion of future research directions in the conclusion section

Discussion of future research directions can be found in Section 5.2 Future Work before Section 6 Conclusion. This decision is made in order to keep the conclusion concise and compact as per another reviewer’s suggestion.

Reviewer 2 Report

Comments and Suggestions for Authors

1) My primary concern is training the NN to accurately estimate the SoC. Since accurate measurement of SoC is not feasible, the training dataset for the NN will likely be unreliable, leading to a low accuracy in the NN-based estimator.

 

2) What is the advantage of using NN to estimate SoC? One of the primary benefits of NNs is their ability to perform real-time implementation. However, it is worth noting that real-time computations are generally not a significant challenge for SoC measurement.

 

3) More details about the training process is required. Platform? Learning rate? Number of epochs? 

 

4) To make sure that the comparison study is fair, the authors need to provide details of other methods. Most of the methods have design parameters that would be used to improve the accuracy. It is not clear if the comparison study has been conducted fairly. 

Author Response

1) My primary concern is training the NN to accurately estimate the SoC. Since accurate measurement of SoC is not feasible, the training dataset for the NN will likely be unreliable, leading to a low accuracy in the NN-based estimator.

As detailed in Section 3.3 Dataset, “Recognising the frequent inaccuracies in the original SoC reported by the vehicles, we carried out a thorough verification process and corrected the SoC whenever a mistake was found. We achieved this by developing precise Open-Circuit Voltage (OCV_ curves, see Figure 3. The curves serve as a benchmark to align the cell voltages of the cars with established lookup tables, thus ensuring the SoC values to be as accurate and reliable as possible.” 

We recognise that accurate measurement of SoC is not always feasible, therefore we have employed methods to make it as accurate and reliable as possible with the dataset that we had during the project.

 

2) What is the advantage of using NN to estimate SoC? One of the primary benefits of NNs is their ability to perform real-time implementation. However, it is worth noting that real-time computations are generally not a significant challenge for SoC measurement.

NNs can learn meaningful representations (features) without manual feature engineering. Neural Networks learn direct mapping from raw input to model prediction, and their ability to process large amounts of data allows them to deal with such redundancy. Therefore NNs work well when the existing information of the battery cell internals cannot be easily attained. This point is explored further in Section 2.2 Battery Modelling. However as the feature selection is internal due the weighting assignment during training, the comprehension of the weighting of features with respect to the output of the model is limited. 

Therefore, we explored within the manuscript how xAI, which are methods dedicated to understanding NN, can be used to improve input profile for NN, as well as further understand the relationship between the signals within a battery cell. This is useful as it can provide information about the battery cell internals when such information is not easily attainable.

 

3) More details about the training process is required. Platform? Learning rate? Number of epochs? 

Details of parameters can be found in Section B Appendix: Hyperparameters. Details of learning rate, number of epochs, etc. can be found in the subsection Training Procedure of each experiment section.

 

4) To make sure that the comparison study is fair, the authors need to provide details of other methods. Most of the methods have design parameters that would be used to improve the accuracy. It is not clear if the comparison study has been conducted fairly. 

It has been explained within the Training Procedure of E1and E2 that the TRESampler of Optuna was used to optimize the set of parameters for each model used within the comparison study. Optuna is a framework that provides automatic optimisation of hyperparameters within a neural network model. The same set of optuna study parameters was used for each model in order to ensure that each model has been optimised to the same degree. The NNs are trained, validated, tested against the outcomes in Open-Circuit Voltage approach as defined above, which offers the best possible approximation to the ground truth.

These outcomes have been obtained in the course of a project and are offered as contribution so other researchers can build on this experience, take up the lessons learned and avoid the pitfalls. The outcomes of the project are offered as is. Further modeling or developments that require extensive pre-processing and/or feature engineering are no longer possible. This is the reason why this is an article of type ‘Project Report’.

Reviewer 3 Report

Comments and Suggestions for Authors

Review the Manuscript ID: applsci-3515416

Title: Explainable Artificial Intelligence for State of Charge Estimation of Lithium-Ion Batteries

The objective of this work is to demonstrate the application of Neural Networks (NN) for State of Charge (SoC) estimation with real-world dataset. Additionally, explainable AI (xAI) is applied upon the trained NN to identify feature importance within the multivariate signals of the electric vehicle battery.

Authors should review the text, as in some cases terms are used that are not widely used from a technical point of view (this work is to showcase), (are Pivotal), etc. 

When indicating references, authors use indications such as: see [1–3]. Review the way in which references are indicated and included in the main text.

Bibliographic references should be included in sections 2.1 and 2.2.

The information included in section 3 should be included in an existing section. A section is not justified.

Adapt the text formatting to better integrate table 3.

Sections 4 and 5 should be revised and rewritten. In some cases the information is very confusing, with a topic-based structure.

Conclusions should be rewritten and presented in a more concise format highlighting the main contributions of the study.

In summary, the topic is interesting and current. However, the way the article is structured is confusing, highlighting generic issues associated with the topic developed. I believe the authors should enhance the discussion of technical aspects by providing a more detailed explanation of the developed methodology. Comments on the Quality of English Language

It should be improved

Author Response

Authors should review the text, as in some cases terms are used that are not widely used from a technical point of view (this work is to showcase), (are Pivotal), etc. 

Some of these terms have been reworded.

 

When indicating references, authors use indications such as: see [1–3]. Review the way in which references are indicated and included in the main text.

The formatting of references adheres to the guideline from MDPI Reference Guide https://mdpi-res.com/data/mdpi_references_guide_v9.pdf. The manuscript was prepared in LaTeX using BibTeX to process references. Please provide the formatting option that BibTeX should be using when compiling if the referencing format is incorrect.

 

Bibliographic references should be included in sections 2.1 and 2.2.

References have been added to Section 2.1 Battery Testing and Section 2.2 Battery Modelling.

 

The information included in section 3 should be included in an existing section. A section is not justified.

This has been merged with Section 3 Methodology.

 

Adapt the text formatting to better integrate table 3.

Table 3 has been better incorporated into the text.

 

Sections 4 and 5 should be revised and rewritten. In some cases the information is very confusing, with a topic-based structure.

Due to the complexity of the work conducted during the project, there is no easy way to present the information from Section 3 Methodology and Section 4 Experiment. Both sections have been reworked to present the information in a more concise and clear manner.

Each subsection within Section 3 contains information regarding the nature of the data or the setting of the experiments. And each subsection within Section 4 outlines each experiment along with their respective procedure, parameter setting, results and findings.

 

Conclusions should be rewritten and presented in a more concise format highlighting the main contributions of the study.

The conclusion has been rewritten in a more concise manner. The main contributions of the study are now shown clearer. 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Tanks for your effort, there is no further comments.

Reviewer 2 Report

Comments and Suggestions for Authors

N/A

Reviewer 3 Report

Comments and Suggestions for Authors

Review the Manuscript ID: applsci-3515416

Title: Explainable Artificial Intelligence for State of Charge Estimation of Lithium-Ion Batteries

 

The objective of this work is to demonstrate the application of Neural Networks (NN) for State of Charge (SoC) estimation with real-world dataset. Additionally, explainable AI (xAI) is applied upon the trained NN to identify feature importance within the multivariate signals of the electric vehicle battery.

The review / correction performed by the authors allowed an improvement in the quality of the paper.