options
search icon
email icon
EN
rrss gif icons
twitter icon
linkedin icon youtube icon
shape

Development and Validation of Li-ion Battery State Algorithms Capable of Adapting to New Chemistries

Markel Azkue Etxeandia


14/12/2023

  • DIRECTORS: Haizea Gaztañaga eta Unai Iraola 
  • UNIVERSITY: Mondragon Unibertsitatea

ABSTRACT

Recognized as the bedrock of modern energy storage, lithium-ion batteries (Li-ion) have ascended to become an integral component in a large number of applications, which span from ubiquitous devices like smartphones and laptops to intricate systems like electric vehicles and large-scale energy storage infrastructures. The surge in demand can be directly attributed to the global paradigm shift towards renewable energy sources and the expanding sphere of electric mobility. These transitions impose an imperative to continuously innovate the underlying technology of Li-ion batteries, with objectives to augment their performance, reinforce safety, and optimize cost.

One of the vibrant research domains in this context is the development of effective State of Charge (SoC) and State of Health (SoH) estimation algorithms. These metrics are crucial to ensure efficient usage, longevity, safety, and performance optimization of li-ion batteries in various applications.

In addition, the need of flexible battery state estimation methods arises from the diverse characteristics of batteries, varying operating conditions, battery aging, mixed usage scenarios, emerging battery technologies, user behaviours, limited data availability, and the continuous evolution of battery research.

In view of this, the present work aims to develop adaptable SoC and SoH estimators. Leveraging the power of NNs, these estimators, complemented by the technique of Transfer Learning (TL), will exploit knowledge gathered from previous tasks or applications to adapt to new battery types or scenarios.

In pursuit of creating these estimators, an exhaustive review of related literature is conducted initially. The review concentrates on existing SoC and SoH estimation algorithms, with a special emphasis on those utilizing NNs. The review offers a comparative study discussing the merits and constraints of each methodology, introduces various types of NNs, such as Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN), and explains the principle of TL, highlighting its potential benefits for SoC and SoH estimation.

After the literature review, a comprehensive methodology is proposed, which forms the backbone of the entire research. This methodology describes a five-stage process:

  • Setup and Preparation (Stage 0): This initial phase involves necessary tasks before starting the training, such as data pre-processing, model selection, and determination of tuneable hyperparameters.
  • Model Creation (Stages 1, 2, and 3): These stages involve creating different SOC and SOH estimation models.
    • Stage 1: Baseline model: A baseline model is created and refined through rigorous training, testing, and hyperparameter tuning.
    • Stage 2: Comparative model: A new model is built from scratch using the data of a new cell, serving as a benchmark to compare the performance of the TL model.
    • Stage 3: TL model: The baseline model is retrained with new data from a different cell, utilizing the concept of TL.
  • Evaluation (Stage 4): This final stage involves comparing the results from the models in stages 2 and 3. By comparing their performances, the effectiveness of the TL approach can be gauged.

Following the outlined methodology, a variety of SoC estimators were developed leveraging Long Short-Term Memory (LSTM) networks. Synthetic data generated from a electrochemical model was employed to formulate a foundational model, onto which TL was subsequently applied to model a real cell. In parallel, to assess the viability and benefits of TL, models were independently constructed from scratch.

The results are compelling: Provided there is an established baseline model for execution of TL, the TL model consistently outperforms its counterparts built from scratch. Specifically, the TL model achieves a Mean Absolute Error (MAE) of a mere 0.88%, in stark contrast to the 1.84% and 5.62% MAE exhibited by the models built from scratch under the same SOC testing profiles. The TL model not only delivers superior results and demonstrates greater robustness, but it also demands substantially less data from the new cell for training - as much as 80% less, to be precise.

In a parallel exercise, similar to the approach employed for the SoC estimator, a series of SoH estimators were also developed. These used Fully Connected Network (FCN)-based NNs following the aforementioned methodology. The TL model once again outshines the models trained from scratch across all metrics, achieving an impressive MAE of 0.7% as opposed to the 1.2% and 1.6% MAE observed for the from-scratch models. Furthermore, echoing the earlier results, the TL model required half the data to train the new estimator compared to the other models.

In conclusion, the study strongly advocates the amalgamation of NNs and TL for adaptable and robust SoC and SoH estimation. The proposed methodology demonstrated that the use of TL models consistently outperforms their counterparts built from scratch, achieving notably lower MAE and demonstrating enhanced robustness. This approach not only enhances accuracy, but it also significantly reduces data requirements, and expedites training. This is particularly valuable in scenarios where data generation is limited or costly, making this method an effective solution for achieving high-quality results under constraints

close overlay