3. Embedded AI — Battery State of Charge using Machine Learning

🌈 Abstract

The article discusses the use of machine learning (ML) models to accurately estimate the state of charge (SOC) of Li-ion batteries, which is crucial for energy management in embedded devices. It explores various ML approaches, including shallow neural networks, deep learning, support vector machines, and Gaussian process regression, and their suitability for implementation in resource-constrained embedded systems.

🙋 Q&A

[01] Battery State of Charge (SOC) Estimation

1. What is the importance of accurately estimating battery SOC?

• Accurately estimating battery SOC is vital for the energy management of embedded devices.

2. What are the challenges in accurately estimating SOC?

• The complex behavior of Li-ion batteries, influenced by factors such as temperature, battery health, and SOC itself, makes accurate SOC estimation challenging.
• Traditional methods like electrochemical models require precise parameters and in-depth understanding of the battery's composition and physical characteristics.

3. How can machine learning (ML) models help in SOC estimation?

• ML models offer a data-driven approach that simplifies SOC estimation by requiring less detailed knowledge about the battery's behavior, making them suitable for implementation in embedded systems.

4. What is the objective of developing an ML model for SOC estimation?

• The objective is to develop an ML model that can accurately predict the SOC of a battery based on its voltage, current, temperature, and time, and can generalize well across different operating conditions and battery states to ensure reliable SOC estimation in real-time applications.

[02] Data Preprocessing

1. What dataset is used for training and testing the ML models?

• The dataset used is the Digital Commons Data for the LG 18650HG2 Li-ion Battery, published by McMaster University.

2. How is the data preprocessed?

• The data is downloaded and extracted from MATLAB format (.mat) files.
• The training and test data are separated based on the ambient temperature conditions.
• The data is resampled by taking every 100th point, and new moving averages are computed for the average voltage and average current predictors using the last 5 samples for each point.
• The preprocessed data is saved in CSV format for easier handling.

3. What are the shapes of the preprocessed training and test datasets?

• The training data shape after resampling is (5, 6700) for X and (1, 6700) for Y.
• The test data shapes after resampling vary for different ambient temperatures, ranging from (5, 393) to (5, 476) for X and (1, 393) to (1, 476) for Y.

[03] Machine Learning Approaches

1. What are the four prominent machine learning approaches discussed for battery SOC and SOH estimation?

• Shallow neural networks (SNN)
• Deep learning (DL) models, such as deep neural networks (DNNs) and convolutional neural networks (CNNs)
• Support vector machines (SVM)
• Gaussian process regression (GPR)

2. What are the key characteristics and suitability of each approach for embedded applications?

• Shallow neural networks (SNN): Relatively simple structure, requiring less computational power and memory, making them well-suited for real-time SOC and SOH estimation in resource-constrained embedded devices.
• Deep learning (DL) models: Can capture complex patterns and interactions, providing highly accurate estimates, but require more computational resources and memory, which can be a constraint for some embedded systems.
• Support vector machines (SVM): Effectively handle high-dimensional data, are robust to overfitting, and have relatively low computational requirements, making them suitable for embedded applications.
• Gaussian process regression (GPR): Provides not only predictions but also uncertainty estimates, making it useful for modeling the underlying uncertainties in battery behavior, but can be computationally intensive, requiring efficient implementations for real-time embedded applications.

3. Why is the GPR model selected for implementation?

• GPR is a flexible, non-parametric, and probabilistic approach that can model the underlying uncertainties in the battery's behavior due to varying operating conditions and aging.
• GPR models can be updated with new data, making them adaptable to changes over the battery's lifecycle.
• While GPR can be computationally intensive, the article mentions that the processing can be done off-device, and the trained model can be deployed once optimized.

[04] Next Steps

1. What is the plan for the next part of the series?

• The next part of the series will focus on training and evaluating the Gaussian process regression (GPR) model for battery SOC prediction.