💻 computer science

Deploying machine learning for remaining useful life prediction of lithium-ion batteries on Coral Dev Board

This paper evaluates and compares the performance of Random Forest, XGBoost, and Temporal Convolutional Network models for predicting the remaining useful life of lithium-ion batteries, demonstrating their practical deployment and effectiveness on the Coral Dev Board for predictive maintenance in electric vehicles.

Original authors: Mariem Boujneh, Nesrine Majdoub, Taoufik Ladhari, Ali Serdar Atalay, Anis Sakly

Published 2026-09-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mariem Boujneh, Nesrine Majdoub, Taoufik Ladhari, Ali Serdar Atalay, Anis Sakly

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Electric vehicles and renewable energy storage systems rely heavily on a single, critical component: the lithium-ion battery. These power sources are not infinite; like any mechanical part, they wear down over time, losing the ability to hold a full charge. To keep these systems running safely and efficiently, engineers need to know exactly how much life remains in a battery before it fails. This concept is known as the "remaining useful life." It is distinct from simply checking how much charge is left at a specific moment; instead, it is a prediction of how many more charging cycles a battery can endure before its capacity drops below a usable limit. Predicting this lifespan is difficult because batteries degrade in complex, non-linear ways that depend on temperature, charging speed, and usage history. Without accurate predictions, users risk sudden failures, while manufacturers might replace batteries too early, wasting resources.

A team of researchers from the National Engineering School of Tunis and Enterprise Estonia has tackled this challenge by testing whether advanced computer programs can predict battery life directly on small, portable devices. They focused on three different types of machine learning algorithms—computer systems that learn from data rather than following rigid, pre-written rules. The first, called Random Forest, works like a committee of decision-makers, each making a guess and then averaging the result. The second, XGBoost, builds a sequence of models where each new one tries to correct the mistakes of the previous one. The third, a Temporal Convolutional Network, is designed specifically to understand patterns that unfold over time, such as the slow decline of a battery's health. The researchers wanted to see which of these methods was best at forecasting battery life and, crucially, whether they could run efficiently on a piece of hardware called the Coral Dev Board. This board is a small computer equipped with a specialized chip designed to run artificial intelligence models quickly and with very little power, making it ideal for use inside vehicles or remote equipment where cloud computing is not an option.

To train these computer models, the team used real-world data collected from twenty electric vehicles over a period of twenty-nine months. These cars, all of the same model, were equipped with lithium-ion battery packs that were monitored continuously. The data recorded every detail of the charging process, including the voltage, the current flowing into the battery, the temperature of the cells, and the total available capacity. Because the raw data from sensors often contains gaps or errors, the researchers first cleaned and smoothed the information, filling in missing values and removing noise to create a clear picture of how the batteries aged. They then calculated the remaining useful life for each data point, creating a target for the computer models to learn from. The dataset was split so that the models could learn from eighty percent of the data and then be tested on the remaining twenty percent to see how well they performed on new, unseen information.

The researchers deployed all three algorithms onto the Coral Dev Board to measure not just their accuracy, but also their energy consumption and speed. The results showed a clear trade-off between complexity and performance. The Temporal Convolutional Network, despite being designed for time-based data, struggled to generalize its findings. While it produced very low error rates for specific short-term predictions, its overall ability to predict the general trend of battery degradation was poor, resulting in a score that indicated it performed no better than random chance for the broader dataset. It also consumed the least amount of energy, but its inconsistency made it unreliable for this specific task. The Random Forest algorithm, known for its ability to explain which factors are most important, performed better than the deep learning network but still lagged behind in accuracy. It also required the most energy to run, making it less suitable for battery-powered edge devices.

The standout performer was the XGBoost algorithm. This method achieved the highest overall accuracy, correctly predicting the remaining life of the batteries with the smallest average error. It balanced high precision with moderate energy use, consuming significantly less power than the Random Forest model while maintaining a strong ability to detect the complex, non-linear patterns of battery aging. The researchers found that this algorithm was particularly effective at handling the messy, real-world data from the electric vehicles without needing the heavy computational resources of a large server. By successfully running this model on the Coral Dev Board, the team demonstrated that sophisticated predictive maintenance is possible directly on the device. This means that in the future, electric vehicles could potentially monitor their own battery health in real-time, alerting drivers to maintenance needs before a failure occurs, all without needing to send data to a distant cloud. The study concludes that while deep learning networks show promise, simpler, well-tuned algorithms like XGBoost are currently the most practical and reliable choice for deploying battery health monitoring on small, energy-efficient hardware.

Technical Summary: Deploying Machine Learning for Remaining Useful Life Prediction of Lithium-Ion Batteries on Coral Dev Board

Problem Statement
Lithium-ion batteries are critical for renewable energy storage and electric vehicles (EVs), yet their degradation over time necessitates precise monitoring to ensure safety and reliability. Key metrics such as State of Health (SOH) and Remaining Useful Life (RUL) are essential for optimizing battery utilization and planning maintenance. While model-based approaches (e.g., electrochemical or equivalent circuit models) exist, they often struggle with the complexity of real-time estimation or lack the ability to represent dynamic responses across diverse working conditions. Consequently, there is a shift toward data-driven machine learning (ML) techniques. However, a significant challenge remains in deploying these complex ML models on resource-constrained edge devices to enable real-time, on-board battery health forecasting without relying on cloud resources.

Methodology
This study proposes a data-driven framework for predicting the RUL of lithium-ion batteries using three distinct ML algorithms: Random Forest (RF), XGBoost, and Temporal Convolutional Networks (TCN).

Dataset: The research utilizes real-world charging data from 20 BAIC EU500 electric vehicles equipped with CATL lithium-ion battery packs, collected over approximately 29 months. The dataset includes features such as State of Charge (SOC), pack voltage, charge current, cell voltage extremes, temperature, and available capacity.
Preprocessing: Raw sensor data underwent rigorous cleaning to remove duplicates and outliers, with missing values imputed via interpolation. A Savitzky-Golay filter was applied to smooth high-frequency noise. Crucially, the RUL target variable was not directly available; it was engineered using the formula $RUL = (C_{available} - C_{EOL}) / (C_{init} - C_{EOL})$ , where $C_{init}$ is the nominal initial capacity (1200 Ah) and $C_{EOL}$ is the end-of-life threshold (70% of initial capacity). Data was normalized using Min-Max scaling. For deep learning models, particularly TCN, the data was segmented into overlapping sequences using sliding windows to capture temporal evolution, while tree-based models (RF and XGBoost) utilized the processed data without this specific sequence generation step.
Algorithms:
- Random Forest (RF): An ensemble of decision trees aggregating regression outputs to improve robustness and handle non-linear correlations.
- XGBoost: A gradient boosting algorithm utilizing regularization (L1/L2) and parallel processing to optimize an objective function, designed to handle missing data and prevent overfitting.
- Temporal Convolutional Networks (TCN): A deep learning architecture employing causal and dilated convolutions to capture long-range temporal dependencies while maintaining strict causality and enabling parallel training.
Deployment Platform: All models were deployed on the Google Coral Dev Board, an edge computing device featuring an NXP i.MX 8M CPU and a dedicated Google Edge TPU (Tensor Processing Unit) capable of up to 4 TOPS. Models were converted to the TensorFlow Lite format to leverage hardware acceleration for efficient inference.

Key Contributions

Comparative Edge Deployment: The study provides a direct comparative analysis of RF, XGBoost, and TCN specifically within the constraints of an edge device (Coral Dev Board), evaluating not just accuracy but also energy consumption and inference speed.
Hardware-Accelerated Inference: It demonstrates the feasibility of running ML-based RUL prediction locally on EV hardware, reducing latency and energy overhead associated with cloud processing.
Performance Benchmarking: The research establishes a baseline for the trade-offs between predictive accuracy (MSE, RMSE, MAE, $R^2$ ) and computational efficiency (energy consumption) for different ML architectures in battery health monitoring.

Experimental Results
The models were evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Squared Error (MSE), Coefficient of Determination ( $R^2$ ), and energy consumption.

XGBoost: Achieved the best overall performance with the lowest MAE (0.1600) and a competitive RMSE (0.4243), resulting in the highest $R^2$ score of 0.65. It demonstrated a balanced energy consumption of 83.6%, making it the most effective strategy for the specific constraints of the Coral Dev Board.
Random Forest: Showed moderate performance with an $R^2$ of 0.50 and higher error values (MAE 0.3500). It exhibited the highest energy consumption (89.1%), suggesting it may be less suitable for resource-constrained edge applications compared to XGBoost.
TCN: Displayed mixed results. While it achieved the lowest MSE (0.1432) and RMSE (0.3784), it suffered from a high MAE (0.5394) and a negative $R^2$ score (-0.002). The authors attribute this instability to the shallow network design required to fit the Coral Dev Board's resource limits and data sparsity. However, it was the most energy-efficient (72.1%).

Significance and Claims
The paper claims that XGBoost is the most promising approach for real-time RUL prediction on edge hardware like the Coral Dev Board, as it offers the most reliable balance between prediction precision and processing efficiency. While TCNs hold potential for time-series modeling, the study concludes that their edge implementation currently requires further optimization (e.g., hybrid techniques or deeper structures) to overcome instability. The work supports the development of predictive maintenance in the battery industry by validating that effective ML-based health forecasting can be executed directly on low-power, embedded devices, thereby enabling localized, real-time battery management systems.

Future Directions
The authors suggest future research should focus on:

Investigating sophisticated hybrid decomposition and feature extraction methods to enhance TCN performance on edge devices.
Exploring deeper neural network topologies combined with transfer learning to improve generalization across different battery types.
Further reducing latency and energy consumption through model quantization and pruning optimizations specifically for the Coral Dev Board deployment pipeline.

Technical Summary: Deploying Machine Learning for Remaining Useful Life Prediction of Lithium-Ion Batteries on Coral Dev Board

More like this