⚡ electrical engineering

A Deep Learning Model for Battery State Prediction towards Intelligent Energy Management

This paper proposes a deep learning-based computational framework that integrates advanced neural network architectures with large-scale datasets to accurately predict battery health indicators, thereby enabling predictive maintenance and optimizing energy resource management for electric vehicles and large-scale storage systems.

Original authors: Athanasios Koukosiasa, Vasileios Tzanidakis, Sotiris Athanasiou, Kostas Kolomvatsos

Published 2026-05-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Athanasios Koukosiasa, Vasileios Tzanidakis, Sotiris Athanasiou, Kostas Kolomvatsos

Original paper licensed under CC BY 4.0 (http://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

Imagine you have a very expensive, high-tech backpack (a battery) that powers your electric vehicle or a giant energy storage system. You want to know exactly how much "fuel" is left inside and how long the backpack will last before it gets worn out.

Traditionally, guessing the fuel level is like trying to guess how much water is in a bucket just by looking at the outside. You might get it wrong if the bucket is tilted, if the water is hot, or if the bucket has a tiny leak you can't see. This paper introduces a smarter way to look inside the bucket using a "super-brain" made of computer code.

Here is a simple breakdown of what the researchers did:

1. The Problem: The "Guessing Game"

Old ways of measuring battery health (called State of Charge, or SoC) are like using a ruler that stretches and shrinks depending on the weather. They work okay in perfect conditions, but if the battery is used quickly, gets hot, or is old, the ruler gives bad readings. This can lead to running out of power unexpectedly or damaging the battery.

2. The Solution: A "Super-Brain" (Deep Learning)

Instead of using a rigid ruler, the authors built a Deep Learning model. Think of this as a student who has read millions of stories about how batteries behave.

The Teacher: They didn't just use textbook theories; they fed this student real-world data from a massive cloud system (Sunlight Group's "GLocal") that monitors thousands of actual industrial batteries.
The Learning: The model learned to spot patterns in voltage, temperature, and current that humans or simple math formulas miss.

3. The Secret Sauce: A Two-Part Team

The researchers didn't just use one type of brain; they built a hybrid team with two special roles:

The Filter (Autoencoder): Imagine you have a messy room full of toys, clothes, and books (raw data). Before you can study the room, you need to tidy it up. The "Autoencoder" is a robot that instantly sorts through the messy data, throws away the junk (noise), and packs the important stuff into a tiny, neat suitcase. This makes the next step much faster and clearer.
The Time Traveler (BiLSTM): This is the main detective. Unlike a normal detective who only looks at clues from the past, this one looks at the past and the future simultaneously.
- Analogy: Imagine watching a movie. A normal detective only watches the scenes that have already played. This "Bidirectional" detective can see the scenes that just played and the scenes that are about to play. This helps it understand the full story of how the battery is behaving, not just what happened a second ago.

4. The "Fuel Gauge" Calculation

Once the Time Traveler predicts how much electricity will flow in and out of the battery, the system uses a classic method called Coulomb Counting.

Analogy: Think of this like a water meter. If you know how much water went in and how much went out, you can calculate exactly how much is left in the tank. The AI predicts the flow, and the water meter does the final math to tell you the exact fuel level.

5. The Results: How Well Did It Work?

The team tested this system on real industrial batteries (the kind used in forklifts and large energy storage).

The Score: They measured success using a score called $R^2$ . A score of 100% is perfect. Their model hit scores between 89% and 94%.
What this means: The computer's guess was almost identical to the actual measurement. It was much better at predicting the battery's future than the old methods.
Reliability: It worked well even when they tested it on different types of batteries, proving it's not just a "one-trick pony."

Summary

The paper claims to have built a smart, hybrid system that cleans up messy battery data, uses a "time-traveling" AI to understand complex patterns, and accurately predicts how much energy is left in industrial batteries. It's a tool that helps battery managers know exactly when to charge, when to stop, and when to replace a battery, making energy systems safer and more efficient.

What the paper does NOT claim:

It does not claim this works for medical devices or human health (it's strictly for industrial and energy storage batteries).
It does not claim to invent a new type of battery chemistry; it only improves how we monitor existing ones.
It does not promise to fix broken batteries; it only predicts their state to help manage them better.

Technical Summary: A Deep Learning Model for Battery State Prediction towards Intelligent Energy Management

Problem Statement
Accurate forecasting of battery health indicators, particularly the State of Charge (SoC), is critical for the reliability, safety, and operational efficiency of electric vehicles and large-scale energy storage systems. Traditional estimation methods, such as Coulomb counting, Open Circuit Voltage (OCV) lookups, and model-based observers (e.g., Kalman filters), often struggle with OCV hysteresis, sensor drift, temperature dependence, and aging-induced parameter shifts. These methods require frequent re-calibration and often degrade in accuracy under dynamic duty cycles, partial cycling, or heterogeneous deployments. While deep learning offers a data-driven alternative, many existing studies rely heavily on controlled laboratory datasets, limiting their direct applicability to real-world industrial environments where battery chemistries, formats, and usage patterns vary significantly.

Methodology
The authors propose a hybrid computational framework that integrates deep learning architectures with fundamental electrochemical principles to predict battery states using real-world data from the Sunlight Group's GLocal cloud platform. The methodology proceeds through several stages:

Data Pre-processing: The study utilizes operational and diagnostic data from industrial traction batteries (51.2V, 360–504 Ah, Lithium Iron Phosphate chemistry). Raw sensor outputs (voltage, current, temperature, and operational modes) are anonymized, cleaned, and normalized via min-max scaling.
Model Comparison: The authors initially evaluated Support Vector Regression (SVR) against Long Short-Term Memory (LSTM) networks. While SVR is effective for non-linear regression, it was found to lack computational scalability for large datasets and was highly sensitive to data quality. Consequently, the research focused on LSTM-based architectures.
Architecture Evolution: Several LSTM variants were explored:
- Standard LSTM: A baseline model achieving $R^2$ scores between 86.7% and 90.3%.
- Autoencoder-LSTM Hybrid: An autoencoder was introduced to compress multi-sensor data into a compact latent space, reducing redundancy and noise before feeding data into the LSTM. This improved generalization across different usage profiles.
- Multi-Output LSTM: Designed to predict individual cell voltages before aggregation to minimize error propagation, achieving $R^2$ scores between 89% and 94%.
- Bidirectional LSTM (BiLSTM): This architecture processes sequences in both forward and backward directions to capture complex temporal dependencies from past and future contexts.
Final Hybrid Framework: The proposed system combines a pre-trained autoencoder for feature compression with a BiLSTM network for temporal modeling. The BiLSTM predicts charge and discharge currents (in Ampere-hours). These predictions are then converted into SoC values using the Coulomb counting methodology, adjusted with practical rounding logic (rounding down during discharge, up during charging) to reflect real-world Battery Management System (BMS) behavior.

Key Contributions
The paper outlines four primary contributions:

Hybrid SoC Estimation Framework: A novel integration of autoencoders, BiLSTM networks, and Coulomb counting, merging data-driven learning with electrochemical consistency.
Efficient Feature Learning: The use of autoencoders to compress multi-sensor data into latent features, enhancing learning efficiency and reducing data redundancy.
Temporal Pattern Extraction: The application of BiLSTM to capture bidirectional temporal dependencies in charge-discharge sequences, improving prediction accuracy.
Experimental Validation: The model was trained and tested on real-world datasets from the Sunlight GLocal platform, demonstrating robustness across multiple battery types and configurations.

Results
The experimental evaluation utilized an 80/20 chronological split for training and validation across various battery types (Traction Battery A, Traction Battery B, and ESS Traction Battery A). Performance was measured using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination ( $R^2$ ).

Prediction Accuracy: The model achieved $R^2$ scores ranging from 89% to 94% for predicting critical parameters such as ampere-hours charged and discharged.
SoC Estimation: The SoC values derived from the LSTM-predicted currents via Coulomb counting showed high consistency with measured field values, with deviations primarily occurring during abrupt transitions.
Forecasting: In the future state forecasting stage (predicting future AH charge/discharge behavior), the model achieved an accuracy of 64.79% for AH charged and 70.14% for AH discharged.
Cross-Configuration Performance: The framework demonstrated strong generalization capabilities, maintaining high performance when tested on battery configurations different from the training set (e.g., training on Type A and testing on Type B).

Significance and Claims
The authors claim that the primary significance of this work lies in the successful fusion of data-driven deep learning with fundamental battery electrochemistry. By leveraging real-world industrial data rather than purely synthetic or laboratory datasets, the framework addresses the gap between academic advancements and practical deployment in Battery Management Systems.

The paper asserts that the hybrid approach not only achieves high prediction accuracy but also maintains physical interpretability, a crucial requirement for real-world applications. The consistent performance across diverse battery types and the system's ability to process multivariate time series data while capturing bidirectional temporal dependencies are highlighted as key factors in its robustness. The authors conclude that this framework provides a solid foundation for developing intelligent, adaptive, and trustworthy battery analytics systems capable of supporting predictive maintenance and efficient energy resource allocation in diverse real-world conditions. Future work is suggested to include the integration of additional health indicators and the extension to probabilistic forecasting.