Trustworthy energy intelligence

Explainable Battery Digital Twins

A battery prediction should arrive with evidence, especially when it informs safety, maintenance, and remaining life.

Published in IEEE Access, 2024

Why I built it

How can deep-learning battery twins remain accurate while making their State of Charge and State of Health estimates understandable?

Battery digital twins can estimate State of Charge and State of Health, but opaque predictions limit their usefulness in safety-critical energy systems. An engineer needs to know which voltage, temperature, capacity, and degradation signals drove a result.

This work develops DNN and LSTM battery twins using the NASA battery dataset, then explains their predictions through SHAP, LIME, and linear surrogate models. It treats interpretability as part of the twin rather than an annotation added after deployment.

Explainable battery digital twin framework for state prediction

The complete framework connects battery measurements to DNN and LSTM digital twins, State of Charge and State of Health predictions, and explanation methods that expose why each estimate was produced.

How the system works

Cross-battery learning

All four batteries in the NASA prognostics dataset were used to build models that learn across varied charge, discharge, impedance, temperature, voltage, and current conditions.

DNN and LSTM twins

Feedforward and recurrent architectures estimate State of Charge and State of Health while capturing nonlinear and temporal battery behavior.

Three explanation paths

SHAP provides global and local attribution, LIME explains individual estimates, and linear surrogate models test whether the learned behavior can be approximated transparently.

2battery states predicted

2deep-learning twin families

3explainability methods

4NASA batteries used

What this work contributes

From a research question to an inspectable system.

An explainable digital-twin framework for State of Charge and State of Health estimation.
Direct comparison of DNN and LSTM prediction behavior across battery cycles.
SHAP, LIME, and surrogate explanations for both global patterns and individual estimates.
Feature analysis linking State of Charge to voltage and temperature and State of Health to capacity fade.
Evidence that predictive performance and interpretability can be evaluated together.

Study design for explainable battery state estimation — The study design traces data preparation, feature selection, digital-twin training, prediction, and comparative explanation through SHAP, LIME, and surrogate models.

How I read the results

What the evidence means

The explanations turn a state estimate into an inspectable engineering claim. For example, they reveal whether a State of Charge prediction follows measured voltage and temperature or whether a health estimate is responding to capacity and capacity fade. That traceability supports debugging, trust, and better maintenance decisions.

Current scope

What this study validates

The study validates the framework on the NASA battery dataset. It provides the explainable prediction foundation that later became part of the broader BatteryMetrix research program.

Inside the system

More evidence from the research.

Explainability workflow for battery digital twin predictions — The explanation workflow translates a model prediction into local and global evidence that engineers can use to inspect the variables driving battery-state estimates.

Battery feature correlation matrix for State of Charge and State of Health — The correlation matrix connects the predicted states to physical evidence, including measured voltage, temperature, capacity, and capacity fade.

Actual and predicted State of Charge across battery cycles using an LSTM digital twin — The LSTM twin follows changing State of Charge across hundreds of cycles, including sharp operating transitions.

Actual and predicted State of Health using a DNN digital twin — The DNN results show how closely the predicted health trajectory follows the battery's measured degradation pattern.

Actual and predicted State of Health using an LSTM digital twin — The recurrent model provides a second view of health estimation, allowing prediction accuracy and explanation quality to be compared across architectures.