⚡ electrical engineering

StateFormer: A Multivariate Transformer for Learning History-Dependent Battery State Dynamics and Long-Horizon Health Forecasting

This paper introduces StateFormer, a novel multivariate Transformer model that unifies short-term thermal and electrochemical dynamics with long-term aging mechanisms to accurately forecast battery states (SOC, SOH, and temperature) across diverse synthetic and real-world datasets, effectively bridging the gap between laboratory simulations and field operations for optimized maintenance and decision-making.

Original authors: Zhe Bai, Stephen Harris

Published 2026-07-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhe Bai, Stephen Harris

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 the power grid as a giant, bustling city where electricity is the traffic. In the past, this traffic flowed one way: from big power plants to your home. But today, the city is changing. We are adding millions of "battery cars" that can store energy and send it back when needed. These aren't just tiny AA batteries; they are massive fleets of energy storage systems, like giant warehouses of power, sitting on rooftops and in industrial yards.

However, just like a car engine, these batteries don't last forever. They get tired. They heat up, they cool down, and they slowly lose their ability to hold a charge. This process is called "degradation." The tricky part is that batteries are complicated. They react to how hot it is outside, how fast you are charging them, and even how long they've been sitting idle. It's a messy, tangled web of chemistry and physics that changes over years. Scientists and engineers really need to know: "How much life is left in this battery?" and "When will it break?" If they can predict this, they can keep the lights on, save money, and make sure the grid doesn't crash. But guessing is hard because every battery is a little different, and the data we have is often noisy or incomplete.

This is where a new tool called StateFormer comes in. Think of StateFormer as a super-smart, time-traveling detective for batteries. Instead of just looking at a battery's current temperature or voltage, it reads the battery's entire diary. It looks at the past few hours, days, or even years of how the battery was used, combined with the weather and the electricity it was asked to carry. Using a type of artificial intelligence called a "Transformer" (the same kind of technology that helps computers understand language), StateFormer learns to spot patterns in the chaos. It figures out how fast a battery is aging based on its history, even if the sensors measuring it are a bit fuzzy or if the battery is facing extreme heat it hasn't seen before.

The researchers behind StateFormer tested this detective in two very different ways. First, they built a giant, virtual fleet of 50 batteries in a computer simulation. They made these virtual batteries age over three years, subjecting them to different temperatures (from a cool 25°C to a scorching 45°C) and adding random "noise" to the data to mimic real-world sensor errors. Even when the data was messy and the temperatures were higher than what the AI had ever seen during its training, StateFormer predicted the batteries' health with incredible accuracy. It was so good that the difference between its guess and the "true" answer was tiny—less than 0.03% in many cases.

Then, they took the detective into the real world. They fed it five years of actual data from a residential home battery system that had been running alongside solar panels. The AI had to predict what the battery's voltage and temperature would do in the future, based only on the current and the room temperature. Just like in the simulation, it nailed the long-term trends. It correctly predicted how the battery would behave over a whole year, even capturing sudden changes in how the battery was used.

What makes StateFormer special is that it doesn't just guess; it learns the rules of the game. It understands that a battery's health depends on a mix of fast things (like a sudden spike in heat) and slow things (like the slow, years-long process of aging). It also handles the fact that no two batteries are exactly alike, thanks to tiny differences in how they were made. The authors show that this model can bridge the gap between clean, perfect lab simulations and the messy, unpredictable reality of real-world batteries. By doing this, StateFormer offers a way to keep a close eye on huge fleets of batteries, helping operators know exactly when to replace them, how to run them safely, and how to keep our future energy grid running smoothly for years to come.

Technical Summary: StateFormer – A Multivariate Transformer for Learning History-Dependent Battery State Dynamics and Long-Horizon Health Forecasting

Problem Statement
The rapid integration of large-scale Battery Energy Storage Systems (BESS) into future grid infrastructures necessitates reliable long-term forecasting of battery degradation. Accurate prediction of states such as State of Charge (SOC) and State of Health (SOH) is critical for system reliability, maintenance optimization, and asset lifetime extension. However, existing challenges include:

Complex Dynamics: Battery degradation involves deeply coupled electrochemical, thermal, and operational feedback loops that evolve over multi-year timescales under diverse conditions.
Data Limitations: Current data-driven approaches are often constrained by limited training data. Laboratory datasets fail to capture real-world heterogeneity, while field data are frequently proprietary, incomplete, or lack ground-truth degradation labels.
Homogeneity Assumptions: Many existing methods assume homogeneous degradation across entire fleets, ignoring significant variability caused by manufacturing tolerances, thermal gradients, and diverse operating conditions.
Robustness Requirements: Models must remain resilient to measurement noise, abrupt operational transitions, and discontinuous system behaviors while maintaining accuracy over long horizons.

Methodology
The authors propose StateFormer, a novel multivariate Transformer architecture designed as a state-space model for long-horizon battery state forecasting. The framework integrates state-space representations with long-range attention mechanisms to model coupled electrochemical and thermal dynamics.

Architecture: StateFormer operates as an autoregressive Transformer. It takes a historical window of battery operating states ( $L_{in}$ ) and the latest operating conditions ( $\Upsilon_{t+1}$ ), which include instantaneous variables like current and ambient temperature.
Mechanism: Through self-attention mechanisms, the model dynamically identifies relevant historical degradation patterns while positional encoding preserves temporal ordering. The model learns a state transition function $\hat{S}_{t+1} = F(\hat{S}_{t-L_{in}+1:t}, \Upsilon_{t-L_{in}+1:t+1})$ , where $S$ represents multivariate outputs (SOC, SOH, temperature) and $\Upsilon$ represents operating inputs.
Training Protocol: The model is trained to learn nonlinear degradation dynamics directly from data without requiring explicit electrochemical or empirical aging equations. Crucially, evaluation is performed via autoregressive rollout without state correction from future measurements. For synthetic data, only the first 50 hours are provided for initialization; for field data, only the first 120 minutes are used. The model then recursively predicts future states based solely on its own previous predictions and measured inputs.

Key Contributions and Results
The paper validates StateFormer using two distinct datasets: a physics-grounded synthetic battery fleet and five years of real-world field data from a residential utility battery system.

Synthetic Fleet Evaluation:
- Dataset: Generated from hour-resolved, multi-decade simulations coupling stochastic environmental conditions, market-driven dispatch, and fleet heterogeneity (50 unique cells across 5 ambient temperatures: 25°C to 45°C).
- Generalization: The model was trained on lower temperatures (25–35°C) and evaluated on unseen elevated temperatures (40°C and 45°C). It successfully captured long-term degradation trajectories, maintaining strong agreement with ground truth.
- Noise Resilience: The model demonstrated robustness against additive Gaussian noise in current and voltage signals ranging from 1% to 10%. Relative $L_2$ errors remained on the order of $2 \times 10^{-2}\%$ , with 95% confidence intervals below $1 \times 10^{-4}\%$ .
- Statistical Accuracy: Predictions for SOC and SOH at the beginning and end of life showed maximum discrepancies of approximately 0.02% compared to reference values.
Field Data Evaluation:
- Dataset: Five years of measurements from a residential home storage system (HSS) integrated with rooftop PV. The dataset was split chronologically: 35 months for training, 7 months for validation, and 10 months for testing.
- Performance: The model accurately reproduced long-term temporal evolution of battery voltage and temperature, including abrupt operational transitions.
- Metrics: Across five independent training realizations, the model achieved an average relative $L_2$ error of approximately 0.05% for voltage and temperature predictions, with measured values consistently remaining within predicted uncertainty bands.

Significance and Claims
The authors claim that StateFormer bridges the "laboratory-to-field" gap by providing a unified framework that captures both fast electrochemical/thermal processes and slow aging mechanisms. The significance of this work lies in:

Robust Long-Horizon Forecasting: The ability to perform accurate, uncertainty-aware predictions over years without access to future ground-truth states.
Generalization: Demonstrated capability to generalize across diverse operating conditions, unseen temperature regimes, and measurement uncertainties.
Practical Application: The resulting model serves as a foundation for predictive intelligence in maintenance planning, operational optimization, and economic decision-making for grid-scale and residential storage.
Digital Twin Capability: Beyond forward prediction, the framework acts as a computationally efficient surrogate for inverse degradation modeling, enabling the synthesis of input profiles to achieve target battery lifetimes.

The paper concludes that next-generation energy management systems require such AI frameworks to move beyond static, single-asset approximations, instead capturing coupled dynamics and fleet-level heterogeneity to ensure the reliability and economic efficiency of large-scale energy storage.

More like this