BatteryMFormer: Multi-level Learning for Battery Degradation Trajectory Forecasting

Technical Summary: BatteryMFormer: Multi-level Learning for Battery Degradation Trajectory Forecasting

Problem Statement
Early battery degradation trajectory forecasting (BDTF) aims to predict the full-life State-of-Health (SOH) trajectory of a battery using only early-stage operational data. This capability is critical for optimizing battery manufacturing, deployment, and maintenance. However, existing approaches face two primary limitations:

1. Failure to Model Multi-level Structure: Current methods often operate at the individual battery level, neglecting the regularities shared within specific aging conditions (e.g., chemical composition, operating protocols) and the shared trajectory patterns (e.g., linear, superlinear, sublinear) that exist across different batteries.
2. Neglect of SOC-Localized Variations: Degradation-related variations in voltage-current profiles are often concentrated within specific State of Charge (SOC) intervals due to underlying electrochemical mechanisms (e.g., phase transitions). Most existing models either treat SOC intervals uniformly or focus solely on temporal dynamics, thereby diluting these localized signals.

Methodology: BatteryMFormer
To address these gaps, the authors propose BatteryMFormer, a multi-level Transformer architecture designed to integrate aging-condition regularities, trajectory patterns, and battery-specific dynamics. The model consists of three core components:

1. Dual-View Encoder: This module processes early operational data (voltage and current time series) into two complementary representations:

   • Temporal View: Captures intra-cycle dynamics by summarizing each cycle into a cycle-level token using a patching mechanism and an intra-cycle encoder. It also incorporates cycle-level descriptors (e.g., Coulombic efficiency).
   • SOC View: Captures SOC-localized degradation signatures by modeling cross-cycle evolution within specific SOC intervals. It reshapes data along the SOC axis, applies 1D convolution, and aggregates information across cycles to produce SOC tokens.

2. Aging-Condition-Aware Decoder (ACDecoder): This component injects aging-condition priors to promote consistent representations for batteries under similar conditions. It utilizes two mechanisms:

   • Aging-Condition-Informed Queries: It converts structured aging-condition metadata (e.g., electrode materials, protocols) into text prompts, encodes them using a language-based embedder (LLM), and adds the resulting embeddings to learnable generic queries.
   • Aging-Condition-Aware Attention: It modulates the attention mechanism by adding the aging-condition embedding to the query vectors, ensuring that attention operations are conditioned on the specific aging context.

3. Meta Degradation Pattern Memory (MDPM): Inspired by memory networks, this module stores and retrieves prototypical trajectory patterns (meta degradation patterns) to guide long-horizon forecasting.

   • Retrieval: The decoder output is used to query the memory slots via cosine similarity, retrieving the top-2 relevant trajectory prototypes.
   • Learning: The model is trained to align retrieved pattern embeddings with full-life trajectory embeddings (using a trajectory encoder/decoder pair) and reconstruct the trajectory, ensuring the memory captures diverse global shapes (e.g., initial capacity rise, capacity regeneration).
   • Fusion: The retrieved pattern embedding is fused with the decoder output via a gated mechanism to produce the final SOH prediction.

Key Contributions

• Formalization of Multi-level Structure: The paper identifies and formalizes the multi-level nature of early BDTF, explicitly modeling aging-condition regularities, shared trajectory patterns, and SOC-localized degradation signatures.
• Novel Architecture: The proposal of BatteryMFormer, which uniquely integrates an aging-condition-aware decoder, a meta degradation pattern memory, and a dual-view encoder (temporal and SOC perspectives).
• Comprehensive Evaluation: Extensive experiments conducted on four distinct battery domains (Li-ion, CALB, Na-ion, and Zn-ion) from the largest public real-world battery lifetime database.

Experimental Results
The authors evaluated BatteryMFormer against state-of-the-art baselines, including battery-specific models (e.g., IC2ML, CPTransformer) and generic time-series forecasters (e.g., Informer, PatchTST, TimeMixer++).

• Performance: BatteryMFormer consistently outperformed all baselines across all four domains and metrics (MAPE and MAE). For instance, on the Li-ion domain, it reduced MAPE by 11.07% and MAE by 10.94% compared to the second-best model.
• Generalizability: The model demonstrated robust performance under "aging-condition-exclusive" testing, where test batteries came from aging conditions unseen during training.
• Data Efficiency: In experiments where only 50% of the training batteries were retained, BatteryMFormer maintained superior performance, indicating its ability to extract informative patterns from limited lifetime data.
• Ablation Studies: Removing any of the three main components (SOC view, MDPM, or ACDecoder) resulted in performance degradation across all domains, confirming the necessity of each module. Case studies further validated that the model successfully retrieves relevant trajectory prototypes and attends to SOC intervals corresponding to known electrochemical degradation features (e.g., differential voltage peaks).

Significance
The paper claims that BatteryMFormer marks a significant step toward reliable early BDTF by explicitly modeling the multi-level structure inherent in battery degradation data. By integrating aging-condition priors and learning from meta degradation patterns, the model overcomes the limitations of treating BDTF as a generic time-series forecasting problem. The authors emphasize that their approach provides a more robust framework for battery optimization and deployment, particularly in scenarios with diverse aging conditions and limited training data. The code is made publicly available to facilitate further research in battery informatics.