Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

NOCE-Net: Representation Learning for Battery Operational Context via Nested Sequence Modelling

NeurIPS 2026 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Zero-shot robustness and transfer

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NOCE-Net, a transformer-GRU nested sequence model trained on FEC-segmented battery telemetry (re-BatteryData, ~700 batteries) to predict voltage responses and learn degradation-related latent states.

Problem

Learning generalisable, control-enabling representations from long battery operational histories without response-dependent features or RPTs, for voltage prediction and downstream SOH estimation.

Approach

Segments operational history into full equivalent cycles defined by absolute charge throughput (2x nominal capacity), then uses a multi-patch transformer encoder per cycle and a GRU across cycles, trained by supervised voltage reconstruction; GRU hidden states are then reused for SOH prediction.

Key results

• Zero-shot unseen-battery voltage prediction: NOCE-Net reaches 13.7% MAPE (MAE 0.0656), about 43.8% better MAPE than the next-best baseline (TTM + Indep.) with ~40x fewer parameters than the average baseline.
• PC1 of GRU hidden states (relativised to initialisation) reaches mean |rho| = 1.000 against FEC index and |rho| >= 0.83 against normalised SOH on monotone-degradation datasets.
• A per-dataset MLP head on frozen PID-R features with 30% of in-domain batteries achieves 5-33% MAPE on well-behaved datasets; ISU-ILCC and NASA-Rand remain hard at 72% and 77%.

Surprising findings

• Raw GRU hidden-state norm gives only |rho| = 0.750 on zero-shot test batteries, but relativising to the task-conditioned initialisation raises it to 0.997.
• Few-shot fine-tuning with 100 cycles performed worse than with 50 cycles for NOCE-Net, attributed to overfitting of the projection head.

Limitations

• Voltage supervision captures degradation direction and shape but not absolute scale, requiring per-dataset calibration for SOH.
• Scalar SOH is ill-defined for evTOL-style partial-discharge profiles with non-monotone capacity trajectories.
• re-BatteryData requires current, temperature and time signals; the sequential GRU is a computational bottleneck.

Experimental setup

Datasets: ISU-ILCC NASA-Rand TJU MATR UMICH XJTU-Partial evTOL NASA-PCoE Stanford-EV HUST XJTU-Full

Benchmarks: NASA-PCoE Stanford-EV HUST XJTU-Full evTOL ISU-ILCC NASA-Rand TJU MATR UMICH-Pouch XJTU-Partial

Metrics: MAE MAPE Spearman correlation Pearson shape correlation

Baselines: TimerXL AutoTimes MAMBA TinyTimeMixers

Builds on: TimerXL AutoTimes Mamba TinyTimeMixers

Closest papers by meaning

• BatteryML: An Open-source Platform for Machine Learning on Battery Degradation (ICLR 2024 · similarity 0.87)
• PowerPM: Foundation Model for Power Systems (NeurIPS 2024 · similarity 0.80)
• Reinforced Fast Weights via Next-Sequence Prediction (NeurIPS 2026 · similarity 0.80)
• MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models (NeurIPS 2026 · similarity 0.80)
• The Hidden Joules: Evaluating the Energy Consumption of Vision Backbones for Progress Towards More Efficient Model Inference (ICML 2025 · similarity 0.80)
• Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors (ICLR 2024 · similarity 0.80)