Hand-crafted features beat raw sequences when battery protocols shift

Every lithium-ion cell in an electric vehicle carries a hidden countdown. Long before capacity visibly fades, subtle electrochemical fingerprints in early charging cycles already encode how quickly the battery will age, when its degradation curve will bend sharply downward at the so-called knee, and when it will reach end of life. Getting that prognosis right from the first few dozen cycles is one of the most consequential prediction problems in energy storage, because it determines warranty costs, second-life decisions, and the safety margins of entire fleets. A new study published in the journal Ionics by Chinedu I. Otuka and colleagues at Chengdu University of Technology asks a deceptively simple question about how these predictions should be framed, and the answer overturns a fashionable assumption in the machine-learning battery community.

The field has split into two camps. The first formulates early-life prognosis as a single-step regression: engineers distill the first hundred or so cycles into hand-crafted macro-features, such as summary statistics of voltage curves, capacity fade slopes, and charge-time descriptors, and a model maps that feature vector directly to a lifetime quantity like end-of-life cycle number. The second camp, inspired by the success of sequence models in language and vision, feeds raw per-cycle descriptors into a recurrent network and lets the model learn temporal structure itself, often adding an adaptive gating mechanism and a temporal-attention layer to squeeze out extra accuracy. Proponents of the multi-step view argue that attention over the cycle sequence should discover degradation patterns that human feature engineering might miss. Until now, however, the two formulations had rarely been compared under one controlled protocol, leaving the field without a rigorous answer to which framing actually earns its complexity.

Otuka’s team ran that comparison on two public datasets with very different characters. The MIT/Severson dataset, the workhorse of data-driven battery prognosis since 2019, contains 138 lithium-ion cells cycled under a range of fast-charging protocols. The CLO dataset contributes 45 validation cells cycled under a different charging protocol altogether, which is precisely the kind of protocol shift that deployed batteries face in the real world: a model trained on one charging regime must generalize to another. The team evaluated eight different recurrent architectures spanning both formulations, reporting end-of-life, knee-onset, and knee-point predictions, the first such benchmark on the CLO validation cells.

The headline result is that the fashionable multi-step reformulation actively hurts under protocol shift. When models trained on the MIT/Severson data were evaluated on the out-of-distribution CLO cells, the multi-step formulation degraded end-of-life accuracy relative to the single-step formulation by between 4.60 plus or minus 4.87 and 7.38 plus or minus 8.47 percentage points of mean absolute percentage error. The authors quantified this penalty with paired t-tests yielding p-values of 0.015 and 0.022, and with Wilcoxon signed-rank tests yielding p-values of 0.010 and 0.027, so the effect is statistically robust rather than a fluke of one unlucky training run. Crucially, the penalty appeared under two different recurrent cell types, meaning it is not an artifact of one particular recurrent implementation.

One might suspect that the multi-step models simply lacked capacity, that their recurrent layers were too narrow to process long raw sequences. The team tested that hypothesis directly by widening the single-step model’s recurrent layers to match the multi-step models. The result: error moved by only 0.05 plus or minus 1.13 percentage points, with a p-value of 0.90, indistinguishable from no change. Model capacity, in other words, does not explain the gap. The problem lies in the formulation itself, not in the size of the network. When a model must learn to extract degradation-relevant structure from raw per-cycle descriptors across a distribution shift, it apparently learns structure that is specific to the training protocol, while hand-crafted macro-features encode physics-informed invariants that transfer more gracefully across charging regimes.

Within the multi-step family, the story gets stranger. The team had hypothesized that an adaptive gate and a temporal-attention mechanism would justify their added complexity by improving sequence modeling. Yet no component effect was resolvable once the analysis accounted for statistical noise. Between-seed variation in the out-of-distribution gap ranged from 4 to 12 percentage points for the multi-step models, which is the same magnitude as the effects attributed to attention, to the gate, and to the recurrent-cell implementation. When the noise floor equals the signal, nothing survives replication. Several components that appeared decisive in single-run comparisons turned out to be indistinguishable from chance once the experiment was properly powered.

That methodological point may be the study’s most viral contribution, because it indicts a practice that pervades the applied machine-learning literature. A single training run gives a single draw of a noisy quantity, and neural network training is notoriously sensitive to random initialization, data shuffling, and other stochastic choices. The authors therefore repeated every architectural comparison over ten random seeds and treated the seed as the unit of analysis, applying paired statistical tests across seeds rather than comparing point estimates. Under this discipline, in-distribution performance told an even starker story: all eight architectures fell within a 11.17 to 11.62 percent end-of-life MAPE range, a spread narrower than the seed-to-seed variation of any single architecture. In other words, the differences celebrated in single-run ablation studies are often smaller than the run-to-run noise of the very models being compared.

The team also delivered a mathematical demolition of temporal attention in this setting. At one time step, attention is a no-op: with a single position to attend over, the softmax weighting collapses to a trivial constant, and the mechanism cannot redistribute information across the sequence. Even at 100 steps, the learned attention weights converged to near-uniform distributions under both additive and multi-head forms, meaning the layer was effectively averaging its inputs rather than selectively focusing on informative cycles. The intuition that attention should discover which early cycles matter most for prognosis, while appealing, found no empirical support in this regime, where the sequence length is modest and the input descriptors are already cycle-level summaries.

Why does this matter beyond the benchmark? Electric vehicle fleets, grid storage installations, and second-life battery markets all depend on early-life prognosis models that must generalize across charging protocols, temperatures, and usage patterns that differ from training data. The study’s evidence says that under such protocol shift, the safer input is the hand-crafted macro-feature vector, not the raw sequence, and that the extra machinery of gates and attention layers buys nothing measurable while adding training variance. This does not mean sequence models are useless in battery prognosis; it means the burden of proof now sits with anyone claiming that architectural complexity improves out-of-distribution generalization, and that claims must survive multi-seed replication rather than a single favorable run.

The study also leaves the community with a concrete gift: the first end-of-life, knee-onset, and knee-point benchmark on the 45 CLO validation cells, establishing reference numbers that future models can be tested against. The authors report that their TensorFlow/Keras code will be released on a public GitHub repository upon publication, and the datasets are publicly available, so the multi-seed evaluation discipline they advocate can be adopted immediately. For a field racing to deploy learned prognostics in safety-critical energy systems, the message is uncomfortable but clarifying: simplicity, statistical rigor, and physics-informed feature engineering remain the most reliable foundation for predicting how long a battery will last, and the flashiest component in the architecture diagram is often just expensive noise.

Subject of Research: Single-step versus multi-step formulations for early-life battery prognosis under protocol shift

Article Title: Single-step versus multi-step formulations for early-life battery prognosis under protocol shift

Article References: Otuka, C. I., Ukwuoma, C. C., Ukwuoma, C. D., Nwosu, C. I., Olughu, D. S., Olugbenel, A. O., Huang, Q., & Cai, D. (2026). Single-step versus multi-step formulations for early-life battery prognosis under protocol shift. Ionics. https://doi.org/10.1007/s11581-026-07578-w

Image Credits: AI Generated

DOI: 10.1007/s11581-026-07578-w

Keywords: lithium-ion battery, machine learning, prognosis, protocol shift, LSTM, attention mechanism, knee-point prediction, statistical replication, Single-step, versus, multi-step, formulations