Dual-Stream AI Network Predicts Battery Heat and Health to Warn of Thermal Risks Early

Lithium-ion batteries power everything from electric vehicles to grid-scale storage, but their safety depends on two intertwined processes that most monitoring systems treat separately: how hot the cell gets and how quickly it ages. A research team from Zhengzhou University, China Resources New Energy Investment, and XJ Electric has now built a single artificial intelligence framework that predicts both at once, and then converts those predictions into early warnings of dangerous thermal behavior. In tests on two public battery datasets, the model, called DS-MTNet, estimated battery health with a mean absolute error of 0.011 and maximum temperature with an error of 0.898 degrees Celsius, and its warning scheme delivered a timely response in 92.6 percent of evaluated cases, flagging problems an average of 48.6 cycles before the defined thermal event began.

The work, published in Frontiers in Energy Research, addresses a persistent weakness in battery management systems. State of Health, or SOH, describes how much of a battery’s original capacity remains, while thermal state describes how the cell’s temperature behaves under load. Existing data-driven approaches usually model these targets independently, which the authors argue leads to insufficient coupling awareness, restricts practical warning decisions, and yields poor generalization when a model trained on one set of conditions is applied to another. Because battery aging and thermal behavior are physically linked, temperature characteristics can even serve as indicators of aging, so separating the two problems throws away information that a joint model can exploit.

The core innovation is a dual-stream temporal encoder that assigns each physical timescale to the architecture best suited to it. Battery temperature responds rapidly to load changes within a single charge-discharge cycle, driven by fast electrochemical and transport processes such as ohmic heating and reaction kinetics. Degradation, by contrast, unfolds slowly and cumulatively over hundreds of cycles through mechanisms like solid-electrolyte-interphase growth, lithium inventory loss, and active-material deterioration. A single model with a fixed inductive bias struggles to capture both. DS-MTNet therefore runs two parallel branches: a Temporal Convolutional Network, or TCN, with stacked dilated causal convolutions captures short-range, high-resolution thermal fluctuations, while a Mamba-based selective state-space model tracks long-range degradation evolution with linear computational complexity.

Before temporal processing begins, a Squeeze-and-Excitation block recalibrates the input features channel by channel, amplifying informative signals and suppressing irrelevant ones. The outputs of the two streams are then combined by an adaptive gated fusion module rather than a simple concatenation or fixed average, because the relative importance of local and global features varies between batteries and across cycling stages. A data-dependent gate vector weighs each stream at every time step, producing a joint representation that feeds four parallel prediction heads, each a small multilayer perceptron. These heads regress maximum temperature, mean temperature, SOH, and an auxiliary temperature-difference target whose non-negativity is enforced by a Softplus activation, encouraging physically consistent thermal outputs.

Two learning strategies sharpen the model where it matters most. First, because battery datasets are dominated by normal-condition samples, the team introduced a High-Temperature Weighting mechanism that increases the loss contribution of samples in elevated-temperature regions, with weights of 1.5 and 2.5 selected by a coarse grid search on the validation set. The thresholds are anchored to recognized safety reference points, marking where aging accelerates and where thermal stress becomes a genuine safety concern. Second, an auxiliary temperature-difference prediction task encourages consistency between the model’s thermal outputs. Together these choices improve reliability in exactly the safety-critical regions where prediction errors carry the highest risk, a point the ablation study confirmed: removing the weighting mechanism notably increased high-temperature errors under dynamic conditions.

Generalization across operating conditions and environments posed a harder challenge. Data-driven models often lose accuracy when temperature, C-rate, or testing environments change, a phenomenon known as domain shift. The team’s answer is a few-shot domain adaptation strategy built on Maximum Mean Discrepancy, or MMD, a measure of the distance between feature distributions in a reproducing kernel Hilbert space. During pre-training, an MMD regularization term aligns the encoder’s representations from a source domain with those of a target domain, encouraging domain-invariant features. After alignment-based pre-training, the model is fine-tuned on a small number of labeled target-domain samples. The authors are careful to note that this makes the framework a few-shot adaptation approach rather than a fully unsupervised one; the alignment reduces but does not eliminate the need for target supervision.

The evaluation used two public datasets. The XJTU dataset contains 18650-format nickel-manganese-cobalt cells, split into 23 cells cycled at a fixed 1.0 C discharge rate and 16 cells discharged at rates varying among 0.5, 1, 2, 3, and 5 C across cycles. The NASA dataset, from the Ames Research Center, provided a harsher cross-environment test: one group of cells cycled at 24 degrees Celsius and another at 4 degrees Celsius, with different discharge cut-off voltages, creating a 20-degree temperature shift. Fourteen physics-based features were extracted per cycle from constant-current charging, constant-voltage charging, and discharging segments, including charging time, input energy, rates of voltage change, Shannon entropies of current and voltage, and discharge statistics. Where the NASA records lacked certain charging-phase details, zero-padding preserved a unified feature dimension, with the MMD alignment module neutralizing the resulting dataset-specific signature in the latent space.

Under fixed and dynamic conditions on the XJTU dataset, DS-MTNet achieved the best overall balance among compared models, including LSTM, GRU, Transformer, CNN-LSTM, BiLSTM, and PatchTST, with an SOH mean absolute error of 0.011 under dynamic conditions and a maximum-temperature error of 0.898 degrees Celsius. The Transformer baseline actually beat DS-MTNet on mean-temperature prediction under dynamic conditions, illustrating that specialized architectures can excel on individual tasks, but the dual-stream design delivered the strongest synergistic performance across all targets. Ablation experiments confirmed that removing the SE block, the fusion gate, or either temporal stream degraded performance, and that the TCN-only variant favored rapid variations while the Mamba-only variant better captured smooth long-term trends, validating the complementarity the architecture was designed around.

Transfer experiments showed the value of explicit distribution alignment. In bidirectional transfers between the fixed and dynamic XJTU subsets, the Transfer-Align strategy achieved the lowest errors in both directions, outperforming not only a no-transfer baseline, which collapsed under domain shift, but also target-only training and standard fine-tuning. The asymmetry of transfer difficulty was informative: adaptation from the richer dynamic domain to the fixed domain benefited more from alignment, suggesting that complex source representations help but also raise the risk of negative transfer. In the NASA cross-environment test, naive transfer failed almost completely, with no-transfer temperature errors above 20 degrees Celsius, while Transfer-Align reached an SOH mean absolute error of 0.022 and a maximum-temperature error of 1.80 degrees Celsius. The authors caution that the cross-environment experiment used a limited number of target cells, so the results demonstrate transfer feasibility under the evaluated conditions rather than exhaustive generalization.

The final stage converted predictions into decisions through a hierarchical early warning system. A temperature baseline is established from the model’s predictions over the first cycles of each battery’s life, and three trigger levels are then evaluated: an absolute trigger when predicted maximum temperature persistently exceeds a critical safety threshold, a strong trigger when the moving average deviates sharply from baseline, and a weak trigger based on a robust Z-score that detects gradual but persistent anomalies. Across 27 evaluated test cases, the system produced 25 timely hits, one slightly late warning within tolerance, and one late warning, with no missed events, yielding the 92.6 percent strict timely response rate and an average lead time of 48.6 cycles, plus or minus 28.4. The authors note these statistics are descriptive rather than population-level probability estimates, given batch-wise correlations within datasets. Delayed warnings tended to occur in batteries with early-onset thermal events, which left too little time for statistical evidence to accumulate, pointing toward adaptive baseline windows as future work alongside unsupervised domain adaptation and model compression for embedded deployment.

Subject of Research: Joint thermal and state-of-health prediction and early warning for lithium-ion batteries using a dual-stream multi-task neural network

Article Title: Cross-domain thermal-health prediction and early warning for lithium-ion batteries via dual-stream network

Article References: Tian, P., Li, X., Yang, X., Wang, B., & Liu, Y. (2026). Cross-domain thermal-health prediction and early warning for lithium-ion batteries via dual-stream network. Frontiers in Energy Research, 14, 1975958. https://doi.org/10.3389/fenrg.2026.1975958

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1975958

Keywords: lithium-ion batteries, state of health, thermal prediction, early warning system, dual-stream network, Mamba, temporal convolutional network, domain adaptation, battery management systems, machine learning, thermal runaway risk, transfer learning