Dual-Timescale Spatiotemporal Deep Learning Network Improves Battery Health Estimation Accuracy

A dual-timescale spatiotemporal deep learning architecture accurately estimates lithium-ion battery state of health by coupling intra-cycle operational dynamics with inter-cycle aging trajectories, according to a study published in Results in Engineering. In systematic benchmark evaluations across standard cycling datasets, the proposed neural network reduced state-of-health estimation errors by 10 to 60 percent compared to conventional Transformer and temporal convolutional network baselines under chronological testing. Built specifically around partial charging and discharging regimes representative of real-world electric vehicle use, the architecture extracts health features from segmented voltage, current, and incremental capacity measurements without requiring full laboratory charge cycles. By combining correlation-guided sequence winnowing with dedicated convolutional, attentional, and recurrent neural branches, the framework tracks progressive capacity loss and electrochemical changes while maintaining stability against sensor noise and systematic measurement perturbations.

Lithium-ion batteries serve as the dominant energy storage technology in electric vehicle traction applications due to their high power and energy density, extended operational lifespan, and cost-effective maintenance. However, operational cycling inevitably causes progressive capacity degradation and internal resistance increases over extended service life. If left unmonitored, this deterioration diminishes vehicle driving range, degrades power delivery, and can escalate into critical safety risks such as toxic gas emissions, combustion hazards, and catastrophic thermal runaway under extreme conditions. Accurate state-of-health monitoring is therefore vital for onboard battery management systems to ensure safe operation, optimize performance management, and establish reliable criteria for repurposing retired automotive batteries into second-life stationary grid storage applications.

Despite its importance, estimating battery state of health under real-world operating conditions presents substantial engineering hurdles. Battery degradation arises from coupled electrochemical, thermal, and mechanical mechanisms that interact across multiple scales. These degradation processes are further complicated by stochastic operating stresses, including ambient temperature fluctuations, dynamic state-of-charge swings, and irregular driver load profiles. Existing research strategies generally rely on physics-based modeling or data-driven computational methods. Physics-based techniques, such as detailed electrochemical models and equivalent circuit models, offer distinct characteristics. Full electrochemical formulations explicitly represent degradation mechanisms, including active lithium inventory loss, side reaction rates, and solid-electrolyte-interphase growth. However, their heavy computational demands and challenging parameter identification make them impractical for real-time onboard estimation. Equivalent circuit models offer lower computational complexity but remain sensitive to initial assumptions and parameterization errors, where minor identification inaccuracies propagate into unacceptable estimation errors.

Data-driven machine learning provides an alternative approach that bypasses full mechanistic modeling by learning directly from operational sensor measurements. However, pure data-driven approaches require massive, diverse training datasets covering extensive operational conditions, which is rarely feasible in automotive development. Feature-based techniques address this limitation by extracting health features correlated with capacity fade prior to model training, enabling simpler architectures and smaller datasets. Even so, conventional feature extraction often assumes access to complete charge or discharge cycles via Coulomb counting. In everyday electric vehicle operations, drivers rarely fully charge or deplete their packs, operating instead within a partial window of approximately 20 to 90 percent state of charge. To make state-of-health monitoring practical for vehicle operations, feature extraction must be adapted to these partial operational ranges without depending on complete cycling records.

To address this practical operating constraint, the authors formulated a correlation-guided characteristic sequence winnowing module operating within the typical 20 to 90 percent electric vehicle state-of-charge window. Rather than relying on fixed, predefined acquisition windows, the researchers established eight candidate ranges for voltage, current, and incremental capacity segments. Candidate ranges were evaluated using a single-cycle diagnostic method trained and validated on subsets of the NASA battery dataset, assessing correlation through root mean square error and mean absolute error metrics. To enrich the feature representations and suppress polarization and resistance effects, the team integrated voltage-to-state-of-charge relationships directly into incremental capacity curves, generating multidimensional incremental capacity sequences. This systematic winnowing isolated near-optimal data capture ranges that eliminate dependence on complete cycling profiles.

The winnowing analysis revealed optimal operational ranges that diverged from conventional engineering assumptions. For voltage segments during constant-current constant-voltage charging, maximum correlation appeared in the 4.0 to 4.2 volt range near the transition point, where progressive leftward shifts across lifespan demonstrated sensitivity to aging. For current sequences, optimal correlation occurred within the 1.0 to 0.3 ampere interval, positioned further from the transition point than initially anticipated. Transition-proximal current ranges performed worse, exceeding the optimal range by 0.14 to 0.27 percent in lowest root mean square error. In contrast, current trajectories between 1.0 and 0.3 amperes converged progressively across battery life, yielding an optimal single-cycle mean root mean square error of 0.65 percent. For incremental capacity sequences, the optimal range was identified between 3.6 and 4.0 volts, capturing the descent of the primary peak, the inter-peak valley, and the secondary peak while excluding the ascending slope.

The core architecture consists of an intra-cycle health feature extraction network, a multi-characteristic feature fusion layer, and an inter-cycle degradation trajectory network. In the intra-cycle stage, winnowed voltage, current, and multidimensional incremental capacity sequences are processed independently by dedicated neural branches. Each branch employs a one-dimensional convolutional neural network with 32 filters of kernel size 5 to extract spatial features and reduce dimensionality, followed by batch normalization and rectified linear unit activations. To preserve sequence ordering within the attention mechanism, sinusoidal positional encodings are added to the normalized feature maps. A multihead self-attention module with four heads computes importance weights across sequence elements, retaining the top 32 highest-weighted features. These selected representations are then passed into an intra-cycle Long Short-Term Memory network with 256 hidden units, from which flattened hidden states are mapped by fully connected layers into branch-specific health feature vectors.

To prevent reliance on any single measurement channel, the architecture implements a multi-characteristic feature fusion strategy that unites complementary degradation cues. Because battery aging manifests differently across voltage shifts, current convergence, and electrochemical peaks, integrating multiple signal sources reduces vulnerability to sensor disturbances affecting any individual channel. The intra-cycle network projects the extracted health features into compact representations: three features from the voltage branch, three from the current branch, and nine from the incremental capacity branch. For each cycle, these branch outputs are concatenated into a unified 15-dimensional per-cycle feature vector. The authors clarify that this concatenation represents feature-level integration rather than conventional machine-learning meta-learning, directly fusing distinct physical degradation perspectives into a consolidated health snapshot.

The consolidated feature vectors across five consecutive cycles form a cycle-related feature matrix that serves as input to the inter-cycle degradation trajectory network. An inter-cycle Long Short-Term Memory network processes this sequence of fused vectors to model multi-cycle aging trajectories, after which a final fully connected layer maps the learned representation to a scalar state-of-health estimate. This hierarchical separation enables the network to capture rapid operational dynamics within individual cycles while filtering out cycle-to-cycle transient fluctuations through long-term temporal modeling. In baseline ablation tests, the dual-timescale multi-cycle approach achieved 20 to 50 percent lower root mean square errors than a single-cycle baseline lacking inter-cycle tracking across all tested signals, reducing voltage estimation root mean square error by 0.006 to 0.011 across experimental battery cells.

The proposed framework was validated using two independent lithium-ion battery cycling repositories sharing nickel manganese cobalt cathode chemistry: the NASA battery dataset and the Oxford battery dataset. The NASA dataset contains 2.0-ampere-hour cells cycled at room temperature under constant-current constant-voltage charging and constant-current discharging. The Oxford dataset comprises eight Kokam 740-milliampere-hour pouch cells cycled at 40 degrees Celsius under drive-cycle discharging, with characterization measurements every 100 cycles. Under a chronological evaluation setting using a 70 percent training split and a held-out 30 percent testing portion, the proposed model consistently outperformed Transformer and temporal convolutional network benchmarks across twelve cells. The model reduced estimation errors by 10 to 60 percent, decreasing root mean square error by approximately 0.002 to 0.015, lowering mean absolute error by 0.002 to 0.013, and increasing coefficients of determination up to values between 0.9985 and 0.9999 while accurately tracking local capacity regeneration fluctuations.

To evaluate cross-cell generalization without model retraining, the researchers performed leave-one-cell-out cross-validation across the eight Oxford cells, training on seven cells and testing on the single unseen cell. The proposed architecture reduced estimation errors by up to 50 percent relative to benchmark models at most test cells, achieving root mean square error reductions between negative 0.002 and 0.006. Although the temporal convolutional network achieved lower root mean square error on two specific cells, the proposed model maintained tighter tracking during late-stage degradation around 80 percent state of health where benchmarks showed notable deviations. Furthermore, under perturbation tests with injected additive white Gaussian noise and five percent systematic sensor bias, the fused model maintained maximum state-of-health errors below 0.02 and root mean square error differences below 0.0053 across tested cells.

The authors noted several important constraints, emphasizing that the framework was evaluated on laboratory datasets characterized by regular cycling protocols and calendar-dominated aging. Practical deployment must still confront cell-to-cell variability from manufacturing tolerances and complex real-world duty cycles. Furthermore, the architecture remains predominantly data-driven, embedding battery physics only implicitly through incremental capacity sequences and correlation-guided window winnowing. Incorporating broader multiphysics measurements, such as operating cell temperature variations and dynamic impedance characterization, will also be pursued to enhance the physical interpretability and practical generalizability of state-of-health estimation in commercial battery management systems.

Subject of Research: Dual-timescale spatiotemporal deep learning for lithium-ion battery state-of-health estimation

Article Title: Dual-timescale spatiotemporal learning architecture for battery state-of-health estimation

Article References: Cai, C., Gong, Y., Fotouhi, A., Auger, D. J., Wang, T., Wang, B., & Zhang, T. (2026). Dual-timescale spatiotemporal learning architecture for battery state-of-health estimation. Results in Engineering, 32, Article 114042. https://doi.org/10.1016/j.rineng.2026.114042

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.114042

Keywords: lithium-ion batteries, state of health, deep learning, dual-timescale network, spatiotemporal modeling, battery management systems, incremental capacity analysis, convolutional neural networks, long short-term memory, self-attention mechanism, feature fusion, degradation trajectory, electric vehicles, NASA battery dataset, Oxford battery dataset, characteristic sequence winnowing, capacity fade