Towards reliable energy storage: a review of machine learning-driven SOH estimation for electrochemical energy storage systems

Data quality and preprocessing

In practical SOH estimation, data often determine the upper bound of model performance before model architecture does. Even a sophisticated learning framework can only be as reliable as the labels it is trained on, the signals that are observable in real operation, and the range of operating conditions represented in the training set. For this reason, recent methodological work has increasingly treated data foundations as a primary technical issue rather than a background condition, emphasizing that label quality, observability, and operating-domain coverage jointly define the credibility of SOH estimation results^[8,41,42]. A core difficulty is that real electrochemical ESS rarely provide laboratory-grade observability. In most deployments, continuously available measurements are limited to voltage, current, and temperature, sometimes supplemented by control commands, operating modes, and fault logs. By contrast, measurements that are highly informative for degradation, such as impedance spectra, pressure, or gas evolution, are usually unavailable online or only obtainable during maintenance. This mismatch between field observability and degradation physics is one of the main reasons why strong performance in laboratory studies does not automatically translate to robust field deployment^[43]. Nonstationary operation further complicates the data problem. Signatures learned under a single temperature band, charging protocol, or duty profile may lose validity when the same asset later participates in another service or operates under a new control policy. Cross-condition robustness therefore cannot be inferred from a single average error metric. Instead, practical SOH estimation requires explicit consideration of domain shift, source-target similarity, and evaluation settings that separate within-domain fit from across-domain generalization^[44,45]. The variables most commonly used as inputs are voltage, current, and temperature, as they can capture major aging-related factors; however, high-frequency sampling across long operations can accumulate measurement errors and propagate uncertainty into the estimator inputs.

These observations lead to two practical requirements for deployable SOH estimation. First, data quality control should be treated as an integral part of the estimation method rather than as a secondary preprocessing step. Second, benchmarking protocols should reflect field constraints instead of assuming laboratory-grade observability, dense labels, and idealized operating regularity.

Field observability

To connect data limitations with method design, field observability can be organized into three practical levels. The first level is routine observability, which includes continuously available measurements such as voltage, current, temperature, timestamps, operating states, and event logs. These signals support online estimation and are the backbone of most deployable SOH methods, but they often provide only indirect access to degradation. The second level is opportunistic observability, which includes informative segments that arise naturally during operation, such as charge segments, discharge segments, rest periods, relaxation windows, and pulse-like transients induced by control actions. These segments are often more informative than routine streams for estimating practical health proxies, particularly resistance-related and dynamic-response-related quantities. However, their occurrence is policy-dependent, and their repeatability can be poor across sites and services. The third level is maintenance observability, which includes capacity checks, standardized pulse tests, and impedance measurements. These measurements are sparse, but they usually support stronger labels and more interpretable health inference. As a result, the same nominal SOH target can have very different ground-truth quality depending on whether it is derived from routine signals, opportunistic segments, or maintenance diagnostics. This difference is often hidden in published studies and is one of the main sources of inconsistency in cross-paper comparison^[17,18,25].

Open datasets

Open datasets play a central role in ML-driven SOH research because they determine which methods can be reproduced, compared, and stress-tested. At the same time, available datasets are unevenly distributed across chemistries or device families. Lithium-ion datasets are the most mature and abundant, while openly accessible datasets for sodium-ion batteries, aqueous batteries, and supercapacitors remain relatively limited. As a result, the evidence base for ML-SOH is still strongly shaped by lithium-ion-centric benchmarks. Accordingly, cross-chemistry ML-SOH should presently be interpreted more as a transferability problem than as a fully benchmarked consensus across all electrochemical storage families, with sodium-ion batteries, aqueous batteries, and supercapacitors serving mainly as comparative or emerging domains rather than equally mature benchmark ecosystems^[35-38].

HI taxonomy and physical meaning

Thermal- and efficiency-related indicators

Thermal and efficiency-related HIs are sometimes underused in SOH studies, but they are increasingly relevant for real systems. Degradation affects Joule heating, polarization heat, round-trip efficiency, and energy throughput under load. As a result, temperature rise, thermal gradients, heat-generation tendency, or energy-efficiency decline can serve as indirect markers of health evolution, especially in applications where operational safety and thermal margin are central concerns. Field-oriented and multimodal studies have increasingly used energy-based and power-based quantities as practical health surrogates, especially in settings where laboratory reference tests are unavailable but operational logs remain accessible^[68]. These indicators are especially valuable because they make clear that HI design should follow the health target rather than the other way around. If the target problem is capacity certification under controlled cycling, curve-derived features may dominate. If the target is thermal-safe dispatch in a real ESS, then temperature- or efficiency-related HIs may be more informative than a single capacity proxy. A related implication is that a useful HI taxonomy is not merely descriptive. It helps align the feature space with the actual decision context of the storage system. This alignment is one of the main differences between benchmark-driven SOH estimation and deployment-oriented SOH estimation^[18].

Figure 3 illustrates how practically available signals are transformed into different HI representations and how these representations acquire distinct health meanings in batteries and supercapacitors.

Figure 3. Health indicators and representation pathways across batteries and supercapacitors. ESR: Equivalent series resistance.

While the examples in this Section focus mainly on batteries and supercapacitors, the same representation logic also applies to other electrochemical storage chemistries, although the observability and physical meaning of specific indicators may differ.

Device-specific signatures

Although the HI taxonomy above provides a common representation framework, the physical meaning and practical value of a given indicator remain device-dependent. This distinction is particularly important for electrochemical ESS, where batteries and supercapacitors may share observable signals but differ substantially in health semantics and service objectives. A device-specific discussion is therefore necessary before moving from representation to model design.

For batteries, HI design is typically organized around four degradation manifestations: capacity fade, resistance growth, polarization increase, and loss of energy or power delivery under load. Curve-derived features are useful because they reflect changes in electrochemical staging and reaction pathways, while resistance- and impedance-related features capture internal transport loss and degraded power capability. In practical terms, battery HIs are meaningful when they link measurable signal distortion to a health consequence that matters for operation - for example, reduced usable energy, increased voltage sag, or stronger heat generation during the same duty cycle. This is why battery SOH estimation literature has progressively shifted from single-capacity proxies toward multi-HI formulations that combine energy, resistance, and temporal descriptors. Interpretable short-window voltage and feature-parameter studies similarly show that these manifestations are often best captured by compact indicators with clear physical meaning rather than by indiscriminate feature accumulation^[69].

Another important feature of battery HIs is that they often evolve across relatively long horizons. Many useful indicators become apparent only after repeated cycling, cumulative throughput, or staged changes in charging behavior. This gives battery health signatures a strong temporal structure and partly explains why sequential learning models perform well once meaningful HIs have been extracted. Yet the longer-timescale nature of battery degradation also means that apparent HI stability on laboratory datasets can be misleading if the same indicator depends strongly on protocol regularity. This point is illustrated in both domain-guided and real-world studies. The feature-selection study in^[58] explicitly evaluates IC features under robustness constraints, while^[70] extracts health-related indicators from online vehicle driving data. Taken together, these studies suggest that physically informed and operationally extractable indicators are often more valuable for deployment than large but fragile feature pools.

For supercapacitors, health semantics are fundamentally different. Their value is tied much more directly to high-power buffering, rapid charge-discharge behavior, and cycling endurance than to high specific energy. Accordingly, the most meaningful HIs are usually capacitance retention, equivalent series resistance (ESR) growth, leakage or self-discharge behavior, and transient power deterioration. Reviews on supercapacitor management and reliability consistently treat these variables as the core observables for monitoring, protection, and lifetime assessment. In other words, a supercapacitor is rarely judged healthy simply because it can still store charge; it must also preserve low-loss, fast-response, and thermally acceptable high-power behavior^[71,72]. This supercapacitor-centered view is also reinforced by recent review work on reliability, management, and deployment constraints in power-buffering applications^[72,73].

This distinction becomes even more important when aging phenomena are considered. During supercapacitor aging, performance degradation is typically observed as a decrease in capacitance and an increase in ESR, but this evolution is not always monotonic. Under power cycling, rest periods can induce partial recovery, especially in apparent capacitance, while ESR may also change with interruption duration^[32]. Temperature and voltage stress are repeatedly identified as dominant drivers of supercapacitor aging, and self-discharge or leakage behavior can become a reliability concern in long-duration idle conditions. Therefore, supercapacitor HIs must be interpreted with attention to duty cycle, rest behavior, and stress condition, rather than borrowed directly from battery aging practice^[35,74]. Batteries and supercapacitors do share observable signals such as voltage, current, temperature, and operational history. They also share a methodological affinity with data-driven learning, because both device classes generate rich time-series information. However, shared signals do not imply shared health semantics. The same voltage response can indicate energy fade in a battery but rising ESR in a supercapacitor. The same current pulse may reveal polarization growth in one device and fast power degradation in the other. The term “capacity” can be misleading if transferred across device classes without qualification: in batteries, it is tied to deliverable charge over a relatively broad energy window; in supercapacitors, capacitance is tied much more directly to charge-voltage proportionality and short-timescale power support. Consequently, the value of a feature depends not only on how easy it is to extract, but also on whether its physical interpretation remains valid for the target device class^[71,72]. Recent degradation studies likewise emphasize that capacitance fade, ESR growth, and lifecycle dispersion depend strongly on temperature, duty profile, and mission structure^[75-77].

For this reason, a battery-centered HI framework should not be extended to supercapacitors by analogy. This is also why the present Section emphasizes feature representation rather than generic feature engineering. A feature should be judged not only by predictive utility, but also by semantic validity within the target electrochemical system.

The same principle extends beyond the battery-supercapacitor contrast: in lead-acid batteries, degradation signatures must be interpreted against sulfation- and corrosion-dominated aging, whereas in flow batteries the meaning of observable signals is more tightly coupled to electrolyte-side imbalance and flow-dependent operating behavior, so even shared measurements remain informative only when read within the health semantics of the target device class^[10-12].

Robustness of HIs under domain shift

A good HI is not merely correlated with SOH; it should remain informative when the operating domain changes. This requirement is difficult to satisfy because many widely used indicators are entangled with temperature, current rate, SOC window, and protocol structure. IC peaks can shift or blur under different charge rates. Voltage-time features can stretch or compress with thermal conditions. Pulse-response features can vary due to excitation design rather than actual degradation. The IC feature-selection study in^[60] makes this point particularly clear: once robustness across charge rate and temperature is included in the assessment, the ranking of candidate features changes accordingly. This means that an HI selected under one protocol may not remain optimal under another.

This issue becomes even more severe in real systems because operating windows are rarely complete or repeatable. Electric-vehicle (EV) charging sessions may terminate before the constant-voltage stage stabilizes, stationary ESS dispatch may involve shallow cycling within restricted SOC regions, and module- or pack-level data may exhibit cell inconsistency that is absent from cell-level laboratory experiments. Recent multimodal and real-world studies emphasize that point features extracted in idealized cell tests may not transfer directly to packs and field assets, precisely because operational context changes the observability of degradation. Therefore, HI robustness should be treated as an empirical, deployment-dependent property rather than an intrinsic property of the feature itself^[78,79].

A second source of HI fragility is that feature usefulness often changes over the aging trajectory itself. Some indicators are highly sensitive in early degradation but saturate later. Others remain nearly flat for a long time and then become informative only near the end of life. This stage dependence is well aligned with the broader probabilistic health literature, which argues that uncertainty and interpretability should be assessed together with the phase of degradation rather than assuming uniform feature quality across the full lifecycle. In practice, this means that HI selection should avoid treating all aging stages as statistically homogeneous.

For supercapacitors, the problem is compounded by recovery behavior. Under cycling tests, apparent capacitance can partially recover during rest, while ESR may also respond to interruption duration. As a result, a feature that appears monotonic under uninterrupted stress may become non-monotonic in realistic mission profiles. Batteries can exhibit analogous, though usually less pronounced, issues through relaxation, hysteresis, or usage-pattern transitions. The broader lesson is that “feature monotonicity” is not a universal property; it is conditional on how the device is stressed, observed, and rested. This is one reason why highly ranked benchmark features can still fail once moved into realistic operating schedules^[80,81].

From a deployment perspective, the best HI is rarely the most elaborate one. A robust indicator should satisfy four practical conditions: it should be physically interpretable, extractable from realistically available signals, reasonably stable under expected operating variation, and informative for the health target that matters to the application. This principle favors compact, observable, and semantically meaningful feature sets over exhaustive but fragile combinations. It also explains why recent domain-guided and real-world studies increasingly prefer a small number of physically grounded indicators to large feature pools whose correlations may collapse under data shift^[82].

This Section therefore suggests a shift in emphasis: HI design should be evaluated by deployment value, not only by benchmark utility. In other words, the question is not simply whether a feature can improve root-mean-square error (RMSE) on a curated dataset, but whether it preserves meaning when labels are sparse, operating profiles are irregular, and uncertainty must be managed explicitly. Once this perspective is adopted, feature representation becomes tightly connected to model trustworthiness. This shift in emphasis has an important consequence: representation design cannot be separated from model trustworthiness. Once HI quality becomes deployment-dependent, model choice, evaluation protocol, uncertainty quantification, and conservative fallback are no longer independent algorithmic decisions. They become distinct aspects of the same deployment problem, which motivates the discussion in Section MODELS UNDER DIFFERENT DATA REGIMES.

Model selection under different data regimes

Adaptive and physics-guided learning

Adaptive and physics-guided learning become important when the training setting no longer matches the target deployment condition. In SOH estimation, this mismatch may arise from sparse target-domain labels, changing operating regimes, cross-dataset differences, or the need to preserve physical consistency under limited supervision. Under such conditions, transfer learning, online updating, and physics-guided strategies should be understood not as isolated algorithmic variants, but as different responses to distribution mismatch and model unreliability^[86].

Transfer-learning-based battery capacity prediction provides a representative example of this logic^[101]. Its significance lies not only in the predictive framework itself, but in the assumption that knowledge learned under one data regime can be transferred to another with limited additional supervision. More recent work has made this idea even more explicit. Domain-adaptive deep learning has been used for battery SOH estimation without requiring additional degradation experiments on the target batteries^[102], showing that transferable health representations can still be learned across manufacturers and datasets when adaptation is handled carefully. More broadly, these developments suggest that adaptation is most useful when it addresses representation mismatch and sparse supervision under changing operating conditions. These studies indicate that transfer learning in SOH estimation is most valuable when it addresses representation mismatch and sparse supervision rather than simply adding another model layer^[81].

A related development is the growing use of physics-guided and hybrid learning. Earlier studies often treated data-driven and physics-based methods as separate categories, whereas more recent work increasingly combines them^[103]. The motivation is straightforward: a purely black-box model may fit the data well, yet remain difficult to validate or interpret under unseen conditions. This issue is especially relevant in battery management, where estimation error can affect maintenance, derating, and safety-related control. Recent studies therefore embed physical priors, degradation constraints, or consistency relations into neural estimators to improve robustness without sacrificing the flexibility of machine learning. Physics-guided battery-health research has argued that combining mechanism awareness with data efficiency is one of the most promising directions for SOH estimation under realistic operating variability^[18].

These developments suggest that transfer learning, online updating, and physics-guided learning are best viewed as different responses to the same underlying problem: the training distribution is incomplete, the target distribution is variable, and a static black-box learner is rarely sufficient on its own. What differs among them is the strategy used to bridge that gap—knowledge transfer across domains, incremental correction during operation, or structural regularization based on physical understanding.

Figure 4 summarizes the correspondence between representative data regimes and the model families most commonly adopted for SOH estimation, highlighting the primary advantage that motivates each choice.

Figure 4. Representative learning strategies for SOH estimation under different data regimes. SOH: State-of-health.

Generalization and evaluation bias

Good performance on a benchmark does not necessarily imply that a model has learned transferable degradation behavior. This problem has become increasingly apparent in SOH estimation because many widely used datasets are generated under clean and highly structured laboratory protocols. The fast-charging benchmark introduced by Severson et al.^[55] is a representative example. It enabled early-life lifetime prediction at an unprecedented scale and greatly accelerated data-driven battery prognostics. At the same time, it exposed a broader issue: when charge protocols, cycling patterns, and measurement structures are highly regular, a model may exploit correlations that are informative for the benchmark but much weaker in practice.

The same concern extends to other widely used datasets, including NASA, CALCE, Oxford, and other laboratory-derived benchmarks discussed in Section PRACTICAL DATA FOUNDATIONS FOR SOH ESTIMATION. Strong benchmark performance is therefore necessary but not sufficient. The more demanding question is whether the learned relationship remains valid when the operating regime, chemistry, manufacturer, or sampling structure changes. Domain shift is central here. In electrochemical ESS, it is rarely a marginal issue; rather, it is one of the defining conditions of practical deployment.

A related difficulty is data leakage. In SOH estimation, leakage often arises not from direct duplication of training and test samples, but from correlations that survive after splitting. Adjacent cycles from the same cell, overlapping temporal windows, decomposition performed before train-test separation, normalization statistics computed on the full dataset, or protocol fragments that are nearly identical on both sides of the split can all make the evaluation appear more realistic than it actually is. This issue is especially serious for expressive sequence models, which can exploit hidden regularities without making the shortcut obvious.

For this reason, split design should be treated as part of the methodological claim rather than as a minor implementation detail. If the intended claim is generalization across assets, cell-wise splits are more meaningful than random sample splits. If the target is robustness across temperatures, protocols, or manufacturers, condition-wise splits become necessary. If the application concerns forecasting or online updating, time-wise splits are essential. Once this principle is accepted, performance numbers can no longer be interpreted in isolation. They must be read together with the split logic, preprocessing scope, and the relationship between source and target domains.

Recent studies make this point increasingly explicit. Domain-adaptive frameworks address cross-dataset and cross-manufacturer mismatches directly rather than remain within a single narrow protocol^[102]. Multimodal and real-vehicle investigations further indicate that feature usefulness itself may change once operating context becomes part of the data-generating process^[104]. Generalization is therefore not only a property of the trained model; it is also a property of the supervision regime and the evaluation design.

Uncertainty and trustworthiness

In practical SOH estimation, a model should do more than return a point estimate. It should also indicate how much confidence can be placed in that estimate. This need is not new. Early probabilistic work, such as the Bayesian battery lifespan model in^[83] and the Gaussian-process-based supercapacitor degradation modeling in^[100], already showed that uncertainty-aware learning had value in lifecycle estimation. What has changed in recent years is that uncertainty is no longer regarded as a secondary refinement. It is increasingly treated as part of the information required for action.

Uncertainty in SOH estimation may arise from noisy measurements, sparse labels, imperfect model structure, incomplete domain coverage, or operating conditions absent from the training data. Trustworthiness should therefore not be reduced to a single confidence interval. In electrochemical ESS, at least three capabilities are central. The first is uncertainty quantification, which attaches confidence information to the estimate. The second is out-of-distribution (OOD) awareness, which detects when the current input no longer resembles the regime represented in training. In this sense, OOD detection is not separate from SOH estimation, but part of determining whether a health estimate remains valid under the current operating regime. The third is conservative output behavior, meaning that the estimator or its supervisory layer should revert to safer actions when confidence is low, or the distribution has shifted.

Recent literature has made these requirements more explicit. Probabilistic battery-health reviews argue that uncertainty quantification should be treated as a necessary component of diagnostics and prognostics rather than as a supplementary layer^[18]. A similar lesson can be drawn from field-oriented supercapacitor studies. The framework based on tram operational data^[37] showed that sparse and incomplete field fragments can still support SOH prediction, but only when fragment selection, label aggregation, and result interpretation are handled with great care. This is important not only because it concerns supercapacitors, but also because it places incomplete observability and low-resolution field data at the center of the estimation problem itself.

Trustworthiness is therefore not an optional layer added after model development. It is the natural culmination of the earlier Sections. Section SOH TARGETS AND DEFINITIONS defined what should be labeled; Section PRACTICAL DATA FOUNDATIONS FOR SOH ESTIMATION clarified what can be observed; Section HEALTH INDICATORS AND FEATURE REPRESENTATION examined what degradation information can be represented; and the present Section shows that model choice is only one component of a broader reliability problem. A trustworthy estimator is not simply the one with the lowest RMSE. It is the one whose representation, supervision, evaluation protocol, and confidence behavior remain aligned with the intended operating context.

Overall, machine-learning-based SOH estimation is better understood as a family of learning strategies matched to different data conditions than as a ranking of algorithm names. Traditional regressors remain effective when informative HIs and limited labels are available. Deep temporal and hybrid sequence models become valuable when degradation information is distributed across long, noisy, and structured trajectories. Transfer learning, online updating, and physics-guided learning become particularly relevant when target-domain data are sparse, distributions shift, or physical consistency must be preserved. Across all model families, the decisive issue is no longer only whether a model fits the data, but whether it generalizes under realistic evaluation settings, resists leakage, communicates uncertainty, and supports reliable decision-making under changing operating conditions.

Benchmarking and evaluation for deployable SOH estimation

A large number of machine learning models have been reported for SOH estimation, but comparison across studies remains difficult because datasets, label definitions, operating conditions, and evaluation protocols are rarely aligned. Recent reviews have therefore emphasized that performance reported on curated datasets should not be interpreted as a direct indicator of deployability in real ESS^[20].

This issue is evident in the way predictive error is commonly reported. Metrics such as RMSE and mean absolute error (MAE) are strongly influenced by the underlying dataset, including cycling protocol, temperature range, sampling structure, and the definition of the health target itself. As a result, a model that performs well under controlled laboratory conditions may not retain the same level of performance once the operating regime becomes less regular or less observable.

For this reason, recent literature increasingly treats benchmarking as a multi-dimensional problem rather than an accuracy-only exercise. In addition to predictive performance, comparison of SOH estimation methods has begun to incorporate data requirements, observability dependence, robustness to domain shift, interpretability, uncertainty awareness, and computational feasibility. These factors are particularly important in electrochemical ESS, where signals are often incomplete, labels are sparse, and operating profiles are non-stationary.

Viewed from this perspective, different model families occupy different positions in the deployment landscape. Regression-based and probabilistic models remain effective when structured health indicators are available and labels are reliable, especially because they retain relatively strong interpretability and data efficiency^[88]. Deep temporal models, including LSTM- and Transformer-based frameworks, are more suitable when degradation information is distributed across long and structured time-series data^[90]. Transfer learning and online updating methods become more attractive when source and target conditions differ, and target-domain supervision is limited. Physics-guided approaches further strengthen consistency and robustness by constraining learning with electrochemical knowledge^[103].

Taken together, these studies do not support a universal ranking of model families. Instead, they indicate that method suitability depends on the interaction among data condition, feature representation, supervision regime, and deployment objective. This perspective underlies the comparison framework summarized in Table 5.

System-level integration of SOH estimation outputs

In practice, the usefulness of an SOH estimator is determined not only by model-level performance, but also by how its outputs are integrated into system operation. In real ESS, health outputs are rarely used in isolation; instead, they are embedded in control, scheduling, protection, and maintenance processes. Recent studies have increasingly emphasized that the value of an SOH estimator is ultimately determined by the way its outputs are translated into system-level decisions^[20,39].

In battery systems, SOH estimation is closely linked to operational limits such as allowable power, thermal margin, and maintenance interval. Capacity fade reduces the usable energy available for dispatch, while resistance growth may tighten power constraints and increase heat generation under the same duty cycle. For this reason, battery SOH is often operationalized not as an isolated scalar, but as a quantity that informs derating, dispatch planning, and risk management under prevailing operating conditions.

The integration logic is different in supercapacitor systems. As discussed in earlier sections, the dominant health variables are more closely associated with capacitance retention, equivalent series resistance, and fast-response capability than with battery-style energy capacity^[71-74]. This means that supercapacitor SOH estimation is more directly connected to transient power support and short-timescale efficiency. Under such conditions, state estimation and health estimation should be interpreted jointly in the context of high-frequency cycling, power buffering, and recovery behavior.

The complexity becomes greater in hybrid battery-supercapacitor systems, where the two devices play complementary but non-interchangeable roles. Batteries provide the main energy support, whereas supercapacitors absorb rapid power fluctuations and short-timescale transients. This division of labor has been repeatedly shown to reduce battery current stress and slow degradation, which explains why hybridization is often introduced as a lifetime-oriented design choice rather than only a power-management option. At the same time, recent reviews show that the hybrid literature remains dominated by sizing and energy-management strategies, while coupled health representation remains comparatively underdeveloped^[105-108]. This gap is important for SOH estimation because battery aging and supercapacitor degradation evolve under different stress channels; system health can no longer be represented adequately by a single scalar metric.

A more consistent interpretation is to distinguish two levels of health reporting. The first is component-wise, in which battery and supercapacitor health are estimated separately and then interpreted jointly at the supervisory level. The second is system-oriented, in which the outputs of lower-level estimators are translated into aggregate quantities such as power capability margin, efficiency loss, thermal burden, or maintenance urgency under specified operating constraints. Health-relevant energy-management studies already point in this direction by optimizing power allocation to explicitly account for battery lifetime extension and supercapacitor utilization under realistic operating conditions^[109,110]. Viewed in this way, SOH estimation in hybrid systems is better treated as a hierarchical problem that links component-level degradation to system-level operating consequences, rather than as a direct extension of single-device diagnostics.

Decision guide for method selection and open directions

Given the diversity of data conditions and application scenarios, the selection of an SOH estimation strategy remains a central methodological issue. Recent studies increasingly suggest that method choice should be guided by data availability, label quality, and deployment requirements rather than by algorithmic preference alone.

When structured health indicators are available and labels are reliable, conventional regression and probabilistic models remain attractive because they offer relatively high interpretability and data efficiency. When degradation information is distributed across long, noisy, and structured time-series data, deep temporal models become more suitable. Under cross-domain conditions with limited target-domain supervision, transfer learning and online updating are more likely to preserve performance. In safety-critical settings, physics-guided and uncertainty-aware learning becomes especially important because reliability depends not only on predictive accuracy, but also on consistency and confidence.

This selection logic is summarized conceptually in Figure 5. The figure links data condition, label regime, and deployment demand to representative modeling strategies, thereby emphasizing that method choice is not a one-dimensional preference for a particular algorithm. Instead, it is a constrained decision problem shaped by observability, supervision, and operational consequence.

Figure 5. Constraint-driven selection logic for SOH estimation methods. SOH: State-of-health.

Despite substantial progress, several challenges remain open. One is the lack of standardized datasets and evaluation protocols, which continues to limit the comparability of existing studies. Another is the limited evidence for robustness under real-world conditions, since a large share of the literature is still validated primarily on laboratory data. Uncertainty quantification and OOD detection also remain underdeveloped in practical deployments. In addition, system-level health estimation for hybrid ESS is still emerging and lacks mature evaluation frameworks.

Future research is therefore likely to focus on four converging directions: the development of more standardized and realistic benchmarks, improved cross-domain generalization, deeper integration of physical knowledge with data-driven learning, and the advancement of trustworthy machine learning frameworks for practical energy storage operation. Progress along these lines is likely to determine whether ML-based SOH estimation can move from promising methodology to robust infrastructure for real electrochemical ESS.