Entropy-Based SOH Prediction for Lithium-Ion Battery Clusters: Real-Time Health Assessment Without Capacity Testing

TL;DR #

Entropy-based analysis of voltage drop and temperature distribution across cell clusters enables real-time health assessment without disassembling battery packs: information entropy values between 1.85–2.00 for voltage features and 1.51–1.53 for temperature features indicate consistent cell aging, while entropy deviations beyond ±0.15 from baseline signal accelerated degradation or individual cell failures. For procurement teams managing grid-scale lithium-ion storage deployments, this method detects cluster-level inconsistencies 40–60 cycles earlier than traditional capacity-based SOH tracking. Require suppliers to provide entropy calculation algorithms in their BMS firmware and validate that their systems can flag entropy anomalies before visible capacity loss occurs.

Overview #

Most procurement teams evaluate battery energy storage system health through cycle count or nominal capacity fade—methods that are too coarse for early fault detection. In qualifying suppliers for a 500 kWh lithium-ion grid storage project last year, we found that three of eight candidate BMS platforms could not detect individual cell degradation until the affected pack had already lost 8–12% usable capacity. By that point, thermal runaway risk had increased measurably and replacement costs were unavoidable. A controlled degradation study conducted at a power systems laboratory in Hunan Province—using 20S1P battery modules cycled under constant-current peak-shaving profiles with 1440 data points per cell per day—demonstrated that information entropy applied to voltage transient and temperature distribution data provides a statistically robust indicator of both aging rate and cell-to-cell consistency. The study monitored 240 cells across multiple clusters through 300+ cycles, recording ohmic voltage drops during discharge initiation and peak cell temperatures throughout operation.

CompactBESS works with global buyers to identify BMS suppliers capable of implementing real-time entropy calculation in their monitoring firmware—a feature rarely included in standard commercial offerings but essential for installations where downtime carries five-figure hourly penalties.

Entropy-Driven Health Assessment vs. Traditional Capacity Metrics #

Traditional SOH estimation relies on cycle counting or periodic capacity verification tests, both of which lag behind actual degradation. The entropy method treats the battery cluster as an information system: each cell’s voltage drop amplitude (Δuohm) during the initial discharge transient reflects its ohmic resistance, while peak operating temperature (Tmax) correlates with combined ohmic and polarization resistance per the Bernardi heat generation model. These parameters are normalized across all cells in the cluster and binned into probability distributions—voltage drops typically segmented into 0.1 V intervals from 0.0 to 1.0, temperatures into 0.25°C steps across the observed range.

Information entropy H is calculated as H = -Σ P(i) ln P(i), where P(i) is the probability of a cell’s normalized parameter falling within bin i. For a healthy, well-matched cluster, both voltage and temperature distributions approximate normal curves with entropy values in narrow ranges: HΔu typically 1.85–2.00 and HT around 1.51–1.53 for clusters in the study. As cells age uniformly, entropy remains stable because the distribution shape is preserved. When one or more cells degrade faster—evidenced by increased internal resistance—the distribution broadens (uniform aging) or develops outliers (localized failure), and entropy shifts accordingly.

Correlation analysis on NASA’s publicly available 18650 cell degradation dataset (B0005 battery) showed Pearson coefficients of -0.9987 between voltage drop amplitude and discharge capacity, and -0.9862 between peak temperature and capacity. These values confirm that Δuohm and Tmax serve as reliable proxies for cell health without requiring offline capacity testing.

The entropy approach detects failures that capacity tests miss. In the Hunan field trial, introducing a deliberately aged cell (80% SOH) into an otherwise healthy cluster caused H_Δu to spike from 1.92 to 2.15 while cluster-average voltage remained within 2% of nominal. Traditional BMS alarms—set to voltage deviation thresholds—did not trigger until the aged cell caused a 6% capacity reduction at the pack level, roughly 40 cycles later. By tracking entropy in real time, operators gain 40–60 cycles of advance warning, enough time to schedule maintenance during low-demand periods rather than responding to emergency shutdowns.

Honestly, most buyers over-specify voltage precision in BMS tenders (±5 mV is common) but under-specify the computational capability needed to run entropy calculations on 200+ cells every charge/discharge cycle. A 100 kW / 200 kWh system with 10S20P module architecture generates 200 voltage samples and 200 temperature samples per cycle; calculating entropy for both parameters requires binning, probability summation, and logarithmic operations—manageable for modern microcontrollers but often omitted from low-cost BMS designs to save 8–12% on controller cost. You’ll pay that back in a single avoided shutdown.

For additional context on internal resistance progression and its impact on cycle life degradation, refer to our technical overview of aging mechanisms in lithium-ion systems.

Feature Extraction from Operational Data #

Battery management systems record voltage (U), current (I), and temperature (T) as direct measurements, then derive quantities like state-of-charge (SOC) and state-of-health (SOH) through estimation algorithms. Most commercial BMS platforms log 5–10 parameters per cell per minute, generating gigabytes of data monthly for a utility-scale installation. The entropy method intentionally discards most of this data and focuses on two feature sets:

Voltage transient features: During constant-current discharge, the cell voltage drops instantaneously by Δuohm ≈ I · Rohm when load current I is applied, where Rohm is the ohmic resistance. This drop occurs within the first sampling interval (typically <1 second) and is visible in high-resolution voltage logs. For a 20S1P module discharging at 1C, individual cell Δuohm values ranged from 0.048 to 0.057 V in new cells and increased to 0.062–0.078 V after 160 cycles in the laboratory study. Normalization into dimensionless ratio λi = (Δuohm,i – Δuohm,min) / (Δuohm,max – Δu_ohm,min) eliminates absolute voltage dependencies and allows cross-cluster comparison.

Temperature distribution features: Peak cell temperature during discharge correlates with total internal resistance (ohmic + polarization) per the Bernardi equation: q = (I²/Vcell) · (Rohm + Rp), where q is heat generation rate. Cells with higher resistance run hotter under identical load conditions. In the field trial, peak temperatures in a 240-cell cluster ranged from 28.3°C to 32.1°C during 1C discharge from an ambient baseline of 25°C. Normalization into μt = (Tt – Tmin) / (Tmax – Tmin) creates a dimensionless temperature distribution for entropy calculation. Temperature data is sampled continuously at 1-minute intervals rather than only at discharge initiation, providing denser statistical coverage.

The binning process divides the normalized parameter range [0, 1] into equal intervals—typically 10 bins of width 0.1 for voltage and 4–5 bins for temperature—and counts how many cells fall into each bin. For a 240-cell cluster with voltage drops distributed as [5, 22, 58, 72, 48, 25, 8, 2, 0, 0] across bins [0–0.1, 0.1–0.2, …, 0.9–1.0], the probability distribution is P(i) = ni / 240, yielding P = [0.021, 0.092, 0.242, 0.300, 0.200, 0.104, 0.033, 0.008, 0, 0]. Entropy follows as H = -Σ P(i) ln P(i) = 1.89 in this example. A perfectly uniform distribution across 10 bins yields maximum entropy Hmax = ln(10) ≈ 2.30, while a distribution concentrated in one bin approaches Hmin = 0. Healthy clusters exhibit entropy between 70–85% of Hmax, reflecting modest variance around a central tendency.

Recent studies confirm that entropy-based clustering metrics outperform standard-deviation-based metrics for fault detection in systems with non-Gaussian noise, which is typical in large battery arrays where sensor calibration drift, cell placement effects, and thermal gradients create measurement asymmetries. Standard deviation treats all deviations equally; entropy penalizes long-tail outliers more heavily, making it sensitive to the single-cell failures that precipitate thermal runaway events. In supplier qualification, we saw three of six BMS samples fail to flag a deliberately introduced outlier cell (R_ohm = 2× cluster median) because their alarm logic used ±3σ voltage thresholds—outliers within 3σ don’t trigger alarms even when they represent real failures.

The IEC 62619:2022 Safety requirements for secondary lithium cells and batteries standard mandates cell-level voltage and temperature monitoring but does not specify statistical methods for interpreting that data. Entropy calculation fills this gap by converting raw monitoring data into a single, comparable health metric.

Neural Network Prediction for Short-Term Health Forecasting #

Entropy values form a time series: HΔu(t), HΔu(t-1), HΔu(t-2), … recorded daily or per cycle. A trained neural network can predict HΔu(t+1) from the previous 4–7 data points, enabling short-term health forecasting without waiting for measurable capacity loss. The Hunan study used a feedforward network with 4 input neurons [H(t), H(t-1), H(t-2), H(t-3)], one hidden layer of 10 neurons with sigmoid activation, and a single output neuron for H(t+1). Training data consisted of 50 cycles from a 100 kW / 200 kWh system, with the model achieving mean absolute error <0.02 on validation data—well within the ±0.10 threshold that separates normal operation from early-stage degradation.

Prediction accuracy depends on training dataset size and operational consistency. Systems with stable daily charge/discharge profiles (e.g., time-of-use arbitrage with fixed on-peak and off-peak windows) produce more predictable entropy trajectories than systems with variable dispatch (frequency regulation, demand response). In the laboratory validation, prediction error remained below 3% for 20 consecutive cycles when discharge current and ambient temperature were held constant, but increased to 8–12% when ambient temperature varied by ±5°C daily—still useful for flagging anomalies but requiring recalibration after significant environmental changes.

For buyers deploying systems in regions with high seasonal temperature variation, this means the BMS must either retrain entropy prediction models quarterly or implement temperature-compensated entropy baselines. Neither capability is standard in off-the-shelf products. Most commercial BMS platforms lack the onboard storage and processing power to run neural network inference in real time, instead relying on cloud-based analytics with 15–60 minute latency. That latency is acceptable for long-term trend monitoring but inadequate for fast-developing faults like thermal runaway precursors, where intervention windows can be measured in minutes.

The UL 9540A Test Method for Evaluating Thermal Runaway Fire Propagation in Battery Energy Storage Systems defines test protocols for worst-case thermal runaway scenarios but does not address predictive monitoring techniques. Entropy-based forecasting combined with IEEE 1679 Recommended Practice for the Characterization and Evaluation of Emerging Energy Storage Technologies provides a framework for integrating real-time risk assessment into BMS architectures.

Simulation Validation and Sensitivity Analysis #

Before field deployment, the entropy method was validated using a 20S1P Simulink battery model with parametric variation to simulate cell-to-cell inconsistency. Starting from a baseline where all 20 cells had identical Rohm = 15 mΩ and Rp = 20 mΩ, the model introduced progressive degradation: one cell’s Rohm increased by 1 mΩ per 10 cycles while others remained constant. Entropy HΔu increased from 1.92 (cycle 0) to 2.08 (cycle 80) and 2.21 (cycle 160), tracking the coefficient of variation in cell resistance with a linear correlation R² = 0.96. Standard deviation of voltage drop amplitudes increased from 0.004 V to 0.011 V over the same period—a 2.75× change—while entropy increased 15%, demonstrating higher sensitivity to distributional changes.

A second simulation scenario introduced random resistance drift: each cell’s R_ohm varied by ±10% cycle-to-cycle, mimicking temperature fluctuations or SOC estimation errors. In this case, entropy fluctuated within ±0.08 of the baseline (1.84–2.00 range), never crossing the ±0.15 alarm threshold, while standard deviation showed larger relative swings (±18%) that would trigger false alarms if used directly. This result confirms that entropy filters high-frequency noise more effectively than first-order statistics, a critical advantage in operational environments where measurement uncertainty is unavoidable.

Field validation at the 100 kW / 200 kWh Hunan site compared entropy-based alarms to the site’s existing capacity-based SOH monitoring over 180 days. The entropy system flagged 4 anomalous events (H_Δu > 2.15) during that period; subsequent offline capacity testing confirmed that 3 of the 4 flags corresponded to cells with >12% capacity loss relative to cluster median, while the fourth was a thermal sensor calibration error (cell temperature read 6°C high). The existing BMS, which triggered alarms only when cluster-average capacity fell below 90% of nameplate, issued zero alarms during the same 180 days despite the presence of degraded cells. Entropy-based monitoring provided a 60–90 day advance warning for 75% of detected faults.

Most procurement teams don’t realize that capacity-based SOH definitions were developed for consumer electronics (smartphones, laptops) where warranty periods are 1–3 years and replacement is economically feasible. For grid-scale systems with 10–15 year design lives and six-figure replacement costs, waiting for 10% capacity loss before intervening is unacceptable. Entropy thresholds can be tuned to match financial risk tolerance: a conservative threshold (±0.10) flags more false positives but catches early-stage failures; a permissive threshold (±0.20) reduces false alarms but accepts higher failure risk. In our experience, ±0.15 provides the best balance for utility-scale applications.

Practical Guidance for Buyers #

When issuing RFQs for lithium-ion storage systems, require that BMS platforms include entropy calculation as a standard feature, not an add-on. Specifically, the BMS should compute and log HΔu and HT at least once per charge/discharge cycle, flag entropy deviations exceeding configurable thresholds, and provide API access to raw entropy time-series data for external analytics. Verify that the supplier’s firmware has sufficient RAM and clock speed to handle entropy computation for your cluster size—200+ cell systems may require dedicated coprocessors or edge computing modules rather than relying on the primary BMS microcontroller.

During supplier qualification, request entropy calculation results from their own test data. If they cannot provide this, it indicates they are not using entropy metrics internally and are unlikely to support it well post-delivery. Ask for the bin width and sample interval they use for probability distribution construction; arbitrary or inconsistent choices here suggest the feature was added as a checkbox item rather than a validated engineering tool.

For retrofit installations where replacing the BMS is not feasible, consider deploying external monitoring hardware that reads voltage and temperature via CAN or Modbus and performs entropy calculations independently. Several industrial IoT platforms now offer battery-specific analytics modules, though integration requires custom scripting and validation. Be cautious of cloud-only solutions that introduce latency—entropy’s value is in real-time fault detection, not post-mortem analysis.

As a Guangzhou-based sourcing service provider connecting global buyers with Chinese BESS manufacturers, we routinely help procurement engineers identify suppliers whose BMS firmware supports advanced health metrics like entropy tracking—a capability that separates Tier 1 from Tier 2 suppliers. If your current RFQ responses lack technical depth on this topic, talk to our sourcing team → and we’ll match you with vendors who have already deployed entropy-based monitoring in commercial installations.

Supplier Qualification Questions #

1. What is the sampling resolution (voltage precision and time interval) during discharge initiation, and can your BMS capture the ohmic voltage drop (Δu_ohm) within the first 500 ms of constant-current discharge with <±2 mV measurement uncertainty?

1. Does your BMS firmware calculate information entropy for voltage drop distribution and temperature distribution in real time, and if so, what bin width and probability calculation method do you use—can you provide entropy time-series logs from a 200+ cell cluster operating under peak-shaving duty?

1. At what entropy deviation threshold (absolute value or percentage above baseline) does your system generate health warnings, and has this threshold been validated against actual cell failure data in systems with >150 cycles of operational history?

1. Can your BMS perform short-term entropy prediction (4-cycle or 7-day lookback) using onboard processing, or does prediction require cloud upload and external analytics—what is the maximum acceptable latency for entropy-based fault detection in your architecture?

1. What is the Pearson correlation coefficient between your BMS’s reported SOH (derived from capacity estimation) and the ohmic resistance measured via EIS or DC pulse test for cells at 50%, 70%, and 90% of rated cycle life—can you provide validation data showing R² > 0.95?

Sourcing Checklist #

• [ ] BMS logs entropy values (HΔu and HT) at least once per discharge cycle with timestamped records accessible via CSV export or API call
• [ ] Voltage measurement precision ≤±5 mV across the full SOC range (10%–100%) at sampling rates ≥1 Hz during transient events
• [ ] Temperature sensor accuracy ±1°C or better with placement within 10 mm of cell terminals for all monitored cells
• [ ] Firmware supports user-configurable entropy alarm thresholds between ±0.05 and ±0.25 from baseline without requiring recompilation
• [ ] System has been field-tested in at least one installation >50 kWh capacity with documented entropy tracking over >100 charge/discharge cycles
• [ ] BMS architecture includes sufficient processing headroom (≤60% CPU utilization during peak data logging) to run entropy calculations without impacting core safety functions
• [ ] Supplier provides entropy baseline values and threshold recommendations specific to your cell chemistry (LFP, NMC, LTO) and expected duty cycle (daily cycling, infrequent backup)
• [ ] Communication protocol (CAN, Modbus, Ethernet) supports real-time streaming of entropy data to external SCADA or analytics platforms with <10 second latency

Key Specifications Table #

Can’t find a supplier meeting these specs? Submit your requirements and we’ll match you within 48 hours.

Frequently Asked Questions #

How does entropy-based SOH differ from capacity-based SOH for large battery clusters?

Capacity-based SOH requires periodic full discharge cycles to measure available amp-hours, which is impractical for grid-scale systems in continuous operation. Entropy-based SOH operates on routine partial-depth discharge data, detecting cell-to-cell inconsistency without interrupting service. A cluster can show 95% capacity-based SOH while entropy reveals that 10% of cells are degrading 2× faster than the median—information lost in aggregate capacity measurements.

What operational conditions affect entropy baseline values—do I need different thresholds for summer vs. winter?

Temperature significantly affects absolute resistance values but entropy is calculated from normalized distributions, which reduces temperature sensitivity. In the field trial, baseline entropy varied by <±0.05 across a 15°C ambient temperature range (10°C to 25°C) without recalibration. If your system operates in environments with >20°C seasonal swings or undergoes frequent rapid temperature transients, implement temperature-compensated baselines by maintaining separate entropy profiles for 5°C ambient bands.

Can entropy calculation detect manufacturing defects in new battery packs before they cause failures?

Yes. New packs that pass initial capacity testing may still contain cells with out-of-spec internal resistance. In acceptance testing of a 500 kWh delivery, entropy analysis identified 12 cells (2% of total) with HΔu contributions >2.5σ above mean despite all cells meeting the supplier’s ±3% capacity tolerance. Offline EIS testing confirmed these cells had 18–24% higher Rohm than pack median. Entropy screening during commissioning prevents “infant mortality” failures that typically manifest within the first 50–100 cycles.

Is real-time entropy calculation computationally intensive—will it overload low-cost BMS hardware?

For a 200-cell cluster with 10-bin voltage distribution and 5-bin temperature distribution, entropy calculation requires ~3,000 floating-point operations per cycle: 200 normalizations, 30 bin counts, 15 probability divisions, and 15 logarithms. A 100 MHz ARM Cortex-M4 microcontroller completes this in <5 ms, consuming <0.5% of available clock cycles. The computational bottleneck is usually data logging and communication overhead (writing results to flash, transmitting via CAN) rather than the entropy calculation itself. Budget an additional 512 KB flash and 64 KB RAM for entropy time-series storage and prediction model weights.

What is the minimum dataset size needed to establish a reliable entropy baseline for a new installation?

At least 20 complete charge/discharge cycles under consistent operating conditions (±10% on discharge current, ±3°C ambient temperature). More cycles improve baseline precision: 50-cycle baselines showed ±0.03 standard error while 20-cycle baselines had ±0.07 standard error in validation testing. For systems with variable dispatch profiles (renewable integration, demand response), collect data from 40+ cycles spanning typical operating modes before finalizing alarm thresholds. If the installation will experience multiple duty cycles (daily peak-shaving plus occasional backup discharge), establish separate baselines for each mode.

---

Published by compactbess.com Technical Team | Request a sourcing quote

---

Data source: Information Entropy-Based Health Assessment and Prediction for Lithium-Ion Battery Clusters in Energy Storage Systems, Y. Xia et al., Journal of the Electrochemical Society, 2024