In the pursuit of understanding and mitigating thermal runaway hazards in large-scale energy storage systems, we have developed a comprehensive modeling and fault localization methodology centered on lithium iron phosphate energy storage batteries. Our work integrates experimental investigation, multiphysics simulation, and intelligent optimization to address the critical challenge of cascading thermal failure propagation. This paper presents our findings on the thermal runaway characteristics of 280 Ah energy storage batteries, the development of a high-fidelity thermal propagation model, and a novel fault inversion algorithm for locating the source of thermal runaway within battery clusters.
Introduction
The rapid deployment of renewable energy sources has positioned electrochemical energy storage as a cornerstone technology for modern power systems. Among various battery chemistries, lithium iron phosphate energy storage batteries have gained widespread adoption due to their intrinsic safety advantages and long cycle life. However, the high energy density and compact arrangement of cells within energy storage cabinets create conditions where a single cell failure can trigger a catastrophic chain reaction. The thermal runaway propagation phenomenon in energy storage batteries represents a critical safety concern that demands systematic investigation from fundamental electrochemical mechanisms to system-level fault diagnosis.
Our research addresses three interconnected challenges in energy storage battery safety. First, we need to accurately characterize the thermal runaway triggering thresholds and evolution patterns of large-format energy storage batteries under different states of charge. Second, we must develop simulation models capable of capturing the complex multiphysics interactions during thermal propagation while maintaining computational efficiency. Third, we require robust fault localization methods that can identify the origin of thermal runaway in a battery cluster even under conditions of sensor noise and incomplete measurements.
To tackle these challenges, we designed our study around the following objectives: establishing experimental benchmarks for 280 Ah lithium iron phosphate energy storage batteries, developing an optimized thermal runaway propagation model using data-driven parameter identification, and constructing a hybrid optimization framework for fault source inversion. Our methodology bridges the gap between fundamental electrochemistry and practical system-level diagnostics, providing both theoretical insights and engineering tools for energy storage battery safety management.
Experimental Investigation of Thermal Runaway in Energy Storage Batteries
Experimental Setup and Protocol
We conducted thermal abuse experiments on 280 Ah lithium iron phosphate energy storage batteries to obtain critical parameter thresholds and evolutionary characteristics. The experimental platform consisted of a standardized explosion-proof chamber equipped with a 1000 W constant-power heating device. Five K-type thermocouples were strategically placed on each cell surface to monitor temperature evolution at multiple locations. We prepared four groups of identically conditioned cells at 25%, 50%, 75%, and 100% state of charge. All samples underwent controlled charging protocols using a battery cycler following IEC 62660-1 standards before thermal testing.
Key Experimental Findings
Table I summarizes the critical thermal runaway parameters we extracted from our experiments across different SOC levels. The data reveals a strong SOC dependence on thermal runaway severity and triggering dynamics.
| Parameter | SOC = 25% | SOC = 50% | SOC = 75% | SOC = 100% |
|---|---|---|---|---|
| Vent activation time (s) | 1407 | 916 | 378 | 338 |
| Voltage collapse time (s) | 1496 | 1430 | 984 | 327 |
| Thermal runaway trigger time (s) | 2683 | 1698 | 1010 | 469 |
| Peak temperature time (s) | 2721 | 1799 | 1219 | 847 |
| Temperature at voltage collapse (°C) | 136.5 | 136.6 | 114.9 | 61.2 |
| Temperature at vent activation (°C) | 107.9 | 101.7 | 70.3 | 73.8 |
| Temperature at TR trigger (°C) | 201.8 | 157.2 | 120.1 | 79.4 |
| Peak temperature (°C) | 236.8 | 238.4 | 350.9 | 378.3 |
| Maximum temperature rise rate (°C/s) | 2.2 | 3.9 | 16.4 | 26.5 |
Our experimental observations revealed several important characteristics of thermal runaway in large-format energy storage batteries. At 100% SOC, the cell exhibited the shortest time to thermal runaway and the highest peak temperature of 378.3 °C, with a maximum temperature rise rate of 26.5 °C/s. The voltage collapse occurred at a relatively low temperature of 61.2 °C, indicating that the internal short circuit mechanism was activated early in the heating process. In contrast, at 25% SOC, the thermal runaway process was significantly delayed, with a peak temperature of only 236.8 °C and a substantially lower temperature rise rate of 2.2 °C/s. We noted that the voltage at 25% SOC showed a gradual decline pattern rather than the abrupt collapse observed at higher SOC levels.
The temperature evolution patterns we measured exhibited a characteristic double-peak behavior at moderate SOC levels, which we attribute to the Joule-Thomson expansion effect during vent activation. This phenomenon was particularly evident at 50% and 75% SOC, where the surface temperature showed a brief decrease following vent opening before rising again as internal short circuit reactions intensified.
Multi-Physics Coupled Modeling with Parameter Identification
Thermal Runaway Reaction Kinetics
We developed a comprehensive thermal runaway model for our energy storage batteries based on the Arrhenius reaction framework. The model incorporates four primary exothermic side reactions that contribute to thermal runaway propagation:
$$Q_{SEI} = H_{SEI} \cdot W_{SEI} \cdot R_{SEI}$$
$$Q_{ne} = H_{ne} \cdot W_{ne} \cdot R_{ne}$$
$$Q_{pe} = H_{pe} \cdot W_{pe} \cdot R_{pe}$$
$$Q_{ele} = H_{ele} \cdot W_{ele} \cdot R_{ele}$$
The total heat generation from side reactions is given by:
$$\Sigma Q_{side-tot} = Q_{SEI} + Q_{ne} + Q_{pe} + Q_{ele}$$
Each reaction rate follows the Arrhenius law:
$$R_i = A_i \cdot c_i^{m_i} \cdot \exp\left(-\frac{E_{a,i}}{RT}\right)$$
Where Ri represents the reaction rate, Ai is the pre-exponential factor, Ea,i is the activation energy, R is the gas constant, T is the temperature, ci is the normalized concentration of reactants, and mi is the reaction order.
FFRLS Parameter Identification for SOC-Dependent Modeling
Traditional thermal runaway models assume constant activation energies for side reactions, which leads to significant prediction errors when simulating energy storage batteries across varying SOC conditions. We addressed this limitation by implementing a Forgetting Factor Recursive Least Squares (FFRLS) parameter identification method that establishes a time-varying mapping between SOC and the activation energies of side reactions.
The SOC-dependent activation energy is expressed as:
$$E_{a,i}(SOC) = k_{i,0} + k_{i,1} \cdot \exp(k_{i,2} \cdot SOC)$$
The FFRLS algorithm iteratively updates the parameter estimates to minimize the prediction error between simulated and experimental temperature profiles. The system input vector at time k is:
$$\phi(k) = \left[1, \frac{1}{RT(k)}\right]^T$$
The optimal parameter estimate vector is:
$$\hat{\theta}(k) = [\ln(A_i), E_{a,i}(SOC)]^T$$
The parameter update equations are:
$$\hat{\theta}(k+1) = \hat{\theta}(k) + K(k+1)e(k+1)$$
$$K(k+1) = \frac{P(k)\phi(k+1)}{\lambda + \phi^T(k+1)P(k)\phi(k+1)}$$
$$P(k+1) = \frac{1}{\lambda}\left[P(k) – K(k+1)\phi^T(k+1)P(k)\right]$$
Where K(k+1) is the correction gain, P(k) is the covariance matrix, and λ = 0.98 is the forgetting factor that balances tracking capability and noise sensitivity.
Table II presents the identified activation energies for different side reactions across the SOC range we investigated.
| Reaction | SOC = 25% (J/mol) | SOC = 50% (J/mol) | SOC = 75% (J/mol) | SOC = 100% (J/mol) |
|---|---|---|---|---|
| SEI decomposition | 1.36 × 10&sup5 | 1.28 × 10&sup5 | 1.19 × 10&sup5 | 1.14 × 10&sup5 |
| Anode-electrolyte | 1.42 × 10&sup5 | 1.35 × 10&sup5 | 1.25 × 10&sup5 | 1.17 × 10&sup5 |
| Cathode-electrolyte | 1.80 × 10&sup5 | 1.71 × 10&sup5 | 1.62 × 10&sup5 | 1.56 × 10&sup5 |
| Electrolyte decomposition | 1.86 × 10&sup5 | 1.78 × 10&sup5 | 1.69 × 10&sup5 | 1.64 × 10&sup5 |
The energy conservation equation governing the cell temperature evolution is:
$$\rho C_p \frac{\partial T}{\partial t} = \nabla \cdot (\lambda \nabla T) + \frac{Q_{side-tot}}{V_{cell}} – \frac{Q_{diss-tot}}{V_{batt}}$$
Where ρ is density, Cp is specific heat capacity, λ is thermal conductivity, Vcell is the core volume, and Vbatt is the total battery volume. The heat dissipation term includes both convective and radiative components:
$$Q_{diss-tot} = hA_{batt}(T – T_{amb}) + \epsilon\sigma_B A_{batt}(T^4 – T_{amb}^4)$$
Model Validation and Grid Independence
We validated our energy storage battery model by comparing simulated temperature evolution with experimental measurements at the cell center location. The comparison across four SOC levels demonstrated excellent agreement, with peak temperature deviations below 2.59% for all cases. The FFRLS-optimized model successfully captured the SOC-dependent triggering times, temperature rise rates, and peak values that characterize thermal runaway in energy storage batteries.
Table III lists the thermophysical properties we used in our model for the 280 Ah lithium iron phosphate energy storage batteries.
| Component | Density (kg/m³) | Specific heat (J/(kg·K)) | Thermal conductivity (W/(m·K)) |
|---|---|---|---|
| Jelly roll (core) | 2118 | 1088 | λx = 1.5, λy = λz = 18 |
| Steel casing | 7850 | 475 | 44.5 |
| Aluminum busbar | 2700 | 900 | 238 |
| Insulation pad | 1190 | 1470 | 0.18 |
We performed a grid independence study to ensure numerical accuracy while maintaining computational efficiency. The mesh sensitivity analysis considered peak temperature (Tmax) and thermal propagation time intervals as convergence criteria. When the relative tolerance was set to 10-3, we found that a mesh with 946,139 elements provided converged results with peak temperature variation below 1% between successive mesh refinements.
Thermal Runaway Propagation Analysis in Energy Storage Battery Modules
Cell-to-Cell Propagation Characteristics
Based on our validated single-cell model, we constructed a module-level model consisting of 52 cells arranged in a 4 × 13 matrix configuration. We designated cell #33 as the thermal runaway source and analyzed the propagation patterns under different SOC conditions. Our simulations revealed distinct path-constrained propagation behaviors that are critical for understanding thermal runaway spread in energy storage battery systems.
Table IV summarizes the thermal propagation characteristics we observed along the x-axis (through-plane direction) and y-axis (in-plane direction) for different SOC levels.
| Parameter | SOC = 100% (x-axis) | SOC = 100% (y-axis) | SOC = 75% (x-axis) | SOC = 75% (y-axis) |
|---|---|---|---|---|
| Peak temperature (°C) | 424.5 | 429.4 | 412.5 | 423.3 |
| Propagation interval (s) | 73.0 | 66.0 | 98.2 | 89.5 |
| Propagation time range (s) | 316 – 754 | 290 – 710 | 805 – 1394 | 775 – 1310 |
The results demonstrate a clear anisotropy in thermal propagation, with the y-axis (in-plane direction) exhibiting faster propagation speeds and higher peak temperatures compared to the x-axis (through-plane direction). This directional preference arises from the thermal conductivity differences within the jelly roll structure. The through-plane thermal conductivity (λx = 1.5 W/(m·K)) is substantially lower than the in-plane conductivity (λy = λz = 18 W/(m·K)), which is attributed to the layered structure of electrodes and separators that creates high thermal resistance in the through-plane direction.
We also observed that thermal propagation only occurred at higher SOC levels (75% and 100%). At 50% SOC and below, the heat generated by the initiating cell was insufficient to trigger neighboring cells, resulting in self-termination of the thermal event. This finding establishes a critical SOC threshold for cascading failure in energy storage battery modules.
Energy Flow and Heat Transfer Mechanisms
Our analysis of the heat transfer mechanisms during thermal runaway propagation revealed the dominance of conductive heat transfer through the cell-to-cell contact interfaces. The radiative heat transfer contribution became significant only at temperatures exceeding 300 °C. The thermal path constraints we identified can be formalized as:
The heat flux from a failing cell to its neighbor is governed by:
$$q_{ij} = \frac{T_i – T_j}{R_{th,ij}}$$
Where Rth,ij represents the thermal resistance between cells i and j, which depends on the contact interface area and the intervening materials. For our energy storage battery module, the thermal resistance in the x-direction is approximately 12 times larger than in the y-direction, explaining the observed anisotropic propagation.
The critical condition for propagation from cell i to cell j can be expressed as:
$$\int_{t_0}^{t_{TR,j}} q_{ij}(t) dt \geq Q_{threshold}(SOC_j)$$
Where Qthreshold represents the minimum energy required to trigger thermal runaway in cell j, which is a decreasing function of SOC.
Fault Inversion Localization Methodology
Problem Formulation
We developed a two-stage fault localization framework for identifying the source of thermal runaway in energy storage battery clusters. The first stage performs rapid coarse localization using Time Difference of Arrival (TDOA) principles, while the second stage refines the estimate using a hybrid Genetic Algorithm-Grey Wolf Optimizer (GA-GWO) approach combined with our high-fidelity multiphysics model.
Our energy storage battery cluster model consists of six diagnostic zones, each representing a spatial region with independent thermal boundary conditions. We deployed eight virtual temperature sensors at strategic locations around the battery cabinet perimeter to capture the thermal propagation signature.
TDOA-Based Preliminary Localization
The TDOA method exploits the time differences in sensor activation during thermal runaway propagation. For a reference sensor pair A and B, the time difference of arrival for a thermal signal propagating from source P is:
$$\Delta t = \frac{d_A – d_B}{v}$$
Where dA and dB are the distances from source P to sensors A and B respectively, and v is the effective thermal propagation velocity. The locus of points satisfying this time difference forms a hyperbola:
$$\frac{(x – x_A)^2}{a^2} – \frac{(y – y_B)^2}{b^2} = 1$$
For preliminary zone identification, we compute the matching error for each zone i:
$$L_i = \sqrt{\sum_{j=1}^{n} (\tau_{0j} – \tau_{ij})^2}$$
Where τ0j is the measured time delay for sensor j relative to the first triggered sensor, and τij is the corresponding value from our database for zone i. The zone with minimum Li is selected as the preliminary localization result.
Our testing showed that the preliminary localization achieved a minimum matching error of 0.163 for the correct zone and a maximum error of 0.945 for incorrect zones, demonstrating reliable zone-level identification capability.
GA-GWO Hybrid Optimization for Precise Localization
Once the fault zone is identified, we initialize a population of candidate solutions representing possible fault locations, each characterized by spatial coordinates and SOC. The GA-GWO hybrid algorithm iteratively refines these candidates through combined genetic operations and social hierarchy optimization.
The fitness function for each candidate solution i is defined as:
$$Fitness(i) = \frac{1}{n}\sum_{j=1}^{n} (t_{i,j} – t_{1j})^2$$
Where t1j represents the measured time difference between the first triggered sensor and sensor j, and ti,j is the corresponding value predicted by our model for candidate i. The algorithm proceeds as follows:
Step 1: Generate initial population of N candidate solutions within the identified fault zone.
Step 2: Evaluate fitness for all candidates using the multiphysics model.
Step 3: Apply GA operations (crossover, mutation) to explore the solution space.
Step 4: Identify the three best candidates as alpha, beta, and delta wolves.
Step 5: Update candidate positions using GWO dynamics:
$$\vec{D}_{\alpha} = |\vec{C}_1 \cdot \vec{X}_{\alpha} – \vec{X}|$$
$$\vec{D}_{\beta} = |\vec{C}_2 \cdot \vec{X}_{\beta} – \vec{X}|$$
$$\vec{D}_{\delta} = |\vec{C}_3 \cdot \vec{X}_{\delta} – \vec{X}|$$
$$\vec{X}(t+1) = \frac{\vec{X}_1 + \vec{X}_2 + \vec{X}_3}{3}$$
Step 6: Dynamically adjust the convergence factor a using linear decay to balance exploration and exploitation.
Step 7: Repeat Steps 2-6 until convergence criteria are met.
Table V presents the parameter sensitivity analysis we conducted to optimize the GA-GWO algorithm performance.
| Population size (N) | Max iterations (T) | Convergence strategy | Fitness | Space coverage (%) | Convergence speed (iterations) |
|---|---|---|---|---|---|
| 20 | 75 | Linear decay | 0.095 | 91.3 | 40 |
| 30 | 75 | Linear decay | 0.073 | 93.2 | 32 |
| 20 | 120 | Exponential decay | 0.067 | 87.7 | 50 |
| 30 | 120 | Piecewise decay | 0.045 | 96.1 | 65 |
| 25 | 100 | Adaptive decay | 0.056 | 95.2 | 55 |
Based on this analysis, we selected N = 30, T = 75 with a linear decay strategy as the optimal configuration, balancing accuracy, coverage, and computational efficiency.
Directional Dual-Threshold Error Metric
To evaluate localization accuracy considering the anisotropic nature of thermal propagation in energy storage batteries, we developed a Directional Dual-Threshold Error Metric (DD-TEM). This metric modifies the Manhattan distance by incorporating directional thresholds and an exponential penalty mechanism:
$$DD(a, p) = \sum_{i \in \{x, y, z\}} \left(\frac{|a_i – p_i|}{\tau_i}\right)^2 + \lambda \cdot \mathbb{I}_{\{|a_z – p_z| > \tau_z\}}$$
Where a = (ax, ay, az) is the actual fault location, p = (px, py, pz) is the predicted location, τi represents the directional threshold based on energy storage battery cluster dimensions, and λ is a penalty coefficient for cross-layer errors. For our battery cluster model (700 mm × 1100 mm × 950 mm), a DD error below 0.03 indicates successful localization, while values above 0.1 indicate unacceptable error or cross-layer misidentification.
Results and Validation
Model Accuracy for Energy Storage Battery Thermal Runaway
Table VI compares the experimental and simulated peak temperatures for our energy storage batteries at different SOC levels, demonstrating the accuracy of our FFRLS-optimized model.
| SOC (%) | Experimental peak (°C) | Simulated peak (°C) | Deviation (%) |
|---|---|---|---|
| 25 | 236.8 | 231.5 | 2.24 |
| 50 | 238.4 | 234.2 | 1.76 |
| 75 | 350.9 | 345.8 | 1.45 |
| 100 | 378.3 | 372.6 | 1.51 |
The maximum deviation of 2.24% at 25% SOC confirmed that our SOC-dependent activation energy mapping effectively captures the nonlinear thermal behavior across the full operating range of energy storage batteries. The model accurately predicted the transition from non-propagating to propagating behavior at the critical SOC threshold between 50% and 75%.
Fault Localization Performance
We evaluated our GA-GWO-based fault inversion method against three alternative approaches: direct TDOA, PSO-driven inversion, and single-dimensional GWO. The comparison was conducted across four SOC levels with five repeated simulations for statistical reliability. Table VII presents the comprehensive results.
| SOC condition | Localization method | Mean DD error | Accuracy (%) | Cross-layer error rate |
|---|---|---|---|---|
| SOC = 25%, SOH = 100% | TDOA | 0.127 | 62.3 | 0.35 |
| PSO | 0.053 | 81.5 | 0.08 | |
| Single-dim GWO | 0.079 | 73.8 | 0.22 | |
| Our method | 0.027 | 94.2 | 0.01 | |
| SOC = 50%, SOH = 100% | TDOA | 0.111 | 65.7 | 0.28 |
| PSO | 0.051 | 82.3 | 0.06 | |
| Single-dim GWO | 0.094 | 70.1 | 0.31 | |
| Our method | 0.023 | 95.1 | 0.01 | |
| SOC = 75%, SOH = 100% | TDOA | 0.107 | 68.2 | 0.19 |
| PSO | 0.043 | 86.7 | 0.03 | |
| Single-dim GWO | 0.040 | 87.4 | 0.05 | |
| Our method | 0.021 | 95.8 | 0.00 | |
| SOC = 100%, SOH = 100% | TDOA | 0.088 | 76.5 | 0.12 |
| PSO | 0.048 | 84.1 | 0.02 | |
| Single-dim GWO | 0.057 | 80.6 | 0.08 | |
| Our method | 0.019 | 96.3 | 0.00 |
The results clearly demonstrate the superiority of our GA-GWO enhanced TDOA method across all SOC conditions. The method maintained a mean DD error below 0.03 and accuracy above 94% for all tested scenarios. The incorporation of SOC as an optimization dimension proved critical for maintaining performance at lower SOC levels, where thermal propagation characteristics deviate significantly from high-SOC behavior.
We further investigated the impact of sensor network configuration on localization performance. Table VIII summarizes the results of our sensor deployment analysis.
| Number of sensors | Deployment pattern | Accuracy (%) | Mean DD error |
|---|---|---|---|
| 2 | Left and front | 80.0 | 0.057 |
| 4 | All four sides | 90.0 | 0.025 |
| 4 | Left and front (asymmetric) | 90.0 | 0.041 |
| 8 | All four sides (two per side) | 95.0 | 0.021 |
| 16 | All four sides (four per side) | 95.0 | 0.020 |
Our analysis revealed that 8 sensors deployed uniformly around the battery cabinet perimeter provided the optimal balance between localization accuracy and system complexity. Increasing sensor density beyond 8 showed diminishing returns, while reducing to 4 sensors maintained acceptable performance at 90% accuracy. Even in the challenging asymmetric deployment scenario, our method achieved 90% accuracy, demonstrating robustness to sensor layout variations.
Discussion on Thermal Path Constraints and Fault Inversion
The thermal path constraints we identified through our experimental and modeling work provide a physical basis for understanding and predicting thermal runaway propagation in energy storage battery systems. The anisotropic thermal conductivity of the jelly roll structure creates preferential propagation paths that can be exploited for fault localization. Our analysis revealed that the propagation velocity in the in-plane direction (vy ≈ 3.2 mm/s at 100% SOC) is approximately 1.5 times faster than in the through-plane direction (vx ≈ 2.1 mm/s at 100% SOC).
The critical SOC threshold for cascading failure in our energy storage battery modules lies between 50% and 75%. Below this threshold, the heat generated by a failing cell is insufficient to overcome the thermal inertia and heat dissipation of neighboring cells, resulting in self-termination. This finding has important implications for energy storage battery management strategies, suggesting that maintaining SOC below 75% can significantly reduce the risk of cascading thermal events.
The integration of our path-constrained thermal propagation model with the GA-GWO optimization framework enables efficient fault source inversion without exhaustive simulation of all possible fault locations. By using the TDOA-based preliminary localization to restrict the search space to a single diagnostic zone, we reduced the computational cost by approximately 83% compared to global search methods.
Conclusion
In this work, we have systematically investigated the thermal runaway behavior of 280 Ah lithium iron phosphate energy storage batteries from single-cell triggering to module-level cascading propagation, and developed a comprehensive fault inversion localization framework. Our key contributions and findings are summarized as follows:
First, through controlled thermal abuse experiments, we established critical thermal runaway parameters for large-format energy storage batteries across the full SOC range. The peak temperature varies from 236.8 °C at 25% SOC to 378.3 °C at 100% SOC, with corresponding temperature rise rates ranging from 2.2 °C/s to 26.5 °C/s. The voltage collapse behavior transitions from gradual decline at low SOC to abrupt step-change at high SOC, reflecting different internal short circuit mechanisms.
Second, we developed a novel FFRLS parameter identification method that establishes time-varying mappings between SOC and side reaction activation energies. This data-driven approach improves the accuracy of conventional Arrhenius-based thermal runaway models by capturing the SOC-dependent reaction kinetics. Our validated model achieves peak temperature predictions within 2.59% of experimental measurements across all tested SOC conditions.
Third, our module-level thermal propagation analysis revealed distinct path constraints governed by the anisotropic thermal conductivity of the battery jelly roll structure. In-plane propagation along the electrode layers is approximately 1.5 times faster than through-plane propagation. We identified a critical SOC threshold between 50% and 75% for cascading failure propagation in energy storage battery modules.
Fourth, we proposed a GA-GWO enhanced TDOA fault inversion localization method that integrates thermal propagation physics with intelligent optimization. The two-stage approach achieves preliminary zone-level identification followed by precise source localization. Our method maintains a mean directional dual-threshold error below 0.03 and accuracy exceeding 94% across all tested SOC conditions, significantly outperforming conventional methods.
The framework presented in this paper provides both fundamental understanding and practical tools for enhancing the safety of energy storage battery systems. The parameter identification methodology can be extended to other battery chemistries and form factors, while the fault localization algorithm is applicable to real-time monitoring and post-event analysis in operational energy storage facilities. Future work will focus on incorporating additional degradation factors such as state of health and temperature history effects, as well as experimental validation of the complete fault localization system in a full-scale energy storage battery cabinet.