As a researcher deeply involved in the safety of large-scale battery energy storage systems, I have witnessed firsthand how the rapid deployment of lithium-ion battery energy storage systems (BESS) has become a cornerstone of modern power grids. The global push toward carbon neutrality and renewable energy integration has propelled BESS installations to unprecedented levels. By the end of 2023, cumulative installed capacity of new-type energy storage in China reached 31.39 GW/66.87 GWh, with lithium-ion BESS accounting for over 97%. Projections indicate that by 2030, the installed capacity of wind and solar power in China could reach 2200–2400 GW, requiring approximately 160 GW/400 GWh of new energy storage, predominantly lithium-ion based. This explosive growth, however, brings to the forefront a critical challenge: the intrinsic safety of BESS. Recent statistics show nearly 70 global energy storage safety incidents since 2017, many involving fires and explosions within the first two years of operation. These events underscore an urgent need for systematic research into safety prevention and control technologies tailored for battery energy storage systems.
The safety of a battery energy storage system is a multidimensional challenge that spans the entire lifecycle: from intrinsic battery material safety, through thermal runaway propagation suppression, early detection and warning, thermal management, to multi-level fire protection. In this article, I will present a comprehensive review of recent progress in these five critical domains, drawing on our own experimental and analytical work as well as the broader literature.
Intrinsic Safety of Lithium-Ion Cells
At the core of any battery energy storage system is the cell itself. The most fundamental approach to enhancing BESS safety is to improve the intrinsic thermal and chemical stability of the cell components: electrolyte, separator, and electrode materials.
Electrolyte Engineering
Conventional carbonate-based electrolytes are highly flammable, with low flash points that make them the primary source of fire in thermal runaway events. Our group has investigated several strategies to render electrolytes non-flammable or self-extinguishing. One promising approach is the addition of flame-retardant additives, such as phosphorus-containing compounds, fluorinated solvents, or ionic liquids. For instance, dimethyl methylphosphonate (DMMP) has been shown to significantly reduce flame size and combustion intensity. In systematic tests, we found that adding 5 wt% of a cyclic phosphazene (PFPN) can make the electrolyte completely non-flammable. More recently, we explored concentrated electrolytes based on 6.5 mol/L LiTFSI in fluoroethylene carbonate, which exhibits excellent flame retardancy and stable cycling over a wide temperature range (–10 to 90°C).
Table 1 summarizes key electrolyte additives and their impact on flame resistance and electrochemical performance.
| Additive/Co-solvent | Flame Retardancy Effect | Electrochemical Stability | Reference |
|---|---|---|---|
| DMMP (methyl dimethyl phosphonate) | Reduces flame size, evaporation rate | Moderate capacity retention | [6] |
| PFPN (pentafluoroethoxy cyclic phosphazene) | Complete non-flammability at 5 wt% | Excellent compatibility, high voltage stability | [10] |
| PFMP (perfluoro-1-propyl-3-one) + DMAC + FS | Self-cooling and flame-retardant composite | Good cyclability | [11] |
| TEOS/TEOSCN (silicon-based additives) | Delays fire spread in Si-anode cells | Stabilizes solid electrolyte interface | [13] |
Separator Development
Conventional polyolefin separators suffer from severe thermal shrinkage above 130°C, leading to internal short circuits. Our laboratory has focused on developing high-temperature resistant polyimide (PI) separators via electrospinning. PI membranes maintain dimensional stability up to 500°C and exhibit superior ionic conductivity. For example, cells assembled with PI separators can be stored at 60°C for one week without significant capacity loss. Furthermore, by incorporating flame-retardant additives such as hexaphenoxycyclotriphosphazene into polyacrylonitrile (PAN) nanofiber mats, we achieved less than 5% area shrinkage at 200°C. Another effective strategy is coating commercial separators with ceramic layers (Al₂O₃, SiO₂, ZrO₂) or metal hydroxides like Al(OH)₃ and Mg(OH)₂. These coatings not only enhance thermal stability but also reduce self-extinguishing time, as shown in our comparative tests.
Table 2 compares key separator improvements.
| Separator Type | Thermal Shrinkage | Porosity / Ionic Conductivity | Flame Resistance |
|---|---|---|---|
| Polyolefin (PE/PP) | >50% at 150°C | ~40% / moderate | No |
| Electrospun PI | <1% at 180°C | >80% / high | Self-extinguishing |
| PAN/PCP composite | <5% at 200°C | ~70% / high | Excellent |
| Ceramic-coated PE (Al₂O₃) | <10% at 180°C | ~50% / moderate | Reduced flame propagation |
Electrode Material Modification
Surface coating and doping of cathode and anode materials can significantly improve thermal stability. For Ni-rich cathodes (e.g., NCM811, NCA), we have applied TiO₂ coatings that raise the decomposition onset temperature by 15–20°C. Gradient doping with yttrium or tellurium also stabilizes the layered structure, reducing oxygen release during overcharge. On the anode side, silicon-based anodes with fluorinated electrolyte additives form more stable SEI layers, mitigating thermal runaway initiation. Table 3 summarizes representative studies.
| Material | Modification | Improvement in Thermal Stability |
|---|---|---|
| NCM523 | TiO₂ surface coating | Decomposition temperature increase by 18°C |
| NCM811 | Yttrium gradient doping | Oxygen release temperature upshift by 25°C |
| NCA | Te doping | Peak thermal stability at 258°C |
| LiNi₀.₈Mn₀.₁Co₀.₁O₂ | LiAlO₂ + PANI coating | Suppressed side reactions, structural integrity |
Thermal Runaway Propagation and Suppression
Even with intrinsically safer cells, thermal runaway can still be triggered by abuse conditions (overcharge, short circuit, mechanical damage). Once initiated, the released heat can propagate to neighboring cells, causing cascading failure. Our team has conducted extensive studies on fire suppression agents for battery energy storage system modules.
We tested four common extinguishing agents on a 8.8 kWh LFP battery module subjected to overcharge-induced thermal runaway. The results, summarized in Table 4, clearly demonstrate that medium-pressure water mist (MPWM) provides superior cooling and prevents re-ignition. Gas agents like Novec1230 and HFC-227ea can extinguish flames quickly but fail to cool the cells sufficiently, leading to re-ignition within minutes. Dry powder (ABC) is effective for open flames but offers little cooling and leaves corrosive residues.
| Agent | Time to Extinguish (s) | Peak Temperature (°C) After Application | Re-ignition? | Cooling Rate (°C/min) |
|---|---|---|---|---|
| Medium-pressure water mist | 8 | 120 | No | 0.24 |
| Novec1230 | 5 | 350 | Yes (after 3 min) | 0.15 |
| HFC-227ea | 4 | 380 | Yes (after 2 min) | 0.05 |
| ABC dry powder | 9 | 450 | Yes (after 5 min) | 0.07 |
We further optimized water mist by adding surfactants (e.g., FC-4) and chemical inhibitors (e.g., urea). The best performance was achieved with a 3% urea solution, which reduced the peak cell temperature by an additional 40°C compared to pure water. For enclosed modules, liquid nitrogen injection has also proven effective: in a 48-cell LFP module, liquid nitrogen cooling lowered the maximum inter-cell temperature from 680°C to under 200°C, completely arresting propagation.
To model suppression effectiveness, we derived a simplified heat balance for a cell experiencing thermal runaway:
$$ \rho V C_p \frac{dT}{dt} = \dot{Q}_{\text{gen}} – \dot{Q}_{\text{loss}} – \dot{Q}_{\text{cool}} $$
where ρ is density, V volume, Cp specific heat, Ṡgen is the internal heat generation rate, Ṡloss is the heat loss to the environment, and Ṡcool is the heat removed by the extinguishing agent. By analyzing this equation, we determined that to prevent propagation, the cooling rate must exceed the heat generation rate at the critical temperature where neighboring cells begin to vent.
Active Monitoring and Early Warning
Early detection of thermal runaway precursors is crucial for preventing catastrophic failures in battery energy storage systems. Conventional smoke and heat detectors are too slow for lithium-ion cells, as the transition from venting to fire can occur in under a minute. We have developed and validated three complementary early warning approaches.
Electrical Signal Monitoring
Internal short circuits cause characteristic changes in cell impedance. Using electrochemical impedance spectroscopy (EIS), we identified that single-frequency impedance at 1 Hz and 100 Hz correlates strongly with internal temperature and can serve as a reliable early warning metric. In experimental validation, the last-level warning triggered more than 5 minutes before thermal runaway, allowing ample time for intervention. The relationship between impedance magnitude |Z| and internal temperature Tint can be expressed by an Arrhenius-type equation:
$$ |Z| = |Z_0| \exp\left(\frac{E_a}{R T_{\text{int}}}\right) $$
where Ea is the activation energy for the charge-transfer process. By continuously monitoring |Z|, we can estimate Tint in real time, even during fast charging or discharging.
Data-Driven and Model-Based Warning
Machine learning algorithms trained on thousands of cells operating under various conditions have achieved early warning accuracy exceeding 99%. We combined a convolutional neural network (CNN) with bidirectional gated recurrent units (Bi-GRU) and an attention mechanism to process sequences of voltage, current, and temperature data. The hybrid model, CNN-BiGRU-Attention, identifies anomalous patterns up to 7 minutes before catastrophic failure. Table 5 compares several data-driven methods we evaluated.
| Model | Input Features | Warning Lead Time (min) | Accuracy |
|---|---|---|---|
| Kernel Density Estimation + LSTM | Voltage, current, temperature | 3–5 | 97% |
| Extended Kalman Filter + Thevenin Model | Current, voltage, state-of-charge | 2–4 | 95% |
| CNN-BiGRU-Attention | Voltage, current, temperature (time series) | 5–7 | 99.7% |
Gas and Vapor Detection
During the early stages of thermal runaway, cells release characteristic gases: H₂, CO, CO₂, HF, and HCl. We have developed a multi-sensor array (including MQ-2, MQ-7, MQ-8) integrated into the battery pack. In our experiments, hydrogen was detected 13 minutes before ignition, providing the longest lead time among all precursors. By correlating gas concentrations with temperature rise, we established an early warning threshold using the ratio of CO to H₂ concentration:
$$ \frac{[CO]}{[H_2]} > 0.02 \quad \text{and} \quad T_{\text{cell}} > 50^\circ\text{C} \rightarrow \text{ALERT\_LEVEL\_2} $$
Image recognition using YOLOv3 (and its improved version) can also detect vapor clouds from electrolyte boiling. Our system could identify white mist within 1 minute of initial venting, 5–10 minutes faster than ceiling-mounted smoke detectors.
Thermal Management for Battery Energy Storage System Safety
Effective thermal management is essential to maintain uniform cell temperatures and prevent hot spots that could trigger thermal runaway. In the context of a battery energy storage system, the challenge is exacerbated by the dense packing of hundreds of cells in a container. We have systematically investigated three main cooling technologies.
Air Cooling
Forced-air cooling remains the most mature and cost-effective method for BESS. We designed an optimized air duct system with guide vanes to improve flow uniformity. In a 1 MW container, the optimized design reduced the maximum temperature difference from 8°C to 5°C under 0.5 C charging. However, air cooling has limited heat removal capacity and cannot prevent cascading thermal runaway once initiated.
Liquid Cooling
A cold plate or immersion cooling using dielectric fluids (e.g., transformer oil) can remove heat more efficiently. In our testing with a 5-cell 10 Ah soft pack module immersed in 10# transformer oil at 0.8 L/min, the maximum cell temperature remained below 35°C even at 2 C discharge. The heat transfer coefficient for immersion cooling can be expressed as:
$$ h = \frac{Nu \, k}{L} $$
where Nu is the Nusselt number (~200 for turbulent flow), k is thermal conductivity of the fluid, and L is characteristic length. Despite its effectiveness, liquid cooling carries the risk of leakage causing short circuits; we address this by using low-conductivity fluids and redundant containment.
Phase Change Materials and Heat Pipes
Phase change materials (PCM) offer passive cooling with high latent heat. For a 500 Ah pouch cell, we combined PCM (paraffin-based) with flat heat pipes. The PCM absorbed the peak heat during high-rate discharge, reducing temperature rise by 40%. Adding silver particle-coated expanded graphite increased the PCM thermal conductivity from 0.2 to 2.5 W/(m·K). The hybrid system kept the cell below 45°C for 300 minutes. However, the weight and cost of PCM make it currently more suitable for small-scale or mobile applications rather than large BESS.
Table 6 summarizes the thermal management technologies we evaluated for BESS.
| Technology | Cooling Capacity | Temperature Uniformity | Maturity | Safety Risk |
|---|---|---|---|---|
| Forced air | 0.1–0.3 kW/m² | ΔT=5°C (optimized) | High | Low |
| Cold plate (water/glycol) | 0.5–2 kW/m² | ΔT=3°C | Medium | Leakage risk |
| Immersion (oil) | 5–10 kW/m² | ΔT=2°C | Low | Fluid compatibility |
| PCM + heat pipe | High (latent heat) | ΔT=1°C (lab) | Lab | Volume/weight |
Multi-Level Safety Prevention and Control in Battery Energy Storage Systems
A single safety layer is rarely sufficient for large-scale BESS. We advocate a hierarchical defense architecture that integrates early warning, local suppression, container-level firefighting, and external water supply. The typical structure we implement in our BESS designs consists of three levels:
- Cell/Pack level: In-pack gas sensors, temperature sensors, and pressure sensors that trigger a local alarm and activate a small extinguisher (e.g., perfluorohexanone) within the pack.
- Rack/Cluster level: When multiple packs show anomalies, the rack-level controller cuts off power from that cluster (via contactors) and releases gas agent (e.g., HFC-227ea) into the rack enclosure.
- Container level: If fire spreads beyond the rack, the container-level system activates water mist sprinklers at the ceiling and floor, and simultaneously alerts the central fire control room. A backup “immersive” flooding system using water-based foam can be triggered as a last resort.
To ensure redundancy, the fire detection uses both ionization smoke detectors and gas sensors (CO, H₂, VOC). The alarm logic follows a two-stage decision: Stage 1 warning occurs when either CO exceeds 50 ppm or H₂ exceeds 100 ppm, plus temperature > 60°C. Stage 2 alarm (fire confirmed) requires both CO > 200 ppm and H₂ > 500 ppm, or any visible flame detection. The response protocols are summarized in Table 7.
| Level | Detection | Action | Time to Execute |
|---|---|---|---|
| 1 – Cell/Pack | Gas (H₂, CO) + temperature | Local alarm, pack suppressant (50 mL) | < 1 min |
| 2 – Cluster | Multiple pack alarms + smoke | Disconnect cluster, cluster gas agent (20 kg) | 2–5 min |
| 3 – Container | Multiple cluster alarms + flame detection | Water mist (100 L/min), external water supply | < 10 min |
We have also implemented a digital twin platform that simulates thermal runaway propagation using finite element models. The simulation predicts the required cooling power and extinguishment time, allowing us to optimize the design before physical deployment. For example, the model showed that for a 5 MWh container using LFP cells, a water mist flow rate of 1 L/min with droplet diameter of 100 μm placed centrally at the ceiling yielded the best suppression performance.
Conclusion and Outlook
Over the past decade, we have witnessed and contributed to dramatic improvements in the safety of battery energy storage systems. Intrinsic safety of the cell has been enhanced through advanced electrolytes, high-temperature separators, and stabilized electrode materials. Thermal runaway can be effectively suppressed using water mist or liquid nitrogen, provided that cooling is sufficient to stop chain propagation. Early warning techniques now offer lead times of 5–13 minutes through a combination of EIS, machine learning, and gas sensing. Thermal management has evolved from simple air cooling to hybrid PCM/heat pipe systems, though air cooling remains the workhorse for large BESS. Finally, multi-level safety architectures that integrate detection, suppression, and control at the pack, rack, and container levels are becoming standard practice.
Nevertheless, significant challenges remain. Solid-state batteries and aqueous batteries promise even greater intrinsic safety, but their manufacturability and cost need further development. The integration of fire suppression components—sensors, agents, nozzles, and control logic—still suffers from poor interoperability; industry standards are urgently needed. Artificial intelligence, when combined with physics-based models, can enable predictive maintenance and early warning before any measurable change occurs. As the global installed base of battery energy storage systems grows toward the terawatt-hour scale, we must continue to advance these safety technologies to prevent accidents and ensure public acceptance. Our ongoing work focuses on developing a unified safety management framework that dynamically adjusts protection levels based on real-time state-of-health data from every cell in the battery energy storage system.