Battery energy storage systems, especially those based on lithium-ion batteries, are critical infrastructure for modern power grids. They enable the integration of renewable energy sources, improve grid stability, and support the transition toward carbon neutrality. However, the frequent occurrence of thermal runaway incidents in battery energy storage systems poses severe safety risks, often leading to fires or explosions. Thermal runaway is the result of a chain of exothermic reactions triggered by internal short circuits, which can be induced by electrical abuse (overcharge, overdischarge), thermal abuse (overheating), or mechanical abuse (penetration, crush). To ensure the safe operation of large-scale battery energy storage systems, early active safety warning techniques that can detect anomalies before thermal runaway becomes irreversible are essential. This paper systematically reviews four categories of early warning methods: signal‑feature‑based, model‑based predictive, data‑driven, and hybrid strategies. We discuss their theoretical foundations, implementation paths, advantages, and limitations. The goal is to provide a comprehensive reference for researchers and engineers working on the safety of battery energy storage systems.
1. Signal‑Feature‑Based Safety Warning Methods
In battery energy storage systems, sensors collect raw physical signals such as voltage, current, temperature, pressure, gas concentration, and acoustic emissions. Feature extraction algorithms then transform these signals into representative characteristics. Anomaly detection algorithms (e.g., threshold comparison, difference calculation) convert the features into anomaly indicators, and when the indicators exceed predefined thresholds, warnings are issued. The general workflow is illustrated in the conceptual diagram below.
Among various signals, expansion force (or pressure) has shown great potential for early detection. During normal operation, reversible volume changes occur due to lithium intercalation/deintercalation. Before thermal runaway, irreversible deformations caused by lithium plating, gas generation, or structural degradation are often observed. Studies have indicated that pressure signals can be detected much earlier than voltage or temperature signals. For example, experiments using hard‑shell batteries under different thermal runaway triggers revealed that the expansion force signal could provide warnings several hundred seconds before catastrophic failure. The expansion force evolution was categorized into three stages, and the force signal was found to be more sensitive than voltage or temperature in the early stage.
| Signal Type | Detection Time Advantage | Limitations |
|---|---|---|
| Voltage | Easily measurable, threshold‑based detection for overcharge/discharge | Weak response in LiFePO₄ cells; masked in series strings |
| Temperature | Directly indicates exothermic reactions | Large internal‑surface gradient; delayed response in large‑format cells |
| Gas (H₂, CO, etc.) | Early detection before venting (e.g., H₂ appears 639 s earlier than smoke) | Sensor cost; diffusion delay; environmental interference |
| Acoustic (valve opening sound) | Low cost, fast (accuracy ~92%) | Occurs after safety valve opens – too late for prevention |
| Expansion force / pressure | Very early (473 s average before thermal runaway in module‑level tests) | Requires force sensors; calibration needed for different cell formats |
In battery energy storage systems, multi‑signal fusion is increasingly advocated to overcome the limitations of single‑signal methods. For instance, a three‑level early warning strategy based on voltage inflection point, temperature, and gas species (H₂, CO) was proposed for LiFePO₄ batteries under overcharge. Another study optimized the placement of H₂ detectors in container‑type battery energy storage systems, showing that three detectors could advance detection by 116 s compared to a single detector. Nonetheless, most signal‑feature methods still rely on fixed thresholds, which suffer from high false‑positive or false‑negative rates when operating conditions vary. To improve reliability, adaptive thresholding or statistical methods are being integrated.
2. Model‑Based Predictive Safety Warning Methods
Model‑based methods utilize equivalent circuit models (ECMs), electrochemical models, or mathematical models to simulate battery behavior and detect anomalies. These approaches often target specific abuse mechanisms, such as internal short circuits, overcharge, or external short circuits.
Internal short circuit (ISC) is a common precursor to thermal runaway. Detection methods typically rely on the correlation between electrical/thermal signals and the short‑circuit state. For example, an ECM‑based online ISC detection algorithm calculates the residual charging capacity difference between cells in a pack. When a cell exhibits a higher self‑discharge rate, it may indicate an internal short. Table 2 summarizes representative model‑based detection strategies.
| Method Basis | Key Principle | Detection Time / Performance |
|---|---|---|
| Electrochemical‑thermal 3D coupled model | Correlates voltage, current, temperature with ISC resistance | Online detection possible |
| Residual charging capacity variation | Compares charge acceptance between cells | Validated with external resistor emulation |
| Symmetric loop topology | Detects imbalance in parallel‑series hybrid packs | Efficient and accurate |
| Self‑discharge estimation via sigmoid model | Reduces evaluation time by 2× | Early evaluation within hours |
| State‑of‑charge (SOC) monitoring | Detects early ISC with >100 Ω shunt resistance | Detected within 20.4 h online |
Although model‑based methods provide a mechanistic understanding, they suffer from several challenges in battery energy storage systems. Model parameters are difficult to converge under varying charge/discharge profiles and coarse data sampling. Most models focus on a single failure mode (e.g., ISC) and cannot cover complex, multi‑factor degradation paths. Moreover, the computational cost of high‑fidelity electrochemical models hinders real‑time deployment. Some efforts use simplified ECMs combined with Kalman filters to estimate states like SOC and SOH, but the accuracy still depends heavily on model fidelity and input data quality.
3. Data‑Driven Safety Warning Methods
Data‑driven methods leverage historical operational data from battery energy storage systems to train models that detect anomalies without explicit knowledge of failure mechanisms. These methods can be divided into statistical approaches and machine learning (ML) approaches.
3.1 Statistical Approaches
Statistical methods extract features from voltage, temperature, or resistance signals using metrics such as Shannon entropy, Z‑score, or clustering. For instance, the modified Shannon entropy of voltage signals is calculated for each cell, and Z‑score thresholds are applied to identify outliers. In a real‑world vehicle dataset, this method successfully warned of thermal runaway hours before the event. Another method uses two‑dimensional clustering (voltage vs. SOC) to find abnormal cells, enabling periodic risk assessment. The entropy calculation can be expressed as:
$$E_j = -\sum_{i=1}^{n} p_i \log(p_i)$$
where \( p_i \) is the probability density of the voltage of the \(i\)‑th cell in a sliding window. A sudden increase in \(E_j\) indicates growing inconsistency among cells.
| Category | Typical Algorithms | Advantages | Disadvantages |
|---|---|---|---|
| Statistical | Shannon entropy, Z‑score, DBSCAN, k‑means | Simple, interpretable, low computational cost | Requires large datasets; threshold tuning sensitive; decoupled from physics |
| Machine Learning | LSTM, CNN, DBN, SVM, Random Forest | Powerful feature extraction; can handle nonlinear relationships | Needs abundant fault data; poor generalizability to unseen scenarios; high computational burden |
3.2 Machine Learning Approaches
ML methods are widely used for state estimation (e.g., SOC, SOH) and fault diagnosis. Long short‑term memory (LSTM) networks predict future voltage profiles; deviations between predicted and measured voltages indicate anomalies. A typical LSTM architecture for voltage forecasting can be represented by:
$$h_t = \text{LSTM}(x_t, h_{t-1}, c_{t-1})$$
$$\hat{v}_{t+\Delta} = W \cdot h_t + b$$
where \( x_t \) is the input vector (voltage, current, temperature) at time \(t\), \(h_t\) and \(c_t\) are hidden and cell states, and \(\hat{v}_{t+\Delta}\) is the predicted voltage after \(\Delta\) steps.
Another approach estimates battery state of health (SOH) using deep neural network ensembles. For example, a model trained on five different cell datasets achieved 98.9% of samples with error less than 5%. However, the lack of labeled fault data in real battery energy storage systems remains a major bottleneck. To address this, some works use an indirect strategy: first train a model on healthy data to predict normal behavior, then monitor the prediction error for anomaly detection. This reduces the need for fault samples but still suffers from distribution shift under varying operating conditions.
4. Hybrid Safety Warning Strategies
To overcome the limitations of individual methods, hybrid strategies combine signal features, model prediction, and data‑driven techniques. The main goals are to improve accuracy, expand applicability, and enhance reliability.
- Improving model performance: Kalman filters are often combined with ECM to reduce noise; particle swarm optimization (PSO) speeds up parameter identification. For remaining useful life (RUL) prediction, relevance vector regression (RVR) is fused with unscented Kalman filter (UKF) to handle regeneration phenomena.
- Expanding application scenarios: Data‑driven models trained for one driving cycle may fail in another. By classifying driving conditions into categories (e.g., urban, highway) and training separate models, the hybrid system adapts to various scenarios. A backpropagation neural network (BPNN) was trained per scenario to predict voltage, and the prediction error served as a safety indicator.
- Enhancing reliability: Multi‑level warning architectures are common. For instance, Stage 1 uses entropy‑based voltage anomaly detection; Stage 2 incorporates gas sensors; Stage 3 triggers emergency protocols. This reduces false alarms and delays unnecessary maintenance.
| Category | Key Features | Suitable Scale | Main Limitations |
|---|---|---|---|
| Signal‑Feature | Uses voltage, temperature, gas, force; threshold‑based | Cell/module level; early R&D | Sensor cost; threshold sensitivity; environmental drift |
| Model‑Based | ECM, electrochemical, mathematical; targets specific failures | Cell level; small‑scale monitoring | Parameter convergence; single‑failure focus; high computational cost |
| Data‑Driven | Statistical or ML; large‑data driven; no physics needed | System level with abundant data | Poor interpretability; data dependency; generalization issues |
| Hybrid | Combines multiple methods; multi‑level alerts | Large‑scale battery energy storage systems | High computational resource; error propagation; deployment complexity |
5. Challenges and Future Directions
Despite the progress, several key challenges remain for early active safety warning in battery energy storage systems:
- Physical level: Internal sensors (e.g., pressure, gas) suffer from corrosion and high cost. External signals have intrinsic delays. Multi‑physical fusion (force, electric, thermal, gas, acoustic, optical) is needed to reduce false alarms, but coupling mechanisms between different signals are not well understood.
- Data level: High‑quality labeled fault data are scarce. Data acquisition and storage strategies must balance cost and reliability. Interpretable AI methods are required to link data‑driven outputs with actual degradation mechanisms.
- System level: Most methods are validated at cell or small module scale. Scaling up to full‑size battery energy storage systems (MWh‑scale) introduces issues of communication latency, heterogeneity among cells, and computational overhead. Customized solutions for grid‑scale, residential, and automotive applications are needed.
Future research should focus on developing smart batteries with embedded multi‑sensor arrays, advancing physics‑informed machine learning that respects electrochemical constraints, and designing modular, scalable warning architectures that can be integrated into existing battery management systems. Only by addressing these challenges can battery energy storage systems achieve the level of safety required for widespread deployment.
6. Conclusion
This paper has provided a comprehensive review of early active safety warning techniques for thermal runaway in battery energy storage systems. Four categories – signal‑feature‑based, model‑based predictive, data‑driven, and hybrid strategies – were systematically analyzed. Each method has its strengths and weaknesses; no single approach is universally optimal. Signal‑feature methods offer the earliest detection but rely on costly sensors; model‑based methods provide mechanistic insight but are narrow in scope; data‑driven methods leverage large data but suffer from interpretability and generalization; hybrid strategies combine advantages but increase complexity. Moving forward, the integration of multi‑physical signals, interpretable AI, and system‑level validation will be key to enabling safe and reliable operation of battery energy storage systems.