Abstract: The development of large-scale and distributed energy storage is of vital importance for achieving the target of carbon peak and carbon neutrality. Maintaining the safety of battery energy storage (BES) is essential for the security of people's lives and the development of society and economy. However, the existing early warning methods for BES face the following two challenges. On the one hand, the operating status considered by the mechanism-based research methods is single, which is difficult to be promoted and applied. On the other hand, the data-driven methods based on the supervised learning cannot work well with the small number of samples. To solve these problems, a data-driven early warning method for thermal runaway of BES is proposed in this paper. First, based on the idea of unsupervised learning, a data-driven early warning framework for thermal runaway of BES is proposed, in which the reconstruction errors are used to measure the difference between BES. Benefiting from unsupervised learning, the proposed framework can work well with limited samples. Besides, ensemble learning is used to improve the stability of the method while quantifying the probability of thermal runaway of BES. Then, to effectively extract the time-varying characteristics of BES including voltage, temperature, current, state of charge(SOC), etc., and discover the difference of BES before and after the thermal runaway, a bi-long short-term memory (Bi-LSTM) network combined with the attention mechanism is used to accurately extract the features from training data. Finally, case studies based on the practical operating data of the electric vehicle BES are conducted, which verifies the effectiveness of the proposed method.