Thermal Runaway Prediction of Lithium-ion Battery based on Dynamic Pruned LSTM

This research raised a dynamic pruned long short-term memory (LSTM) architecture for predicting thermal runaway of lithium-ion battery in real time, and thus dealt with the key demand for high-precision anomaly detection under strict latency limits. Our method incorporated an over-parameterized LSTM network into a lightweight policy network to prune redundant calculations adaptively in the process of inference. Hence, computing efficiency was refined without decreasing detection performance. The policy network assessed input data features, like temperature variations and voltage fluctuations, to yield binary masks which selectively motivated LSTM cells. Compared with static models or rule-based detectors, our method employed dynamic sparsity and input-dependent thresholds, which were learned end-to-end to safeguard robustness under various working conditions. Results proved a recall rate of 98.7% for thermal runaway cases. The novelty lies in the synergy of adaptive computing and temporal modeling. Our work provides an effective and scalable solution for detecting early thermal runaway, and therefore advancing the domain of battery safety.