A data-driven method for predicting abnormal operating states in zinc-based flow batteries

Zinc–bromine flow batteries are prone to zinc-deposition induced instability during long-term cycling, which can trigger performance degradation. How to leverage observable signals such as voltage and current to achieve cross-cycle state prediction and early anomaly warning remains an urgent challenge. To address this issue, this study proposes a joint prediction framework, PatchTST_BiLSTM, which integrates a patch-based time series Transformer with a bidirectional long short-term memory network. The framework targets multi-cycle sequence modeling by learning cross-cycle evolution patterns from historical cycles, which are then used to constrain and correct the voltage and multi-channel current forecasts, thereby improving cross-cycle prediction stability and increasing sensitivity to abnormal evolution trends. Experiments were conducted using multi-condition cyclic voltage–current data collected from a self-developed automated charge–discharge platform. The results demonstrate that the proposed method outperforms multiple baseline models in both voltage and current prediction tasks, with R 2 values reaching 0.99 across multiple test cycles, while reducing the mean squared error of current and voltage prediction by approximately 45% and 85%. In addition, the model can accurately predict degradation-related descriptors such as voltage-step and the cycle-wise mean-square deviation of multi-channel currents, enabling automated identification and early warning of abnormal operating states in zinc–bromine flow batteries. • A data-driven framework predicts voltage and four-channel currents in Zn–Br flow batteries. • Cycle-level voltage step strength and current MSD are extracted as degradation indicators. • The method enables early detection of abnormal operating states related to zinc deposition. • Long-cycle experiments demonstrate high accuracy and stable cross-cycle predictive behavior.