Remaining useful life prediction of lithium-ion batteries using a causal STL-based multiscale CNN-BiLSTM-attention network

As a key energy storage component in high-reliability power systems, such as electric transportation and spacecraft, lithium-ion batteries necessitate precise remaining useful life (RUL) estimation to guarantee safe operation and facilitate early fault detection. However, battery capacity degradation is jointly affected by complex electrochemical mechanisms, external environmental disturbances, and cell-to-cell variability. As a result, it is highly nonlinear, shows pronounced local fluctuations, and remains difficult to generalize across different batteries. This paper presents an STL- convolutional neural network (CNN)-bidirectional long short-term memory (BiLSTM)-Attention framework for lithium-ion battery RUL estimation, leveraging recursive multi-step capacity trajectory prediction. First, at each prediction point, rolling STL decomposition is performed using only the historical capacity sequence. The resulting trend, seasonal, and residual components are then integrated with multidimensional features, including cycle number, capacity, state of health, internal resistance, and charging-process parameters, to prevent information leakage from future data. Second, a sliding-window approach is employed to construct time-series samples, enabling the network to simultaneously identify short-term variations and extended degradation trends. Furthermore, a difference-based label learning strategy is introduced to convert absolute capacity prediction into capacity-change prediction, thereby improving the representation of local degradation variations and short-term capacity prediction. By combining the local feature extraction ability of CNN, the temporal dependency modeling ability of BiLSTM, and the attention mechanism’s ability to weight key degradation information, the proposed model enables more accurate prediction of complex nonlinear capacity fading. The proposed framework differs from conventional one-step capacity prediction approaches by performing recursive multi-step capacity trajectory prediction from a predefined starting point, where the RUL is determined as the first cycle at which the predicted capacity reaches the failure threshold. The experimental findings demonstrate that the STL-CNN-BiLSTM-Attention model achieves better performance than CNN, BiLSTM, CNN-BiLSTM, and other decomposition-based models across root mean square error, mean absolute error, and R 2 metrics. The model effectively reduces prediction errors and improves the accuracy and robustness of RUL prediction for unseen batteries, demonstrating its practical potential in complex degradation scenarios.