Prediction of the Remaining Useful Life of Lithium-Ion Batteries Based on the Optimized TTAO-VMD-BiLSTM

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries is critical for ensuring the safe operation of equipment, optimizing industrial cost management, and promoting the sustainable development of the renewable energy sector. Although various deep learning-based approaches for RUL prediction have been proposed, their performance is highly dependent on the availability of large training datasets. As a result, these methods generally achieve satisfactory accuracy only when sufficient training samples are available. To address this limitation, this study proposes a novel hybrid strategy that combines a parameter-optimized signal decomposition algorithm with an enhanced neural network architecture, aiming to improve RUL prediction reliability under small-sample conditions. Specifically, we develop a lithium-ion battery capacity prediction method that integrates the Triangle Topology Aggregation Optimizer (TTAO), Variational Mode Decomposition (VMD), and a Bidirectional Long Short-Term Memory (BiLSTM) network. First, the TTAO algorithm is used to optimize the number of modes and the quadratic penalty factor in VMD, enabling the decomposition of battery capacity data into multiple intrinsic mode functions (IMFs) while minimizing the impact of phenomena such as capacity regeneration. Key features highly correlated with battery life are then extracted as inputs for prediction. Subsequently, a BiLSTM network is employed to capture subtle variations in the capacity degradation process and to predict capacity based on the decomposed sequences. The prediction results are effectively integrated, and comprehensive experiments are conducted on the NASA and CALCE lithium-ion battery aging datasets. The results show that the proposed TTAO-VMD-BiLSTM model exhibits a small number of parameters, low memory consumption, high prediction accuracy, and fast convergence. The root mean square error (RMSE) does not exceed 0.8%, and the maximum mean absolute error (MAE) is less than 0.5%.