The remaining useful life (RUL) of lithium-ion batteries is a concern for users, as it relates to the timing of battery replacement and safety. Addressing the non-linear variation trend in the capacity of lithium-ion batteries, a method for predicting the RUL is proposed based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and a bidirectional long short-term memory (BiLSTM)-Transformer. First, the lithium-ion battery capacity data is decomposed using CEEMDAN method. Subsequently, a concatenated model consisting of BiLSTM neural networks and a Transformer network is employed to model and predict the residual sequences obtained from the decomposition and the intrinsic mode component sequences. Finally, the predicted intrinsic mode component sequences and residual sequences are summed, and the RUL is forecast by comparing the final data after summation with the original data. The proposed method is validated using NASA’s publicly available battery dataset. Experimental results demonstrate that the mean absolute, root mean square, mean absolute percentage errors and absolute errors are controlled within 0.0173, 0.0231, 1.2084% and 3 cycles, respectively. The proposed approach effectively enhances the accuracy of RUL prediction for lithium-ion batteries.