<p>Accurately predicting the remaining useful life (RUL) of lithium-ion batteries (LiBs) is paramount for optimizing maintenance schedules and ensuring the reliability of energy storage systems. However, achieving high-precision RUL prediction remains critically dependent on the selection of features extracted from the data and the efficacy of model training strategies. To address these challenges, this paper proposes a novel RUL prediction method based on feature optimization and an ensemble deep learning model, CGLA (CNN-GRU-LSTM-AM). Initially, a systematic health features (HFs) extraction and correlation analysis is conducted. The minimum redundancy-maximum relevance (MRMR) algorithm is then employed to select representative HFs, ensuring both strong correlation with battery capacity degradation and minimal inter-feature redundancy. Subsequently, to enhance the capability of latent information extraction, both manually engineered HFs and features automatically learned by a convolutional neural network (CNN) are fused, significantly improving the relevance and quality of the input features for the subsequent RUL prediction model. Furthermore, to achieve high-accuracy RUL prediction, a novel CGLA ensemble model is proposed, combining CNN, gated recurrent unit (GRU), long short-term memory (LSTM) networks, and an attention mechanism (AM) to capture complex temporal dependencies and focus on critical degradation patterns. Finally, the proposed method is rigorously validated using the CALCE, MIT, and NASA datasets across three representative stages of LiBs’ lifespan (early, middle, and late). Experimental results demonstrate exceptional prediction accuracy, with MAE, RMSE, and MAPE consistently maintained below 0.0064, 0.0082, and 0.0036, respectively. These findings underscore that the proposed method substantially improves both the accuracy and generalization capability of LiBs RUL prediction.</p>