Large capacity retired battery packs contain extensive high-dimensional state monitoring data, presenting a technical challenge for accurately predicting their remaining service life (RUL). To enhance prediction accuracy and convergence efficiency of retired batteries' RUL, this digest proposes a two-stage selective deep neural network ensemble method. The first stage involves generating a candidate set under joint perturbation of multiple methods. Strengthening the diversity of candidate deep neural network sets is achieved by adopting a heterogeneous neural network structure, designing multiple time scales, and randomizing algorithm parameters to eliminate internal coupling relationships within the model. The second stage employs a genetic algorithm to integrate and eliminate redundant models, effectively removing underperforming redundant learners. This stage also acquires a diversified optimal candidate subset, which outputs prediction results based on the average ensemble. Comparison with individual model data experiments demonstrates the proposed method concurrently enhances the accuracy and diversity of the integrated model, improving RUL prediction accuracy by nearly 20%. This support significantly benefits operation and maintenance decision-making.