Abstract: With the global expansion of the new energy vehicle industry, a large number of power lithium-ion batteries are reaching retirement, positioning cascade utilization as a highly sustainable pathway for resource reuse. However, achieving precise state of charge (SOC) estimation remains a major technical bottleneck in ensuring the operational safety and longevity of these degraded batteries, particularly within energy storage systems. This challenge primarily stems from complex internal degradation mechanisms, severe inconsistency among retired cells, and excessive redundancy present in multi-source operational data. Furthermore, traditional time-series neural networks frequently suffer from limited nonlinear fitting capabilities and vanishing gradients when processing dynamic, highly degraded battery datasets. To address these limitations, this study proposes a novel SOC estimation methodology for cascade-utilized lithium-ion batteries based on an integrated feature engineering (FE) framework and a residual-connected long short-term memory (RES-LSTM) network. In the feature engineering phase, high-dimensional temporal features are extracted by mining multi-parameter dynamic measurements, including voltage, current, and temperature. A hybrid feature selection strategy, incorporating linear regression analysis and decision tree-based feature importance evaluation, is subsequently applied to eliminate redundant data and isolate a refined core feature subset strongly correlated with battery SOC. To enhance deep feature representation, residual connections are structurally incorporated into the LSTM architecture, mitigating the vanishing gradient phenomenon and enabling the network to capture intricate, long-term temporal dependencies. The complete FE-RES-LSTM model was constructed and rigorously validated across multiple degradation stages using the public NASA lithium-ion battery dataset. Experimental results demonstrate that the proposed FE-RES-LSTM model achieves superior estimation precision and stability compared with conventional baseline models. Specifically, the root mean square error (RMSE) decreases from 4.39% (for a standard LSTM) to 0.13%, while the mean absolute error (MAE) drops from 3.60% to 0.10%. These quantitative findings confirm that the integration of targeted feature selection and residual deep architectures effectively addresses the trade-off between computational efficiency and accuracy, offering a robust and scalable solution for the intelligent management and safe deployment of retired energy storage batteries.