Accurate State-of-Charge (SOC) estimation is pivotal for safeguarding lithium-ion batteries in distributed energy storage systems, especially under dynamic operational scenarios involving fluctuating temperatures, varying charge/discharge rates, and progressive battery aging. To address the limitations of existing methods in handling multi-scale temporal dependencies and cross-condition generalization, this study proposes an innovative hybrid architecture that uniquely combines BiLSTM and TCN. The BiLSTM bidirectionally captures electrochemical dynamics, while the TCN employs dilated convolutions to hierarchically model local temporal patterns. An attention mechanism further enhances robustness by adaptively weighting sensor signals under thermal and aging variations. The BiLSTM module bidirectionally captures electrochemical dynamics across charge/discharge cycles, while the TCN hierarchically extracts local temporal patterns through dilated causal convolutions, ensuring strict causality. The proposed BiLSTM-TCN model demonstrates superior estimation accuracy and faster convergence compared to conventional BiLSTM models. It further exhibits robust generalizability to untrained ambient temperatures and battery aging levels. We validate the constructed model using lithium-ion battery cycling data collected under diverse charging and discharging rates and ambient temperatures. With a more accurate initial SOC, the BiLSTM-TCN model converges to the true SOC faster. The model has a mean absolute error (MAE) within 1.5% and a root-mean-square error (RMSE) within 2%, which compares favorably with traditional forward machine learning approaches for SOC estimation.