Reliable forecasting of lithium-ion battery capacity is essential for ensuring safety, efficiency, a long-term performance in applications such as electric vehicles, renewable energy systems, aerospace, and portable electronics. Traditional physics-based and data-driven approaches often face limitations in generalizing across diverse chemistries and degradation pathways. A hybrid deep learning framework is proposed that integrates Residual Convolutional Neural Networks, Transformer encoders, Squeeze-and-Excitation (SE) blocks, and Gated Attention Pooling to effectively capture both short-term degradation signatures and long-term dependencies. Evaluation on the NASA battery degradation dataset demonstrates superior predictive performance, achieving R2 = 0.984 and surpassing established benchmarks across MAE, RMSE, and tolerance accuracy metrics. The architecture enhances robustness against measurement noise and improves interpretability through attention-based mechanisms. Cross-dataset tests from NASA to CALCE shows without any fine-tuning confirm that the proposed model generalizes reliably beyond the training domain. These outcomes highlight the framework’s potential as a scalable and reliable solution for real-time battery health forecasting and predictive maintenance in Battery Management Systems (BMS).