<p>High-precision forecasting of the state of health (SOH) for lithium-ion batteries is a critical technology for ensuring the safe and reliable operation of energy storage systems. To address the limitations of existing deep learning methods for SOH prediction—such as the difficulty in simultaneously capturing local temporal patterns and long-range dependencies, as well as the heavy reliance on manual hyperparameter tuning—this study proposes a novel hybrid multi-step forecasting model. The model organically integrates the Dream Optimization Algorithm (DOA), Bidirectional Long Short-Term Memory (BiLSTM), and an improved Transformer architecture. In the data preprocessing stage, Z-score standardization and Lagrange interpolation were employed to detect and correct abnormal data among multiple extracted characteristic parameters. On this basis, this paper improves the standard Transformer architecture by replacing the original single encoder with a dual-encoder structure. Specifically, the extracted features are fed into one encoder, while historical SOH values are input to the other encoder. This design leverages the strong correlation between features and SOH to predict future SOH values. Since the standard Transformer is insensitive to local features, a BiLSTM is introduced to address this limitation. Furthermore, considering the critical role of hyperparameter selection, this study employs the recently proposed metaheuristic algorithm—Dream Optimization Algorithm (DOA)—to determine optimal hyperparameters, thereby further enhancing prediction accuracy. Experimental results demonstrate that the proposed model significantly outperforms comparative models across different multi-step prediction horizons, with the MAE consistently remaining below 1.8%.</p>