Accurately predicting battery target variables is essential for reliable battery management, but the nonlinear and time-varying nature of batteries makes this task difficult. This study introduces GAN-LION, a conditional adversarial regression framework designed for multi-target battery prediction using historical operating sequences. Unlike traditional noise-driven GANs, the proposed model formulates prediction as a deterministic sequence-to-one regression task, where the generator predicts the next target value from a lookback window and the discriminator judges whether the input-target pair aligns with the real data distribution. To specifically examine the impact of optimization, the same adversarial architecture is trained with either the Lion or Adam optimizer and is also compared with CNN-LSTM-GRU and Autoencoder-LSTM baselines. A single end-to-end pipeline is used for preprocessing, normalization, temporal sequence construction, and chronological train-validation-test splitting to avoid information leakage. Results from two benchmark battery datasets indicate that GAN-LION achieves the strongest overall performance. Its advantage is most evident on the more challenging dataset, where it yields lower prediction errors, smoother convergence, and more stable gradient behavior. On the simpler dataset, the performance differences are less pronounced, but GAN-LION still remains competitive and stable. Overall, these results suggest that conditional adversarial regression with Lion-based optimization is an effective and technically sound method for battery target prediction.
Alatawi et al. (Fri,) studied this question.