In the context of large-scale applications of lithium-ion batteries, accurate prediction of battery surface temperature—a core technology for thermal runaway warning—remains a critical challenge due to insufficient prediction accuracy, which severely limits the efficacy of safety management systems. To address this, this paper proposes a hybrid architecture integrating Empirical Mode Decomposition (EMD) and Bayesian Optimization (BO) with a Convolutional Neural Network (CNN)-Bidirectional Long Short-Term Memory (BiLSTM) neural network, achieving a breakthrough in temperature prediction accuracy through innovative technological integration. First, Empirical Mode Decomposition is applied to decompose raw temperature signals into multimodal components, effectively resolving their nonlinear fluctuation characteristics. Subsequently, a cascaded CNN-BiLSTM model is developed, where the CNN extracts local spatial features, while the BiLSTM captures bidirectional temporal dependencies. On this basis, a Bayesian Optimization algorithm is further introduced to adaptively tune hyperparameters, minimizing prediction errors by constructing a Gaussian process surrogate model. To validate the effectiveness and generalizability of the proposed algorithm, comparative experiments were conducted not only against baseline but also under varying temperature conditions. The results demonstrate that the proposed method outperforms common models across all evaluation metrics and achieves robust prediction performance under different temperature scenarios.