Refining units in southern coastal China operate in complex service environments where multiple environmental factors are strongly coupled, making metallic equipment prone to accelerated corrosion during long-term operation. High-accuracy prediction models are therefore needed to support equipment health management. To address the limited availability of corrosion data in refining units (target-domain sample size N=12), this study proposes an atmospheric corrosion-rate prediction method termed TL-BiLSTM-Attention, which integrates transfer learning with a BiLSTM-Attention network. First, a source-domain dataset is constructed from the public MICAT corrosion dataset. Then, orthogonal experiments are conducted to optimize the hyperparameters of the BiLSTM-Attention network to improve the performance of the pre-trained model. Finally, the optimal pre-trained model is transferred to the local corrosion dataset of the refining units, and fine-tuning is performed under different layer-freezing strategies. Results show that, in the source domain, the BiLSTM-Attention model achieves stronger fitting capability and lower errors than LSTM and BiLSTM. In the target domain, TL-BiLSTM-Attention improves the prediction accuracy of atmospheric corrosion rates for materials such as Q235, 304, 316L, 5A06 and different metals require differentiated freezing strategies to obtain the best performance. This study verifies the predictive effectiveness and engineering applicability of TL-BiLSTM-Attention under few-shot conditions and provides a data-driven tool for corrosion assessment in refining units.