Since traditional data-driven methods are not effective in cross-domain lithology prediction, researches recently focus on improving the model through learning the feature distribution across the datasets of different wells. However, the sequential relation in logging curves has not been fully extracted in logging data processing, and the distribution discrepancy of lithology class proportion is neglected by the classification model. In this paper, a sequential logging data-driven cross-domain lithology identification model is proposed, which mines spatial neighborhood information while learning data distribution discrepancy of both features and classes. To extract variation trend with depth of logging curves, a temporal convolutional network is designed for lithology prediction. An improved class-balanced self-training algorithm is developed based on semi-supervised learning, where the data distribution in target well is shared with source well. Besides, a model ensemble method is adopted in lithology identification by conducting results averaging to promote stability. Verification experiments on logging datasets of a shale oil field show that the proposed method can learn the sequential characteristics as well as data distribution and realize accurate cross-domain lithology identification.