Abstract: Longitudinal cracks are common and serious quality defects in slab production. Accurate prediction of longitudinal cracks is of great significance to improving slab quality and production efficiency. However, in actual production, the number of longitudinal fissure samples is far less than that of normal samples, resulting in extremely unbalanced data distribution, which affects the effect of model training and brings huge challenges to model construction. How to improve the generalization ability and prediction accuracy of the model has become a key problem. To this end, firstly based on a large amount of measured temperature data of thermocouples on the copper plate of the mold, the sliding window technology was applied to extract temperature samples of longitudinal cracks′ and other conditions. Then, a slab surface depression-type longitudinal crack prediction method based on grid search optimized Convolutional Neural Network(CNN) and Bidirectional long Short-Term Memory(BiLSTM) network was proposed. The samples were input into the CNN-BiLSTM network, CNN was used to obtain the local features of the time series, and BiLSTM was used to obtain the long-term dependency features. Finally, the slab longitudinal crack prediction output was performed through the fully connected layer. Experimental results show that the proposed grid search optimized CNN-BiLSTM model performs significantly better than other models on the test set, with a prediction hit rate of 99.29% for longitudinal crack temperature waveforms, a false alarm rate of 0.71%, and a Matthews Correlation Coefficient(MCC) as high as 0.96, and the training and prediction speeds of the model are relatively fast. The research results provide a reliable theoretical basis for the identification of longitudinal cracks on the slab surface.