In order to improve the prediction accuracy of slurry pressure balance shield attitude in complex strata and ensure tunneling safety and quality, a shield attitude prediction method was proposed based on a convolutional neural network-bidirectional long short-term memory-Attention (CNN-BiLSTM-Attention). Twenty-one key tunneling parameters were selected as inputs to predict six attitude parameters, including the horizontal and vertical deviations of the shield's head and tail, as well as roll and pitch angles. The raw data was preprocessed using boxplot analysis, random forest imputation, discrete wavelet transform, Pearson correlation analysis, and normalization methods. Model hyperparameters were optimized through grid search to achieve improved predictive performance. Ablation experiments and comparative studies were conducted to validate the model's superiority in shield attitude prediction and analyze the contribution of each component to overall prediction performance. The results show that the CNN-BiLSTM-Attention-based method effectively improves the attitude prediction accuracy of slurry balance shield tunneling, showing promising potential for engineering applications.