Abstract: The width of strip steel after continuous annealing and leveling directly affects production costs. To understand the variation in strip width after continuous annealing and leveling, a prediction model based on a convolutional neural network combined with a bidirectional gated recurrent unit (CNN+BiGRU) neural network was developed. First, the mechanism of width variation during the continuous annealing and leveling process was analyzed. During continuous annealing, the strip width is primarily influenced by factors such as temperature, tension, and speed, whereas during leveling, it is mainly affected by parameters such as rolling pressure and front/back tension. Subsequently, data feature selection and preprocessing were conducted. Based on the necking prediction model for the continuous annealing process and the Zaretsky formula, 17 parameters, including temperatures and tensions in different process sections of the furnace as well as process parameters in the leveling section, were selected as input features for the model. To ensure prediction accuracy, Min-Max normalization was applied to standardize the data, followed by data denoising using Tukey's fences method. Next, a CNN and BiGRU combined neural network structure was constructed, comprising two convolutional layers and two gated recurrent unit (GRU) layers. The ReLU activation function was used in the convolutional process, with each of the two GRU layers containing four gated units and connected in series for output. The model was then trained and validated. By comparing the performance of standalone CNN, standalone GRU, CNN+GRU, Random Forest, and LightGBM models, the CNN+BiGRU structure was found to outperform the others in overall performance. During training, the CNN+BiGRU model achieved a coefficient of determination R² of 0.983 9, with mean absolute error (E_MA) of 3.492 66, root mean square error (E_RMS) of 11.616 7, and mean absolute percentage error (E_MAP) of 0.29%. After training, the Shapley additive explanations (SHAP) values of each feature parameter were calculated. It was found that rolling pressure is the most important feature influencing width, with a SHAP value close to 7, significantly higher than other parameters. The tensions in the soaking section and within the furnace ranked second and third in importance, respectively. This finding provides a width adjustment strategy where leveling control serves as the coarse adjustment, and the continuous annealing process is used for fine-tuning. Finally, the trained model was applied in a real production environment, and a self-learning function was developed to enable periodic model updates. By collecting the prediction accuracy of different steel grades, it is found that the predicted value can be controlled within ±3 mm. Comparing width variations across different steel grades revealed that for higher-strength steel grades, the continuous annealing process should be the primary means of width adjustment, whereas for lower-strength steel grades, leveling adjustments should take precedence. The research results realize the accurate prediction of strip width after continuous annealing and leveling, which provides an effective scheme for width adjustment and has good application value.