Abstract
As the casting industry transitions towards digitalization and intelligent systems, precise control of the return water temperature in the cooling system during deep well casting has become a critical factor in enhancing casting quality. However, traditional numerical simulation methods struggle to handle the complex, dynamically coupled scenarios involving multiple process parameters. This paper proposes a hybrid model, MShOA-CNN–LSTM-Attention, which integrates convolutional neural networks (CNN), long short-term memory networks (LSTM), and an attention mechanism for high-precision prediction of return water temperature. First, an input matrix is constructed using cooling system parameters, melt state data, and control instructions (valve opening) collected from industrial sites. The model employs CNN to capture spatial correlations among the input features, LSTM to identify long-term dependencies in the time series, and the attention mechanism to assign weights, focusing on critical stages of the process. Additionally, the mantis shrimp optimization algorithm (MShOA) is applied to optimize the key hyperparameters in the model. Validation with actual production data shows that the proposed model significantly outperforms comparison models in predicting return water temperature, with the coefficient of determination (R(2)) improving by 4.5% and the mean absolute error decreasing by 1.17. The model also demonstrates superior performance in time series prediction, and its results provide valuable guidance for the dynamic regulation of subsequent cooling water control.