In the grinding system of heavy calcium carbonate, particle size was esteemed as a crucial metric for assessing product quality. The precise prediction of particle size was deemed instrumental in controlling product quality and informing the adjustments of vertical roller mill (VRM) parameters. Therefore, the study proposed a model for predicting the particle size of heavy calcium powder, which combined the bidirectional time convolutional network (BiTCN) based on the ivy algorithm （IVYA） with the bidirectional gated recurrent unit (BiGRU). Initially, feature and label data were subjected to preprocessing, employing time-varying filtering empirical mode decomposition in conjunction with wavelet thresholding to eliminate high-frequency noise from the primary current of the vertical mill. Then, the correlations between multidimensional features in the time series were explored from both forward and backward directions through BiTCN. At the output end of BiGRU, an attention module was incorporated to assign different weights to each position, thereby effectively focusing on the key data in the sequence. Finally, the IVYA was integrated into the overall model to ascertain the optimal solutions for critical hyperparameters within the neural network. Actual measured data from a carbonate powder mill factory were subsequently employed to simulate particle size prediction. The experimentation indicates that the optimized model using the IVYA exhibits superior predictive performance compared to other single and combined models. Specifically, it achieves the following metrics: root mean square error of 0.824 4, mean absolute error of 0.423 0, mean relative percentage error of 1.295 4%, and coefficient of determination of 98.95%.