Accurate water quality prediction is crucial for the sustainable management of mariculture, yet it faces challenges in modeling high-dimensional nonlinear data and optimizing model hyperparameters efficiently. To address this, a novel deep learning model named ICPO-ACNN-BiGRU-CA is proposed, which integrates an Improved Crested Porcupine Optimizer (ICPO) with a cross-attention (CA) mechanism. The model utilizes an Attention-based Convolutional Neural Network (ACNN) and a Bidirectional Gated Recurrent Unit (BiGRU) to extract spatio-temporal features, while the CA mechanism dynamically fuses these features. The ICPO algorithm is applied to optimize the model’s hyperparameters automatically. Validation on a dataset from a Zhanjiang aquaculture zone shows that the model achieves high prediction accuracy for key parameters, including water temperature (R² = 0.939), dissolved oxygen (R² = 0.935), pH (R² = 0.927), and electrical conductivity (R² = 0.930), while reducing the mean absolute error by 41.4–62.8%. These results demonstrate that the proposed model significantly enhances prediction accuracy and robustness, providing a reliable tool for real-time water quality warning and precision management in mariculture systems.