Abstract: Abstract: The water environment can be closely related to the living, food, and oxygen for seawater in aquaculture. However, the rapid increase of inorganic nitrogen and active phosphate phosphorus content has indirectly led to the abnormally severe eutrophication of seawater in recent years, due to human activities, environmental pollution, and agricultural production. Water eutrophication has been found in Zhanjiang Bay in China, which is extremely harmful to mariculture. There is some change in the pH and other indicators in the cultured water, which in turn affects the growth of aquatic organisms. Therefore, it is necessary to use real-time monitoring and predict the water quality parameters in the aquaculture process, and then evaluate the quality of aquatic products. This study aims to accurately reveal the mariculture water quality in three-dimensional space for each time series of different water layers. A three-dimensional prediction model was constructed to optimize the water quality time series. Principal component analysis (PCA) was first conducted to screen out the key parameters of water quality. The mutual information (MI) was combined to analyze the non-linear relationship between the key parameters of water quality and output parameters. The water quality parameters were selected with the large correlation as the input of bidirectional long short-term memory (BLSTM) in the feature extraction module using the improved fireworks on BLSTM in hyperparameters for optimization. The learned network was hidden the state vector input into the transformer to focus on more water quality information. The prediction model was trained for the water quality inorganic nitrogen concentration, active phosphate concentration, and pH value for the three-dimensional prediction. The research subject was taken as the sea area in Xuwen County, Zhanjiang City, Guangdong Province, China. The improved model was validated using 13 typical online monitoring stations from June to August 2022, as well as three handheld detection stations with hourly collected data sets. The experimental results show that: 1) The water quality prediction model with the MIPCA-BLSTM-transformer-IFAW outperformed the SC-K-means-RBF, RS-GBRT, 3D cyclic, BRT, M-ConvLSTM, and IPSO-KLSTM prediction model in four metrics of MAPE, RMSE, R and D, where the ICEEMD-BLSTM prediction models were four metrics of 0.075, 0.116, 0.96 and 0.997, respectively. (2) IFAW improved the model prediction performance better than the CS, GA, PSO, and Jaya algorithm. (3) The content of inorganic nitrogen and reactive phosphate increased with the increase in the water depth, whereas, the pH value decreased. The hybrid model can be expected to accurately reveal the trends of inorganic nitrogen concentration, reactive phosphate concentration, and the pH value in different water layers in seawater as well. The next work can continue to simplify the network structure for better model robustness.