In order to solve the problem that the traditional statistical methods and machine learning methods cannot predict the expected state parameters of marine variable frequency seawater cooling system, a time series prediction model based on bidirectional gated recurrent unit (BiGRU) was proposed, and the data normalization method was used to process the characteristic data of high span order of magnitude to realize the parameter optimization of hidden layer neurons by combining with sensitivity analysis method. Taking the parameter data set of variable frequency seawater cooling system generated by MATLAB Simulink simulation platform as the characteristic sample data for training, the evaluation indexes of mean square error (MSE) and adjusted coefficient of determination (Adjusted R²) were used to evaluate the prediction performance of the model, and the recurrent neural network (RNN) and unidirectional gated recurrent unit (GRU) models were established for comparison to analyze the robustness of different time series data prediction algorithm models. The results show that compared with RNN model, the value of MSE of BiGRU model based on parameter optimization is reduced by at least 73.13%, and the value of Adjusted R² is increased by at least 6.44%; compared with GRU model, MSE value decreases by at least 67.86%, and Adjusted R² value increases by at least 3.35%. Compared with GRU and RNN models, BiGRU model can more accurately predict the expected state parameters of the system with excellent prediction accuracy and stability, which can provide accurate data support for the ship safety assessment system, and has reference value for the parameter prediction of ship variable frequency seawater cooling system.