To overcome the limitations of the standard hippopotamus optimization algorithm (HOA), such as inadequate global exploration and limited information sharing in complex optimization problems, an improved hippopotamus optimization algorithm (IHOA) was developed. The global search capability was enhanced by introducing a dynamic exploration factor, the efficiency of information sharing was improved through a multi-information interaction strategy, and the individual position update mechanism was optimized to increase both the accuracy of optimization and the convergence speed. The effectiveness of IHOA was verified through validation on nine benchmark test functions, where superior optimization performance was demonstrated. When applied to the task of liquid loading prediction using multilayer perceptron (MLP) neural networks, and compared with the unoptimized model, the coefficient of determination(R²) was increased by 6%, the root mean square error (RMSE) was reduced by 7%, and the mean absolute percentage error(MAPE) was decreased by 2.21%. These results confirmed that the prediction accuracy and optimization efficiency of the model were significantly improved. The proposed method provides a practical and effective solution for hyperparameter tuning and predictive modeling, offering high potential for engineering application and theoretical research value.