Aiming at the problems of poor real-time, small storage data and low quality of ship power load data prediction, this paper proposes a new short-term load prediction method that combines data interpolation, wavelet threshold denoising and ProbSparse Self-Attention mechanism. Firstly, in data preprocessing stage, the database is expanded by interpolation to meet the model training requirements without affecting the characteristics and trends of the original data; at the same time, taking into account noise disturbance in the original ship power load data, in order to reduce its impact on the model prediction effect, a new wavelet threshold denoising method is adopted to process the original signals and to improve the quality of the data. Secondly, by introducing probabilistic sparse self-attention mechanism in the forecasting model. While effectively capturing the dependencies and important features in the time series power data, it reduces the memory resource consumption and model complexity. It meets the real-time requirements of ship power load forecasting and realizes the double optimization of forecasting accuracy and efficiency. In the comparison experiments with other models, this paper's model reduces at least 13.1% and 18.6% on average in the two indicators of root mean square error and average absolute percentage error, respectively, and improves the efficiency by more than 24.0% on average, and the results show that the method has obvious advantages in the accuracy and efficiency of the ship power load data forecasting model.