A wave forecasting algorithm is established based on the stacking ensemble machine learning method. The algorithm includes two sublayers. The multi-layer perceptron model and the random forest model are used in the first sublayer, and the extreme randomized tree model is used in the second sublayer. To suppress the numerical oscillation of the predicted results, the neighborhood averaging method is introduced into the algorithm. Toestablish relations between the wind speed and the significant wave height, the wind speed and China wave(CWAVE)data from January to September 2016 in the offshore area of the Yangtze River Estuary are used. By using the wind data from October to November 2016, the corresponding significant wave height is predicted by the wave forecasting algorithm. The results show that the values of R² between the predicted wave height and CWAVE data are larger than 0. 97 in the 45 days from 1 October, the mean error is less than 0. 08 m, and the mean relative error is less than 0. 05. The trends of the variations of the significant wave height can be accurately predicted by the algorithm. However, the error increases in the last 15 day, which is due to lack of cold wave data in the training set.