Abstract: Nearshore wave height exhibits significant difference in different sea areas, and most of the time-series prediction models lack the adaptability to wave height prediction in different regions (i.e., multi-source wave height prediction). To solve this problem, we propose a wave height prediction model based on multi-period trend decomposition (STL) with locally weighted regression and two-level fusion strategy (MSTL-WH). Combining the multi-periodicity, nonlinearity and non-stationarity of nearshore wave height time series, periodogram method is used to extract four major periods of the wave height, and multiple STL decomposition was performed based on the major periods, decomposing the original wave height series into seasonal, trend and residual parts. Then, we used the Long Short-Term Memory Network (LSTM) combined with two-level fusion strategy build a nearshore wave height prediction network. Finally, self-attention mechanism was used to readjust the weights and the wave height was calculated for the next 12 hours. Compared with the existing time-series forecasting methods, the method we proposed has better practicability and smaller errors in nearshore wave height forecasting.