基于LSTM神经网络的我国典型试航海域环境短期预报方法研究

Research on Short-Term Prediction of Typical Trial Sea Environment in China Based on LSTM Neural Network

提出了一种基于LSTM(Long-Short Term Memory)神经网络的海洋环境短期预报模型。鉴于传统的梯度优化网络参数通常倾向于收敛到较差的局部解。为避免训练网络陷入局部解的困境,论文首先采用自动编码器和解码器对网络权重参数进行初始化。其次,在网络的训练过程中,利用改进的粒子群算法优化网络权重参数。最后,以我国东海、南海和黄海典型试航海域的风、浪时间序列数据为研究对象进行试验。试验结果表明,该模型在短期范围预报取得了较好的预报精度。

In this paper, a short-term prediction model of the marine environment based on LSTM （long-short term memory） neural network is proposed. However, training network parameters with traditional gradient optimization will often converge to a poor solution. In order to get better prediction results, we take two stages of optimization of network parameters. Firstly, auto-encoder （AE） is used to extract features automatically for the initial LSTM neural network model, and expression characteristics of AE are taken as initial parameters of LSTM neural network for unsupervised learning. Secondly, in order to improve the prediction accuracy in supervised learning phase, .the particle swarm algorithm （PSO） is used to optimize the parameters of the network. Data on wind speed and wave height in three sea areas （the East China Sea, the South China Sea and the Yellow Sea） are used to test the model, and test results show that the model has good prediction accuracy in short-range forecast.