A Deep Learning Method With Merged LSTM Neural Networks for SSHA Prediction

Abstract

Sea surface height anomaly (SSHA) is an elemental factor in ocean environment and marine engineering. Oceanography models can forecast SSH by data simulation, but the accuracy decreases heavily when it predicts a little long time ahead. In this article, a deep learning method, named merged-long short term memory (LSTM), is proposed to predict SSHA. Specifically, SSHA prediction is treated as a time series forecasting problem, and our merged-LSTM can mine the discipline hidden in short time series, and tackle long-term dependence of series changes. Data experiments conducted on SSHA dataset of China Ocean Reanalysis in the South China Sea show that our method achieves average predicting accuracy plus/minus standard deviation of coming 24 h, 48 h, 72 h, 96 h, and 120 h by 90.99±10.56%, 85.49±13.93%, 79.99±16.08%, 74.23±18.05%, 68.15±18.84%, respectively. The proposed method performs better than several state-of-the-art machine learning methods, including artificial neural network, merged-recurrent neural network, time convolutional network, merged-gate recurrent unit, and one-dimensional convolutional neural network in predicting SSHA.