Spatio-Temporal ESN-based Model for Predicting Water Level in Yangtze River

Water level prediction plays a vital role in navigation, particularly for ports and harbors, which attracts more attention in the academic domain. The conventional machine learning algorithms, such as support vector regression and artificial neural network, are commonly employed for predicting water level. However, these algorithms have the high computational cost, which directly impacts on their performance and implementation. Recent, a typical randomization-based algorithm called echo state network obtains the good performance with high efficiency in aspect of water level prediction. To enhance forecasting performance and overcome its own limitations, this study proposes three approaches for enriching features in the conventional echo state network. Firstly, we propose the random bar features selection approach to select useful features. Secondly, a novel structure of ESN is designed to capture much richer temporal features. Next, a spatial feature extraction method is proposed based on a latent correlation and graph neural network to capture the spatial features. Based on these methods, this study proposes a Spatio-temporal ESN-based model (ST-ESN). The experimental results and statistical analysis represent that our proposed model has superior forecasting ability in water level prediction rather than other compared models. It not only overcomes the limitation of single features consideration, but it also provides the data support from the high accuracy water level prediction data, which is benefited in managing water-land transportation, flood protection, and ship route management.