Flow prediction in the lower Yellow River based on CEEMDAN-BILSTM coupled model

Abstract

As one of the important hydrological elements of rivers, flow is of great significance to the development and utilization of water resources and the ecological environment. Based on the excellent nonlinear processing capability of CEEMDAN and the advantages of BILSTM in time-series data modeling, a coupled CEEMDAN-BILSTM model is constructed for flow prediction, and the i-month flows from 1951 to 2016 are used to predict the i-month flows from 2017 to 2021. The results show that the CEEMDAN-BILSTM coupled model predicts the trend more closely with the actual data variation, and the minimum relative error is 0.56 and maximum 9.48, which are maintained within 10%, and the deterministic coefficients are all greater than 0.9, so the prediction accuracy is high. The flow in month i of 5 years was picked up by monthly predictions for 66 consecutive years, which provides a new way of thinking about the prediction of river flow. HIGHLIGHTS Predicting i-months in the 5 years of 2017–2021 by i-months in the 66 years of 1951–2016 can reduce the large flow volatility caused by the abundant and dry periods due to external conditions such as rainfall.; CEEMDAN can effectively improve the modal mixing problem of EMD method, and the decomposition process of CEEMDAN is complete.;

Keywords

• ceemdan-bilstm
• flow prediction
• middle and lower reaches of the yellow river
• neural network