A Hybrid Machine Learning Framework for Real-Time Water Level Prediction in High Sediment Load Reaches

Abstract Accurate prediction of water level is a challenging task in high sediment load reaches because of complicated flood-sediment routing mechanisms and fluctuating reach terrain. We propose a hybrid machine learning framework to predict the water level and validate the framework in five representative high sediment load reaches of the Lower Yellow River (LYR) in China. This framework can automatically generate ensemble prediction models for different flood-sediment scenarios by integrating clustering, classification, and regression models. We compared the proposed framework with two benchmark frameworks and evaluated performances using Nash–Sutcliffe efficiency (NSE) and Pearson correlation coefficient (PCC). The results demonstrated the following: (1) The water level cannot be accurately predicted (NSE 0.85) by adding the time factor in model development. (2) The proposed framework can successfully identify different flood-sediment combinations with a high classification accuracy of 98% by adopting K-means clustering and support vector machine (SVM) classification models. (3) The proposed framework can provide reliable water level prediction for flood early warming in the LYR with an average NSE of 0.83 and 0.89 in the training and testing periods, respectively. It outperformed traditional single and rule-based model frameworks in all reaches with different degrees of improvement between 3% and 13%. The best performance indicated that the proposed framework can take advantages of different type of machine learning models and provide reliable prediction for flood mitigation in high sediment load reaches.