A runoff prediction approach based on machine learning, ensemble forecasting and error correction: A case study of source area of Yellow River

Abstract

Accurately predicting river runoff holds substantial significance to meet several objectives, including basin water resource allocation, flood control, and drought relief. In this study, we introduced a novel multistep runoff prediction model featuring machine learning, ensemble forecasting, and error correction (EC). Five machine learning models served as base learners, providing multistep forecasting results. We introduced an ensemble forecasting strategy (EFS) to effectively combine the forecast results of the base learners. The EFS included scenario segmentation, which divided the runoff time series into flood and nonflood seasons and leveraged the strengths of different predictors to optimize weights individually, and weight combination optimization, which utilized nine optimization methods to fine-tune the weight combinations. Additionally, to further reduce predictable components in the error series, we constructed an EC model to limit forecasting errors. We conducted a series of experiments on the three hydrological stations (Jimai, Maqu, and Tangnaihai) in the source area of the Yellow River to comprehensively evaluate the effectiveness of the proposed runoff prediction approach. The 7-day experimental results demonstrated the following: (1) the EFS integrated base learners to achieve improved performance at each step compared with single models; (2) mean absolute error was notably produced following the EC, with a higher degree of improvement for longer lead days; and (3) the proposed model outperformed other existing models across all experimental stations and lead times.