Application of a Multi-Model Fusion Forecasting Approach in Runoff Prediction: A Case Study of the Yangtze River Source Region

Abstract

Runoff forecasting is crucial for sustainable water resource management. Despite the widespread application of deep learning methods in this field, there is still a need for improvement in the modeling and utilization of multi-scale information. For the first time, we introduce the Neural Basis Expansion Analysis with Exogenous Variable (NBEATSx) model to perform runoff prediction for a full exploration in rich temporal characteristics of runoff sequences. To harness wavelet transform (WT) multi-scale information capabilities, we developed the WT-NBEATSx forecasting model, integrating WT and NBEATSx. This model was further enhanced by incorporating a Long Short-Term Memory (LSTM) model for superior long-term dependency detection and a Random Forest (RF) model as a meta-model. The result is the advanced multi-model fusion forecasting model WT-NBEATSx-LSTM-RF (WNLR). This approach significantly enhances performance in runoff prediction. Utilizing a daily scale runoff and meteorological dataset from the Yangtze River Source region in China from 2006 to 2018, we systematically evaluated the performance of the WNLR model in runoff prediction tasks. Compared with LSTM, Gated Recurrent Units (GRUs), and NBEATSx models, the WNLR model not only significantly outperforms the original NBEATSx model but also surpasses other comparison models, particularly in accurately extracting cyclical change patterns, with NSE scores of 0.986, 0.974, and 0.973 for 5-, 10-, and 15-day forecasts, respectively. Additionally, compared to the standalone LSTM and GRU models, the introduction of wavelet transforms to form WT-LSTM and WT-GRU notably improved prediction performance and robustness, especially in long-term forecasts, where NSE increased by 32% and 1.5%, respectively. This study preliminarily proves the effectiveness of combining the cyclical characteristics of NBEATSx and wavelet transforms and creatively proposes a new deep learning model integrating LSTM and RF, providing new insights for further considering multi-scale features of complex runoff time series, thereby enhancing runoff prediction effectiveness.