Improved Transformer Model for Enhanced Monthly Streamflow Predictions of the Yangtze River

Over the past few decades, floods have severely damaged production and daily life, causing enormous economic losses. Streamflow forecasts prepare us to fight floods ahead of time and mitigate the disasters arising from them. Streamflow forecasting demands a high-capacity model that can make precise long-term predictions. Traditional physics-based hydrological models can only make short-term predictions for streamflow, while current machine learning methods can only obtain acceptable results in normal years without floods. Previous studies have demonstrated a close relation between El Niño-Southern Oscillation (ENSO) and the streamflow of the Yangtze River. However, traditional models, holding the encoder–decoder architecture, only have one encoder block that can not support bivariate time series forecasting. In this study, a transformer-based double-encoder-enabled model was proposed, called the double-encoder Transformer, with a distinctive characteristic: “cross-attention” mechanism that can capture the relation between two time series sequences. Using river flow observation collected by the Yangtze River Water Resources Commission and El Niño-Southern Oscillation (ENSO) observation collected by the National Oceanic and Atmospheric Administration, the model can achieve better performance. By using variational mode decomposition (VMD) technique for preprocessing, the model can make precise long-term predictions for the river flow of the Yangtze River. A monthly prediction of 21 years (from January 1998 to December 2018) was made, and the results indicate that the double-encoder Transformer outperforms mainstream time series models.