Improving the prediction of streamflow in large watersheds based on seasonal trend decomposition and vectorized deep learning models

Abstract

Accurate streamflow prediction is essential for water resource management and ecological conservation. With climate change and human activities intensifying extreme weather events, the risks associated with floods have grown, threatening both socioeconomic robustness and ecological integrity. Conventional prediction methods, such as physical and statistical models, often struggle to capture the complex nonlinear and nonstationary characteristics of streamflow. To address this challenge, this study presents a vectorized hybrid STL-LSTM-GRU-Transformer model designed to enhance prediction accuracy and stability. The approach begins by applying seasonal-trend decomposition using loess (STL) to separate streamflow data into trend, seasonal, and residual components. These components are then modeled independently: long short-term memory (LSTM) and convolutional neural networks (CNN) handle trend and seasonal patterns, while gated recurrent units (GRU) and Transformer process residual fluctuations. Furthermore, the model incorporates the Runoff process vectorization (RPV) method alongside vectorization techniques to improve sensitivity to extreme events. Evaluated on 2010–2022 data from six Jialing River stations, the model achieves 0.9991 (NSE), outperforming 12 benchmarks. SHAP analysis identifies dew point temperature (26.7 % contribution) and solar radiation (15.7 %) as key drivers, while kernel density estimation provides reliable probabilistic forecasts (PICP = 0.90 at 95 % CI). Demonstrating robust performance in flood-drought transition prediction (NSE > 0.9983), this approach contributes valuable insights for advancing flood early warning systems and hydro-ecological security.