Research on the potential of the deep learning-based "decomposition-optimization-reconstruction" method in runoff prediction for typical climate- and human-regulated basins in northern China.

River runoff may be affected mainly by the natural climate or human activities, and runoff series present complex characteristics, such as non-stationarity, which makes accurate prediction of runoff challenging. To address the problem that the prediction accuracy of the traditional deep learning methods is affected by the non-stationarity of runoff, which is based on the idea of "decomposition - optimization - reconstruction", this paper constructs a combination model that introduces variational mode decomposition (VMD) and the whale optimization algorithm (WOA) to optimize a bidirectional long short-term memory neural network (BiLSTM) (VMD-WOA-BiLSTM). The combination model is applied to runoff prediction in typical climate- and human-regulated watersheds in northern China, specifically in the semi-arid regions of the Hailar River Basin and the Dahei River Basin. The results show that the "decomposition-optimization-reconstruction" model significantly improves the prediction accuracy. The model excels in upstream runoff prediction because there are fewer human activities in those areas compared to the downstream areas. When applied to rivers, it more accurately forecasts climate-driven runoff changes and performs better for rivers with relatively large total runoff, which may be because they are less impacted by extreme precipitation events compared with rivers with small total runoff. The model's prediction performance varies across different seasons, which may be related to the seasonal characteristics of runoff and the model's inherent predictive capabilities. The combined model achieves excellent runoff prediction results across various river segments and basins, demonstrating its wide applicability for climate- and human-regulated basins in northern China.