An Adaptive Rolling Runoff Forecasting Framework Based on Decomposition Methods

Runoff forecasting is crucial for water resources management. However, due to numerous complex factors, the characteristics of runoff sequences are highly variable and difficult to accurately predict, posing a significant challenge. Currently, numerous studies have attempted to integrate different decomposition methods into runoff forecasting to enhance prediction accuracy while analyzing runoff components. Unfortunately, these methods are limited to retrospective forecasting and cannot be applied in practice, and there is a lack of comparative quantitative analysis of different decomposition methods. Therefore, this study proposes an adaptive rolling forecasting framework based on decomposition. At the same time, the results of three decomposition methods, namely empirical mode decomposition, variational mode decomposition, and singular spectrum analysis, were compared and analyzed, and the following conclusions were drawn: (1)Signal decomposition methods like Singular spectrum Analysis(SSA) and Variational mode decomposition(VMD) significantly enhance forecasting performance of Back-propagation neural network(BP), Convolutional neural network(CNN), and Long Short-Term Memory(LSTM), outperforming empirical mode decomposition(EMD).(2)The decomposition principles and characteristics of EMD, VMD, and SSA are different: VMD and SSA preferentially decompose low-frequency and high amplitude components, while EMD is the opposite.(3)Intrinsic mode functions(IMFs) from VMD and SSA show higher sensitivity to the original sequence, retaining more predictive information and improving forecasting accuracy compared to EMD.