A hybrid model for monthly runoff forecasting based on mixed signal processing and machine learning
Shu Chen,Wei Sun,Miaomiao Ren,Yutong Xie,Decheng Zeng
DOI: https://doi.org/10.1007/s11356-024-35528-4
IF: 5.8
2024-11-29
Environmental Science and Pollution Research
Abstract:Monthly runoff forecasting plays a critically supportive role in water resources planning and management. Various signal decomposition techniques have been widely applied to enhance the accuracy of monthly runoff forecasting. However, the forecasting of different components, generated through the runoff decomposition, often relies on homogeneous models that utilize identical algorithms or similar structures. The use of a homogeneous model to forecast all components may result in low forecasting accuracy for individual components, which, in turn, impacts the overall forecasting performance negatively. To address this issue, we propose a mixed signal processing model for monthly runoff forecasting, which combines signal processing with heterogeneous machine learning methods that employ different algorithms or structures. Specifically, the SVM and LSTM models are utilized to forecast the original monthly runoff and all components of the monthly runoff decomposed by the Variational Mode Decomposition (VMD), or each component individually. We compare the forecasting models without signal processing and those with either homogeneous or heterogeneous forecasting models that incorporate signal processing. For validation, the Pingshi Hydrological Station in the Lechangxia Basin was selected as the target station. The results demonstrate that the optimal hybrid model, based on mixed signal processing, exhibits a superior performance when compared with the optimal SVM, LSTM, VMD-SVM, and VMD-LSTM models. Specifically, its validation R avg values increased by 3.2%, 3.5%, 0.9%, and 1.2%, respectively, while its validation RMSE avg values decreased by 4.7%, 3%, 1%, and 1%, respectively. The input variables of the optimal hybrid model primarily include sea surface temperature and geopotential height at 500 hPa, suggesting that these factors have a more impact on the monthly runoff in the Lechangxia Basin. This study underscores the importance of selecting a suitable forecasting model for the different characteristics of components, which aids in improving the overall performance of monthly runoff forecasting with signal processing. Moreover, it highlights that reliance solely on teleconnection factors as input variables may not be sufficient for ensuring the accuracy of monthly runoff prediction models.
environmental sciences