Improved Streamflow Forecasting Using Variational Mode Decomposition and Extreme Gradient Boosting

Abstract:

Numerous studies have recently shown the potential for Machine Learning (ML) to enhance the prediction of hydrological processes such as runoff and evaporation (Moosavi et al., 2017; Rezaie-Balf et al., 2019). Studies investigating the prediction of streamflow have explored a wide variety of ML models, along with different preprocessing and decomposition techniques (Wang et al., 2018; Zuo et al., 2020). More specifically, extreme gradient boosting (XGBoost) has been shown to be an effective approach for streamflow forecasting (Ni et al., 2020). Many preprocessing techniques, including various forms of time series decomposition such as wavelet transform, empirical mode decomposition (EMD) and Gaussian mixture models, have been proposed to improve upon ML streamflow prediction models (Moosavi et al., 2017; Ni et al., 2020), but while many provide good hindcasting results, they produce poor forecasting results (Zuo et al., 2020), which commonly occurs due to the boundary effect. Variational Mode Decomposition (VMD) was recently proposed (Dragomiretskiy & Zosso, 2014) as an alternative approach to EMD (shown to effectively isolate underlying cyclical patters related to external factors) that minimizes the boundary effect when producing Intrinsic Mode Functions (IMFs) from complex time series data (Zaji et al., 2019). VMD has recently been shown to provide better streamflow decomposition prior to applying ML models (Zuo et al., 2020). While other studies have developed a forecasting model to predict each IMF (Ni et al., 2020; Wang et al., 2018), this approach uses VMD as a decomposition technique to be directly input into a single XGBoost model to predict daily streamflow, reducing the overall computational demand. This VMD-XGBoost model has been shown to accurately predict stream flow in the Bow River (Alberta, Canada), with r², RMS and MAPE of 0.83, 33.0, and 2.66, respectively. Notably, a Horizontal Error of 0.39 shows the model’s resiliency to imitation error, a common limitation of similar models (Zaji et al., 2019).