Two-stage Variational Mode Decomposition and Support Vector Regression for Streamflow Forecasting

Abstract

Abstract. Streamflow forecasting is a crucial component in the management and control of water resources. Decomposition-based approaches have particularly demonstrated improved streamflow forecasting performance. However, it is not practical to firstly decompose the entire streamflow into several signal components and then divide the data samples of each component into training and validation sets for signal component prediction. This impracticality is due to the fact that some validation information, that is not available in practical streamflow forecasting, is used in that training process. Unfortunately, firstly dividing the entire streamflow into training and validation sets and then decomposing each set separately lead to undesirable boundary effects and complicated forecasting. Moreover, establishing a model for each signal component is quite laborious and summing the component predictions may lead to error accumulation. In addition, summing the decomposition results may sometimes lead to inaccurate reconstruction of the original streamflow. In order to address these shortcomings of decomposition-based models and improve the forecasting performance in basins lacking meteorological observations (e.g., precipitation and temperature), we propose a two-stage decomposition prediction (TSDP) framework, realize this framework using variational mode decomposition (VMD) and support vector machines (SVR), and refer to this realization as VMD-SVR. In the first stage of the TSDP framework, the entire streamflow data was divided into training and validation sets, each of which was then separately decomposed to avoid the influence of validation information on training. In the second stage, a single model for streamflow prediction was established using a set of mixed shuffled samples. This scheme saves the modelling time and reduces the influence of the boundary effects. We demonstrate experimentally the effectiveness, efficiency and reliability of the TSDP framework and its VMD-SVR realization in terms of the boundary effect reduction, decomposition performance, prediction outcomes, time consumption, overfitting, and forecasting capability for long leading times. Specifically, five comparative experiments were conducted based on the ensemble empirical mode decomposition (EEMD), singular spectrum analysis (SSA), discrete wavelet transform (DWT) and SVR. The experimental results on monthly runoff collected from three stations at the Wei River show the superiority of the TSDP framework compared to benchmark models.

Daily streamflow forecasting using hybrid long short-term memory model

“…As we all know, the reason for the successful promotion and application of the data-driven models is the model's powerful mining of the internal connections of historical data [18]. Continuously improving the data mining capabilities of the data-driven based streamflow forecasting models is the key to ensuring the accuracy of predictions.…”

“…The first way is to improve the data-driven model structure itself. Deep learning techniques developed in recent few years are confirmed to outperform traditional machine learning methods in numerous streamflow forecasting applications owing to its ''deeper" representations [18,19]. When solving time-correlated input signals, the traditional ANNs' straight-forward and straight-out structural characteristics make it process only the current input information, cannot use the previous information.…”

A 500-year runoff reconstruction for European catchments

“…This can be achieved through a process-based model of varying complexity, with the advantage of following general physical laws -e.g., preserving mass balance, etc. Physical based models: MIKE SHE (Im et al, 2009) and VELMA (Laaha et al, 2017) or data-driven methods, such as support vector machines (Ji et al, 2021;Zuo et al, 2020), artificial neural networks (ANNs; Kwak et al, 2020;Hu et al, 2018;Senthil Kumar et al, 2005), random forests (Breiman, 2001;Contreras et al, 2021), and Shannon entropy (Thiesen et al, 2019) The objective of the present study is to provide a long-term, hydrological reconstruction for the Central European catchments, utilizing the available gridded precipitation (Pauling et al, 2006) and temperature (Luterbacher et al, 2004) reconstructions, natural proxies (Ljungqvist et al, 2016) and other long-term historical data sources. Specifically, we use a combination of a conceptual hydrological model (GR1A; Mouelhi et al, 2006) and two data-driven models (Chen et al, 2020;Okut, 2016) to simulate the annual evolution of runoff over the period 1500-2000.…”