Monthly runoff time series prediction by variational mode decomposition and support vector machine based on quantum-behaved particle swarm optimization

Abstract

Accurate monthly runoff prediction plays an important role in the planning and management of water resources. However, owing to climate changes and human activities, natural runoff often contains a variety of frequency components, and existing monthly runoff estimation techniques may fail to capture potential change processes effectively. To overcome this problem, we have developed a hybrid model for monthly runoff prediction. First, observed runoff is decomposed into several subcomponents via variational mode decomposition. Second, support vector machine models based on quantum-behaved particle swarm optimization are adopted to identify the input-output relationships hidden in each subcomponent. Finally, the total output of all submodules is treated as the final forecasting result for the original runoff. Three quantitative indexes are considered to test the performance of the developed models. The monthly streamflow of two reservoirs in China’s Yangtze Valley is considered as the survey target. This area contains the world’s largest hydropower project (Three Gorges Reservoir) and the waterhead of the middle line of Asia's largest inter-basin water transfer project (Danjiangkou Reservoir). Test results indicate that the hybrid model provides better forecasting accuracy compared to several traditional methods (artificial neural networks and extreme learning machines), making it an effective tool for the scientific operation of hydropower reservoirs.