A classification-based deep belief networks model framework for daily streamflow forecasting

Data-driven models can achieve high accuracy and low computational cost without a priori knowledge of hydrological system, which have been successfully applied in streamflow forecasting for decades. However, it is still a challenging task to improve the performance of these models, especially for the rivers with dramatic flow changes. In this paper, a new integrated framework for daily streamflow forecasting based on different flow regimes is developed. The framework integrates a Fuzzy C-means (FCM) clustering for streamflow regime identification, a partial mutual information (PMI) for input selection, Deep Belief Networks (DBN) for mapping the nonlinear relationships between the selected inputs and streamflow during different streamflow dynamics processes, and a rigorous validation process considering structure validity to interpret the physical processes simulated using the DBN models. The framework was applied to three streamflow stations with different climate conditions in USA, and the results show that the framework has significantly improved modelling performance (approximately 12%) compared to single data-driven models. The integration of data-driven models and physical process classification also leads to improved integration of physical understanding of the complex characteristic of different flow regimes into the modelling process, leading to overall improved confidence in the developed daily streamflow forecasting models.