Abstract: To improve the mid- and long-term runoff forecasting of a watershed, this paper develops a new method integrating a comprehensive runoff index, factor reduction, and an improved deep belief networks model. First, we examine the consistency of runoff at different hydrological stations and construct a comprehensive runoff index to characterize the abundance and drought of runoff in the watershed. And we apply a partial mutual information approach to select key factors from the multiple factors, and the selected key factors are taken as inputs of deep learning. Then, an improved deep belief networks (IDBN) model is developed for mid- and long-term runoff forecasting. In a case study of Yalong River basin, this model is compared with several state-of-the-art forecasting models: multivariable linear regression (MLR), autoregressive integrated moving average (ARIMA) model, backpropagation neural networks (BPNN), support vector machines (SVM), and typical deep belief networks models. Results demonstrate our method can significantly reduce computational cost and improve forecasting accuracy, thus helping the mid- and long-term runoff forecasting of watersheds.

Key words: hydrological forecasting, mid- and long-term runoff forecasting, comprehensive runoff index, partial mutual information, deep belief networks