Daily runoff forecasting by deep recursive neural network

Abstract

In recent years, deep Recurrent Neural Network (RNN) has been applied to predict daily runoff, as its wonderful ability of dealing with the high nonlinear interactions among the complex hydrology factors. However, most of the existing studies focused on the model structure and the computational load, without considering the impact from the selection of multiple input variables on the model prediction. This article presents a study to evaluate this influence and provides a method of identifying the best meteorological input variables for a run off model. Rainfall data and multiple meteorological data have been considered as input to the model. Principal Component Analysis (PCA) has been applied to the data as a contrast, to reduce dimensionality and redundancy within this input data. Two different deep RNN models, a long-short term memory (LSTM) model and a gated recurrent unit (GRU) model, were comparatively applied to predict runoff with these inputs. In this study, the Muskegon river and the Pearl river were taken as examples. The results demonstrate that the selection of input variables have a great influence on the predictions made using the RNN while the RNN model with multiple meteorological input data is shown to achieve higher accuracy than rainfall data alone. PCA method can improve the accuracy of deep RNN model effectively as it can reflect core information by classifying the original data information into several comprehensive variables.