Short-term runoff prediction with GRU and LSTM networks without requiring time step optimization during sample generation

Abstract

Runoff forecasting is an important approach for flood mitigation. Many machine learning models have been proposed for runoff forecasting in recent years. To reconstruct the time series of runoff data into a standard machine learning dataset, a sliding window method is usually used to pre-process the data, with the size of the window as a variable parameter which is commonly referred to as the time step. Conventional machine learning methods, such as artificial neural network models (ANN), require optimization of the time step because both too small and too large time steps reduce prediction accuracy. In this work two popular variants of Recurrent Neural Network (RNN) named Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks were employed to develop new data-driven flood forecasting models. GRU and LSTM models are in theory able to filter redundant information automatically, and therefore a large time step is expected to not reduce prediction accuracy. The three models (LSTM, GRU, and ANN) were applied to simulate runoff in the Yutan station control catchment, Fujian Province, Southeast China, using hourly discharge measurements of one runoff station and hourly rainfall of four rainfall stations from 2000 to 2014. Results show that the prediction accuracy of LSTM and GRU models increases with increasing time step, and eventually stabilizes. This allows selection of a relatively large time step in practical runoff prediction without first evaluating and optimizing the time step required by conventional machine learning models. We also show that LSTM and GRU models perform better than ANN models when the time step is optimized. GRU models have fewer parameters and less complicated structures compared to LSTM models, and our results show that GRU models perform equally well as LSTM models. GRU may be the preferred method in short term runoff predictions since it requires less time for model training.