Runoff Prediction using Long-Short Term Memory Model

Recurrent Neural Networks (RNNs) are known to solve complex non-linear time series problems. In last decades, due to the limitations of computing power, previous RNN models normally have only one or two hidden layers. Today, with the algorithmic advances (i.e., dropout method and ReLU activation function) and accelerated GPU computing, RNNs with multiple layers can be used to solve more complex problems. Long-Short Term Memory Model (LSTM) is a type of RNNs which can deal with time series problems with the considering of long-term dependencies, and researchers has been trying to use LSTM model in hydrologic modeling in recent years. This study developed LSTM-based deep learning models for hourly rainfall-runoff prediction for the first time, focusing on a typical Midwest watershed, Clear Creek, in Iowa. The models can generate the outputs of runoff for predicted hours with the inputs including runoff from previous hours and the meteorological data (i.e., precipitation and temperature) from both the previous and forecasted hours. The LSTM models show comparable results compared to physical models, which can be used to increase the forecast accuracy and increase the efficiency with standardized machine learning platform for short-term real-time flood forecast.