Research on Streamflow Forecast Based on EEMD and Long Short-Term Memory

摘要

In this study, two used data preprocessing techniques consisting of discrete wavelet transform (DWT) and ensemble empirical mode decomposition (EEMD), and machine learning models consisting of a deep learning model, i.e. long short-term memory (LSTM), were used to investigated the historical daily streamflow series (from 1/1/1997 to 31/12/2014) for Yangxian station, Han River, China. Data preprocessing techniques were used to decompose the training-development set and test set, respectively. The sub-signals obtained by these preprocessing techniques are modeled using LSTM. The results indicate that the EEMD-LSTM model overwhelms DWT-LSTM model in terms of the root mean square error (RMSE=278.1274), determination coefficient (R2=0.3587), mean absolute error (MAE=217.7862) and peak percentage threshold statistics (PPTS(5)= 2.0150%). The comparisons of DWT-LSTM model and EEMD-LSTM denote that EEMD outperform DWT for LSTM. In generally, EEMD-LSTM is a useful tool for predicting highly non-stationary and non-linearity daily streamflow in a real forecasting application.