Predicting Surface Water Level based on Wavelet Transform and Long Short-term Memory Network

Abstract

Water level prediction work is crucial to provide timely water information for flood control and drought relief and shipping. In this study, a hybrid surface water level prediction model (W-LSTM) based on wavelet transform and deep learning model LSTM was constructed to provide an accurate surface water level modeling prediction method. First, to extract the hidden features behind the original water level time series, it is decomposed into multiple sub-series utilizing wavelet transform. Then, combine with a deep learning algorithm to build an LSTM model for each component to make the prediction. Finally, the predicted values of the individual components are summed to produce the ultimate prediction results. The results show that after pre-processing by wavelet transform and then combining the advantages of the LSTM algorithm can significantly improve the prediction accuracy, and W-LSTM demonstrates better prediction power than the other comparative models (BPNN, ELM, LSTM, W-NN, and W-ELM). The evaluation criteria R², RMSE, and MAE of the model are 0.9575, 0.1465, and 0.0737, respectively. The W-LSTM model can achieve accurate and effective water level prediction, which can provide a scientific reference basis for water environment management.