Abstract: Given the challenges associated with predicting water level time series, attributed to their mixed linear and nonlinear characteristics and high uncertainty, we propose a combined model, termed EMD-DELM-LSTM, integrating empirical mode decomposition (EMD), long-short-term memory network (LSTM), and deep extreme learning machine (DELM). In this framework, DELM and LSTM operate in parallel and in series with EMD. Initially, the original signal is decomposed into distinct intrinsic mode functions (IMFs) via EMD, categorizing them into high, medium, and low frequency signals. These signals are then fed into the DELM-LSTM parallel structure for prediction and reconstruction. To validate the efficacy of the model, we utilize data from a lake at a university in Guangzhou. Results indicate superior performance compared to EMD-LSTM, EMD-DELM, LSTM, DELM, and BiLSTM models across various time scales, with the most pronounced enhancement observed at the 40-minute scale. Notably, performance improves by 43.08%, 22.92%, 45.79%, 30.92%, and 47.31% when compared to the respective reference models. These findings underscore the reliability and stability of our proposed model for water level prediction across different temporal scales.

Key words: water level prediction, EMD-DELM-LSTM, empirical mode decomposition, multi-time scale analysis, artificial neural network