Temporal prediction of dissolved oxygen based on CEEMDAN and multi-strategy LSTM hybrid model

Dissolved oxygen (DO) is a vital water quality parameter with significant implications for aquatic life, serving as a key indicator of water pollution. The current research introduces two hybrid models utilizing Long Short-Term Memory (LSTM) networks, integrating the preprocessing methods of Discrete Wavelet Transform (DWT) and Complementary Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). These models aim to efficiently model the DO process across five consecutive hydrologic stations situated along the Savannah River. Initially, a wavelet coherence (WTC) analysis was conducted to identify influential parameters for modeling, revealing that water temperature, discharge, mean water velocity, and turbidity exhibited the strongest correlations with dissolved oxygen. In single-gauge temporal modeling, the outcomes demonstrated that the T (IV) model, incorporating all the specified input parameters, achieved superior performance across all stations. Within all gauges, the fourth gauge exhibited the most favorable results using the hybrid CEEMDAN-LSTM method, attaining evaluation criteria of R = 0.991, DC = 0.976, and RMSE = 0.016. Furthermore, it was observed that the dissolved oxygen values from one day before the upstream gauge had a significant impact on predicting the dissolved oxygen concentration in the downstream gauge. In this strategy, the implementation of the CEEMDAN method results in a 35% increase in the modeling accuracy.