Deep learning-based dissolved oxygen prediction using encoder–decoder architectures and signal decomposit

DO measurement plays a key role in water quality assessment, as it has a significant impact on the survival of aquatic organisms and is closely related to several other factors. The prominent aim of this article is to accurately predict DO values at three specific stations in California. To accomplish this goal, the study will analyze the variables of temperature, pH level, salinity concentration and discharge in detail. In addition, this research uses various specific models, LSTM, Bi-LSTM, auto-encoder-LSTM, attention-based LSTM-encoder–decoder, VMD-LSTM (variational mode decomposition), WD-LSTM (wavlet decomposition), ADIPLS-LSTM (adaptive dynamic iterative partial least squares) and EMD-DFA-LSTM (empirical mode decomposition with detrended fluctuation analysis) to predict DO values. According to results can be note that EMD-DFA-LSTM model demonstrate superior performance in Do predicrion. Furthermore, the findings demonstrate that data preprocessing significantly enhances prediction accuracy. For example, the highest CC value in evaluated station for LSTM and Bi-LSTM models was obtained at station 2 of the Bi-LSTM model, approximately equal to 0.92, while for pre-processed models, for example, in auto-encoder-LSTM and attention-based LSTM-encoder–decoder, the highest value was obtained at the same station 2, equal to 0.965 for the attention-based LSTM-encoder–decoder model. A Performance comparison of the encoder–decoder-based models and the preprocessed models revealed that the performance of the encoder-based model was more effective. Notably, the EMD-DFA-LSTM model achieved correlation values exceeding 0.97 at each monitoring station, and it substantially decreased the root mean square error in value predictions.