A new strategy for prediction of water qualitative and quantitative parameters by deep learning-based models with determination of modelling uncertainties
Mojtaba Poursaeid, Amir Hossein Poursaeed
Abstract
This study presents a new method based on three types of deep learning-based models (DLM) for estimation of water parameters. The DLM models were recurrent neural networks (RNN), long short-term memory (LSTM), and bidirectional long short-term memory (BiLSTM). The study areas were the Colorado River basin in the United States and the Mighan Wetland in Iran. The electrical conductivity (EC), dissolved oxygen (DO), total dissolved solids (TDS), chloride ions (Cl), and river flow rate (debi) were simulated by the DLM models. The Wilson score (WS) uncertainty analysis results for Colorado modelling showed that LSTMdebi, RNNDO, and RNNEC were the best models in simulating due to having the lowest errors (Mean ei equal to 0.36, −1.50, and −0.59), respectively. Finally, the highest value of the R2 index, 0.998, was achieved by the LSTM model in modelling the debi parameter, and 0.996 in EC modelling, in the Mighan Wetland.