A Machine learning framework to predict reverse flow and water level: A case study of Tonle Sap Lake

Abstract

Reliable assessment of the natural interactions in large river–lake systems is vital for water supply planning, flood regulation, and ecosystem services. The existing interaction between the Mekong River (MR), Tonle Sap River (TSR), and Tonle Sap Lake (TSL) has been further complicated due to the reverse flow phenomenon (RF) beyond ongoing anthropogenic activities and climate change. While continuous observations, insufficient data, and event-driven measurement have remained a challenge for such a large basin, accurate prediction of the RF period and water level influenced would pave the way in providing effective remedies for side effects in this area for future scenarios. In this study, the RF periods were investigated employing the K-Nearest Neighbors (KNN) and Ensemble Bagged Tree (EBT) classification algorithms that showed comparable results with observed periods. The predicted periods were then used to predict the daily water level at the confluence where the MR and TSR join using Long Short-Term Memory (LSTM) and Adaptive Neuro-Fuzzy Inference System (ANFIS) methods. Although both models yielded highly accurate water level predictions, a slightly better performance was attributed to the LSTM model. Regarding the TSR’s role in the RF period, results highlighted its significant role in extending the RF period in some years, despite its much lower streamflow than the MR. The developed model could be used as a reliable machine learning (ML) framework where there is a need for highly accurate data regarding RF and water levels for TSL and its environs.