Accurate and efficient runoff simulations are crucial for water management in basins. Rainfall-runoff simulation approaches range between physical, conceptual, and data-driven models. With the recent development of machine-learning techniques, machine learning methods have been widely applied in the field of hydrology. Existing studies show that such methods can achieve comparable or even better performances than conventional hydrological models in runoff simulation. In particular, long short-term memory (LSTM) neural networks are able to overcome the shortcomings of traditional neural network methods in handling time series data. However, the impacts of the time memory on rainfall-runoff simulation are rarely studied. In this study, hysteresis effects in hydrology were investigated and the performances of machine learning methods and traditional hydrological models were assessed. The results show that the ANN model is more suitable for monthly scale simulation, while the LSTM model performs better at daily scale. Hydrological hysteresis is important for runoff simulations when using machine learning methods, especially at daily scale. By considering hysteresis in the simulation, the RMSE is significantly improved by 27% (21%) for LSTM (ANN). In addition, LSTM is more robust for time series handling, while the ANN is easier to be overfitted due to the limitation of neural network structure. This study provides new insights into the potential use of machine learning in hydrological simulations.
Physics and Chemistry of the Earth, Volume 123, October 2021,
Artificial Neural Network; Brain; Comparative Study; Comprehensive Comparisons; Data Handling; Hydrological Model; Hydrological Models; Hydrological Simulations; Hydrology; Hysteresis; Long Short-term Memory; Machine Learning; Machine Learning Methods; Machine Learning Techniques; Neural Network Method; Neural Network Structures; Rain; Rainfall-runoff Modeling; Rainfall-runoff Simulation; Rainfall-runoff Simulations; Runoff; Time Series; Water Management; Asia