[A-HW20] Integrated Watershed Management under the Future Extreme Disaster

[AHW20-01] Discussing the impact of input factors on predicting river levels in data-driven models: using the Long Short Term Memory model as an example

*Wei-Ting Lee¹、Tsunhua Yang¹ (1.National Yang Ming Chiao Tung University)

キーワード：real-time water-level prediction, LSTM, hydrological model, machine learning, disaster risk reduction

Predicting river water levels is critical to reducing disasters caused by flooding. Decision-makers can take precautionary measures accordingly to reduce the impact of disasters. The prediction of river water levels is based on the establishment of numerical models. Theoretically, there are two types of numerical models: physically based models and data-driven models. The first type is built based on the physical connection, such as the rainfall-runoff relationship, and it has the best performance in prediction accuracy in comparison with the second type. However, building a physically based model is time-consuming and challenging, especially for areas with limited data. Establishing data-driven models is relatively fast, but the selection of temporal and spatial input factors to the models and their causes for predicting outcomes has not been systematically studied. Therefore, the purpose of this study is to propose reference criteria for factor selection in data-driven models. Considering the geographical and hydrological parameters that would affect water level prediction, this research selected hydrological factors such as rainfall and river water level and basin characteristic factors such as watershed area, circularity ratio, and stream length as input factors to train the data-driven model to produce water level prediction up to 5 hours in advance (i.e., t+5). In this research, a synthetic hydrologic basin was constructed using the HEC-HMS hydrologic and HEC-RAS hydraulic models to generate the data needed for model training and verification. The data-driven model used is the Long Short -Term Memory (LSTM) model, whih has proven its ability to efficiently predict time series data. Preliminary results have shown that the selection of input factors should be determined in advance according to the watershed characteristics in order to achieve maximum prediction efficiency. For example, the length of the concentration time can affect the weighting of rainfall in the LSTM model, thus affecting the water level prediction results. The best performance was obtained with only up to 2 hours ahead rainfall data as input factors.