a Model for Predicting River Flooding Using Relatively Small Data Sets
Abstract
Flood damage has become a serious public safety issue in the face of rapid meteorological or climate pattern change and rapid urbanization in recent years. Installation of water gauges should be accelerated as a matter of crisis management as heretofore unseen levels of flood damage from overflowing small- and medium-sized rivers have increased. Various methods and models for predicting floods exist. Storage function models, in particular, need to setparameter values for simulating flood events. While this model can predict river discharge even with a relatively small data set, it is necessary to acquire stage-discharge rating curve (H-Q curve) at a high cost in order to obtain the water levels. Models using artificial intelligence (AI), such as a Neural Network (NN), are used to predict water levels directly using only past rainfall and water levels. However, this model is not feasible when past water level data is insufficient.
This study proposes a model for predicting river flooding that is low-cost and has predictive skill. Our proposed model is a hybrid, incorporating hydrological physics and machine learning. We show that, by optimizing function parameters using machine learning, an appropriate flood forecasting model can be constructed even if the accumulation of water level data is insufficient. Additionally, we show that, with our model, unlike those using AI such as NN, physical interpretation of the predicted results is also possible. We used our proposed model to predict water levels using rainfall and water level data from the previous three years in selected small-sized river with drainage areas of 11.4 km2. Our results showed a predictive skill for events exceeding the dangerous water level (Thread Score) of 86%. As a comparison, we used storage function model and a NN model by using the same data for simulating flood events. The NN model consists of single hidden layer with sigmoid function at the hidden layer and the output layer using mini-batch learning method. We outline the advantages of our proposed method compared with storage function and NN models especially in the high water level region.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFM.H43J2603S
- Keywords:
-
- 1821 Floods;
- HYDROLOGYDE: 1834 Human impacts;
- HYDROLOGYDE: 1840 Hydrometeorology;
- HYDROLOGYDE: 1847 Modeling;
- HYDROLOGY