Examining practical applications of a neural network model coupled with a physical model and transfer learning for predicting an unprecedented flood at a lowland drainage pumping station

• Published in: Paddy and water environment 2023-10, Vol.21 (4), p.509-521
• Main Authors: Kimura, Nobuaki, Minakawa, Hiroki, Kimura, Masaomi, Fukushige, Yudai, Baba, Daichi
• Format: Article
• Language: English
• Subjects: Accuracy; Agriculture; Artificial neural networks; Biomedical and Life Sciences; Datasets; Drainage; Ecotoxicology; Flood data; Flood predictions; Floods; Geoecology/Natural Processes; Hydrogeology; Hydrology/Water Resources; Life Sciences; Machine learning; Manpower; Neural networks; Performance prediction; Pumping; Pumping stations; Rainfall; Soil Science & Conservation; Transfer learning; Water levels
• ISSN: 1611-2490; 1611-2504
• DOI: 10.1007/s10333-023-00944-8

A novel and practical approach for predicting unprecedented flood events caused by heavy rainfalls at drainage pumping stations in lowland areas is proposed and examined. This method combines a deep neural network (DNN) coupled with a physical model and transfer learning (TL). To predict unprecedented flood events, the DNN was initially pre-trained using virtual flood data generated by a physical model. Subsequently, the pre-trained DNN was fine-tuned using observed data from the target field. Using approximately 7.5 years of data at 1 h intervals, including large flood events, we investigated three cases with 1000 virtual flood datasets generated by the pseudo-rainfalls of 100, 300, and 500 mm/72 h. After sufficient fine-tuning, the case of rainfall of 300 mm/72 h demonstrated superior performance in predicting the water level for the largest flood event (caused by approximately 300 mm/72 h), which was not included in the training data during the fine-tuning process. Additionally, we evaluated the impact of the amount of the virtual flood data on the prediction accuracy was evaluated to reduce computational costs and manpower. The results indicate that the fine-tuned DNN requires at least 250 virtual flood datasets to achieve a higher accuracy than a simple DNN (i.e., without the physical model and TL) that relies solely on observed data.