WATER-LEVEL PREDICTIONS IN A DRAINAGE PUMPING STATION USING A DEEP LEARNING MODEL, COUPLED WITH A PHYSICAL MODEL AND A TRANSFER LEARNING APPROACH

Our study aims to build a deep learning model that can predict even flood events caused by unprecedented rainfall events to improve the accuracy of water-level predictions at a regulation pond in a drainage pumping station. To improve the accuracy of the deep learning model, a physical model created a large amount of pseudo-flood data using virtual rainfall of 100, 300, and 500 mm for 72 h. The pretrained model of these flood data was applied to the observed data using transfer learning. After extracting only the flood events from the observed data, the deep learning model predicted the largest flood event (TOP 1). The pretrained model with 300 mm/72 h produced more accurate results than those of the model without transfer learning. We performed another prediction using the same pretrained model for the observed data, including dailydrainage operations. The results showed that the prediction accuracy was similar to that of the non-transfer-learning model as well as a good reproduction of the peak water level for TOP 1.