Attention in MLP: A new architecture for urban sewer overflow and flood depth prediction

Abstract

Study region

This research focuses on the vicinity of the A8 Metro Station in Guishan District, Taoyuan City, Taiwan, an area prone to frequent urban flooding. With storm sewer water level and surface flood depth data available, the region offers diverse rainfall conditions and topographical variations. This enables a thorough assessment of model performance for managing overflow risks and inundation.

Study focus

We propose an innovative Attentive Multilayer Perceptron (AM-MLP) architecture, comparing it against widely used sequence models (long short-term memory (LSTM), gated recurrent unit (GRU), and bidirectional LSTM (BiLSTM)). We systematically evaluate sewer water level and flood depth forecasts to test whether attention mechanisms can compensate for MLP’s weak sequence handling. A unified experimental setup ensures fair baseline comparisons, highlighting each model’s strengths and weaknesses.

New hydrological insights for the region

This study provides valuable hydrological insights for the study area around the A8 Metro Station in Guishan District, Taoyuan City, Taiwan. The results demonstrate how the AM-MLP model improves urban flood and sewer overflow predictions in regions with limited or discontinuous data. The model’s ability to capture key hydrological factors, such as variations in rainfall and drainage system limitations, allows for more accurate flood depth and sewer water level forecasts. These insights contribute to better flood risk management and urban resilience planning in regions facing extreme rainfall events.