Yang et al. (2026) Attention in MLP: A new architecture for urban sewer overflow and flood depth prediction

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-18
• Authors: Song-Yue Yang, Bing-Chen Jhong, Rui-Wen Lin, Min-Chien Tsai
• DOI: 10.1016/j.ejrh.2025.103088

Research Groups

• Department of Urban Planning and Spatial Information, Feng Chia University, Taichung, Taiwan
• Department of Marine Environment and Engineering, National Sun Yat-sen University, Kaohsiung, Taiwan
• Geographic Information System Research Center, Feng Chia University, Taichung, Taiwan

Short Summary

This study proposes a novel Attentive Multilayer Perceptron (AM-MLP) architecture for urban sewer overflow and flood depth prediction, demonstrating that the attention mechanism significantly improves MLP's predictive accuracy, especially in regions with limited or discontinuous data, making it competitive with sequence models.

Objective

• To propose and evaluate a novel Attentive Multilayer Perceptron (AM-MLP) architecture for urban sewer overflow and flood depth prediction, specifically investigating if an attention mechanism can compensate for the traditional MLP's limitations in handling sequential hydrological data.
• To compare the AM-MLP's performance against established sequence models (LSTM, GRU, BiLSTM) for both sewer water level and surface flood depth forecasting.
• To analyze the discrepancies between sewer water level and surface flood depth forecasts and discuss their practical management implications.

Study Configuration

• Spatial Scale: Vicinity of the A8 Metro Station in Guishan District, Taoyuan City, Taiwan.
• Temporal Scale: Data collected from July 23, 2019, to May 22, 2021 (310 rainfall events, 12,684 records). Forecasts were generated for lead times up to 60 minutes (6 steps, each 10 minutes).

Methodology and Data

• Models used: Attentive Multilayer Perceptron (AM-MLP), Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (BiLSTM). Seq2Seq framework was used with LSTM and GRU as encoder/decoder.
• Data sources:
  • Rainfall: Linkou observation station (Taoyuan City Government), collected at 10-minute intervals.
  • Sewer water levels (H1) and road flood depths (D1): Measured using Series 26X pressure transducers.

Main Results

• Sewer Water Level Forecasting: The H-AM-MLP model achieved a Root Mean Square Error (RMSE) of 0.095 meters, an r² of 0.645, and a Nash–Sutcliffe Efficiency (NSE) of 0.631 on the testing set, showing an improvement over the baseline H-MLP (RMSE 0.099 meters). While sequence models (H-LSTM, H-GRU, H-BiLSTM) performed well, adding the attention mechanism to them sometimes slightly degraded performance.
• Flood Depth Forecasting: The D-AM-MLP model significantly improved performance, reducing RMSE from 0.058 meters (D-MLP) to 0.047 meters (a 19% reduction), increasing r² from 0.393 to 0.688, and NSE from 0.336 to 0.645 on the testing set. This demonstrates the substantial benefit of attention for MLP in flood depth prediction.
• General Findings: AM-MLP proved effective in compensating for MLP's limitations in handling sequential data, particularly for flood depth forecasting where data might be sparse. Sewer water level forecasting generally achieved higher accuracy than surface flood depth predictions, attributed to more varied training data for sewer levels. The proposed AM-MLP framework uses a single-model, multi-output structure for multi-step forecasting (up to 60 minutes), reducing model complexity and computational burden.

Contributions

• Introduction of a novel Attentive Multilayer Perceptron (AM-MLP) architecture specifically designed for urban sewer overflow and flood depth prediction, addressing the limitations of traditional MLPs with sequential hydrological data.
• Empirical demonstration that an attention mechanism can significantly enhance the predictive accuracy and interpretability of MLPs, making them competitive with recurrent neural networks (LSTM, GRU, BiLSTM) in scenarios with weak-sequence or low-frequency data.
• Development of a single-model, multi-output structure for multi-step-ahead forecasting (up to 60 minutes), offering reduced model complexity and computational burden compared to existing multi-stage or recursive methods.
• Improved model robustness and reduced false alarms in operational early warning systems by training the AM-MLP on a comprehensive dataset including both flooding and non-flooding rainfall events.
• Provision of valuable hydrological insights for the study area, contributing to enhanced flood risk management and urban resilience planning in regions facing diverse and unpredictable rainfall patterns.

Funding

• National Science and Technology Council of Taiwan (Grant number: NSTC 113–2221-E-035–021-MY3)
• National Science and Technology Council of Taiwan (Grant number: NSTC 114–2625-M-110–001)

Citation

@article{Yang2026Attention,
  author = {Yang, Song-Yue and Jhong, Bing-Chen and Lin, Rui-Wen and Tsai, Min-Chien},
  title = {Attention in MLP: A new architecture for urban sewer overflow and flood depth prediction},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2025.103088},
  url = {https://doi.org/10.1016/j.ejrh.2025.103088}
}