Cabila Subramaniyam, Yoshiyuki Imamura, Hideo Amaguchi
Received 1 February, 2024
Accepted 14 April, 2025
Published online 21 June, 2025

Cabila Subramaniyam¹⁾, Yoshiyuki Imamura¹⁾, Hideo Amaguchi¹⁾

1) Department of Civil Engineering, Tokyo Metropolitan University, Japan

Neural networks (NNs) have recently gained attention for establishing Rainfall-runoff Predictive Models (RPMs) to receive accurate upcoming hydrographs. This study established Bidirectional Long-Short-Term Memory Model-based RPMs (BiLSTM-RPMs) for a small-to-medium-scale urban watershed, the Upper Kanda basin, occupying nearly 11 km². The average rainfall of six stations and streamflow of a water level gauge were considered to gather a hundred events with minute-to-minute intervals for 12 years from 1999. The influence of batch-wise shuffling was investigated by developing BiLSTM-RPMs with a sliding window generator (SWG) technique for target lengths (TLs) such as 10 minutes (TL10), TL20, TL30, and TL60. Batch-wise shuffling was supported to predict seamless hydrographs, where all TLs achieved Nash–Sutcliffe model efficiency (NSE) above 0.8, Root Mean Square Error (RMSE) below 0.01 mm/min, and coefficient of determination (R²) for peak alignment above 0.95. Long-Short-Term Memory Model-based RPMs (LSTM-RPMs) derived appreciable forecasted hydrographs for TL10 and TL20; however, they performed poorly for TL30 and TL60, where the R² was 0.88 and 0.62, respectively.

[Full Text]

Copyright (c) 2025 The Author(s) CC-BY 4.0