A STUDY ON DEEP LEARNING FOR RIVER WATER LEVEL PREDICTION IN URBAN WATERSHEDS WITH A DISCUSSION ON UNDERGROUND REGULATING RESERVOIR GATE OPERATIONS

抄録

 The study was focused on developing a conceptualization for the preliminary prediction of gate operations (GOs) using the forecasted water levels (WLs) of upstream and downstream gauges of the underground regulating reservoir (URR) in the Kanda River Basin. The rainfall and WL minute-by-minute data of Zenpukuji and Upper Kanda watersheds were gathered from the last 13 years up to 2024. The deep learning model was built by considering preceding flood events for the input time steps, while the targeted time steps were organized with the upcoming WLs of seven gauges. The precision of WL forecasting was evaluated by adjusting the target lengths (TLs) from 10 minutes (TL10) to 90 minutes (TL90) with an increment of 10 minutes. The prediction accuracy worsened when moving from TL10 to TL90, and even the Nash–Sutcliffe model efficiency coefficient (NSE) remained at or above 0.96 until TL60, with root mean square (RMSE) consistently below 0.1 m. The GO influenced the accuracy of the longer TLs compared to shorter TLs, where almost all WL gauges achieved above 0.9 NSE for TLs from TL30 to TL60. The temporal exactness and peak alignments of forecasted hydrographs were examined to derive the longest TL for anticipating the initiation (Iɢᴏ) and termination (Tɢᴏ) of GOs in advance. The longest TL obtained Iɢᴏ and Tɢᴏ in advance was 40 minutes, while 60 minutes was the longest TL identified only the Iɢᴏ in advance.