2024 Volume 12 Issue 2 Article ID: 24-27025
Neural network-based rainfall-runoff predictive modeling (RPM) is an emerging field for implementing early warning systems to mitigate urban flooding. This research executed real-time runoff forecasting with Bidirectional Long Short-Term Memory RPM (BiLSTM-RPM) for a small-to-medium scale river basin in Tokyo. The focused basin was the Zenpukuji urban watershed, which is about 22.5 km2 in area and has minute-to-minute data. The modeling proceeded with the preceding rainfall gathered from seven rainfall locations and a preceding discharge at Aioi bridge to forecast upcoming runoffs of the same location. One hundred rainfall-runoff events were organized to train BiLSTM-RPMs designed with six target lengths from 10 to 60 minutes with an increment of 10 minutes, and the models were tested with 17 untrained test events. It was found that an equal duration of input and target lengths provided desirable accuracy. The temporal exactness and peak alignment accuracy of BiLSTM-RPMs were derived with satisfactory results for all target lengths, where the models excellently performed real-time forecasting even for an uncommon event of the test set. Upcoming hydrographs could be plotted with excellent temporal exactness and peak alignment for extensive forecasting spans by integrating each TL's last 10-minute predicted runoffs.
