Comparative Study of Deep Learning LSTM and 1D-CNN Models for Real-time Flood Prediction in Red River of the North, USA

• Published in: 2023 IEEE International Conference on Electro Information Technology (eIT) pp. 022 - 028
• Main Authors: Atashi, Vida, Kardan, Ramtin, Gorji, Hamed Taheri, Lim, Yeo Howe
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18.05.2023
• Subjects: 1D-CNN model; Analytical models; Data models; Deep learning; flood prediction; LSTM model; Numerical models; Predictive models; Red River of the North; Rivers; Time series analysis
• ISSN: 2154-0373
• DOI: 10.1109/eIT57321.2023.10187358

The Red River of the North has a history of flooding, dating back to the late 1800s. Flooding in the Red River is caused by a combination of factors, including heavy snowfall, heavy spring rainfall, and poor drainage in the flat terrain of the Red River basin. In recent years, efforts have been made to improve flood forecasting accuracy through numerical and experimental models. An LSTM (Long Short-Term Memory) model and the most popular 1D convolutional neural network (1D-CNN) are introduced in this study to predict floods in the Red River of the North by analyzing past streamflow data and identifying patterns and trends. Based on a fitted model using trained data from 1994 to 2021 at the Grand Forks USGS station on the Red River, USA, the streamflow for 2022 was predicted, including a 16.5-year return period flood. According to the study's findings, the RMSE for the 1D-CNN method is 193.09 cfs, whereas the RMSE for the LSTM method is 76.23 cfs for the Grand Forks station which demonstrates that shows LSTM outperforms 1D-CNN in predicting flood events at the USGS Grand Forks Station.