Feasibility Study Regarding the Use of a Conformer Model for Rainfall-Runoff Modeling

• Published in: Water (Basel) 2024-11, Vol.16 (21), p.3125
• Main Authors: Lo, WeiCheng, Wang, Wei-Jin, Chen, Hsin-Yu, Lee, Jhe-Wei, Vojinovic, Zoran
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial intelligence; Chatbots; Deep learning; Hydraulic measurements; Hydrology; Neural networks; Rain and rainfall; Rivers; Runoff; Simulation methods; Stream flow; Voice recognition; Water; Watersheds; Wind
• ISSN: 2073-4441; 2073-4441
• DOI: 10.3390/w16213125

Flood disasters often result in significant losses of life and property, making them among the most devastating natural hazards. Therefore, reliable and accurate water level forecasting is critically important. Rainfall-runoff modeling, which is a complex and nonlinear time series process, plays a key role in this endeavor. Numerous studies have demonstrated that data-driven methods, particularly deep learning approaches such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and transformers, have shown promising performance in water level prediction tasks. This study introduces the Conformer, a novel deep learning architecture that integrates the strengths of CNNs and transformers for rainfall-runoff modeling. The framework uses self-attention mechanisms combined with convolutional computations to extract essential features—such as water levels, precipitation, and meteorological data—from multiple stations, which are then aggregated to predict subsequent water level series. This study utilized data spanning from 1 April 2006 to 25 July 2021, totaling 5595 days (134,280 h), which were divided into training, validation, and test sets in an 8:1:1 ratio to train the model, adjust parameters, and evaluate performance, respectively. The effectiveness and feasibility of the proposed model are evaluated in the Lanyang River Basin, with a focus on predicting 7-day-ahead water levels. The results obtained from ablation experiments indicate that convolutional computations significantly enhance the ability of the model to capture the local relationships between water levels and other parameters. Additionally, performing convolution computations after executing self-attention operations yields even better results. Compared with other models in simulations, the Conformer model markedly outperforms the CNN, LSTM, and traditional transformer models in terms of the coefficient of determination (R2) and Nash–Sutcliffe efficiency (NSE) indicators. These findings highlight the potential of the Conformer model to replace the commonly used deep learning methods in the field of hydrology.