|
Application of deep transformer neural networks for short-term river discharge prediction using hydrological data; case study: the Gorganrud river
|
Mohammad Hossein Naseri  , Hamid Ebadi  , Abbas Kiani * |
| Department of Surveying Engineering, Faculty of Civil Engineering, Noshirvani University of Technology, Babol, Mazandaran, Iran. |
|
|
Abstract: (1 Views) |
Objective: This study presents a deep learning–based approach for short-term discharge prediction at the Aq Qala station on the Gorganrud River and evaluates the Transformer model's ability to reproduce short-term trends and critical conditions. The necessity of this research arises from the increasing frequency and intensity of flood events, which underscore the need for rapid, accurate forecasting tools to support water resource management decisions and early warning systems. To address this need, a Long Short-Term Memory (LSTM) neural network was used as a comparative benchmark to evaluate the proposed method's performance.
Method: Daily discharge, precipitation, and evaporation data were used as model inputs. After normalization and transformation into ten-day sequences, the data were fed into the model. The temporal segmentation included approximately ten years for training, one year for testing, and one year for evaluation (including a flood event) to assess model stability under both normal and extreme conditions. The Transformer model, with its attention mechanism, captures long-term dependencies and complex relationships among input variables. Its parallel processing capability enables faster, more stable training for time-series problems. Both models were trained for one-, two-, and three-day forecasting horizons and evaluated using the Root Mean Square Error (RMSE) and Nash–Sutcliffe Efficiency (NSE) criteria.
Results: For the one-day horizon, the Transformer model outperformed the LSTM, achieving an RMSE of 0.03 and an NSE of 0.92. As the forecasting horizon increased to two and three days, the accuracy of both models decreased, with RMSE values of approximately 0.064 and 0.068, respectively, reflecting the nonlinear and complex nature of hydrological systems. Quantitative and graphical comparisons indicate that the Transformer model performs better in reproducing overall discharge trends and capturing peak flows during critical periods.
Conclusion: The results demonstrate that the Transformer model is an effective tool for short-term river discharge prediction and can support early warning systems and decision-making in water resource management. This approach can serve as an auxiliary tool for water managers, rapid warning systems, and flood risk reduction planners. |
|
| Keywords: Discharge prediction, Transformer, Long Short-Term Memory (LSTM), Flood, Early warning. |
|
|
Full-Text [PDF 1709 kb]
(4 Downloads)
|
Type of Study: Research |
Subject:
General
Received: 2025/10/11 | Accepted: 2025/10/28 | Published: 2025/10/28
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
