Scholarlink

<p>Artificial Intelligence (AI) is increasingly used to support accurate reservoir inflow modeling, which is essential for effective water resource management. Aligned with this objective, this research aims to compare a process-based model (SWAT) with a data-driven flow simulation model (Deep Learning), specifically utilizing the Long Short-Term Memory (LSTM) method. Three scenarios, labeled S1, S2, and S3, were assumed as inputs for the LSTM network. In the first scenario (S1), all nine features, including date, meteorological data, and discharge data, were included. Notably, S1 uniquely incorporated date (Year and Month) as a distinguishing factor. Conversely, in the second scenario (S2), the absence of river flow features (one-month delayed and two-month delayed) was notable. Scenario 3 (S3) is similar to S1, but it replaces the two-month delayed precipitation feature with one-month and two-month delayed flow features. Quantitatively, the SWAT model achieved an RMSE of 5.153 and a correlation coefficient (CC) of 0.923, whereas the LSTM model in scenario S1 achieved an RMSE of 5.630 and a CC of 0.876, indicating the comparable performance of both approaches. Additionally, incorporating date as a feature was found to improve correlations between simulated and observed data. The similar performance of the process-based (SWAT) and data-driven models suggests that, especially when the detailed input data required by SWAT are unavailable, data-driven models could serve as a viable alternative to process-based models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparison of Process-Based Hydrological Modeling and Deep Learning Approaches for Streamflow Simulation

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Jingyao Wang,
  • Hossein Yousefi,
  • Mojtaba Shourian

摘要

Artificial Intelligence (AI) is increasingly used to support accurate reservoir inflow modeling, which is essential for effective water resource management. Aligned with this objective, this research aims to compare a process-based model (SWAT) with a data-driven flow simulation model (Deep Learning), specifically utilizing the Long Short-Term Memory (LSTM) method. Three scenarios, labeled S1, S2, and S3, were assumed as inputs for the LSTM network. In the first scenario (S1), all nine features, including date, meteorological data, and discharge data, were included. Notably, S1 uniquely incorporated date (Year and Month) as a distinguishing factor. Conversely, in the second scenario (S2), the absence of river flow features (one-month delayed and two-month delayed) was notable. Scenario 3 (S3) is similar to S1, but it replaces the two-month delayed precipitation feature with one-month and two-month delayed flow features. Quantitatively, the SWAT model achieved an RMSE of 5.153 and a correlation coefficient (CC) of 0.923, whereas the LSTM model in scenario S1 achieved an RMSE of 5.630 and a CC of 0.876, indicating the comparable performance of both approaches. Additionally, incorporating date as a feature was found to improve correlations between simulated and observed data. The similar performance of the process-based (SWAT) and data-driven models suggests that, especially when the detailed input data required by SWAT are unavailable, data-driven models could serve as a viable alternative to process-based models.