<p>Hybrid modeling techniques have become increasingly important in the field of hydrology, particularly for water resource management, where accurate predictions of precipitation and runoff are critical. Hybrid models, such as the Convolutional Neural Network – Q-learning (CNN-Qlearning) model, combine the strengths of machine learning and optimization algorithms, offering improved accuracy and reliability in simulating complex hydrological processes. Effective water resource management relies on these models to address challenges like flood prediction, reservoir operation, and climate variability, ensuring sustainable water use and risk mitigation. The Gandoman-Boldaji watershed, located in Chaharmahal-e-Bakhtiari province and spanning approximately 95,735 hectares, represents a critical hydrological system characterized by its cold semi-arid climate. The region receives an average annual precipitation of 400 to 600 millimeters, primarily in the form of snow and rain during colder months. Key rivers such as Aq Balagh and Solgan converge at the Solgan Dam. Data from meteorological stations, including Brojen, Gandoman, and Solgan, play a pivotal role in monitoring precipitation and understanding the watershed’s dynamics for effective water resource management. This study evaluates the hybrid CNN-Qlearning model’s ability to simulate precipitation-runoff dynamics using both noisy and denoised datasets. By applying Variational Mode Decomposition (VMD) for noise reduction, significant improvements in data quality and model performance were observed. Comparisons between noisy and denoised datasets revealed reduced variability, narrower uncertainty bounds, and more stable statistical properties, making it easier to identify hydrological patterns and trends. The denoised data notably enhanced the CNN-Qlearning model’s predictive accuracy. Performance metrics improved significantly: the Root Mean Square Error (RMSE) reduced from 120.85 m³/s to 93.97 m³/s, the Mean Absolute Error (MAE) decreased from 68.77 m³/s to 59.33 m³/s, and the Coefficient of Determination (R²) increased from 0.82 to 0.85. These results demonstrate the critical role of noise reduction in refining input data, leading to more precise and reliable hydrological simulations. The uncertainty analysis further showed that the denoised datasets resulted in significantly narrower confidence intervals, with Min-Max coverage increasing to 94.69% for runoff predictions, compared to 93.45% in noisy data. Overall, this study emphasizes the importance of noise reduction techniques, such as VMD, in improving the accuracy and reliability of hydrological models. This is particularly important in regions where effective water resource management relies on robust hydrological predictions.</p>

The Impact of Denoised Data on the Performance of the Hybrid Model (CNN-Qlearning)

期刊论文
发表日期：

2025年9月15日

查看全文

Hamid Ebrahimi

摘要

Hybrid modeling techniques have become increasingly important in the field of hydrology, particularly for water resource management, where accurate predictions of precipitation and runoff are critical. Hybrid models, such as the Convolutional Neural Network – Q-learning (CNN-Qlearning) model, combine the strengths of machine learning and optimization algorithms, offering improved accuracy and reliability in simulating complex hydrological processes. Effective water resource management relies on these models to address challenges like flood prediction, reservoir operation, and climate variability, ensuring sustainable water use and risk mitigation. The Gandoman-Boldaji watershed, located in Chaharmahal-e-Bakhtiari province and spanning approximately 95,735 hectares, represents a critical hydrological system characterized by its cold semi-arid climate. The region receives an average annual precipitation of 400 to 600 millimeters, primarily in the form of snow and rain during colder months. Key rivers such as Aq Balagh and Solgan converge at the Solgan Dam. Data from meteorological stations, including Brojen, Gandoman, and Solgan, play a pivotal role in monitoring precipitation and understanding the watershed’s dynamics for effective water resource management. This study evaluates the hybrid CNN-Qlearning model’s ability to simulate precipitation-runoff dynamics using both noisy and denoised datasets. By applying Variational Mode Decomposition (VMD) for noise reduction, significant improvements in data quality and model performance were observed. Comparisons between noisy and denoised datasets revealed reduced variability, narrower uncertainty bounds, and more stable statistical properties, making it easier to identify hydrological patterns and trends. The denoised data notably enhanced the CNN-Qlearning model’s predictive accuracy. Performance metrics improved significantly: the Root Mean Square Error (RMSE) reduced from 120.85 m³/s to 93.97 m³/s, the Mean Absolute Error (MAE) decreased from 68.77 m³/s to 59.33 m³/s, and the Coefficient of Determination (R²) increased from 0.82 to 0.85. These results demonstrate the critical role of noise reduction in refining input data, leading to more precise and reliable hydrological simulations. The uncertainty analysis further showed that the denoised datasets resulted in significantly narrower confidence intervals, with Min-Max coverage increasing to 94.69% for runoff predictions, compared to 93.45% in noisy data. Overall, this study emphasizes the importance of noise reduction techniques, such as VMD, in improving the accuracy and reliability of hydrological models. This is particularly important in regions where effective water resource management relies on robust hydrological predictions.