Scholarlink

<p>The nonlinear characteristics and complexity of streamflow process prediction significantly influence water resource allocation decisions for local government planning on supply and demand. This study utilized 20 years of daily rainfall and discharge data to predict streamflow. Data pre-processing procedures (including missing data reconstruction, outlier elimination and normalization), were executed, with 70% and 30% of the data allocated for training and testing, respectively. Hybrid models integrating ANNs and SVR with Ali baba and the forty thieves (AFT) and Fire hawk optimizer (FHO) were utilized for prediction. The model scenarios included Rainfall variables with two lags and discharge with three lags, and Pearson coefficient was used to select the optimal model combination. Various criteria (such as MAE, RMSE, R<sup>2</sup> and EVS) were used to evaluate the model. Various objectives were investigated in this research (comparing the performance of standalone and hybrid models, evaluating the effects of changes in the neurons contained hidden layers, selecting the type of training functions and different transfer functions in ANNs, and selecting the optimal hyperparameters of SVR based on previous studies). The modeling results showed that the performance of different models was very close, and all models provided good predictions. Accordingly, hybrid models SVR-AFT improved the results of Standalone models by approximately 47%. Also, among all the models used, the model SVR-AFT had the best performance (R<sup>2</sup> = 0.9695, RMSE = 0.0813 m<sup>3</sup>/sec). Furthermore, other results indicate that a lower number of neurons (between 10 and 20) performed better than a higher number. Additionally, among the various training models, the Levenberg-Marquardt backpropagation model (trainlm) and the gradient descent backpropagation model (traingd) performed better than other training functions.</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

Long term stream flow for enhanced accuracy prediction through machine learning models (Ali Baba and the forty thieves vs. Fire Hawk Optimizer)

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Edris Merufinia,
  • Ahmad Sharafati,
  • Hirad Abghari,
  • Yousef Hassanzadeh

摘要

The nonlinear characteristics and complexity of streamflow process prediction significantly influence water resource allocation decisions for local government planning on supply and demand. This study utilized 20 years of daily rainfall and discharge data to predict streamflow. Data pre-processing procedures (including missing data reconstruction, outlier elimination and normalization), were executed, with 70% and 30% of the data allocated for training and testing, respectively. Hybrid models integrating ANNs and SVR with Ali baba and the forty thieves (AFT) and Fire hawk optimizer (FHO) were utilized for prediction. The model scenarios included Rainfall variables with two lags and discharge with three lags, and Pearson coefficient was used to select the optimal model combination. Various criteria (such as MAE, RMSE, R2 and EVS) were used to evaluate the model. Various objectives were investigated in this research (comparing the performance of standalone and hybrid models, evaluating the effects of changes in the neurons contained hidden layers, selecting the type of training functions and different transfer functions in ANNs, and selecting the optimal hyperparameters of SVR based on previous studies). The modeling results showed that the performance of different models was very close, and all models provided good predictions. Accordingly, hybrid models SVR-AFT improved the results of Standalone models by approximately 47%. Also, among all the models used, the model SVR-AFT had the best performance (R2 = 0.9695, RMSE = 0.0813 m3/sec). Furthermore, other results indicate that a lower number of neurons (between 10 and 20) performed better than a higher number. Additionally, among the various training models, the Levenberg-Marquardt backpropagation model (trainlm) and the gradient descent backpropagation model (traingd) performed better than other training functions.