Scholarlink

<p>Seasonal prediction of streamflow plays a crucial role in water allocation in southeastern Australia, where hydroclimatic variability is modulated by large-scale ocean-atmosphere interactions. Among them, the El Niño Southern Oscillation (ENSO) is recognised as a dominant driver, commonly represented by the Niño 3.4 index anomalies. This study examines the Niño 3.4 index to predict spring streamflow across Victoria. Two modelling strategies were employed. A conventional linear regression (LR) model as a baseline, compared with a nonlinear Artificial Intelligence neural network (AINN) framework based on a fully connected neural network (FCNN). The FCNN corresponds to a multilayer perceptron trained using the Levenberg–Marquardt algorithm. Antecedent Niño 3.4 values were used as predictors. Model performance was assessed using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and the Pearson correlation coefficient (R). The results demonstrate that the FCNN substantially improves predictive capability compared with the LR benchmark. The FCNN achieved R values between 0.70 and 0.96, whereas the LR model produced considerably weaker relationships, with R ranging from 0.01 to 0.45. The highest accuracy was obtained at the 3-month lag, where the FCNN produced a maximum correlation of 0.96 and a corresponding RMSE of 0.051. To further investigate temporal dependency, a Long Short-Term Memory (LSTM) network was also implemented. Although the LSTM captured certain temporal patterns and produced moderate correlations between 0.30 and 0.78, its performance was less consistent than FCNN with the limited dataset. These findings demonstrate that neural network models can effectively exploit climate-streamflow teleconnections for seasonal prediction.</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

AI-driven seasonal streamflow prediction in Victoria: a focus on ENSO climate predictor

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月3日

查看全文
  • Sabrina Bani,
  • Iqbal Hossain,
  • Monzur Imteaz,
  • Patrick Ryan Morrison

摘要

Seasonal prediction of streamflow plays a crucial role in water allocation in southeastern Australia, where hydroclimatic variability is modulated by large-scale ocean-atmosphere interactions. Among them, the El Niño Southern Oscillation (ENSO) is recognised as a dominant driver, commonly represented by the Niño 3.4 index anomalies. This study examines the Niño 3.4 index to predict spring streamflow across Victoria. Two modelling strategies were employed. A conventional linear regression (LR) model as a baseline, compared with a nonlinear Artificial Intelligence neural network (AINN) framework based on a fully connected neural network (FCNN). The FCNN corresponds to a multilayer perceptron trained using the Levenberg–Marquardt algorithm. Antecedent Niño 3.4 values were used as predictors. Model performance was assessed using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and the Pearson correlation coefficient (R). The results demonstrate that the FCNN substantially improves predictive capability compared with the LR benchmark. The FCNN achieved R values between 0.70 and 0.96, whereas the LR model produced considerably weaker relationships, with R ranging from 0.01 to 0.45. The highest accuracy was obtained at the 3-month lag, where the FCNN produced a maximum correlation of 0.96 and a corresponding RMSE of 0.051. To further investigate temporal dependency, a Long Short-Term Memory (LSTM) network was also implemented. Although the LSTM captured certain temporal patterns and produced moderate correlations between 0.30 and 0.78, its performance was less consistent than FCNN with the limited dataset. These findings demonstrate that neural network models can effectively exploit climate-streamflow teleconnections for seasonal prediction.