Scholarlink

<p>Reliable streamflow forecasting remains a fundamental challenge in hydrology, particularly under data-scarce conditions where access to comprehensive hydro-meteorological predictors is limited. This study examines the performance of machine learning (ML) algorithms in univariate daily streamflow forecasting, utilizing historical discharge values as the sole input source. Four ML models, Random Forest (RF), Gradient Boosting (GB), K-Nearest Neighbors (KNN), and Support Vector Regression with Radial Basis Function kernel (SVR-RBF), were evaluated under two input configurations: (1) models trained exclusively on temporal calendar features, and (2) models incorporated short-term autoregressive inputs, specifically two-day lagged discharge values. The models were trained and tested on a ~ 20-year daily streamflow dataset, with predictive performance assessed using multiple metrics. Results show that without lagged inputs, tree-based models RF and GB outperformed KNN and SVR, achieving Coefficient of Determination (R²) values of 0.75–0.78 and moderate Root Mean Squared Error (RMSE) (~ 29–31&#xa0;m³/s), demonstrating reasonable ability to capture broad seasonal patterns. SVR-RBF performed the weakest under this configuration (R² ≈ 0.41). Including lagged inputs led to significant improvements across all models, particularly for RF and GB, which achieved R² &gt; 0.98 and RMSE &lt; 10&#xa0;m³/s, reductions of over 65% compared to their non-lag counterparts. KNN and SVR-RBF showed significant improvements, with R² values increasing to 0.97 and 0.94, respectively. The predicted discharge distribution further validated that lag-augmented models accounted for seasonal flow variability and extreme events more effectively, closely matching the observed values. The findings highlighted the importance of autoregressive feature engineering in univariate ML-based streamflow forecasting methodologies. They demonstrated that accurate predictions can be made even without exogenous climatic or catchment variables, providing a practical framework for operational hydrological forecasting in data-limited environments.</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

Evaluating the Accuracy of Machine Learning Models with Short-Term Lag Inputs for Univariate Daily Streamflow time Series Forecasting

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Saad Ashfaq,
  • Safia Waseem,
  • Muhammad Tufail,
  • Aqil Tariq

摘要

Reliable streamflow forecasting remains a fundamental challenge in hydrology, particularly under data-scarce conditions where access to comprehensive hydro-meteorological predictors is limited. This study examines the performance of machine learning (ML) algorithms in univariate daily streamflow forecasting, utilizing historical discharge values as the sole input source. Four ML models, Random Forest (RF), Gradient Boosting (GB), K-Nearest Neighbors (KNN), and Support Vector Regression with Radial Basis Function kernel (SVR-RBF), were evaluated under two input configurations: (1) models trained exclusively on temporal calendar features, and (2) models incorporated short-term autoregressive inputs, specifically two-day lagged discharge values. The models were trained and tested on a ~ 20-year daily streamflow dataset, with predictive performance assessed using multiple metrics. Results show that without lagged inputs, tree-based models RF and GB outperformed KNN and SVR, achieving Coefficient of Determination (R²) values of 0.75–0.78 and moderate Root Mean Squared Error (RMSE) (~ 29–31 m³/s), demonstrating reasonable ability to capture broad seasonal patterns. SVR-RBF performed the weakest under this configuration (R² ≈ 0.41). Including lagged inputs led to significant improvements across all models, particularly for RF and GB, which achieved R² > 0.98 and RMSE < 10 m³/s, reductions of over 65% compared to their non-lag counterparts. KNN and SVR-RBF showed significant improvements, with R² values increasing to 0.97 and 0.94, respectively. The predicted discharge distribution further validated that lag-augmented models accounted for seasonal flow variability and extreme events more effectively, closely matching the observed values. The findings highlighted the importance of autoregressive feature engineering in univariate ML-based streamflow forecasting methodologies. They demonstrated that accurate predictions can be made even without exogenous climatic or catchment variables, providing a practical framework for operational hydrological forecasting in data-limited environments.