Presentation: 2025 ND EPSCoR Annual conference 

October 21, 2025, NDSU Memorial Union, Fargo, North Dakota

Optimizing Daily Streamflow Forecasting through Tree-Based Feature Selection and Machine Learning in Cold Regions

The techniques of machine leaning (ML) applied for streamflow prediction need a choice of suitable predictors. The selection of predictors in a misconstrued manner can lead to a poor forecast for independent events. The main objective of this research is to develop two ML models that utilize higher quality and less noisy input data derived from tree-based feature selection. Finally, the modeling framework optimizes streamflow forecasts based on discharge, precipitation, and snow cover area datasets. To realize this, the Boruta-Shap method is utilized to evaluate all input datasets and select the most significant features to feed ML models to predict daily discharge. Specifically, two ML models, including LGBM and DRVFL, are integrated with a signal decomposition method to predict daily streamflow. The proposed modeling methodology was applied to a watershed in North Dakota to demonstrate its extensive ability. The effectiveness of the ML models was evaluated using various statistical measures and visual tools. The comparative analyses of these novel modeling approaches showcased their enhanced computing efficiency and accuracy.