Appraisal of data-driven techniques for predicting short-term streamflow in tropical catchment

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Short-term streamflow prediction is crucial for flood warnings and water resource management.
• Traditional numerical models are data-intensive and require complex calibration.
• Data-driven approaches offer a promising alternative for streamflow prediction.

Purpose of the Study:

• To assess the performance of Multiple Linear Regression (MLR) and Random Forest (RF) models for multi-step ahead short-term streamflow prediction.
• To evaluate the impact of different input combinations, including historical streamflow data, on model accuracy.
• To compare the effectiveness of MLR and RF models in predicting streamflow, particularly peak flows.

Main Methods:

• Developed and evaluated MLR and RF models using 14 years of hydrological data from the Kulim River catchment, Malaysia.
• Tested three distinct input combinations for each model.
• Assessed model performance using metrics such as Nash-Sutcliffe efficiency (NSE).

Main Results:

• Random Forest models demonstrated superior prediction accuracy (NSE: 0.599-0.962) compared to MLR models (NSE: 0.584-0.963).
• Incorporating antecedent streamflow events significantly improved model performance, especially for predicting peak streamflow during high-flow events.
• Prediction accuracy for both arrival time and magnitude of peak streamflow decreased as the prediction lead time increased.

Conclusions:

• Random Forest models, a type of decision tree-based model, are highly effective for short-term streamflow forecasting.
• The inclusion of historical streamflow data is a key factor in enhancing the accuracy of data-driven streamflow prediction models.
• Further research can explore strategies to mitigate the decline in accuracy with longer prediction lead times.