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

Short-term streamflow prediction is essential for managing flood early warning and water resources systems. Although numerical models are widely used for this purpose, they require various types of data and experience to operate the model and often tedious calibration processes. Under the digital revolution, the application of data-driven approaches to predict streamflow has increased in recent decades. In this work, multiple linear regression (MLR) and random forest (RF) models with three different input combinations are developed and assessed for multi-step ahead short-term streamflow predictions, using 14 years of hydrological datasets from the Kulim River catchment, Malaysia. Introducing more precedent streamflow events as predictor improves the performance of these data-driven models, especially in predicting peak streamflow during the high-flow event. The RF model (Nash-Sutcliffe efficiency (NSE): 0.599-0.962) outperforms the MLR model (NSE: 0.584-0.963) in terms of overall prediction accuracy. However, with the increasing lead-time length, the models' overall prediction accuracy on the arrival time and magnitude of peak streamflow decrease. These findings demonstrate the potential of decision tree-based models, such as RF, for short-term streamflow prediction and offer insights into enhancing the accuracy of these data-driven models.

Automated summary

Short-term streamflow prediction plays a crucial role in flood early warning and water resources management. Data-driven approaches have gained popularity in recent decades due to the digital revolution. In this study, multiple linear regression (MLR) and random forest (RF) models were developed and evaluated for short-term streamflow prediction using hydrological datasets from the Kulim River catchment in Malaysia. The results showed that including more precedent streamflow events as predictors improved the performance of data-driven models, especially in predicting peak streamflow during high-flow events. The RF model outperformed the MLR model in overall prediction accuracy but showed decreased accuracy in predicting the arrival time and magnitude of peak streamflow with increasing lead-time length.