Implementation of a machine learning approach to predict streamflows in watersheds: a case study of the Leaf River catchment

Abstract

Accurate prediction of streamflows is crucial for managing water resources. Machine learning approaches have gained popularity for their ability to handle noisy and non-linear data and develop models that are capable of detecting relationships from the data they are provided with. This study was conducted to compare the performance of three machine learning algorithms (including extreme learning machine (ELM), random forest (RF), and gene expression programming (GEP)) and their hybrid versions in predicting the monthly streamflow of the Leaf River catchment. The models were tested with three scenarios and the most accurate scenario has been selected for the implementation of hybrid models. Results of all the models have been examined with a new evaluation index called general index (GI), which is calculated based on the three error statistical indices. Finally, the GEP model outperformed the other models, all the scenarios with GI = 11.268 in the M3 scenario and later, the ELM algorithm presented the best performance with GI = 12.811 in the M2 scenario, while the RF model had the worst overall performance. Regarding the hybrid models, using the EMD and principal components analysis (PCA) methods decreased the precisions of the models with the GI values fluctuating around 35.