Developing Novel Robust Models to Improve the Accuracy of Daily Streamflow Modeling

Abstract

Streamflow plays a major role in the optimal management and allocation of available water resources in each region. Reliable techniques are therefore needed to be developed for streamflow modeling. In the present study, the performance of streamflow modeling is improved via developing novel boosted models. The daily streamflows of four hydrometric stations comprising of the Brantford and Galt stations located on the Grand River, Canada, as well as Macon and Elkton stations respectively, located on the Ocmulgee and Umpqua rivers, United States, are used. Three different types of boosted models are implemented and proposed by coupling the classical multi-layer perceptron (MLP) with the optimization algorithms, including particle swarm optimization (PSO) and coupled particle swarm optimization-multi-verse optimizer (PSOMVO) and a time series model, namely the bi-linear (BL). So, the boosted MLP-PSO, MLP-PSOMVO, and MLP-BL models are developed. The accuracy of all the boosted models is compared with the classical MLP and BL by the statistical metrics used. It is concluded that all the boosted models developed at the studied stations lead to superior modeling results of the daily streamflows to the classical MLP; however, the boosted MLP-BL models generally outperformed the MLP-PSO and MLP-PSOMVO ones.