Hybrid models to improve the monthly river flow prediction: Integrating artificial intelligence and non-linear time series models

Abstract

Prediction of river flow as a fundamental source of hydrological information plays a crucial role in various fields of water projects. In this study, at first, the capabilities of two time series analysis approaches, namely self-exciting threshold autoregressive (SETAR) and generalized autoregressive conditional heteroscedasticity (GARCH) models, then three artificial intelligence approaches including artificial neural networks (ANN), multivariate adaptive regression splines (MARS), and random forests (RF) models were investigated to predict monthly river flow. For this purpose, monthly river flow data of Brantford and Galt stations on Grand River, Canada, for the period from October 1948 to September 2017 were used and their performances were evaluated based on various evaluation criteria. The SETAR model showed better performance than the GARCH one in prediction of river flows at the stations of study. Additionally, the stand-alone MARS and RF models performed slightly better than the ANN. Next, hybrid models were developed by coupling the used ANN, MARS, and RF models with SETAR and GARCH models as the non-linear time series models. The performance of various models presented in this study indicated that the new hybrid models demonstrated a much better performance compared with the stand-alone ones at both stations. Among the developed hybrid models, the RF-SETAR models generally had the best accuracy to improve the river flows modeling. As a result, it can be concluded that the presented methodology can be used to predict hydrological time series such as river flow with a high level of accuracy.