Hybrid artificial intelligence-time series models for monthly streamflow modeling

Abstract

In recent years, many efforts have been made to develop hybrid models for hydrological time series modeling. The present study introduces novel hybrid models by hybridizing time series models with artificial intelligence (AI) methods. These hybrid models were used to model monthly streamflow time series for the Ocmulgee and Umpqua River stations, located in Georgia and Oregon, USA, respectively. AI-based approaches, including gene expression programming (GEP), multivariate adaptive regression splines (MARS), and simple multiple linear regression (MLR), were integrated with fractionally auto-regressive integrated moving average (FARIMA) and self-exciting threshold auto-regressive (SETAR) time series models. Accordingly, six novel hybrid models, namely GEP-FARIMA, MARS-FARIMA, MLR-FARIMA, GEP-SETAR, MARS-SETAR, and MLR-SETAR, were developed. The accuracy of the proposed models was then compared to the standalone GEP, MARS, MLR, FARIMA, and SETAR models. Results showed that time series models were superior to AI and MLR approaches. In addition, the hybrid models offered more accurate results than the standalone models when modeling monthly streamflow time series.