Association between forecasting models' precision and nonlinear patterns of daily river flow time series

The river flow forecasting is of significance to water resources management and risk assessment. Therefore, this study explored the influences of long-term daily river flow nonlinear patterns (> 3500 datapoints) on forecasting models' efficiency. The novelty of the study lies in focusing on the effect of time series' nonlinear dynamic patterns and elements on models' accuracy instead of models' structure. The daily flow data of ten hydrometric stations (in the UK) were collected for 10 years (2009-2019). The nonlinear patterns were evaluated through fractal and multifractal applications, Lyapunov exponent, D-s factor, and correlation dimension techniques. Subsequently, the forecasting was performed using six models, including three hybrid models, two based artificial neural network models, and one linear model, and then the models' accuracy was evaluated. It was established that the RBF and ARIMA models had the inadequate capacity to capture nonlinear dynamic patterns and elements in the daily river flow time series. The ANFIS models had the acceptable qualification of modeling daily river flow with nonlinear elements and patterns. However, the ANFIS-FCM model produced results with insignificant errors. The ANN-MLP model excellently captured the nonlinear dynamic patterns and elements of the daily river flow time series and outperformed other models. The findings of the present study contribute to the understanding of physical forecasting and the effect of nonlinear patterns of time series on the models' accuracy and also furnish references for forecasting daily river flow.

Automated summary

This study investigates the influence of long-term daily river flow nonlinear patterns on the efficiency of forecasting models, and finds that artificial neural network models can better capture the nonlinear dynamic patterns and elements of daily river flow time series.