Improving fuzzy-based model for seasonal river flow forecasting

Abstract

Accurate river flow forecasts play a key role in sustainable water resources and environmental management. Recently, computational intelligence approaches have become increasingly popular due to minimum information requirements and their ability to simulate nonlinear and non-stationary characteristics of hydrological process. In this paper, the performance of seasonal river flow forecasting model is improved when different input combinations and data-preprocessing techniques are applied on adaptive neuro-fuzzy inference system (ANFIS). Wavelet multi-resolution analysis is coupled with ANFIS model to develop hybrid wavelet neuro-fuzzy model (WNF). Different models with different input selection and structure are developed for daily river flow forecasting in Railway Parade station on Ellen Brook River, Western Australia. The River flow time series is decomposed into multi-frequency time series by discrete wavelet transform (DWT) using the Haar, Coiflet number 1 and Daubechies number 5 mother wavelets, then the wavelet coefficients are imposed as input data to the neuro-fuzzy model. Models are developed based on Takagi-Sugeno-Kang fuzzy model with the grid partitioning approach for initializing the fuzzy rule-based structure. Mean square error and coefficients of determination are chosen as the performance criteria. Results show that the right selection of the inputs with high autocorrelation function (ACF) improves the accuracy of forecasting. However, comparing the performance of the hybrid WNF model with those of the original ANFIS models, indicates that the hybrid wavelet neuro-fuzzy models produce significantly better results.