Impact of multi-resolution analysis of artificial intelligence models inputs on multi-step ahead river flow forecasting

• Discrete wavelet transform was applied to decomposed ANN and ANFIS inputs.
• Novel approach of WNF with subtractive clustering applied for flow forecasting.
• Forecasting was performed in 1–5 step ahead, using multi-variate inputs.
• Forecasting accuracy of peak values and longer lead-time significantly improved.

SummaryIn this paper an attempt is made to show that the performance of daily river flow forecasting is improved when data-preprocessing techniques are used with computational intelligence methods. Especially for forecasting longer lead-times, one of the inherent problems in all forecasting methods is that the reliability of forecasting decreases with increasing the lead-time. Therefore wavelet multi-resolution analysis is coupled with artificial neural networks (ANN) and adaptive neuro-fuzzy interface system (ANFIS) which are two promising methods for forecasting nonlinear and non-stationary time series. Different models with a combination of the different input data sets are developed for 1, 2, 3, 4 and 5 days ahead forecasting in Harvey River, Western Australia. Daubechies and Symlet wavelets are used to decompose river flow time series to different levels. Hybrid wavelet neural networks (WNN) models are trained using Levenberg–Marquart (LM) algorithm and the wavelet neuro-fuzzy (WNF) models with subtractive clustering method to find the optimum number of fuzzy rules. Comparing the results with those of the original ANN and ANFIS models indicates that the hybrid models produce significantly better results, especially for the peak values and longer lead-times.