Noise injection - denoising techniques to improve artificial intelligence -based rainfall - runoff modeling

Accurate modeling of hydrological processes such as rainfall-runoff can provide important information for water resources management of a watershed. Consequently, various black box models have been used recently to simulate such a complex phenomenon. Efficiency of any data driven model largely depends on quantity and quality of available data and noisy data may create negative impact on the performance of the model. In this way, noise reduction of data using an appropriate denoising scheme may lead to a better performance in the use of the data-driven model. Therefore in this paper, first wavelet-denoising method was applied to denoise daily time series and then by adding noises to this denoised data and forming different training sets with the denoised- jittered input data, simulation of rainfall – runoff process for Pole Anyan station in Zarrineh River drainage basin in upstream of Bookan dam, was done by both ANN and ANFIS models. To evaluate the model accuracy, the proposed model was compared with MLR and ARIMA models.
Comparison of the obtained results via the trained ANN and ANFIS using denoised-jittered data revealed that the outcome of the this model for runoff forecasting is improved when the proposed approach, as a pre-processing method, is applied to the used data. The results show that the proposed data processing which serves both denoising and jittering approaches could improve performance of the ANN and ANFIS-based rainfall-runoff modeling of the case study respectively up to 23% and 14% in verification phase.