Conjugation of deep learning and de noising data methods for short-term water turbidity forecasting

Abstract

Water turbidity is a critical index of water quality due to its high correlation with the five main water quality parameters (electrical conductivity, nitrogen, dissolved oxygen, phosphorus, and pH). The exact measurement of water turbidity is a difficult process because many conditions affect the reading of turbidity. Although many researchers applied decomposition-based techniques for preprocessing, it is difficult to use these approaches in real estimation because the newly acquired data greatly affect the initial decomposed subsequent values. In this study, the threshold-based wavelet denoising method, as a data pre-processing, coupled with the deep learning models (i.e., ANN and ANFIS) was employed to enhance the performance of the water turbidity modeling. The results showed that deep learning techniques in temporal modeling of water turbidity have good accuracy and can be used with reasonable confidence. Also, data denoising increases the accuracy of deep learning methods in estimating the amount of water turbidity. ANFIS method is more accurate in both calibration and validation modes as well as in noisy and denoised conditions. Based on the results, data denoising in the ANN method has a more significant impact than in the ANFIS technique. For example, in Comb. 5, which is the best case, the improvement rate of the results in the ANN is 12% and in the ANFIS method is 4%. This could be due to the fuzzy system in handling uncertainties in the model.