Evaluation of wavelet-GEP and wavelet-ANN hybrid models for prediction of total nitrogen concentration in coastal marine waters

Abstract

Because of increasing worldwide contamination of coastal marine water in the last decades, the exact prediction of water quality parameters in these areas is an important factor in coastal management. In this study, the evaluation of wavelet-gene expression programing (WGEP) and wavelet-artificial neural network (WANN) hybrid model was assessed in prediction of total nitrogen concentration (TN) in Charlotte harbor marine waters. The WANN and WGEP results were compared with traditional predictive models such as ANN, GEP, and multi-linear regression (MLR) methods. The TN monthly time series for 13 years were applied as inputs, and the TN values of the next month for two stations were simulated and predicted with different models. The comparison results of the wavelet hybrid models with others using statistical criteria (E and RMSE) exhibited the best performance of the wavelet conjunction models for prediction of TN in coastal waters. The E values of WGEP and WANN models with respect to the optimal GEP and ANN models increased to 0.858–0.879 and 0.840–0.857 for the first and second station, respectively. The selection process of suitable model indicated that the wavelet hybrid models have good results also in the prediction of maximum and minimum values of TN time series. Using wavelet transforms, different time-frequencies of TN changes of coastal marine water are extracted and sub-time series and sub-signal changes of TN as monthly, seasonally, 6 monthly and yearly can be recognized; thus, ANN and GEP model are improved.