JWSS - Journal of Water and Soil Science

In this study, the groundwater level (GWL) of the Sarab Qanbar region located in the south of Kermanshah, Iran, was estimated using the Wavelet- Self- Adaptive Extreme Learning Machine (WA- SAELM) model. An artificial intelligence method called “Self- Adaptive Extreme Learning Machine” and the “Wavelet transform” method were implemented for developing the numerical model. First, by using the autocorrelation function (ACF), the partial autocorrelation function (PACF) and the effective lags in estimating GWL, eight distinctive SAELM and WA- SAELM models were developed. Later, the values of the observational well were normalized for estimating GWL. Next, the most optimized mother wavelet was chosen for the modeling. By evaluating the results of SAELM and WA- SAELM, it was concluded that the WA- SAELM models could estimate the values of the objective function with higher accuracy. Then, the superior model was introduced, showing that it could be very accurate in forecasting the GWL. In the test mode, for example, the values of R (correlation coefficient), Main absolute error (MAE) and the NSC- Sutcliffe efficiency coefficient (NSC) for the superior model were calculated to be 0.995, 0.988 and 0.990, respectively. Furthermore, an uncertainty analysis was conducted for the numerical models, proving that the superior model had an underestimated performance.

Rainfall may be considered as the most important source of drinking water and watering land in different areas all over the world. Therefore, simulation and estimation of the hydrological phenomenon is of paramount importance. In this study, for the first time, the long-term rainfall in Rasht city was simulated using an optimum hybrid artificial intelligence (AI) model over a 62 year period from 1956 to 2017. The gene expression programming (GEP) and wavelet transform (WT) were combined to develop the hybrid AI model (WGEP). Firstly, the most effective lags of time series data were identified by means of the autocorrelation function (ACF); then eight various GEP and WGEP models were defined. Next, the GEP models were analyzed and the superior GEP model as well as the most influenced lags was detected. For instance, the variance accounting for (VAF), correlation coefficient (R) and scatter index (SI) for the superior GEP model was calculated to be 0.765, 0.508 and 0.709, respectively. Additionally, lags (t-1), (t-2), (t-3) and (t-12) were the most influenced. Then, the different mother wavelets were examined, indicating that the demy mother wavelet was the most optimal one. Moreover, analyzing the numerical simulations showed that the mother wavelet enhanced the performance of the GEP model significantly. For example, the VAF index for the superior WGEP model was increased almost three times after using the mother wavelet. Furthermore, the R and MARE statistical indices for the WGEP model were computed to be 0.935 and 0.862, respectively.

Unpredictable settlement of earth dams has led researchers to develop new methods such as artificial neural networks, wavelet theory, fuzzy logic, and a combination of them. These methods do not require time-consuming analyses for estimation. In this research, the amount of settlement in rockfill dams with a central core has been estimated using artificial intelligence methods. The data of 35 rockfill dams with a central core were used to train and validate the models. The artificial neural network, wavelet transform model, and fuzzy-neural adaptive inference system are the proposed models which were used in the present study. According to the results, the best model for an artificial neural network had two hidden layers, the first layer of 18 neurons and the second layer of 7 neurons, with the Tansig-Tansig activation function, with a coefficient of determination R2=0.4969. The best model for the fuzzy-neural inference system had the ring function (Dsigmoid) as a membership function, with three membership functions and 142 repetitions with a coefficient of determination R2=0.2860. Also, combining wavelet-neural network conversion with the coif2 wavelet function due to the more adaptation this function has to the input variables, the better the performance, and this function, with a coefficient of determination R2=0.9447, had the highest accuracy compared to other models.