Drought is one of the main natural factors influencing different aspects of human life. Over the decades, intelligent techniques have proven to be very capable of modeling and predicting nonlinear and dynamic time series. Therefore, the present study aims to predict drought by using and comparing neuro-fuzzy adaptive inference systems (ANFIS), artificial neural network of multilayered perceptron (ANN-MLP) and the support vector model (SVR). For this purpose, the precipitation data obtained from the Ain Bittit station were used for a statistical period of 34 years. In addition, the short-term (3 and 6 months) and long-term (9 and 12 months) time scales were calculated using the standardized precipitation index (SPI). Then, depending on the results of the calculations, the period 1979-2000 was selected as a control group and the period 2003-2012 was selected as an experimental group. In order to predict the SPI for the (t + 1) period, SPI values, precipitation from previous months were used. The results indicated that, in the majority of time scales, the ANFIS model gives SPI values and predictive dryness more accurately than the SVR, and ANN models.

This contribution will verify the effectiveness of formal neural networks for predicting drought in a semiarid region using a hybrid model of formal neural networks (ANN-MLP) and the standardized precipitation index ( SPI). Three types of models have been optimized to achieve this objective. A database consisting of SPI values, rain, temperature and potential evapotranspiration (PET) at the monthly time step was used as input for these models. These data have been standardized between 0 and 1 and subdivided into two blocks: a first block composed of 2/3 of the data for learning and a second block composed of 1/3 of the data for the test and the validation of the models. These models have been optimized with supervised learning. The activation function chosen is the logistic variant of the type sigmoid. The mean square error (RMSE), the correlation coefficient (R), the criterion of Nash-Sutcliffe (Nash) and the absolute mean error (MAE) were used to test the performance of these models. The results obtained show that the 3rd model is the most efficient. The application of neural networks for the estimation of the dryness of the Saïss Plain yielded quite good results. Indeed, the coefficients of correlation between the predicted and the measured values range from 0.63 to 0.97. It is therefore noted that the performances obtained are relatively good and could be improved by using a larger database.

Abstract:A new method based on the coupling of discrete wavelets (DWT) and artificial neural networks with perceptron multilayers (ANN-PMC) is proposed to predict the groundwater level. The relative performance of the DWT-ANN-PMC model has been regularly compared to artificial neural network (ANN-PMC) and multiple linear regression (MLR) models. Precipitation, temperature and average groundwater level are the variables introduced to explain and validate the models, with a monthly time step for the period March 1980 to March 2014 at two sites in the Plain of Saïss. The results of the study indicate the potential of DWT-ANN-PMC models in the prediction of groundwater levels. The forecast results indicate that the coupled wavelet neural network (WN) models were the best models for forecasting SPI values over multiple lead times in the Saïss Plain. It is recommended that further studies should explore this proposed methodology, which may in turn be used to facilitate the development and implementation of more effective strategies for the sustainable management of groundwater.