Previsão dos níveis de água subterrânea em poços do aquífero Ogallala utilizando redes neurais artificiais

Abstract

The Ogallala Aquifer, located in the United States of America, is the largest in North America and serves as the main source of agricultural and public water supply. However, due to decades of pumping for irrigation combined with limited recharge, the Ogallala Aquifer has been facing a decline in its groundwater levels (GWLs). The prediction of GWL appears to be an alternative to enable more efficient planning to better use water resources. In this regard, several authors have been using ANNs for modeling groundwater. Therefore, this study aims to use artificial neural networks to predict the GWLs of several wells in the Ogallala Aquifer located in Texas. For this purpose, groundwater data were obtained from the Texas Water Development Board (TWDB) platform, and 403 wells were chosen to perform the GWLs prediction after a selection process. Also, hierarchical cluster analysis was performed in the time series of the 403 wells studied to classify them in groups according to their similarities in the water level fluctuations. In this study, thirty groups were selected from cluster analysis, and it was developed three ANN models to forecast the GWLs of the wells. The first model consisted of developing 403 ANNs, one for each analyzed well, to predict the water level from one to ten years ahead. For the second model, it was determined an ANN for each of the selected clusters to forecast the GWLs of the wells belonging to each cluster for one year ahead. For the last model proposed, it is used a single ANN, to predict the GWLs for a year ahead, to all studied wells. The performance of the ANN models developed in this work was evaluated using the following statistical indices: Pearson's linear correlation coefficient (R), Nash-Sutcliffe efficiency (NSE), mean absolute percentage error (MAPE), the root of the mean square error (RMSE) and the Bias. By analyzing the results, it is possible to conclude that all the predictions for the proposed models are satisfactory, since it obtained a correlation coefficient and NSE close to 1, and MAPE and RMSE close to zero. Additionally, the first model presented the best performance, and the model that defined a single ANN had the lowest performance when compared to the others.