GROUNDWATER LEVEL FORECAST USING COMBINED TIME SERIES MODELS
Combination of predictors. Machine learning. Hydrogeology. Sustainable development.
The modeling and prediction of groundwater levels are fundamental for the assessment of water availability. Therefore, time series formalisms are a robust alternative for inference about the state of this important resource. However, the use of predictor combination techniques for the modeling and prediction of time series of groundwater levels is still a little explored topic. The objective of this work is to develop combined models of time series prediction of groundwater levels. The Araripe Sedimentary Basin, in the State of Ceará, in the semi-arid region of Brazil, was adopted as a case study. The individual models used were classical Artificial Neural Networks (ANN), Extreme Learning Machines (ELM), Long Short-Term Memory (LSTM) networks and Support Vector Regression (SVR). The combination was performed using techniques such as Simple Mean (cSA), Simple Median (cSM), Minimum Variance (cMV), Copulas (cCP), in addition to the combination via ANN (cANN) and SVR (cSVR). The performance of the models was compared and classified according to error metrics. Most of the series studied showed significant lowering of levels. However, the models were able to represent and predict the behavior of groundwater in the aquifers of the Araripe basin in a satisfactory way. The ARIMA and LSTM models were superior, but the combined models showed attractive results, outperforming most of the individual alternatives. The combined approaches proposed can be improved.