Abstract

Water resources have a vital role in maintaining crops' survival in agricultural activities. Still, this resource's availability is limited and strongly affected by the climatic conditions of the area where the crops are located. Because of this situation, Energy-Water Management Systems have been implemented to optimize the use of resources for operating an irrigation system, while avoiding the over-extraction of water from the aquifer. However, these controllers require accurate future predictions of the climatic variables, which usually have a strong stochastic behavior. Therefore, this work proposes a Robust Energy-Water Management System based on prediction interval models to handle the uncertainty generated by the behavior of these climatic signals and the scarcity of accurate data for training the models. This work first analyzes available data from different sources to select the proper dataset for each climatic variable. Then, prediction intervals based on deep learning and fuzzy models are constructed for modeling solar radiation, air temperature, and precipitations. Using the information provided by the intervals, the proposed robust predictive controller is implemented and compared for two different cases, a conventional controller that uses the expected values of the climatic variables and a hypothetical ideal case where the controller knows the future with exact precision. Simulation results show that the prediction intervals based on deep learning and fuzzy models can reach good results for modeling these meteorological signals. Additionally, the proposed controller succeeds in maintaining the crops alive with the water available while getting a crop's profit close to the value achieved in the hypothetical ideal case. One of the main conclusions of this work is the importance that has the process of adequately analyzing different interval methods and data sources for achieving accurate models for climatic variables. Also, the quality of the interval models has an essential role in the controller for reaching results close to the ideal case, despite the stochasticity and uncertainty of the predicted signals.