abstract

This thesis presents the application of artificial neural networks (ANNs) for the estimation of two relevant variables of irrigation engineering: reference evapotranspiration and integrated emitter local losses. On the one hand, one ANN model has been proposed for prediction of pressure losses due to emitter insertions in lateral lines. No previous research has mapped the capability of ANNs to estimate local pressure losses in microirrigation laterals. On the other hand, the applicability of a 4-input ANN model for ETo prediction has been studied in different continental contexts of the Valencia region. Furthermore, a new ANN model for ETo prediction has been proposed to improve the performance quality of the predictions.

For these purposes, multilayer feedforward networks with backpropagation, commonly known as multilayer perceptrons, were used under the supervisión of the Levenberg-Marquardt algorithm. In all cases, several ANN configurations were proposed and tested. Moreover, the process was repeated several times to mind the effect derived form the random assignment of the synaptic weights when the training algorithm is initialized. On the other hand, different strategies were followed to define the matrices for training, cross-validating and testing.

In comparison to traditional regression models, the performance quality of  the proposed ANN model for integrated emitter local loss prediction is referred to an independent test set. This fact allows to evaluate the real generalization potential of the model. For different cross-validation combinations, with data sets from at least three emitters, performance indexes over 0.85 were obtained. In relationship to ETo prediction models, the performance of the 4-input ANN model depends on the thermic oscillation range of the location where the model is created and the validity of the model outside the training location is limited. The new ANN model introduces in general two novelties to improve the performance quality in the training location and outisde. These ones are the consideration of relative humidity and the importation of climatic data from secondary ancillary weather stations which are similar from a continental point of view to the test location. This way, more accurate ETo predictions can be obtained.