Institutional Repository
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Grounwater level estimation in the area of Danube River, using artificial neural networks (ANNs)

Landros Ilias

Full record


URI: http://purl.tuc.gr/dl/dias/B15F7DAA-F087-4E59-A549-427F32CD1B03
Year 2019
Type of Item Diploma Work
License
Details
Appears in Collections

Summary

In the context of this thesis, the use and training of artificial neural networks is examined to simulate the underground level of the wells in the wider Danube River region. The Danube River extends into 10 different countries, but in our case the accidental observation wells for research are in Austria, Bulgaria, Germany, Croatia, Hungary, Romania, Serbia and Slovenia.Artificial neural networks try to simulate the underground level of the wells through the water balance. Initially, a proper pre-processing of the data needed to create the input and target vectors in the neural network was needed. The input vector consists of meteorological data, well coordinates, chronology, potential evapotranspiration, evaporation from water surfaces and evapotranspiration from soil and plants. Τhe target vector contains the actual values of the underground level of the wells. The vectors contain values from 01/01/2000-31/10/2014 for a total of 128 observation wells.After the data pre-processing was completed, artificial neural networks were trained with Neural Fitting tool (nftool) and Neural Network tool (nntool). The two training algorithms used are Levenberg-Marquardt and Bayesian Regularization. The training of artificial neural networks was based on the above for different parameters each time in terms of hidden nodes, training percentages and training algorithms.During the training of artificial neural networks, we try to locate the model with the parameters from which the best results will emerge. Selection criteria for selecting the optimal model were the square root of the mean square error, the correlation coefficient as well as some additional indicators commonly used in hydrology.In this dissertation, artificial neural networks were trained based on all observation wells. After choosing the optimal artificial neural network, from those trained with the data of all wells we also studied the possibility of training an artificial neural network for a single random observation well to compare the behavior of the artificial neural network.Finally, summarizing our results, a fault of the order 〖10〗^(-1) m was achieved using the Bayesian Regularization algorithm based on all observation wells. As far as the results from the training of an artificial neural network for a single well and not for all wells together, it gave better results in simulation of the underground level as the values are closer to the observed values but the artificial neural network can not simulate the extreme values.

Available Files

Services

Statistics