Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Titre : Forecasting monthly precipitation for arid regions using conditional artificial neural networks combined with Markov chain

Kurak bölge aylık yağışlarının Markov zinciri eklenmiş koşullu yapay sinir ağları ile tahmini

Auteur : DAHAMSHEH Y. Müh. Ahmad

Université de soutenance : İstanbul Teknik Üniversitesi

Grade : Doktora Tezi 2008

Résumé
Precipitation amount should be predicted accurately for an affective water resources management and planning. Prediction of hydrometeorological time series is difficult because of uncertainty in the parameters which affect the time series. In literature, the applications of artificial neural networks to forecast arid-region precipitation are limited.In this study, the potential of feed-forward backpropagation, radial basis function, generalized regression artificial neural networks and multiple linear regressions are used.New models are developed combination of Markov chain with artificial neural networks and multiple linear regression models to increase model performance.Synthetic data sets are generated to train artificial neural networks and multiple linear regressions. Thomas-Fiering models are used to generate synthetic series.Models developed in this study are all found not adequate particularly for the prediction of maximum precipitation. For this reason conditional neural networks and conditional multiple linear regression models are introduced to better estimate the observed precipitation. Results of the conditional models were found satisfactory.Three meteorological stations (Amman, Baqura and Safawi) from different geographical regions in Jordan are selected. Various homogeneity tests are employed for the data used in this study. Precipitation data from the three stations are found homogeneous according to the results of the tests. Antecedent monthly precipitation data are used as input to models to predict the total precipitation of next month.In conclusion, it is seen that the results of artificial neural networks and multiple linear regression models are not adequate. Artificial neural networks and multiple linear regressions combined with Markov chains are found successful in the prediction of dry months. Also artificial neural networks and multiple linear regressions trained with synthetic series do not perform efficiently. It is finally seen that the conditional neural networks and conditional multiple linear regressions considerably improve the accuracy of the one month ahead precipitation forecasting

Présentation

Version intégrale